1 // SPDX-License-Identifier: GPL-2.0-only
3 * VFIO: IOMMU DMA mapping support for Type1 IOMMU
5 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
6 * Author: Alex Williamson <alex.williamson@redhat.com>
8 * Derived from original vfio:
9 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
10 * Author: Tom Lyon, pugs@cisco.com
12 * We arbitrarily define a Type1 IOMMU as one matching the below code.
13 * It could be called the x86 IOMMU as it's designed for AMD-Vi & Intel
14 * VT-d, but that makes it harder to re-use as theoretically anyone
15 * implementing a similar IOMMU could make use of this. We expect the
16 * IOMMU to support the IOMMU API and have few to no restrictions around
17 * the IOVA range that can be mapped. The Type1 IOMMU is currently
18 * optimized for relatively static mappings of a userspace process with
19 * userspace pages pinned into memory. We also assume devices and IOMMU
20 * domains are PCI based as the IOMMU API is still centered around a
21 * device/bus interface rather than a group interface.
24 #include <linux/compat.h>
25 #include <linux/device.h>
27 #include <linux/highmem.h>
28 #include <linux/iommu.h>
29 #include <linux/module.h>
31 #include <linux/kthread.h>
32 #include <linux/rbtree.h>
33 #include <linux/sched/signal.h>
34 #include <linux/sched/mm.h>
35 #include <linux/slab.h>
36 #include <linux/uaccess.h>
37 #include <linux/vfio.h>
38 #include <linux/workqueue.h>
39 #include <linux/notifier.h>
40 #include <linux/dma-iommu.h>
41 #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;
69 struct rb_root dma_list;
70 struct blocking_notifier_head notifier;
71 unsigned int dma_avail;
72 unsigned int vaddr_invalid_count;
73 uint64_t pgsize_bitmap;
74 uint64_t num_non_pinned_groups;
75 wait_queue_head_t vaddr_wait;
78 bool dirty_page_tracking;
80 struct list_head emulated_iommu_groups;
84 struct iommu_domain *domain;
85 struct list_head next;
86 struct list_head group_list;
87 bool fgsp : 1; /* Fine-grained super pages */
88 bool enforce_cache_coherency : 1;
93 dma_addr_t iova; /* Device address */
94 unsigned long vaddr; /* Process virtual addr */
95 size_t size; /* Map size (bytes) */
96 int prot; /* IOMMU_READ/WRITE */
98 bool lock_cap; /* capable(CAP_IPC_LOCK) */
100 struct task_struct *task;
101 struct rb_root pfn_list; /* Ex-user pinned pfn list */
102 unsigned long *bitmap;
106 struct page **pages; /* for pin_user_pages_remote */
107 struct page *fallback_page; /* if pages alloc fails */
108 int capacity; /* length of pages array */
109 int size; /* of batch currently */
110 int offset; /* of next entry in pages */
113 struct vfio_iommu_group {
114 struct iommu_group *iommu_group;
115 struct list_head next;
116 bool pinned_page_dirty_scope;
120 struct list_head list;
126 * Guest RAM pinning working set or DMA target
130 dma_addr_t iova; /* Device address */
131 unsigned long pfn; /* Host pfn */
132 unsigned int ref_count;
135 struct vfio_regions {
136 struct list_head list;
142 #define DIRTY_BITMAP_BYTES(n) (ALIGN(n, BITS_PER_TYPE(u64)) / BITS_PER_BYTE)
145 * Input argument of number of bits to bitmap_set() is unsigned integer, which
146 * further casts to signed integer for unaligned multi-bit operation,
148 * Then maximum bitmap size supported is 2^31 bits divided by 2^3 bits/byte,
149 * that is 2^28 (256 MB) which maps to 2^31 * 2^12 = 2^43 (8TB) on 4K page
152 #define DIRTY_BITMAP_PAGES_MAX ((u64)INT_MAX)
153 #define DIRTY_BITMAP_SIZE_MAX DIRTY_BITMAP_BYTES(DIRTY_BITMAP_PAGES_MAX)
157 static int put_pfn(unsigned long pfn, int prot);
159 static struct vfio_iommu_group*
160 vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
161 struct iommu_group *iommu_group);
164 * This code handles mapping and unmapping of user data buffers
165 * into DMA'ble space using the IOMMU
168 static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
169 dma_addr_t start, size_t size)
171 struct rb_node *node = iommu->dma_list.rb_node;
174 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
176 if (start + size <= dma->iova)
177 node = node->rb_left;
178 else if (start >= dma->iova + dma->size)
179 node = node->rb_right;
187 static struct rb_node *vfio_find_dma_first_node(struct vfio_iommu *iommu,
188 dma_addr_t start, u64 size)
190 struct rb_node *res = NULL;
191 struct rb_node *node = iommu->dma_list.rb_node;
192 struct vfio_dma *dma_res = NULL;
195 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
197 if (start < dma->iova + dma->size) {
200 if (start >= dma->iova)
202 node = node->rb_left;
204 node = node->rb_right;
207 if (res && size && dma_res->iova >= start + size)
212 static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new)
214 struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL;
215 struct vfio_dma *dma;
219 dma = rb_entry(parent, struct vfio_dma, node);
221 if (new->iova + new->size <= dma->iova)
222 link = &(*link)->rb_left;
224 link = &(*link)->rb_right;
227 rb_link_node(&new->node, parent, link);
228 rb_insert_color(&new->node, &iommu->dma_list);
231 static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
233 rb_erase(&old->node, &iommu->dma_list);
237 static int vfio_dma_bitmap_alloc(struct vfio_dma *dma, size_t pgsize)
239 uint64_t npages = dma->size / pgsize;
241 if (npages > DIRTY_BITMAP_PAGES_MAX)
245 * Allocate extra 64 bits that are used to calculate shift required for
246 * bitmap_shift_left() to manipulate and club unaligned number of pages
247 * in adjacent vfio_dma ranges.
249 dma->bitmap = kvzalloc(DIRTY_BITMAP_BYTES(npages) + sizeof(u64),
257 static void vfio_dma_bitmap_free(struct vfio_dma *dma)
263 static void vfio_dma_populate_bitmap(struct vfio_dma *dma, size_t pgsize)
266 unsigned long pgshift = __ffs(pgsize);
268 for (p = rb_first(&dma->pfn_list); p; p = rb_next(p)) {
269 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn, node);
271 bitmap_set(dma->bitmap, (vpfn->iova - dma->iova) >> pgshift, 1);
275 static void vfio_iommu_populate_bitmap_full(struct vfio_iommu *iommu)
278 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
280 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
281 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
283 bitmap_set(dma->bitmap, 0, dma->size >> pgshift);
287 static int vfio_dma_bitmap_alloc_all(struct vfio_iommu *iommu, size_t pgsize)
291 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
292 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
295 ret = vfio_dma_bitmap_alloc(dma, pgsize);
299 for (p = rb_prev(n); p; p = rb_prev(p)) {
300 struct vfio_dma *dma = rb_entry(n,
301 struct vfio_dma, node);
303 vfio_dma_bitmap_free(dma);
307 vfio_dma_populate_bitmap(dma, pgsize);
312 static void vfio_dma_bitmap_free_all(struct vfio_iommu *iommu)
316 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
317 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
319 vfio_dma_bitmap_free(dma);
324 * Helper Functions for host iova-pfn list
326 static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
328 struct vfio_pfn *vpfn;
329 struct rb_node *node = dma->pfn_list.rb_node;
332 vpfn = rb_entry(node, struct vfio_pfn, node);
334 if (iova < vpfn->iova)
335 node = node->rb_left;
336 else if (iova > vpfn->iova)
337 node = node->rb_right;
344 static void vfio_link_pfn(struct vfio_dma *dma,
345 struct vfio_pfn *new)
347 struct rb_node **link, *parent = NULL;
348 struct vfio_pfn *vpfn;
350 link = &dma->pfn_list.rb_node;
353 vpfn = rb_entry(parent, struct vfio_pfn, node);
355 if (new->iova < vpfn->iova)
356 link = &(*link)->rb_left;
358 link = &(*link)->rb_right;
361 rb_link_node(&new->node, parent, link);
362 rb_insert_color(&new->node, &dma->pfn_list);
365 static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old)
367 rb_erase(&old->node, &dma->pfn_list);
370 static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova,
373 struct vfio_pfn *vpfn;
375 vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL);
382 vfio_link_pfn(dma, vpfn);
386 static void vfio_remove_from_pfn_list(struct vfio_dma *dma,
387 struct vfio_pfn *vpfn)
389 vfio_unlink_pfn(dma, vpfn);
393 static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma,
396 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
403 static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn)
408 if (!vpfn->ref_count) {
409 ret = put_pfn(vpfn->pfn, dma->prot);
410 vfio_remove_from_pfn_list(dma, vpfn);
415 static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
417 struct mm_struct *mm;
423 mm = async ? get_task_mm(dma->task) : dma->task->mm;
425 return -ESRCH; /* process exited */
427 ret = mmap_write_lock_killable(mm);
429 ret = __account_locked_vm(mm, abs(npage), npage > 0, dma->task,
431 mmap_write_unlock(mm);
441 * Some mappings aren't backed by a struct page, for example an mmap'd
442 * MMIO range for our own or another device. These use a different
443 * pfn conversion and shouldn't be tracked as locked pages.
444 * For compound pages, any driver that sets the reserved bit in head
445 * page needs to set the reserved bit in all subpages to be safe.
447 static bool is_invalid_reserved_pfn(unsigned long pfn)
450 return PageReserved(pfn_to_page(pfn));
455 static int put_pfn(unsigned long pfn, int prot)
457 if (!is_invalid_reserved_pfn(pfn)) {
458 struct page *page = pfn_to_page(pfn);
460 unpin_user_pages_dirty_lock(&page, 1, prot & IOMMU_WRITE);
466 #define VFIO_BATCH_MAX_CAPACITY (PAGE_SIZE / sizeof(struct page *))
468 static void vfio_batch_init(struct vfio_batch *batch)
473 if (unlikely(disable_hugepages))
476 batch->pages = (struct page **) __get_free_page(GFP_KERNEL);
480 batch->capacity = VFIO_BATCH_MAX_CAPACITY;
484 batch->pages = &batch->fallback_page;
488 static void vfio_batch_unpin(struct vfio_batch *batch, struct vfio_dma *dma)
490 while (batch->size) {
491 unsigned long pfn = page_to_pfn(batch->pages[batch->offset]);
493 put_pfn(pfn, dma->prot);
499 static void vfio_batch_fini(struct vfio_batch *batch)
501 if (batch->capacity == VFIO_BATCH_MAX_CAPACITY)
502 free_page((unsigned long)batch->pages);
505 static int follow_fault_pfn(struct vm_area_struct *vma, struct mm_struct *mm,
506 unsigned long vaddr, unsigned long *pfn,
513 ret = follow_pte(vma->vm_mm, vaddr, &ptep, &ptl);
515 bool unlocked = false;
517 ret = fixup_user_fault(mm, vaddr,
519 (write_fault ? FAULT_FLAG_WRITE : 0),
527 ret = follow_pte(vma->vm_mm, vaddr, &ptep, &ptl);
532 if (write_fault && !pte_write(*ptep))
535 *pfn = pte_pfn(*ptep);
537 pte_unmap_unlock(ptep, ptl);
542 * Returns the positive number of pfns successfully obtained or a negative
545 static int vaddr_get_pfns(struct mm_struct *mm, unsigned long vaddr,
546 long npages, int prot, unsigned long *pfn,
549 struct vm_area_struct *vma;
550 unsigned int flags = 0;
553 if (prot & IOMMU_WRITE)
557 ret = pin_user_pages_remote(mm, vaddr, npages, flags | FOLL_LONGTERM,
560 *pfn = page_to_pfn(pages[0]);
564 vaddr = untagged_addr(vaddr);
567 vma = vma_lookup(mm, vaddr);
569 if (vma && vma->vm_flags & VM_PFNMAP) {
570 ret = follow_fault_pfn(vma, mm, vaddr, pfn, prot & IOMMU_WRITE);
575 if (is_invalid_reserved_pfn(*pfn))
582 mmap_read_unlock(mm);
586 static int vfio_wait(struct vfio_iommu *iommu)
590 prepare_to_wait(&iommu->vaddr_wait, &wait, TASK_KILLABLE);
591 mutex_unlock(&iommu->lock);
593 mutex_lock(&iommu->lock);
594 finish_wait(&iommu->vaddr_wait, &wait);
595 if (kthread_should_stop() || !iommu->container_open ||
596 fatal_signal_pending(current)) {
603 * Find dma struct and wait for its vaddr to be valid. iommu lock is dropped
604 * if the task waits, but is re-locked on return. Return result in *dma_p.
605 * Return 0 on success with no waiting, WAITED on success if waited, and -errno
608 static int vfio_find_dma_valid(struct vfio_iommu *iommu, dma_addr_t start,
609 size_t size, struct vfio_dma **dma_p)
614 *dma_p = vfio_find_dma(iommu, start, size);
617 else if (!(*dma_p)->vaddr_invalid)
620 ret = vfio_wait(iommu);
621 } while (ret == WAITED);
627 * Wait for all vaddr in the dma_list to become valid. iommu lock is dropped
628 * if the task waits, but is re-locked on return. Return 0 on success with no
629 * waiting, WAITED on success if waited, and -errno on error.
631 static int vfio_wait_all_valid(struct vfio_iommu *iommu)
635 while (iommu->vaddr_invalid_count && ret >= 0)
636 ret = vfio_wait(iommu);
642 * Attempt to pin pages. We really don't want to track all the pfns and
643 * the iommu can only map chunks of consecutive pfns anyway, so get the
644 * first page and all consecutive pages with the same locking.
646 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
647 long npage, unsigned long *pfn_base,
648 unsigned long limit, struct vfio_batch *batch)
651 struct mm_struct *mm = current->mm;
652 long ret, pinned = 0, lock_acct = 0;
654 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
656 /* This code path is only user initiated */
661 /* Leftover pages in batch from an earlier call. */
662 *pfn_base = page_to_pfn(batch->pages[batch->offset]);
664 rsvd = is_invalid_reserved_pfn(*pfn_base);
671 /* Empty batch, so refill it. */
672 long req_pages = min_t(long, npage, batch->capacity);
674 ret = vaddr_get_pfns(mm, vaddr, req_pages, dma->prot,
684 rsvd = is_invalid_reserved_pfn(*pfn_base);
689 * pfn is preset for the first iteration of this inner loop and
690 * updated at the end to handle a VM_PFNMAP pfn. In that case,
691 * batch->pages isn't valid (there's no struct page), so allow
692 * batch->pages to be touched only when there's more than one
693 * pfn to check, which guarantees the pfns are from a
697 if (pfn != *pfn_base + pinned ||
698 rsvd != is_invalid_reserved_pfn(pfn))
702 * Reserved pages aren't counted against the user,
703 * externally pinned pages are already counted against
706 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
707 if (!dma->lock_cap &&
708 mm->locked_vm + lock_acct + 1 > limit) {
709 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
710 __func__, limit << PAGE_SHIFT);
727 pfn = page_to_pfn(batch->pages[batch->offset]);
730 if (unlikely(disable_hugepages))
735 ret = vfio_lock_acct(dma, lock_acct, false);
738 if (batch->size == 1 && !batch->offset) {
739 /* May be a VM_PFNMAP pfn, which the batch can't remember. */
740 put_pfn(pfn, dma->prot);
745 if (pinned && !rsvd) {
746 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
747 put_pfn(pfn, dma->prot);
749 vfio_batch_unpin(batch, dma);
757 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
758 unsigned long pfn, long npage,
761 long unlocked = 0, locked = 0;
764 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
765 if (put_pfn(pfn++, dma->prot)) {
767 if (vfio_find_vpfn(dma, iova))
773 vfio_lock_acct(dma, locked - unlocked, true);
778 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
779 unsigned long *pfn_base, bool do_accounting)
781 struct page *pages[1];
782 struct mm_struct *mm;
785 mm = get_task_mm(dma->task);
789 ret = vaddr_get_pfns(mm, vaddr, 1, dma->prot, pfn_base, pages);
795 if (do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
796 ret = vfio_lock_acct(dma, 1, true);
798 put_pfn(*pfn_base, dma->prot);
800 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
801 "(%ld) exceeded\n", __func__,
802 dma->task->comm, task_pid_nr(dma->task),
803 task_rlimit(dma->task, RLIMIT_MEMLOCK));
812 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
816 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
821 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
824 vfio_lock_acct(dma, -unlocked, true);
829 static int vfio_iommu_type1_pin_pages(void *iommu_data,
830 struct iommu_group *iommu_group,
831 unsigned long *user_pfn,
833 unsigned long *phys_pfn)
835 struct vfio_iommu *iommu = iommu_data;
836 struct vfio_iommu_group *group;
838 unsigned long remote_vaddr;
839 struct vfio_dma *dma;
843 if (!iommu || !user_pfn || !phys_pfn)
846 /* Supported for v2 version only */
850 mutex_lock(&iommu->lock);
853 * Wait for all necessary vaddr's to be valid so they can be used in
854 * the main loop without dropping the lock, to avoid racing vs unmap.
857 if (iommu->vaddr_invalid_count) {
858 for (i = 0; i < npage; i++) {
859 iova = user_pfn[i] << PAGE_SHIFT;
860 ret = vfio_find_dma_valid(iommu, iova, PAGE_SIZE, &dma);
868 /* Fail if notifier list is empty */
869 if (!iommu->notifier.head) {
875 * If iommu capable domain exist in the container then all pages are
876 * already pinned and accounted. Accounting should be done if there is no
877 * iommu capable domain in the container.
879 do_accounting = list_empty(&iommu->domain_list);
881 for (i = 0; i < npage; i++) {
882 struct vfio_pfn *vpfn;
884 iova = user_pfn[i] << PAGE_SHIFT;
885 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
891 if ((dma->prot & prot) != prot) {
896 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
898 phys_pfn[i] = vpfn->pfn;
902 remote_vaddr = dma->vaddr + (iova - dma->iova);
903 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
908 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
910 if (put_pfn(phys_pfn[i], dma->prot) && do_accounting)
911 vfio_lock_acct(dma, -1, true);
915 if (iommu->dirty_page_tracking) {
916 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
919 * Bitmap populated with the smallest supported page
922 bitmap_set(dma->bitmap,
923 (iova - dma->iova) >> pgshift, 1);
928 group = vfio_iommu_find_iommu_group(iommu, iommu_group);
929 if (!group->pinned_page_dirty_scope) {
930 group->pinned_page_dirty_scope = true;
931 iommu->num_non_pinned_groups--;
938 for (j = 0; j < i; j++) {
941 iova = user_pfn[j] << PAGE_SHIFT;
942 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
943 vfio_unpin_page_external(dma, iova, do_accounting);
947 mutex_unlock(&iommu->lock);
951 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
952 unsigned long *user_pfn,
955 struct vfio_iommu *iommu = iommu_data;
959 if (!iommu || !user_pfn || npage <= 0)
962 /* Supported for v2 version only */
966 mutex_lock(&iommu->lock);
968 do_accounting = list_empty(&iommu->domain_list);
969 for (i = 0; i < npage; i++) {
970 struct vfio_dma *dma;
973 iova = user_pfn[i] << PAGE_SHIFT;
974 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
978 vfio_unpin_page_external(dma, iova, do_accounting);
981 mutex_unlock(&iommu->lock);
982 return i > 0 ? i : -EINVAL;
985 static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
986 struct list_head *regions,
987 struct iommu_iotlb_gather *iotlb_gather)
990 struct vfio_regions *entry, *next;
992 iommu_iotlb_sync(domain->domain, iotlb_gather);
994 list_for_each_entry_safe(entry, next, regions, list) {
995 unlocked += vfio_unpin_pages_remote(dma,
997 entry->phys >> PAGE_SHIFT,
998 entry->len >> PAGE_SHIFT,
1000 list_del(&entry->list);
1010 * Generally, VFIO needs to unpin remote pages after each IOTLB flush.
1011 * Therefore, when using IOTLB flush sync interface, VFIO need to keep track
1012 * of these regions (currently using a list).
1014 * This value specifies maximum number of regions for each IOTLB flush sync.
1016 #define VFIO_IOMMU_TLB_SYNC_MAX 512
1018 static size_t unmap_unpin_fast(struct vfio_domain *domain,
1019 struct vfio_dma *dma, dma_addr_t *iova,
1020 size_t len, phys_addr_t phys, long *unlocked,
1021 struct list_head *unmapped_list,
1023 struct iommu_iotlb_gather *iotlb_gather)
1025 size_t unmapped = 0;
1026 struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
1029 unmapped = iommu_unmap_fast(domain->domain, *iova, len,
1035 entry->iova = *iova;
1037 entry->len = unmapped;
1038 list_add_tail(&entry->list, unmapped_list);
1046 * Sync if the number of fast-unmap regions hits the limit
1047 * or in case of errors.
1049 if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
1050 *unlocked += vfio_sync_unpin(dma, domain, unmapped_list,
1058 static size_t unmap_unpin_slow(struct vfio_domain *domain,
1059 struct vfio_dma *dma, dma_addr_t *iova,
1060 size_t len, phys_addr_t phys,
1063 size_t unmapped = iommu_unmap(domain->domain, *iova, len);
1066 *unlocked += vfio_unpin_pages_remote(dma, *iova,
1068 unmapped >> PAGE_SHIFT,
1076 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
1079 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
1080 struct vfio_domain *domain, *d;
1081 LIST_HEAD(unmapped_region_list);
1082 struct iommu_iotlb_gather iotlb_gather;
1083 int unmapped_region_cnt = 0;
1089 if (list_empty(&iommu->domain_list))
1093 * We use the IOMMU to track the physical addresses, otherwise we'd
1094 * need a much more complicated tracking system. Unfortunately that
1095 * means we need to use one of the iommu domains to figure out the
1096 * pfns to unpin. The rest need to be unmapped in advance so we have
1097 * no iommu translations remaining when the pages are unpinned.
1099 domain = d = list_first_entry(&iommu->domain_list,
1100 struct vfio_domain, next);
1102 list_for_each_entry_continue(d, &iommu->domain_list, next) {
1103 iommu_unmap(d->domain, dma->iova, dma->size);
1107 iommu_iotlb_gather_init(&iotlb_gather);
1108 while (iova < end) {
1109 size_t unmapped, len;
1110 phys_addr_t phys, next;
1112 phys = iommu_iova_to_phys(domain->domain, iova);
1113 if (WARN_ON(!phys)) {
1119 * To optimize for fewer iommu_unmap() calls, each of which
1120 * may require hardware cache flushing, try to find the
1121 * largest contiguous physical memory chunk to unmap.
1123 for (len = PAGE_SIZE;
1124 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
1125 next = iommu_iova_to_phys(domain->domain, iova + len);
1126 if (next != phys + len)
1131 * First, try to use fast unmap/unpin. In case of failure,
1132 * switch to slow unmap/unpin path.
1134 unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys,
1135 &unlocked, &unmapped_region_list,
1136 &unmapped_region_cnt,
1139 unmapped = unmap_unpin_slow(domain, dma, &iova, len,
1141 if (WARN_ON(!unmapped))
1146 dma->iommu_mapped = false;
1148 if (unmapped_region_cnt) {
1149 unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list,
1153 if (do_accounting) {
1154 vfio_lock_acct(dma, -unlocked, true);
1160 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
1162 WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list));
1163 vfio_unmap_unpin(iommu, dma, true);
1164 vfio_unlink_dma(iommu, dma);
1165 put_task_struct(dma->task);
1166 vfio_dma_bitmap_free(dma);
1167 if (dma->vaddr_invalid) {
1168 iommu->vaddr_invalid_count--;
1169 wake_up_all(&iommu->vaddr_wait);
1175 static void vfio_update_pgsize_bitmap(struct vfio_iommu *iommu)
1177 struct vfio_domain *domain;
1179 iommu->pgsize_bitmap = ULONG_MAX;
1181 list_for_each_entry(domain, &iommu->domain_list, next)
1182 iommu->pgsize_bitmap &= domain->domain->pgsize_bitmap;
1185 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
1186 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
1187 * That way the user will be able to map/unmap buffers whose size/
1188 * start address is aligned with PAGE_SIZE. Pinning code uses that
1189 * granularity while iommu driver can use the sub-PAGE_SIZE size
1190 * to map the buffer.
1192 if (iommu->pgsize_bitmap & ~PAGE_MASK) {
1193 iommu->pgsize_bitmap &= PAGE_MASK;
1194 iommu->pgsize_bitmap |= PAGE_SIZE;
1198 static int update_user_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1199 struct vfio_dma *dma, dma_addr_t base_iova,
1202 unsigned long pgshift = __ffs(pgsize);
1203 unsigned long nbits = dma->size >> pgshift;
1204 unsigned long bit_offset = (dma->iova - base_iova) >> pgshift;
1205 unsigned long copy_offset = bit_offset / BITS_PER_LONG;
1206 unsigned long shift = bit_offset % BITS_PER_LONG;
1207 unsigned long leftover;
1210 * mark all pages dirty if any IOMMU capable device is not able
1211 * to report dirty pages and all pages are pinned and mapped.
1213 if (iommu->num_non_pinned_groups && dma->iommu_mapped)
1214 bitmap_set(dma->bitmap, 0, nbits);
1217 bitmap_shift_left(dma->bitmap, dma->bitmap, shift,
1220 if (copy_from_user(&leftover,
1221 (void __user *)(bitmap + copy_offset),
1225 bitmap_or(dma->bitmap, dma->bitmap, &leftover, shift);
1228 if (copy_to_user((void __user *)(bitmap + copy_offset), dma->bitmap,
1229 DIRTY_BITMAP_BYTES(nbits + shift)))
1235 static int vfio_iova_dirty_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1236 dma_addr_t iova, size_t size, size_t pgsize)
1238 struct vfio_dma *dma;
1240 unsigned long pgshift = __ffs(pgsize);
1244 * GET_BITMAP request must fully cover vfio_dma mappings. Multiple
1245 * vfio_dma mappings may be clubbed by specifying large ranges, but
1246 * there must not be any previous mappings bisected by the range.
1247 * An error will be returned if these conditions are not met.
1249 dma = vfio_find_dma(iommu, iova, 1);
1250 if (dma && dma->iova != iova)
1253 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1254 if (dma && dma->iova + dma->size != iova + size)
1257 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1258 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1260 if (dma->iova < iova)
1263 if (dma->iova > iova + size - 1)
1266 ret = update_user_bitmap(bitmap, iommu, dma, iova, pgsize);
1271 * Re-populate bitmap to include all pinned pages which are
1272 * considered as dirty but exclude pages which are unpinned and
1273 * pages which are marked dirty by vfio_dma_rw()
1275 bitmap_clear(dma->bitmap, 0, dma->size >> pgshift);
1276 vfio_dma_populate_bitmap(dma, pgsize);
1281 static int verify_bitmap_size(uint64_t npages, uint64_t bitmap_size)
1283 if (!npages || !bitmap_size || (bitmap_size > DIRTY_BITMAP_SIZE_MAX) ||
1284 (bitmap_size < DIRTY_BITMAP_BYTES(npages)))
1290 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
1291 struct vfio_iommu_type1_dma_unmap *unmap,
1292 struct vfio_bitmap *bitmap)
1294 struct vfio_dma *dma, *dma_last = NULL;
1295 size_t unmapped = 0, pgsize;
1296 int ret = -EINVAL, retries = 0;
1297 unsigned long pgshift;
1298 dma_addr_t iova = unmap->iova;
1299 u64 size = unmap->size;
1300 bool unmap_all = unmap->flags & VFIO_DMA_UNMAP_FLAG_ALL;
1301 bool invalidate_vaddr = unmap->flags & VFIO_DMA_UNMAP_FLAG_VADDR;
1302 struct rb_node *n, *first_n;
1304 mutex_lock(&iommu->lock);
1306 pgshift = __ffs(iommu->pgsize_bitmap);
1307 pgsize = (size_t)1 << pgshift;
1309 if (iova & (pgsize - 1))
1316 } else if (!size || size & (pgsize - 1) ||
1317 iova + size - 1 < iova || size > SIZE_MAX) {
1321 /* When dirty tracking is enabled, allow only min supported pgsize */
1322 if ((unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
1323 (!iommu->dirty_page_tracking || (bitmap->pgsize != pgsize))) {
1327 WARN_ON((pgsize - 1) & PAGE_MASK);
1330 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
1331 * avoid tracking individual mappings. This means that the granularity
1332 * of the original mapping was lost and the user was allowed to attempt
1333 * to unmap any range. Depending on the contiguousness of physical
1334 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
1335 * or may not have worked. We only guaranteed unmap granularity
1336 * matching the original mapping; even though it was untracked here,
1337 * the original mappings are reflected in IOMMU mappings. This
1338 * resulted in a couple unusual behaviors. First, if a range is not
1339 * able to be unmapped, ex. a set of 4k pages that was mapped as a
1340 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
1341 * a zero sized unmap. Also, if an unmap request overlaps the first
1342 * address of a hugepage, the IOMMU will unmap the entire hugepage.
1343 * This also returns success and the returned unmap size reflects the
1344 * actual size unmapped.
1346 * We attempt to maintain compatibility with this "v1" interface, but
1347 * we take control out of the hands of the IOMMU. Therefore, an unmap
1348 * request offset from the beginning of the original mapping will
1349 * return success with zero sized unmap. And an unmap request covering
1350 * the first iova of mapping will unmap the entire range.
1352 * The v2 version of this interface intends to be more deterministic.
1353 * Unmap requests must fully cover previous mappings. Multiple
1354 * mappings may still be unmaped by specifying large ranges, but there
1355 * must not be any previous mappings bisected by the range. An error
1356 * will be returned if these conditions are not met. The v2 interface
1357 * will only return success and a size of zero if there were no
1358 * mappings within the range.
1360 if (iommu->v2 && !unmap_all) {
1361 dma = vfio_find_dma(iommu, iova, 1);
1362 if (dma && dma->iova != iova)
1365 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1366 if (dma && dma->iova + dma->size != iova + size)
1371 n = first_n = vfio_find_dma_first_node(iommu, iova, size);
1374 dma = rb_entry(n, struct vfio_dma, node);
1375 if (dma->iova >= iova + size)
1378 if (!iommu->v2 && iova > dma->iova)
1381 if (invalidate_vaddr) {
1382 if (dma->vaddr_invalid) {
1383 struct rb_node *last_n = n;
1385 for (n = first_n; n != last_n; n = rb_next(n)) {
1387 struct vfio_dma, node);
1388 dma->vaddr_invalid = false;
1389 iommu->vaddr_invalid_count--;
1395 dma->vaddr_invalid = true;
1396 iommu->vaddr_invalid_count++;
1397 unmapped += dma->size;
1402 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
1403 struct vfio_iommu_type1_dma_unmap nb_unmap;
1405 if (dma_last == dma) {
1406 BUG_ON(++retries > 10);
1412 nb_unmap.iova = dma->iova;
1413 nb_unmap.size = dma->size;
1416 * Notify anyone (mdev vendor drivers) to invalidate and
1417 * unmap iovas within the range we're about to unmap.
1418 * Vendor drivers MUST unpin pages in response to an
1421 mutex_unlock(&iommu->lock);
1422 blocking_notifier_call_chain(&iommu->notifier,
1423 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
1425 mutex_lock(&iommu->lock);
1429 if (unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
1430 ret = update_user_bitmap(bitmap->data, iommu, dma,
1436 unmapped += dma->size;
1438 vfio_remove_dma(iommu, dma);
1442 mutex_unlock(&iommu->lock);
1444 /* Report how much was unmapped */
1445 unmap->size = unmapped;
1450 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
1451 unsigned long pfn, long npage, int prot)
1453 struct vfio_domain *d;
1456 list_for_each_entry(d, &iommu->domain_list, next) {
1457 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
1458 npage << PAGE_SHIFT, prot | IOMMU_CACHE);
1468 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) {
1469 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
1476 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
1479 dma_addr_t iova = dma->iova;
1480 unsigned long vaddr = dma->vaddr;
1481 struct vfio_batch batch;
1482 size_t size = map_size;
1484 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1487 vfio_batch_init(&batch);
1490 /* Pin a contiguous chunk of memory */
1491 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
1492 size >> PAGE_SHIFT, &pfn, limit,
1501 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
1504 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
1506 vfio_batch_unpin(&batch, dma);
1510 size -= npage << PAGE_SHIFT;
1511 dma->size += npage << PAGE_SHIFT;
1514 vfio_batch_fini(&batch);
1515 dma->iommu_mapped = true;
1518 vfio_remove_dma(iommu, dma);
1524 * Check dma map request is within a valid iova range
1526 static bool vfio_iommu_iova_dma_valid(struct vfio_iommu *iommu,
1527 dma_addr_t start, dma_addr_t end)
1529 struct list_head *iova = &iommu->iova_list;
1530 struct vfio_iova *node;
1532 list_for_each_entry(node, iova, list) {
1533 if (start >= node->start && end <= node->end)
1538 * Check for list_empty() as well since a container with
1539 * a single mdev device will have an empty list.
1541 return list_empty(iova);
1544 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1545 struct vfio_iommu_type1_dma_map *map)
1547 bool set_vaddr = map->flags & VFIO_DMA_MAP_FLAG_VADDR;
1548 dma_addr_t iova = map->iova;
1549 unsigned long vaddr = map->vaddr;
1550 size_t size = map->size;
1551 int ret = 0, prot = 0;
1553 struct vfio_dma *dma;
1555 /* Verify that none of our __u64 fields overflow */
1556 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1559 /* READ/WRITE from device perspective */
1560 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1561 prot |= IOMMU_WRITE;
1562 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1565 if ((prot && set_vaddr) || (!prot && !set_vaddr))
1568 mutex_lock(&iommu->lock);
1570 pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
1572 WARN_ON((pgsize - 1) & PAGE_MASK);
1574 if (!size || (size | iova | vaddr) & (pgsize - 1)) {
1579 /* Don't allow IOVA or virtual address wrap */
1580 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) {
1585 dma = vfio_find_dma(iommu, iova, size);
1589 } else if (!dma->vaddr_invalid || dma->iova != iova ||
1590 dma->size != size) {
1594 dma->vaddr_invalid = false;
1595 iommu->vaddr_invalid_count--;
1596 wake_up_all(&iommu->vaddr_wait);
1604 if (!iommu->dma_avail) {
1609 if (!vfio_iommu_iova_dma_valid(iommu, iova, iova + size - 1)) {
1614 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1626 * We need to be able to both add to a task's locked memory and test
1627 * against the locked memory limit and we need to be able to do both
1628 * outside of this call path as pinning can be asynchronous via the
1629 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1630 * task_struct and VM locked pages requires an mm_struct, however
1631 * holding an indefinite mm reference is not recommended, therefore we
1632 * only hold a reference to a task. We could hold a reference to
1633 * current, however QEMU uses this call path through vCPU threads,
1634 * which can be killed resulting in a NULL mm and failure in the unmap
1635 * path when called via a different thread. Avoid this problem by
1636 * using the group_leader as threads within the same group require
1637 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1640 * Previously we also used the task for testing CAP_IPC_LOCK at the
1641 * time of pinning and accounting, however has_capability() makes use
1642 * of real_cred, a copy-on-write field, so we can't guarantee that it
1643 * matches group_leader, or in fact that it might not change by the
1644 * time it's evaluated. If a process were to call MAP_DMA with
1645 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1646 * possibly see different results for an iommu_mapped vfio_dma vs
1647 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1648 * time of calling MAP_DMA.
1650 get_task_struct(current->group_leader);
1651 dma->task = current->group_leader;
1652 dma->lock_cap = capable(CAP_IPC_LOCK);
1654 dma->pfn_list = RB_ROOT;
1656 /* Insert zero-sized and grow as we map chunks of it */
1657 vfio_link_dma(iommu, dma);
1659 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1660 if (list_empty(&iommu->domain_list))
1663 ret = vfio_pin_map_dma(iommu, dma, size);
1665 if (!ret && iommu->dirty_page_tracking) {
1666 ret = vfio_dma_bitmap_alloc(dma, pgsize);
1668 vfio_remove_dma(iommu, dma);
1672 mutex_unlock(&iommu->lock);
1676 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1677 struct vfio_domain *domain)
1679 struct vfio_batch batch;
1680 struct vfio_domain *d = NULL;
1682 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1685 ret = vfio_wait_all_valid(iommu);
1689 /* Arbitrarily pick the first domain in the list for lookups */
1690 if (!list_empty(&iommu->domain_list))
1691 d = list_first_entry(&iommu->domain_list,
1692 struct vfio_domain, next);
1694 vfio_batch_init(&batch);
1696 n = rb_first(&iommu->dma_list);
1698 for (; n; n = rb_next(n)) {
1699 struct vfio_dma *dma;
1702 dma = rb_entry(n, struct vfio_dma, node);
1705 while (iova < dma->iova + dma->size) {
1709 if (dma->iommu_mapped) {
1713 if (WARN_ON(!d)) { /* mapped w/o a domain?! */
1718 phys = iommu_iova_to_phys(d->domain, iova);
1720 if (WARN_ON(!phys)) {
1728 while (i < dma->iova + dma->size &&
1729 p == iommu_iova_to_phys(d->domain, i)) {
1736 unsigned long vaddr = dma->vaddr +
1738 size_t n = dma->iova + dma->size - iova;
1741 npage = vfio_pin_pages_remote(dma, vaddr,
1751 phys = pfn << PAGE_SHIFT;
1752 size = npage << PAGE_SHIFT;
1755 ret = iommu_map(domain->domain, iova, phys,
1756 size, dma->prot | IOMMU_CACHE);
1758 if (!dma->iommu_mapped) {
1759 vfio_unpin_pages_remote(dma, iova,
1763 vfio_batch_unpin(&batch, dma);
1772 /* All dmas are now mapped, defer to second tree walk for unwind */
1773 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1774 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1776 dma->iommu_mapped = true;
1779 vfio_batch_fini(&batch);
1783 for (; n; n = rb_prev(n)) {
1784 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1787 if (dma->iommu_mapped) {
1788 iommu_unmap(domain->domain, dma->iova, dma->size);
1793 while (iova < dma->iova + dma->size) {
1794 phys_addr_t phys, p;
1798 phys = iommu_iova_to_phys(domain->domain, iova);
1807 while (i < dma->iova + dma->size &&
1808 p == iommu_iova_to_phys(domain->domain, i)) {
1814 iommu_unmap(domain->domain, iova, size);
1815 vfio_unpin_pages_remote(dma, iova, phys >> PAGE_SHIFT,
1816 size >> PAGE_SHIFT, true);
1820 vfio_batch_fini(&batch);
1825 * We change our unmap behavior slightly depending on whether the IOMMU
1826 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1827 * for practically any contiguous power-of-two mapping we give it. This means
1828 * we don't need to look for contiguous chunks ourselves to make unmapping
1829 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1830 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1831 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1832 * hugetlbfs is in use.
1834 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1837 int ret, order = get_order(PAGE_SIZE * 2);
1839 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1843 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1844 IOMMU_READ | IOMMU_WRITE | IOMMU_CACHE);
1846 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1848 if (unmapped == PAGE_SIZE)
1849 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1851 domain->fgsp = true;
1854 __free_pages(pages, order);
1857 static struct vfio_iommu_group *find_iommu_group(struct vfio_domain *domain,
1858 struct iommu_group *iommu_group)
1860 struct vfio_iommu_group *g;
1862 list_for_each_entry(g, &domain->group_list, next) {
1863 if (g->iommu_group == iommu_group)
1870 static struct vfio_iommu_group*
1871 vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
1872 struct iommu_group *iommu_group)
1874 struct vfio_iommu_group *group;
1875 struct vfio_domain *domain;
1877 list_for_each_entry(domain, &iommu->domain_list, next) {
1878 group = find_iommu_group(domain, iommu_group);
1883 list_for_each_entry(group, &iommu->emulated_iommu_groups, next)
1884 if (group->iommu_group == iommu_group)
1889 static bool vfio_iommu_has_sw_msi(struct list_head *group_resv_regions,
1892 struct iommu_resv_region *region;
1895 list_for_each_entry(region, group_resv_regions, list) {
1897 * The presence of any 'real' MSI regions should take
1898 * precedence over the software-managed one if the
1899 * IOMMU driver happens to advertise both types.
1901 if (region->type == IOMMU_RESV_MSI) {
1906 if (region->type == IOMMU_RESV_SW_MSI) {
1907 *base = region->start;
1916 * This is a helper function to insert an address range to iova list.
1917 * The list is initially created with a single entry corresponding to
1918 * the IOMMU domain geometry to which the device group is attached.
1919 * The list aperture gets modified when a new domain is added to the
1920 * container if the new aperture doesn't conflict with the current one
1921 * or with any existing dma mappings. The list is also modified to
1922 * exclude any reserved regions associated with the device group.
1924 static int vfio_iommu_iova_insert(struct list_head *head,
1925 dma_addr_t start, dma_addr_t end)
1927 struct vfio_iova *region;
1929 region = kmalloc(sizeof(*region), GFP_KERNEL);
1933 INIT_LIST_HEAD(®ion->list);
1934 region->start = start;
1937 list_add_tail(®ion->list, head);
1942 * Check the new iommu aperture conflicts with existing aper or with any
1943 * existing dma mappings.
1945 static bool vfio_iommu_aper_conflict(struct vfio_iommu *iommu,
1946 dma_addr_t start, dma_addr_t end)
1948 struct vfio_iova *first, *last;
1949 struct list_head *iova = &iommu->iova_list;
1951 if (list_empty(iova))
1954 /* Disjoint sets, return conflict */
1955 first = list_first_entry(iova, struct vfio_iova, list);
1956 last = list_last_entry(iova, struct vfio_iova, list);
1957 if (start > last->end || end < first->start)
1960 /* Check for any existing dma mappings below the new start */
1961 if (start > first->start) {
1962 if (vfio_find_dma(iommu, first->start, start - first->start))
1966 /* Check for any existing dma mappings beyond the new end */
1967 if (end < last->end) {
1968 if (vfio_find_dma(iommu, end + 1, last->end - end))
1976 * Resize iommu iova aperture window. This is called only if the new
1977 * aperture has no conflict with existing aperture and dma mappings.
1979 static int vfio_iommu_aper_resize(struct list_head *iova,
1980 dma_addr_t start, dma_addr_t end)
1982 struct vfio_iova *node, *next;
1984 if (list_empty(iova))
1985 return vfio_iommu_iova_insert(iova, start, end);
1987 /* Adjust iova list start */
1988 list_for_each_entry_safe(node, next, iova, list) {
1989 if (start < node->start)
1991 if (start >= node->start && start < node->end) {
1992 node->start = start;
1995 /* Delete nodes before new start */
1996 list_del(&node->list);
2000 /* Adjust iova list end */
2001 list_for_each_entry_safe(node, next, iova, list) {
2002 if (end > node->end)
2004 if (end > node->start && end <= node->end) {
2008 /* Delete nodes after new end */
2009 list_del(&node->list);
2017 * Check reserved region conflicts with existing dma mappings
2019 static bool vfio_iommu_resv_conflict(struct vfio_iommu *iommu,
2020 struct list_head *resv_regions)
2022 struct iommu_resv_region *region;
2024 /* Check for conflict with existing dma mappings */
2025 list_for_each_entry(region, resv_regions, list) {
2026 if (region->type == IOMMU_RESV_DIRECT_RELAXABLE)
2029 if (vfio_find_dma(iommu, region->start, region->length))
2037 * Check iova region overlap with reserved regions and
2038 * exclude them from the iommu iova range
2040 static int vfio_iommu_resv_exclude(struct list_head *iova,
2041 struct list_head *resv_regions)
2043 struct iommu_resv_region *resv;
2044 struct vfio_iova *n, *next;
2046 list_for_each_entry(resv, resv_regions, list) {
2047 phys_addr_t start, end;
2049 if (resv->type == IOMMU_RESV_DIRECT_RELAXABLE)
2052 start = resv->start;
2053 end = resv->start + resv->length - 1;
2055 list_for_each_entry_safe(n, next, iova, list) {
2059 if (start > n->end || end < n->start)
2062 * Insert a new node if current node overlaps with the
2063 * reserve region to exclude that from valid iova range.
2064 * Note that, new node is inserted before the current
2065 * node and finally the current node is deleted keeping
2066 * the list updated and sorted.
2068 if (start > n->start)
2069 ret = vfio_iommu_iova_insert(&n->list, n->start,
2071 if (!ret && end < n->end)
2072 ret = vfio_iommu_iova_insert(&n->list, end + 1,
2082 if (list_empty(iova))
2088 static void vfio_iommu_resv_free(struct list_head *resv_regions)
2090 struct iommu_resv_region *n, *next;
2092 list_for_each_entry_safe(n, next, resv_regions, list) {
2098 static void vfio_iommu_iova_free(struct list_head *iova)
2100 struct vfio_iova *n, *next;
2102 list_for_each_entry_safe(n, next, iova, list) {
2108 static int vfio_iommu_iova_get_copy(struct vfio_iommu *iommu,
2109 struct list_head *iova_copy)
2111 struct list_head *iova = &iommu->iova_list;
2112 struct vfio_iova *n;
2115 list_for_each_entry(n, iova, list) {
2116 ret = vfio_iommu_iova_insert(iova_copy, n->start, n->end);
2124 vfio_iommu_iova_free(iova_copy);
2128 static void vfio_iommu_iova_insert_copy(struct vfio_iommu *iommu,
2129 struct list_head *iova_copy)
2131 struct list_head *iova = &iommu->iova_list;
2133 vfio_iommu_iova_free(iova);
2135 list_splice_tail(iova_copy, iova);
2138 /* Redundantly walks non-present capabilities to simplify caller */
2139 static int vfio_iommu_device_capable(struct device *dev, void *data)
2141 return device_iommu_capable(dev, (enum iommu_cap)data);
2144 static int vfio_iommu_domain_alloc(struct device *dev, void *data)
2146 struct iommu_domain **domain = data;
2148 *domain = iommu_domain_alloc(dev->bus);
2149 return 1; /* Don't iterate */
2152 static int vfio_iommu_type1_attach_group(void *iommu_data,
2153 struct iommu_group *iommu_group, enum vfio_group_type type)
2155 struct vfio_iommu *iommu = iommu_data;
2156 struct vfio_iommu_group *group;
2157 struct vfio_domain *domain, *d;
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);
2165 mutex_lock(&iommu->lock);
2167 /* Check for duplicates */
2168 if (vfio_iommu_find_iommu_group(iommu, iommu_group))
2172 group = kzalloc(sizeof(*group), GFP_KERNEL);
2175 group->iommu_group = iommu_group;
2177 if (type == VFIO_EMULATED_IOMMU) {
2178 list_add(&group->next, &iommu->emulated_iommu_groups);
2180 * An emulated IOMMU group cannot dirty memory directly, it can
2181 * only use interfaces that provide dirty tracking.
2182 * The iommu scope can only be promoted with the addition of a
2183 * dirty tracking group.
2185 group->pinned_page_dirty_scope = true;
2191 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2193 goto out_free_group;
2196 * Going via the iommu_group iterator avoids races, and trivially gives
2197 * us a representative device for the IOMMU API call. We don't actually
2198 * want to iterate beyond the first device (if any).
2201 iommu_group_for_each_dev(iommu_group, &domain->domain,
2202 vfio_iommu_domain_alloc);
2203 if (!domain->domain)
2204 goto out_free_domain;
2206 if (iommu->nesting) {
2207 ret = iommu_enable_nesting(domain->domain);
2212 ret = iommu_attach_group(domain->domain, group->iommu_group);
2216 /* Get aperture info */
2217 geo = &domain->domain->geometry;
2218 if (vfio_iommu_aper_conflict(iommu, geo->aperture_start,
2219 geo->aperture_end)) {
2224 ret = iommu_get_group_resv_regions(iommu_group, &group_resv_regions);
2228 if (vfio_iommu_resv_conflict(iommu, &group_resv_regions)) {
2234 * We don't want to work on the original iova list as the list
2235 * gets modified and in case of failure we have to retain the
2236 * original list. Get a copy here.
2238 ret = vfio_iommu_iova_get_copy(iommu, &iova_copy);
2242 ret = vfio_iommu_aper_resize(&iova_copy, geo->aperture_start,
2247 ret = vfio_iommu_resv_exclude(&iova_copy, &group_resv_regions);
2251 resv_msi = vfio_iommu_has_sw_msi(&group_resv_regions, &resv_msi_base);
2253 INIT_LIST_HEAD(&domain->group_list);
2254 list_add(&group->next, &domain->group_list);
2256 msi_remap = irq_domain_check_msi_remap() ||
2257 iommu_group_for_each_dev(iommu_group, (void *)IOMMU_CAP_INTR_REMAP,
2258 vfio_iommu_device_capable);
2260 if (!allow_unsafe_interrupts && !msi_remap) {
2261 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
2268 * If the IOMMU can block non-coherent operations (ie PCIe TLPs with
2269 * no-snoop set) then VFIO always turns this feature on because on Intel
2270 * platforms it optimizes KVM to disable wbinvd emulation.
2272 if (domain->domain->ops->enforce_cache_coherency)
2273 domain->enforce_cache_coherency =
2274 domain->domain->ops->enforce_cache_coherency(
2278 * Try to match an existing compatible domain. We don't want to
2279 * preclude an IOMMU driver supporting multiple bus_types and being
2280 * able to include different bus_types in the same IOMMU domain, so
2281 * we test whether the domains use the same iommu_ops rather than
2282 * testing if they're on the same bus_type.
2284 list_for_each_entry(d, &iommu->domain_list, next) {
2285 if (d->domain->ops == domain->domain->ops &&
2286 d->enforce_cache_coherency ==
2287 domain->enforce_cache_coherency) {
2288 iommu_detach_group(domain->domain, group->iommu_group);
2289 if (!iommu_attach_group(d->domain,
2290 group->iommu_group)) {
2291 list_add(&group->next, &d->group_list);
2292 iommu_domain_free(domain->domain);
2297 ret = iommu_attach_group(domain->domain,
2298 group->iommu_group);
2304 vfio_test_domain_fgsp(domain);
2306 /* replay mappings on new domains */
2307 ret = vfio_iommu_replay(iommu, domain);
2312 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
2313 if (ret && ret != -ENODEV)
2317 list_add(&domain->next, &iommu->domain_list);
2318 vfio_update_pgsize_bitmap(iommu);
2320 /* Delete the old one and insert new iova list */
2321 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2324 * An iommu backed group can dirty memory directly and therefore
2325 * demotes the iommu scope until it declares itself dirty tracking
2326 * capable via the page pinning interface.
2328 iommu->num_non_pinned_groups++;
2329 mutex_unlock(&iommu->lock);
2330 vfio_iommu_resv_free(&group_resv_regions);
2335 iommu_detach_group(domain->domain, group->iommu_group);
2337 iommu_domain_free(domain->domain);
2338 vfio_iommu_iova_free(&iova_copy);
2339 vfio_iommu_resv_free(&group_resv_regions);
2345 mutex_unlock(&iommu->lock);
2349 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
2351 struct rb_node *node;
2353 while ((node = rb_first(&iommu->dma_list)))
2354 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
2357 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
2359 struct rb_node *n, *p;
2361 n = rb_first(&iommu->dma_list);
2362 for (; n; n = rb_next(n)) {
2363 struct vfio_dma *dma;
2364 long locked = 0, unlocked = 0;
2366 dma = rb_entry(n, struct vfio_dma, node);
2367 unlocked += vfio_unmap_unpin(iommu, dma, false);
2368 p = rb_first(&dma->pfn_list);
2369 for (; p; p = rb_next(p)) {
2370 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
2373 if (!is_invalid_reserved_pfn(vpfn->pfn))
2376 vfio_lock_acct(dma, locked - unlocked, true);
2381 * Called when a domain is removed in detach. It is possible that
2382 * the removed domain decided the iova aperture window. Modify the
2383 * iova aperture with the smallest window among existing domains.
2385 static void vfio_iommu_aper_expand(struct vfio_iommu *iommu,
2386 struct list_head *iova_copy)
2388 struct vfio_domain *domain;
2389 struct vfio_iova *node;
2390 dma_addr_t start = 0;
2391 dma_addr_t end = (dma_addr_t)~0;
2393 if (list_empty(iova_copy))
2396 list_for_each_entry(domain, &iommu->domain_list, next) {
2397 struct iommu_domain_geometry *geo = &domain->domain->geometry;
2399 if (geo->aperture_start > start)
2400 start = geo->aperture_start;
2401 if (geo->aperture_end < end)
2402 end = geo->aperture_end;
2405 /* Modify aperture limits. The new aper is either same or bigger */
2406 node = list_first_entry(iova_copy, struct vfio_iova, list);
2407 node->start = start;
2408 node = list_last_entry(iova_copy, struct vfio_iova, list);
2413 * Called when a group is detached. The reserved regions for that
2414 * group can be part of valid iova now. But since reserved regions
2415 * may be duplicated among groups, populate the iova valid regions
2418 static int vfio_iommu_resv_refresh(struct vfio_iommu *iommu,
2419 struct list_head *iova_copy)
2421 struct vfio_domain *d;
2422 struct vfio_iommu_group *g;
2423 struct vfio_iova *node;
2424 dma_addr_t start, end;
2425 LIST_HEAD(resv_regions);
2428 if (list_empty(iova_copy))
2431 list_for_each_entry(d, &iommu->domain_list, next) {
2432 list_for_each_entry(g, &d->group_list, next) {
2433 ret = iommu_get_group_resv_regions(g->iommu_group,
2440 node = list_first_entry(iova_copy, struct vfio_iova, list);
2441 start = node->start;
2442 node = list_last_entry(iova_copy, struct vfio_iova, list);
2445 /* purge the iova list and create new one */
2446 vfio_iommu_iova_free(iova_copy);
2448 ret = vfio_iommu_aper_resize(iova_copy, start, end);
2452 /* Exclude current reserved regions from iova ranges */
2453 ret = vfio_iommu_resv_exclude(iova_copy, &resv_regions);
2455 vfio_iommu_resv_free(&resv_regions);
2459 static void vfio_iommu_type1_detach_group(void *iommu_data,
2460 struct iommu_group *iommu_group)
2462 struct vfio_iommu *iommu = iommu_data;
2463 struct vfio_domain *domain;
2464 struct vfio_iommu_group *group;
2465 bool update_dirty_scope = false;
2466 LIST_HEAD(iova_copy);
2468 mutex_lock(&iommu->lock);
2469 list_for_each_entry(group, &iommu->emulated_iommu_groups, next) {
2470 if (group->iommu_group != iommu_group)
2472 update_dirty_scope = !group->pinned_page_dirty_scope;
2473 list_del(&group->next);
2476 if (list_empty(&iommu->emulated_iommu_groups) &&
2477 list_empty(&iommu->domain_list)) {
2478 WARN_ON(iommu->notifier.head);
2479 vfio_iommu_unmap_unpin_all(iommu);
2481 goto detach_group_done;
2485 * Get a copy of iova list. This will be used to update
2486 * and to replace the current one later. Please note that
2487 * we will leave the original list as it is if update fails.
2489 vfio_iommu_iova_get_copy(iommu, &iova_copy);
2491 list_for_each_entry(domain, &iommu->domain_list, next) {
2492 group = find_iommu_group(domain, iommu_group);
2496 iommu_detach_group(domain->domain, group->iommu_group);
2497 update_dirty_scope = !group->pinned_page_dirty_scope;
2498 list_del(&group->next);
2501 * Group ownership provides privilege, if the group list is
2502 * empty, the domain goes away. If it's the last domain with
2503 * iommu and external domain doesn't exist, then all the
2504 * mappings go away too. If it's the last domain with iommu and
2505 * external domain exist, update accounting
2507 if (list_empty(&domain->group_list)) {
2508 if (list_is_singular(&iommu->domain_list)) {
2509 if (list_empty(&iommu->emulated_iommu_groups)) {
2510 WARN_ON(iommu->notifier.head);
2511 vfio_iommu_unmap_unpin_all(iommu);
2513 vfio_iommu_unmap_unpin_reaccount(iommu);
2516 iommu_domain_free(domain->domain);
2517 list_del(&domain->next);
2519 vfio_iommu_aper_expand(iommu, &iova_copy);
2520 vfio_update_pgsize_bitmap(iommu);
2525 if (!vfio_iommu_resv_refresh(iommu, &iova_copy))
2526 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2528 vfio_iommu_iova_free(&iova_copy);
2532 * Removal of a group without dirty tracking may allow the iommu scope
2535 if (update_dirty_scope) {
2536 iommu->num_non_pinned_groups--;
2537 if (iommu->dirty_page_tracking)
2538 vfio_iommu_populate_bitmap_full(iommu);
2540 mutex_unlock(&iommu->lock);
2543 static void *vfio_iommu_type1_open(unsigned long arg)
2545 struct vfio_iommu *iommu;
2547 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
2549 return ERR_PTR(-ENOMEM);
2552 case VFIO_TYPE1_IOMMU:
2554 case VFIO_TYPE1_NESTING_IOMMU:
2555 iommu->nesting = true;
2557 case VFIO_TYPE1v2_IOMMU:
2562 return ERR_PTR(-EINVAL);
2565 INIT_LIST_HEAD(&iommu->domain_list);
2566 INIT_LIST_HEAD(&iommu->iova_list);
2567 iommu->dma_list = RB_ROOT;
2568 iommu->dma_avail = dma_entry_limit;
2569 iommu->container_open = true;
2570 mutex_init(&iommu->lock);
2571 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
2572 init_waitqueue_head(&iommu->vaddr_wait);
2573 iommu->pgsize_bitmap = PAGE_MASK;
2574 INIT_LIST_HEAD(&iommu->emulated_iommu_groups);
2579 static void vfio_release_domain(struct vfio_domain *domain)
2581 struct vfio_iommu_group *group, *group_tmp;
2583 list_for_each_entry_safe(group, group_tmp,
2584 &domain->group_list, next) {
2585 iommu_detach_group(domain->domain, group->iommu_group);
2586 list_del(&group->next);
2590 iommu_domain_free(domain->domain);
2593 static void vfio_iommu_type1_release(void *iommu_data)
2595 struct vfio_iommu *iommu = iommu_data;
2596 struct vfio_domain *domain, *domain_tmp;
2597 struct vfio_iommu_group *group, *next_group;
2599 list_for_each_entry_safe(group, next_group,
2600 &iommu->emulated_iommu_groups, next) {
2601 list_del(&group->next);
2605 vfio_iommu_unmap_unpin_all(iommu);
2607 list_for_each_entry_safe(domain, domain_tmp,
2608 &iommu->domain_list, next) {
2609 vfio_release_domain(domain);
2610 list_del(&domain->next);
2614 vfio_iommu_iova_free(&iommu->iova_list);
2619 static int vfio_domains_have_enforce_cache_coherency(struct vfio_iommu *iommu)
2621 struct vfio_domain *domain;
2624 mutex_lock(&iommu->lock);
2625 list_for_each_entry(domain, &iommu->domain_list, next) {
2626 if (!(domain->enforce_cache_coherency)) {
2631 mutex_unlock(&iommu->lock);
2636 static int vfio_iommu_type1_check_extension(struct vfio_iommu *iommu,
2640 case VFIO_TYPE1_IOMMU:
2641 case VFIO_TYPE1v2_IOMMU:
2642 case VFIO_TYPE1_NESTING_IOMMU:
2643 case VFIO_UNMAP_ALL:
2644 case VFIO_UPDATE_VADDR:
2646 case VFIO_DMA_CC_IOMMU:
2649 return vfio_domains_have_enforce_cache_coherency(iommu);
2655 static int vfio_iommu_iova_add_cap(struct vfio_info_cap *caps,
2656 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas,
2659 struct vfio_info_cap_header *header;
2660 struct vfio_iommu_type1_info_cap_iova_range *iova_cap;
2662 header = vfio_info_cap_add(caps, size,
2663 VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE, 1);
2665 return PTR_ERR(header);
2667 iova_cap = container_of(header,
2668 struct vfio_iommu_type1_info_cap_iova_range,
2670 iova_cap->nr_iovas = cap_iovas->nr_iovas;
2671 memcpy(iova_cap->iova_ranges, cap_iovas->iova_ranges,
2672 cap_iovas->nr_iovas * sizeof(*cap_iovas->iova_ranges));
2676 static int vfio_iommu_iova_build_caps(struct vfio_iommu *iommu,
2677 struct vfio_info_cap *caps)
2679 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas;
2680 struct vfio_iova *iova;
2682 int iovas = 0, i = 0, ret;
2684 list_for_each_entry(iova, &iommu->iova_list, list)
2689 * Return 0 as a container with a single mdev device
2690 * will have an empty list
2695 size = struct_size(cap_iovas, iova_ranges, iovas);
2697 cap_iovas = kzalloc(size, GFP_KERNEL);
2701 cap_iovas->nr_iovas = iovas;
2703 list_for_each_entry(iova, &iommu->iova_list, list) {
2704 cap_iovas->iova_ranges[i].start = iova->start;
2705 cap_iovas->iova_ranges[i].end = iova->end;
2709 ret = vfio_iommu_iova_add_cap(caps, cap_iovas, size);
2715 static int vfio_iommu_migration_build_caps(struct vfio_iommu *iommu,
2716 struct vfio_info_cap *caps)
2718 struct vfio_iommu_type1_info_cap_migration cap_mig;
2720 cap_mig.header.id = VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION;
2721 cap_mig.header.version = 1;
2724 /* support minimum pgsize */
2725 cap_mig.pgsize_bitmap = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2726 cap_mig.max_dirty_bitmap_size = DIRTY_BITMAP_SIZE_MAX;
2728 return vfio_info_add_capability(caps, &cap_mig.header, sizeof(cap_mig));
2731 static int vfio_iommu_dma_avail_build_caps(struct vfio_iommu *iommu,
2732 struct vfio_info_cap *caps)
2734 struct vfio_iommu_type1_info_dma_avail cap_dma_avail;
2736 cap_dma_avail.header.id = VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL;
2737 cap_dma_avail.header.version = 1;
2739 cap_dma_avail.avail = iommu->dma_avail;
2741 return vfio_info_add_capability(caps, &cap_dma_avail.header,
2742 sizeof(cap_dma_avail));
2745 static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu,
2748 struct vfio_iommu_type1_info info;
2749 unsigned long minsz;
2750 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
2751 unsigned long capsz;
2754 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
2756 /* For backward compatibility, cannot require this */
2757 capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
2759 if (copy_from_user(&info, (void __user *)arg, minsz))
2762 if (info.argsz < minsz)
2765 if (info.argsz >= capsz) {
2767 info.cap_offset = 0; /* output, no-recopy necessary */
2770 mutex_lock(&iommu->lock);
2771 info.flags = VFIO_IOMMU_INFO_PGSIZES;
2773 info.iova_pgsizes = iommu->pgsize_bitmap;
2775 ret = vfio_iommu_migration_build_caps(iommu, &caps);
2778 ret = vfio_iommu_dma_avail_build_caps(iommu, &caps);
2781 ret = vfio_iommu_iova_build_caps(iommu, &caps);
2783 mutex_unlock(&iommu->lock);
2789 info.flags |= VFIO_IOMMU_INFO_CAPS;
2791 if (info.argsz < sizeof(info) + caps.size) {
2792 info.argsz = sizeof(info) + caps.size;
2794 vfio_info_cap_shift(&caps, sizeof(info));
2795 if (copy_to_user((void __user *)arg +
2796 sizeof(info), caps.buf,
2801 info.cap_offset = sizeof(info);
2807 return copy_to_user((void __user *)arg, &info, minsz) ?
2811 static int vfio_iommu_type1_map_dma(struct vfio_iommu *iommu,
2814 struct vfio_iommu_type1_dma_map map;
2815 unsigned long minsz;
2816 uint32_t mask = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE |
2817 VFIO_DMA_MAP_FLAG_VADDR;
2819 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
2821 if (copy_from_user(&map, (void __user *)arg, minsz))
2824 if (map.argsz < minsz || map.flags & ~mask)
2827 return vfio_dma_do_map(iommu, &map);
2830 static int vfio_iommu_type1_unmap_dma(struct vfio_iommu *iommu,
2833 struct vfio_iommu_type1_dma_unmap unmap;
2834 struct vfio_bitmap bitmap = { 0 };
2835 uint32_t mask = VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP |
2836 VFIO_DMA_UNMAP_FLAG_VADDR |
2837 VFIO_DMA_UNMAP_FLAG_ALL;
2838 unsigned long minsz;
2841 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
2843 if (copy_from_user(&unmap, (void __user *)arg, minsz))
2846 if (unmap.argsz < minsz || unmap.flags & ~mask)
2849 if ((unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
2850 (unmap.flags & (VFIO_DMA_UNMAP_FLAG_ALL |
2851 VFIO_DMA_UNMAP_FLAG_VADDR)))
2854 if (unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
2855 unsigned long pgshift;
2857 if (unmap.argsz < (minsz + sizeof(bitmap)))
2860 if (copy_from_user(&bitmap,
2861 (void __user *)(arg + minsz),
2865 if (!access_ok((void __user *)bitmap.data, bitmap.size))
2868 pgshift = __ffs(bitmap.pgsize);
2869 ret = verify_bitmap_size(unmap.size >> pgshift,
2875 ret = vfio_dma_do_unmap(iommu, &unmap, &bitmap);
2879 return copy_to_user((void __user *)arg, &unmap, minsz) ?
2883 static int vfio_iommu_type1_dirty_pages(struct vfio_iommu *iommu,
2886 struct vfio_iommu_type1_dirty_bitmap dirty;
2887 uint32_t mask = VFIO_IOMMU_DIRTY_PAGES_FLAG_START |
2888 VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP |
2889 VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP;
2890 unsigned long minsz;
2896 minsz = offsetofend(struct vfio_iommu_type1_dirty_bitmap, flags);
2898 if (copy_from_user(&dirty, (void __user *)arg, minsz))
2901 if (dirty.argsz < minsz || dirty.flags & ~mask)
2904 /* only one flag should be set at a time */
2905 if (__ffs(dirty.flags) != __fls(dirty.flags))
2908 if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_START) {
2911 mutex_lock(&iommu->lock);
2912 pgsize = 1 << __ffs(iommu->pgsize_bitmap);
2913 if (!iommu->dirty_page_tracking) {
2914 ret = vfio_dma_bitmap_alloc_all(iommu, pgsize);
2916 iommu->dirty_page_tracking = true;
2918 mutex_unlock(&iommu->lock);
2920 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP) {
2921 mutex_lock(&iommu->lock);
2922 if (iommu->dirty_page_tracking) {
2923 iommu->dirty_page_tracking = false;
2924 vfio_dma_bitmap_free_all(iommu);
2926 mutex_unlock(&iommu->lock);
2928 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP) {
2929 struct vfio_iommu_type1_dirty_bitmap_get range;
2930 unsigned long pgshift;
2931 size_t data_size = dirty.argsz - minsz;
2932 size_t iommu_pgsize;
2934 if (!data_size || data_size < sizeof(range))
2937 if (copy_from_user(&range, (void __user *)(arg + minsz),
2941 if (range.iova + range.size < range.iova)
2943 if (!access_ok((void __user *)range.bitmap.data,
2947 pgshift = __ffs(range.bitmap.pgsize);
2948 ret = verify_bitmap_size(range.size >> pgshift,
2953 mutex_lock(&iommu->lock);
2955 iommu_pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2957 /* allow only smallest supported pgsize */
2958 if (range.bitmap.pgsize != iommu_pgsize) {
2962 if (range.iova & (iommu_pgsize - 1)) {
2966 if (!range.size || range.size & (iommu_pgsize - 1)) {
2971 if (iommu->dirty_page_tracking)
2972 ret = vfio_iova_dirty_bitmap(range.bitmap.data,
2975 range.bitmap.pgsize);
2979 mutex_unlock(&iommu->lock);
2987 static long vfio_iommu_type1_ioctl(void *iommu_data,
2988 unsigned int cmd, unsigned long arg)
2990 struct vfio_iommu *iommu = iommu_data;
2993 case VFIO_CHECK_EXTENSION:
2994 return vfio_iommu_type1_check_extension(iommu, arg);
2995 case VFIO_IOMMU_GET_INFO:
2996 return vfio_iommu_type1_get_info(iommu, arg);
2997 case VFIO_IOMMU_MAP_DMA:
2998 return vfio_iommu_type1_map_dma(iommu, arg);
2999 case VFIO_IOMMU_UNMAP_DMA:
3000 return vfio_iommu_type1_unmap_dma(iommu, arg);
3001 case VFIO_IOMMU_DIRTY_PAGES:
3002 return vfio_iommu_type1_dirty_pages(iommu, arg);
3008 static int vfio_iommu_type1_register_notifier(void *iommu_data,
3009 unsigned long *events,
3010 struct notifier_block *nb)
3012 struct vfio_iommu *iommu = iommu_data;
3014 /* clear known events */
3015 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
3017 /* refuse to register if still events remaining */
3021 return blocking_notifier_chain_register(&iommu->notifier, nb);
3024 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
3025 struct notifier_block *nb)
3027 struct vfio_iommu *iommu = iommu_data;
3029 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
3032 static int vfio_iommu_type1_dma_rw_chunk(struct vfio_iommu *iommu,
3033 dma_addr_t user_iova, void *data,
3034 size_t count, bool write,
3037 struct mm_struct *mm;
3038 unsigned long vaddr;
3039 struct vfio_dma *dma;
3040 bool kthread = current->mm == NULL;
3046 ret = vfio_find_dma_valid(iommu, user_iova, 1, &dma);
3050 if ((write && !(dma->prot & IOMMU_WRITE)) ||
3051 !(dma->prot & IOMMU_READ))
3054 mm = get_task_mm(dma->task);
3061 else if (current->mm != mm)
3064 offset = user_iova - dma->iova;
3066 if (count > dma->size - offset)
3067 count = dma->size - offset;
3069 vaddr = dma->vaddr + offset;
3072 *copied = copy_to_user((void __user *)vaddr, data,
3074 if (*copied && iommu->dirty_page_tracking) {
3075 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
3077 * Bitmap populated with the smallest supported page
3080 bitmap_set(dma->bitmap, offset >> pgshift,
3081 ((offset + *copied - 1) >> pgshift) -
3082 (offset >> pgshift) + 1);
3085 *copied = copy_from_user(data, (void __user *)vaddr,
3088 kthread_unuse_mm(mm);
3091 return *copied ? 0 : -EFAULT;
3094 static int vfio_iommu_type1_dma_rw(void *iommu_data, dma_addr_t user_iova,
3095 void *data, size_t count, bool write)
3097 struct vfio_iommu *iommu = iommu_data;
3101 mutex_lock(&iommu->lock);
3103 ret = vfio_iommu_type1_dma_rw_chunk(iommu, user_iova, data,
3104 count, write, &done);
3113 mutex_unlock(&iommu->lock);
3117 static struct iommu_domain *
3118 vfio_iommu_type1_group_iommu_domain(void *iommu_data,
3119 struct iommu_group *iommu_group)
3121 struct iommu_domain *domain = ERR_PTR(-ENODEV);
3122 struct vfio_iommu *iommu = iommu_data;
3123 struct vfio_domain *d;
3125 if (!iommu || !iommu_group)
3126 return ERR_PTR(-EINVAL);
3128 mutex_lock(&iommu->lock);
3129 list_for_each_entry(d, &iommu->domain_list, next) {
3130 if (find_iommu_group(d, iommu_group)) {
3135 mutex_unlock(&iommu->lock);
3140 static void vfio_iommu_type1_notify(void *iommu_data,
3141 enum vfio_iommu_notify_type event)
3143 struct vfio_iommu *iommu = iommu_data;
3145 if (event != VFIO_IOMMU_CONTAINER_CLOSE)
3147 mutex_lock(&iommu->lock);
3148 iommu->container_open = false;
3149 mutex_unlock(&iommu->lock);
3150 wake_up_all(&iommu->vaddr_wait);
3153 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
3154 .name = "vfio-iommu-type1",
3155 .owner = THIS_MODULE,
3156 .open = vfio_iommu_type1_open,
3157 .release = vfio_iommu_type1_release,
3158 .ioctl = vfio_iommu_type1_ioctl,
3159 .attach_group = vfio_iommu_type1_attach_group,
3160 .detach_group = vfio_iommu_type1_detach_group,
3161 .pin_pages = vfio_iommu_type1_pin_pages,
3162 .unpin_pages = vfio_iommu_type1_unpin_pages,
3163 .register_notifier = vfio_iommu_type1_register_notifier,
3164 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
3165 .dma_rw = vfio_iommu_type1_dma_rw,
3166 .group_iommu_domain = vfio_iommu_type1_group_iommu_domain,
3167 .notify = vfio_iommu_type1_notify,
3170 static int __init vfio_iommu_type1_init(void)
3172 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
3175 static void __exit vfio_iommu_type1_cleanup(void)
3177 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
3180 module_init(vfio_iommu_type1_init);
3181 module_exit(vfio_iommu_type1_cleanup);
3183 MODULE_VERSION(DRIVER_VERSION);
3184 MODULE_LICENSE("GPL v2");
3185 MODULE_AUTHOR(DRIVER_AUTHOR);
3186 MODULE_DESCRIPTION(DRIVER_DESC);