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 list_head device_list;
71 struct mutex device_list_lock;
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;
81 struct list_head emulated_iommu_groups;
85 struct iommu_domain *domain;
86 struct list_head next;
87 struct list_head group_list;
88 bool fgsp : 1; /* Fine-grained super pages */
89 bool enforce_cache_coherency : 1;
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 page **pages; /* for pin_user_pages_remote */
108 struct page *fallback_page; /* if pages alloc fails */
109 int capacity; /* length of pages array */
110 int size; /* of batch currently */
111 int offset; /* of next entry in pages */
114 struct vfio_iommu_group {
115 struct iommu_group *iommu_group;
116 struct list_head next;
117 bool pinned_page_dirty_scope;
121 struct list_head list;
127 * Guest RAM pinning working set or DMA target
131 dma_addr_t iova; /* Device address */
132 unsigned long pfn; /* Host pfn */
133 unsigned int ref_count;
136 struct vfio_regions {
137 struct list_head list;
143 #define DIRTY_BITMAP_BYTES(n) (ALIGN(n, BITS_PER_TYPE(u64)) / BITS_PER_BYTE)
146 * Input argument of number of bits to bitmap_set() is unsigned integer, which
147 * further casts to signed integer for unaligned multi-bit operation,
149 * Then maximum bitmap size supported is 2^31 bits divided by 2^3 bits/byte,
150 * that is 2^28 (256 MB) which maps to 2^31 * 2^12 = 2^43 (8TB) on 4K page
153 #define DIRTY_BITMAP_PAGES_MAX ((u64)INT_MAX)
154 #define DIRTY_BITMAP_SIZE_MAX DIRTY_BITMAP_BYTES(DIRTY_BITMAP_PAGES_MAX)
158 static int put_pfn(unsigned long pfn, int prot);
160 static struct vfio_iommu_group*
161 vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
162 struct iommu_group *iommu_group);
165 * This code handles mapping and unmapping of user data buffers
166 * into DMA'ble space using the IOMMU
169 static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
170 dma_addr_t start, size_t size)
172 struct rb_node *node = iommu->dma_list.rb_node;
175 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
177 if (start + size <= dma->iova)
178 node = node->rb_left;
179 else if (start >= dma->iova + dma->size)
180 node = node->rb_right;
188 static struct rb_node *vfio_find_dma_first_node(struct vfio_iommu *iommu,
189 dma_addr_t start, u64 size)
191 struct rb_node *res = NULL;
192 struct rb_node *node = iommu->dma_list.rb_node;
193 struct vfio_dma *dma_res = NULL;
196 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
198 if (start < dma->iova + dma->size) {
201 if (start >= dma->iova)
203 node = node->rb_left;
205 node = node->rb_right;
208 if (res && size && dma_res->iova >= start + size)
213 static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new)
215 struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL;
216 struct vfio_dma *dma;
220 dma = rb_entry(parent, struct vfio_dma, node);
222 if (new->iova + new->size <= dma->iova)
223 link = &(*link)->rb_left;
225 link = &(*link)->rb_right;
228 rb_link_node(&new->node, parent, link);
229 rb_insert_color(&new->node, &iommu->dma_list);
232 static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
234 rb_erase(&old->node, &iommu->dma_list);
238 static int vfio_dma_bitmap_alloc(struct vfio_dma *dma, size_t pgsize)
240 uint64_t npages = dma->size / pgsize;
242 if (npages > DIRTY_BITMAP_PAGES_MAX)
246 * Allocate extra 64 bits that are used to calculate shift required for
247 * bitmap_shift_left() to manipulate and club unaligned number of pages
248 * in adjacent vfio_dma ranges.
250 dma->bitmap = kvzalloc(DIRTY_BITMAP_BYTES(npages) + sizeof(u64),
258 static void vfio_dma_bitmap_free(struct vfio_dma *dma)
264 static void vfio_dma_populate_bitmap(struct vfio_dma *dma, size_t pgsize)
267 unsigned long pgshift = __ffs(pgsize);
269 for (p = rb_first(&dma->pfn_list); p; p = rb_next(p)) {
270 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn, node);
272 bitmap_set(dma->bitmap, (vpfn->iova - dma->iova) >> pgshift, 1);
276 static void vfio_iommu_populate_bitmap_full(struct vfio_iommu *iommu)
279 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
281 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
282 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
284 bitmap_set(dma->bitmap, 0, dma->size >> pgshift);
288 static int vfio_dma_bitmap_alloc_all(struct vfio_iommu *iommu, size_t pgsize)
292 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
293 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
296 ret = vfio_dma_bitmap_alloc(dma, pgsize);
300 for (p = rb_prev(n); p; p = rb_prev(p)) {
301 struct vfio_dma *dma = rb_entry(n,
302 struct vfio_dma, node);
304 vfio_dma_bitmap_free(dma);
308 vfio_dma_populate_bitmap(dma, pgsize);
313 static void vfio_dma_bitmap_free_all(struct vfio_iommu *iommu)
317 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
318 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
320 vfio_dma_bitmap_free(dma);
325 * Helper Functions for host iova-pfn list
327 static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
329 struct vfio_pfn *vpfn;
330 struct rb_node *node = dma->pfn_list.rb_node;
333 vpfn = rb_entry(node, struct vfio_pfn, node);
335 if (iova < vpfn->iova)
336 node = node->rb_left;
337 else if (iova > vpfn->iova)
338 node = node->rb_right;
345 static void vfio_link_pfn(struct vfio_dma *dma,
346 struct vfio_pfn *new)
348 struct rb_node **link, *parent = NULL;
349 struct vfio_pfn *vpfn;
351 link = &dma->pfn_list.rb_node;
354 vpfn = rb_entry(parent, struct vfio_pfn, node);
356 if (new->iova < vpfn->iova)
357 link = &(*link)->rb_left;
359 link = &(*link)->rb_right;
362 rb_link_node(&new->node, parent, link);
363 rb_insert_color(&new->node, &dma->pfn_list);
366 static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old)
368 rb_erase(&old->node, &dma->pfn_list);
371 static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova,
374 struct vfio_pfn *vpfn;
376 vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL);
383 vfio_link_pfn(dma, vpfn);
387 static void vfio_remove_from_pfn_list(struct vfio_dma *dma,
388 struct vfio_pfn *vpfn)
390 vfio_unlink_pfn(dma, vpfn);
394 static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma,
397 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
404 static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn)
409 if (!vpfn->ref_count) {
410 ret = put_pfn(vpfn->pfn, dma->prot);
411 vfio_remove_from_pfn_list(dma, vpfn);
416 static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
418 struct mm_struct *mm;
424 mm = async ? get_task_mm(dma->task) : dma->task->mm;
426 return -ESRCH; /* process exited */
428 ret = mmap_write_lock_killable(mm);
430 ret = __account_locked_vm(mm, abs(npage), npage > 0, dma->task,
432 mmap_write_unlock(mm);
442 * Some mappings aren't backed by a struct page, for example an mmap'd
443 * MMIO range for our own or another device. These use a different
444 * pfn conversion and shouldn't be tracked as locked pages.
445 * For compound pages, any driver that sets the reserved bit in head
446 * page needs to set the reserved bit in all subpages to be safe.
448 static bool is_invalid_reserved_pfn(unsigned long pfn)
451 return PageReserved(pfn_to_page(pfn));
456 static int put_pfn(unsigned long pfn, int prot)
458 if (!is_invalid_reserved_pfn(pfn)) {
459 struct page *page = pfn_to_page(pfn);
461 unpin_user_pages_dirty_lock(&page, 1, prot & IOMMU_WRITE);
467 #define VFIO_BATCH_MAX_CAPACITY (PAGE_SIZE / sizeof(struct page *))
469 static void vfio_batch_init(struct vfio_batch *batch)
474 if (unlikely(disable_hugepages))
477 batch->pages = (struct page **) __get_free_page(GFP_KERNEL);
481 batch->capacity = VFIO_BATCH_MAX_CAPACITY;
485 batch->pages = &batch->fallback_page;
489 static void vfio_batch_unpin(struct vfio_batch *batch, struct vfio_dma *dma)
491 while (batch->size) {
492 unsigned long pfn = page_to_pfn(batch->pages[batch->offset]);
494 put_pfn(pfn, dma->prot);
500 static void vfio_batch_fini(struct vfio_batch *batch)
502 if (batch->capacity == VFIO_BATCH_MAX_CAPACITY)
503 free_page((unsigned long)batch->pages);
506 static int follow_fault_pfn(struct vm_area_struct *vma, struct mm_struct *mm,
507 unsigned long vaddr, unsigned long *pfn,
514 ret = follow_pte(vma->vm_mm, vaddr, &ptep, &ptl);
516 bool unlocked = false;
518 ret = fixup_user_fault(mm, vaddr,
520 (write_fault ? FAULT_FLAG_WRITE : 0),
528 ret = follow_pte(vma->vm_mm, vaddr, &ptep, &ptl);
533 if (write_fault && !pte_write(*ptep))
536 *pfn = pte_pfn(*ptep);
538 pte_unmap_unlock(ptep, ptl);
543 * Returns the positive number of pfns successfully obtained or a negative
546 static int vaddr_get_pfns(struct mm_struct *mm, unsigned long vaddr,
547 long npages, int prot, unsigned long *pfn,
550 struct vm_area_struct *vma;
551 unsigned int flags = 0;
554 if (prot & IOMMU_WRITE)
558 ret = pin_user_pages_remote(mm, vaddr, npages, flags | FOLL_LONGTERM,
561 *pfn = page_to_pfn(pages[0]);
565 vaddr = untagged_addr(vaddr);
568 vma = vma_lookup(mm, vaddr);
570 if (vma && vma->vm_flags & VM_PFNMAP) {
571 ret = follow_fault_pfn(vma, mm, vaddr, pfn, prot & IOMMU_WRITE);
576 if (is_invalid_reserved_pfn(*pfn))
583 mmap_read_unlock(mm);
587 static int vfio_wait(struct vfio_iommu *iommu)
591 prepare_to_wait(&iommu->vaddr_wait, &wait, TASK_KILLABLE);
592 mutex_unlock(&iommu->lock);
594 mutex_lock(&iommu->lock);
595 finish_wait(&iommu->vaddr_wait, &wait);
596 if (kthread_should_stop() || !iommu->container_open ||
597 fatal_signal_pending(current)) {
604 * Find dma struct and wait for its vaddr to be valid. iommu lock is dropped
605 * if the task waits, but is re-locked on return. Return result in *dma_p.
606 * Return 0 on success with no waiting, WAITED on success if waited, and -errno
609 static int vfio_find_dma_valid(struct vfio_iommu *iommu, dma_addr_t start,
610 size_t size, struct vfio_dma **dma_p)
615 *dma_p = vfio_find_dma(iommu, start, size);
618 else if (!(*dma_p)->vaddr_invalid)
621 ret = vfio_wait(iommu);
622 } while (ret == WAITED);
628 * Wait for all vaddr in the dma_list to become valid. iommu lock is dropped
629 * if the task waits, but is re-locked on return. Return 0 on success with no
630 * waiting, WAITED on success if waited, and -errno on error.
632 static int vfio_wait_all_valid(struct vfio_iommu *iommu)
636 while (iommu->vaddr_invalid_count && ret >= 0)
637 ret = vfio_wait(iommu);
643 * Attempt to pin pages. We really don't want to track all the pfns and
644 * the iommu can only map chunks of consecutive pfns anyway, so get the
645 * first page and all consecutive pages with the same locking.
647 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
648 long npage, unsigned long *pfn_base,
649 unsigned long limit, struct vfio_batch *batch)
652 struct mm_struct *mm = current->mm;
653 long ret, pinned = 0, lock_acct = 0;
655 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
657 /* This code path is only user initiated */
662 /* Leftover pages in batch from an earlier call. */
663 *pfn_base = page_to_pfn(batch->pages[batch->offset]);
665 rsvd = is_invalid_reserved_pfn(*pfn_base);
672 /* Empty batch, so refill it. */
673 long req_pages = min_t(long, npage, batch->capacity);
675 ret = vaddr_get_pfns(mm, vaddr, req_pages, dma->prot,
685 rsvd = is_invalid_reserved_pfn(*pfn_base);
690 * pfn is preset for the first iteration of this inner loop and
691 * updated at the end to handle a VM_PFNMAP pfn. In that case,
692 * batch->pages isn't valid (there's no struct page), so allow
693 * batch->pages to be touched only when there's more than one
694 * pfn to check, which guarantees the pfns are from a
698 if (pfn != *pfn_base + pinned ||
699 rsvd != is_invalid_reserved_pfn(pfn))
703 * Reserved pages aren't counted against the user,
704 * externally pinned pages are already counted against
707 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
708 if (!dma->lock_cap &&
709 mm->locked_vm + lock_acct + 1 > limit) {
710 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
711 __func__, limit << PAGE_SHIFT);
728 pfn = page_to_pfn(batch->pages[batch->offset]);
731 if (unlikely(disable_hugepages))
736 ret = vfio_lock_acct(dma, lock_acct, false);
739 if (batch->size == 1 && !batch->offset) {
740 /* May be a VM_PFNMAP pfn, which the batch can't remember. */
741 put_pfn(pfn, dma->prot);
746 if (pinned && !rsvd) {
747 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
748 put_pfn(pfn, dma->prot);
750 vfio_batch_unpin(batch, dma);
758 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
759 unsigned long pfn, long npage,
762 long unlocked = 0, locked = 0;
765 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
766 if (put_pfn(pfn++, dma->prot)) {
768 if (vfio_find_vpfn(dma, iova))
774 vfio_lock_acct(dma, locked - unlocked, true);
779 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
780 unsigned long *pfn_base, bool do_accounting)
782 struct page *pages[1];
783 struct mm_struct *mm;
786 mm = get_task_mm(dma->task);
790 ret = vaddr_get_pfns(mm, vaddr, 1, dma->prot, pfn_base, pages);
796 if (do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
797 ret = vfio_lock_acct(dma, 1, true);
799 put_pfn(*pfn_base, dma->prot);
801 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
802 "(%ld) exceeded\n", __func__,
803 dma->task->comm, task_pid_nr(dma->task),
804 task_rlimit(dma->task, RLIMIT_MEMLOCK));
813 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
817 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
822 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
825 vfio_lock_acct(dma, -unlocked, true);
830 static int vfio_iommu_type1_pin_pages(void *iommu_data,
831 struct iommu_group *iommu_group,
832 dma_addr_t user_iova,
836 struct vfio_iommu *iommu = iommu_data;
837 struct vfio_iommu_group *group;
839 unsigned long remote_vaddr;
840 struct vfio_dma *dma;
844 if (!iommu || !pages)
847 /* Supported for v2 version only */
851 mutex_lock(&iommu->lock);
854 * Wait for all necessary vaddr's to be valid so they can be used in
855 * the main loop without dropping the lock, to avoid racing vs unmap.
858 if (iommu->vaddr_invalid_count) {
859 for (i = 0; i < npage; i++) {
860 iova = user_iova + PAGE_SIZE * i;
861 ret = vfio_find_dma_valid(iommu, iova, PAGE_SIZE, &dma);
869 /* Fail if no dma_umap notifier is registered */
870 if (list_empty(&iommu->device_list)) {
876 * If iommu capable domain exist in the container then all pages are
877 * already pinned and accounted. Accounting should be done if there is no
878 * iommu capable domain in the container.
880 do_accounting = list_empty(&iommu->domain_list);
882 for (i = 0; i < npage; i++) {
883 unsigned long phys_pfn;
884 struct vfio_pfn *vpfn;
886 iova = user_iova + PAGE_SIZE * i;
887 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
893 if ((dma->prot & prot) != prot) {
898 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
900 pages[i] = pfn_to_page(vpfn->pfn);
904 remote_vaddr = dma->vaddr + (iova - dma->iova);
905 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn,
910 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn);
912 if (put_pfn(phys_pfn, dma->prot) && do_accounting)
913 vfio_lock_acct(dma, -1, true);
917 pages[i] = pfn_to_page(phys_pfn);
919 if (iommu->dirty_page_tracking) {
920 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
923 * Bitmap populated with the smallest supported page
926 bitmap_set(dma->bitmap,
927 (iova - dma->iova) >> pgshift, 1);
932 group = vfio_iommu_find_iommu_group(iommu, iommu_group);
933 if (!group->pinned_page_dirty_scope) {
934 group->pinned_page_dirty_scope = true;
935 iommu->num_non_pinned_groups--;
942 for (j = 0; j < i; j++) {
945 iova = user_iova + PAGE_SIZE * j;
946 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
947 vfio_unpin_page_external(dma, iova, do_accounting);
951 mutex_unlock(&iommu->lock);
955 static void vfio_iommu_type1_unpin_pages(void *iommu_data,
956 dma_addr_t user_iova, int npage)
958 struct vfio_iommu *iommu = iommu_data;
962 /* Supported for v2 version only */
963 if (WARN_ON(!iommu->v2))
966 mutex_lock(&iommu->lock);
968 do_accounting = list_empty(&iommu->domain_list);
969 for (i = 0; i < npage; i++) {
970 dma_addr_t iova = user_iova + PAGE_SIZE * i;
971 struct vfio_dma *dma;
973 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
977 vfio_unpin_page_external(dma, iova, do_accounting);
980 mutex_unlock(&iommu->lock);
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)))
1291 * Notify VFIO drivers using vfio_register_emulated_iommu_dev() to invalidate
1292 * and unmap iovas within the range we're about to unmap. Drivers MUST unpin
1293 * pages in response to an invalidation.
1295 static void vfio_notify_dma_unmap(struct vfio_iommu *iommu,
1296 struct vfio_dma *dma)
1298 struct vfio_device *device;
1300 if (list_empty(&iommu->device_list))
1304 * The device is expected to call vfio_unpin_pages() for any IOVA it has
1305 * pinned within the range. Since vfio_unpin_pages() will eventually
1306 * call back down to this code and try to obtain the iommu->lock we must
1309 mutex_lock(&iommu->device_list_lock);
1310 mutex_unlock(&iommu->lock);
1312 list_for_each_entry(device, &iommu->device_list, iommu_entry)
1313 device->ops->dma_unmap(device, dma->iova, dma->size);
1315 mutex_unlock(&iommu->device_list_lock);
1316 mutex_lock(&iommu->lock);
1319 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
1320 struct vfio_iommu_type1_dma_unmap *unmap,
1321 struct vfio_bitmap *bitmap)
1323 struct vfio_dma *dma, *dma_last = NULL;
1324 size_t unmapped = 0, pgsize;
1325 int ret = -EINVAL, retries = 0;
1326 unsigned long pgshift;
1327 dma_addr_t iova = unmap->iova;
1328 u64 size = unmap->size;
1329 bool unmap_all = unmap->flags & VFIO_DMA_UNMAP_FLAG_ALL;
1330 bool invalidate_vaddr = unmap->flags & VFIO_DMA_UNMAP_FLAG_VADDR;
1331 struct rb_node *n, *first_n;
1333 mutex_lock(&iommu->lock);
1335 pgshift = __ffs(iommu->pgsize_bitmap);
1336 pgsize = (size_t)1 << pgshift;
1338 if (iova & (pgsize - 1))
1345 } else if (!size || size & (pgsize - 1) ||
1346 iova + size - 1 < iova || size > SIZE_MAX) {
1350 /* When dirty tracking is enabled, allow only min supported pgsize */
1351 if ((unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
1352 (!iommu->dirty_page_tracking || (bitmap->pgsize != pgsize))) {
1356 WARN_ON((pgsize - 1) & PAGE_MASK);
1359 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
1360 * avoid tracking individual mappings. This means that the granularity
1361 * of the original mapping was lost and the user was allowed to attempt
1362 * to unmap any range. Depending on the contiguousness of physical
1363 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
1364 * or may not have worked. We only guaranteed unmap granularity
1365 * matching the original mapping; even though it was untracked here,
1366 * the original mappings are reflected in IOMMU mappings. This
1367 * resulted in a couple unusual behaviors. First, if a range is not
1368 * able to be unmapped, ex. a set of 4k pages that was mapped as a
1369 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
1370 * a zero sized unmap. Also, if an unmap request overlaps the first
1371 * address of a hugepage, the IOMMU will unmap the entire hugepage.
1372 * This also returns success and the returned unmap size reflects the
1373 * actual size unmapped.
1375 * We attempt to maintain compatibility with this "v1" interface, but
1376 * we take control out of the hands of the IOMMU. Therefore, an unmap
1377 * request offset from the beginning of the original mapping will
1378 * return success with zero sized unmap. And an unmap request covering
1379 * the first iova of mapping will unmap the entire range.
1381 * The v2 version of this interface intends to be more deterministic.
1382 * Unmap requests must fully cover previous mappings. Multiple
1383 * mappings may still be unmaped by specifying large ranges, but there
1384 * must not be any previous mappings bisected by the range. An error
1385 * will be returned if these conditions are not met. The v2 interface
1386 * will only return success and a size of zero if there were no
1387 * mappings within the range.
1389 if (iommu->v2 && !unmap_all) {
1390 dma = vfio_find_dma(iommu, iova, 1);
1391 if (dma && dma->iova != iova)
1394 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1395 if (dma && dma->iova + dma->size != iova + size)
1400 n = first_n = vfio_find_dma_first_node(iommu, iova, size);
1403 dma = rb_entry(n, struct vfio_dma, node);
1404 if (dma->iova >= iova + size)
1407 if (!iommu->v2 && iova > dma->iova)
1410 if (invalidate_vaddr) {
1411 if (dma->vaddr_invalid) {
1412 struct rb_node *last_n = n;
1414 for (n = first_n; n != last_n; n = rb_next(n)) {
1416 struct vfio_dma, node);
1417 dma->vaddr_invalid = false;
1418 iommu->vaddr_invalid_count--;
1424 dma->vaddr_invalid = true;
1425 iommu->vaddr_invalid_count++;
1426 unmapped += dma->size;
1431 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
1432 if (dma_last == dma) {
1433 BUG_ON(++retries > 10);
1439 vfio_notify_dma_unmap(iommu, dma);
1443 if (unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
1444 ret = update_user_bitmap(bitmap->data, iommu, dma,
1450 unmapped += dma->size;
1452 vfio_remove_dma(iommu, dma);
1456 mutex_unlock(&iommu->lock);
1458 /* Report how much was unmapped */
1459 unmap->size = unmapped;
1464 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
1465 unsigned long pfn, long npage, int prot)
1467 struct vfio_domain *d;
1470 list_for_each_entry(d, &iommu->domain_list, next) {
1471 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
1472 npage << PAGE_SHIFT, prot | IOMMU_CACHE);
1482 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) {
1483 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
1490 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
1493 dma_addr_t iova = dma->iova;
1494 unsigned long vaddr = dma->vaddr;
1495 struct vfio_batch batch;
1496 size_t size = map_size;
1498 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1501 vfio_batch_init(&batch);
1504 /* Pin a contiguous chunk of memory */
1505 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
1506 size >> PAGE_SHIFT, &pfn, limit,
1515 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
1518 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
1520 vfio_batch_unpin(&batch, dma);
1524 size -= npage << PAGE_SHIFT;
1525 dma->size += npage << PAGE_SHIFT;
1528 vfio_batch_fini(&batch);
1529 dma->iommu_mapped = true;
1532 vfio_remove_dma(iommu, dma);
1538 * Check dma map request is within a valid iova range
1540 static bool vfio_iommu_iova_dma_valid(struct vfio_iommu *iommu,
1541 dma_addr_t start, dma_addr_t end)
1543 struct list_head *iova = &iommu->iova_list;
1544 struct vfio_iova *node;
1546 list_for_each_entry(node, iova, list) {
1547 if (start >= node->start && end <= node->end)
1552 * Check for list_empty() as well since a container with
1553 * a single mdev device will have an empty list.
1555 return list_empty(iova);
1558 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1559 struct vfio_iommu_type1_dma_map *map)
1561 bool set_vaddr = map->flags & VFIO_DMA_MAP_FLAG_VADDR;
1562 dma_addr_t iova = map->iova;
1563 unsigned long vaddr = map->vaddr;
1564 size_t size = map->size;
1565 int ret = 0, prot = 0;
1567 struct vfio_dma *dma;
1569 /* Verify that none of our __u64 fields overflow */
1570 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1573 /* READ/WRITE from device perspective */
1574 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1575 prot |= IOMMU_WRITE;
1576 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1579 if ((prot && set_vaddr) || (!prot && !set_vaddr))
1582 mutex_lock(&iommu->lock);
1584 pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
1586 WARN_ON((pgsize - 1) & PAGE_MASK);
1588 if (!size || (size | iova | vaddr) & (pgsize - 1)) {
1593 /* Don't allow IOVA or virtual address wrap */
1594 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) {
1599 dma = vfio_find_dma(iommu, iova, size);
1603 } else if (!dma->vaddr_invalid || dma->iova != iova ||
1604 dma->size != size) {
1608 dma->vaddr_invalid = false;
1609 iommu->vaddr_invalid_count--;
1610 wake_up_all(&iommu->vaddr_wait);
1618 if (!iommu->dma_avail) {
1623 if (!vfio_iommu_iova_dma_valid(iommu, iova, iova + size - 1)) {
1628 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1640 * We need to be able to both add to a task's locked memory and test
1641 * against the locked memory limit and we need to be able to do both
1642 * outside of this call path as pinning can be asynchronous via the
1643 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1644 * task_struct and VM locked pages requires an mm_struct, however
1645 * holding an indefinite mm reference is not recommended, therefore we
1646 * only hold a reference to a task. We could hold a reference to
1647 * current, however QEMU uses this call path through vCPU threads,
1648 * which can be killed resulting in a NULL mm and failure in the unmap
1649 * path when called via a different thread. Avoid this problem by
1650 * using the group_leader as threads within the same group require
1651 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1654 * Previously we also used the task for testing CAP_IPC_LOCK at the
1655 * time of pinning and accounting, however has_capability() makes use
1656 * of real_cred, a copy-on-write field, so we can't guarantee that it
1657 * matches group_leader, or in fact that it might not change by the
1658 * time it's evaluated. If a process were to call MAP_DMA with
1659 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1660 * possibly see different results for an iommu_mapped vfio_dma vs
1661 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1662 * time of calling MAP_DMA.
1664 get_task_struct(current->group_leader);
1665 dma->task = current->group_leader;
1666 dma->lock_cap = capable(CAP_IPC_LOCK);
1668 dma->pfn_list = RB_ROOT;
1670 /* Insert zero-sized and grow as we map chunks of it */
1671 vfio_link_dma(iommu, dma);
1673 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1674 if (list_empty(&iommu->domain_list))
1677 ret = vfio_pin_map_dma(iommu, dma, size);
1679 if (!ret && iommu->dirty_page_tracking) {
1680 ret = vfio_dma_bitmap_alloc(dma, pgsize);
1682 vfio_remove_dma(iommu, dma);
1686 mutex_unlock(&iommu->lock);
1690 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1691 struct vfio_domain *domain)
1693 struct vfio_batch batch;
1694 struct vfio_domain *d = NULL;
1696 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1699 ret = vfio_wait_all_valid(iommu);
1703 /* Arbitrarily pick the first domain in the list for lookups */
1704 if (!list_empty(&iommu->domain_list))
1705 d = list_first_entry(&iommu->domain_list,
1706 struct vfio_domain, next);
1708 vfio_batch_init(&batch);
1710 n = rb_first(&iommu->dma_list);
1712 for (; n; n = rb_next(n)) {
1713 struct vfio_dma *dma;
1716 dma = rb_entry(n, struct vfio_dma, node);
1719 while (iova < dma->iova + dma->size) {
1723 if (dma->iommu_mapped) {
1727 if (WARN_ON(!d)) { /* mapped w/o a domain?! */
1732 phys = iommu_iova_to_phys(d->domain, iova);
1734 if (WARN_ON(!phys)) {
1742 while (i < dma->iova + dma->size &&
1743 p == iommu_iova_to_phys(d->domain, i)) {
1750 unsigned long vaddr = dma->vaddr +
1752 size_t n = dma->iova + dma->size - iova;
1755 npage = vfio_pin_pages_remote(dma, vaddr,
1765 phys = pfn << PAGE_SHIFT;
1766 size = npage << PAGE_SHIFT;
1769 ret = iommu_map(domain->domain, iova, phys,
1770 size, dma->prot | IOMMU_CACHE);
1772 if (!dma->iommu_mapped) {
1773 vfio_unpin_pages_remote(dma, iova,
1777 vfio_batch_unpin(&batch, dma);
1786 /* All dmas are now mapped, defer to second tree walk for unwind */
1787 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1788 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1790 dma->iommu_mapped = true;
1793 vfio_batch_fini(&batch);
1797 for (; n; n = rb_prev(n)) {
1798 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1801 if (dma->iommu_mapped) {
1802 iommu_unmap(domain->domain, dma->iova, dma->size);
1807 while (iova < dma->iova + dma->size) {
1808 phys_addr_t phys, p;
1812 phys = iommu_iova_to_phys(domain->domain, iova);
1821 while (i < dma->iova + dma->size &&
1822 p == iommu_iova_to_phys(domain->domain, i)) {
1828 iommu_unmap(domain->domain, iova, size);
1829 vfio_unpin_pages_remote(dma, iova, phys >> PAGE_SHIFT,
1830 size >> PAGE_SHIFT, true);
1834 vfio_batch_fini(&batch);
1839 * We change our unmap behavior slightly depending on whether the IOMMU
1840 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1841 * for practically any contiguous power-of-two mapping we give it. This means
1842 * we don't need to look for contiguous chunks ourselves to make unmapping
1843 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1844 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1845 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1846 * hugetlbfs is in use.
1848 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1851 int ret, order = get_order(PAGE_SIZE * 2);
1853 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1857 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1858 IOMMU_READ | IOMMU_WRITE | IOMMU_CACHE);
1860 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1862 if (unmapped == PAGE_SIZE)
1863 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1865 domain->fgsp = true;
1868 __free_pages(pages, order);
1871 static struct vfio_iommu_group *find_iommu_group(struct vfio_domain *domain,
1872 struct iommu_group *iommu_group)
1874 struct vfio_iommu_group *g;
1876 list_for_each_entry(g, &domain->group_list, next) {
1877 if (g->iommu_group == iommu_group)
1884 static struct vfio_iommu_group*
1885 vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
1886 struct iommu_group *iommu_group)
1888 struct vfio_iommu_group *group;
1889 struct vfio_domain *domain;
1891 list_for_each_entry(domain, &iommu->domain_list, next) {
1892 group = find_iommu_group(domain, iommu_group);
1897 list_for_each_entry(group, &iommu->emulated_iommu_groups, next)
1898 if (group->iommu_group == iommu_group)
1903 static bool vfio_iommu_has_sw_msi(struct list_head *group_resv_regions,
1906 struct iommu_resv_region *region;
1909 list_for_each_entry(region, group_resv_regions, list) {
1911 * The presence of any 'real' MSI regions should take
1912 * precedence over the software-managed one if the
1913 * IOMMU driver happens to advertise both types.
1915 if (region->type == IOMMU_RESV_MSI) {
1920 if (region->type == IOMMU_RESV_SW_MSI) {
1921 *base = region->start;
1930 * This is a helper function to insert an address range to iova list.
1931 * The list is initially created with a single entry corresponding to
1932 * the IOMMU domain geometry to which the device group is attached.
1933 * The list aperture gets modified when a new domain is added to the
1934 * container if the new aperture doesn't conflict with the current one
1935 * or with any existing dma mappings. The list is also modified to
1936 * exclude any reserved regions associated with the device group.
1938 static int vfio_iommu_iova_insert(struct list_head *head,
1939 dma_addr_t start, dma_addr_t end)
1941 struct vfio_iova *region;
1943 region = kmalloc(sizeof(*region), GFP_KERNEL);
1947 INIT_LIST_HEAD(®ion->list);
1948 region->start = start;
1951 list_add_tail(®ion->list, head);
1956 * Check the new iommu aperture conflicts with existing aper or with any
1957 * existing dma mappings.
1959 static bool vfio_iommu_aper_conflict(struct vfio_iommu *iommu,
1960 dma_addr_t start, dma_addr_t end)
1962 struct vfio_iova *first, *last;
1963 struct list_head *iova = &iommu->iova_list;
1965 if (list_empty(iova))
1968 /* Disjoint sets, return conflict */
1969 first = list_first_entry(iova, struct vfio_iova, list);
1970 last = list_last_entry(iova, struct vfio_iova, list);
1971 if (start > last->end || end < first->start)
1974 /* Check for any existing dma mappings below the new start */
1975 if (start > first->start) {
1976 if (vfio_find_dma(iommu, first->start, start - first->start))
1980 /* Check for any existing dma mappings beyond the new end */
1981 if (end < last->end) {
1982 if (vfio_find_dma(iommu, end + 1, last->end - end))
1990 * Resize iommu iova aperture window. This is called only if the new
1991 * aperture has no conflict with existing aperture and dma mappings.
1993 static int vfio_iommu_aper_resize(struct list_head *iova,
1994 dma_addr_t start, dma_addr_t end)
1996 struct vfio_iova *node, *next;
1998 if (list_empty(iova))
1999 return vfio_iommu_iova_insert(iova, start, end);
2001 /* Adjust iova list start */
2002 list_for_each_entry_safe(node, next, iova, list) {
2003 if (start < node->start)
2005 if (start >= node->start && start < node->end) {
2006 node->start = start;
2009 /* Delete nodes before new start */
2010 list_del(&node->list);
2014 /* Adjust iova list end */
2015 list_for_each_entry_safe(node, next, iova, list) {
2016 if (end > node->end)
2018 if (end > node->start && end <= node->end) {
2022 /* Delete nodes after new end */
2023 list_del(&node->list);
2031 * Check reserved region conflicts with existing dma mappings
2033 static bool vfio_iommu_resv_conflict(struct vfio_iommu *iommu,
2034 struct list_head *resv_regions)
2036 struct iommu_resv_region *region;
2038 /* Check for conflict with existing dma mappings */
2039 list_for_each_entry(region, resv_regions, list) {
2040 if (region->type == IOMMU_RESV_DIRECT_RELAXABLE)
2043 if (vfio_find_dma(iommu, region->start, region->length))
2051 * Check iova region overlap with reserved regions and
2052 * exclude them from the iommu iova range
2054 static int vfio_iommu_resv_exclude(struct list_head *iova,
2055 struct list_head *resv_regions)
2057 struct iommu_resv_region *resv;
2058 struct vfio_iova *n, *next;
2060 list_for_each_entry(resv, resv_regions, list) {
2061 phys_addr_t start, end;
2063 if (resv->type == IOMMU_RESV_DIRECT_RELAXABLE)
2066 start = resv->start;
2067 end = resv->start + resv->length - 1;
2069 list_for_each_entry_safe(n, next, iova, list) {
2073 if (start > n->end || end < n->start)
2076 * Insert a new node if current node overlaps with the
2077 * reserve region to exclude that from valid iova range.
2078 * Note that, new node is inserted before the current
2079 * node and finally the current node is deleted keeping
2080 * the list updated and sorted.
2082 if (start > n->start)
2083 ret = vfio_iommu_iova_insert(&n->list, n->start,
2085 if (!ret && end < n->end)
2086 ret = vfio_iommu_iova_insert(&n->list, end + 1,
2096 if (list_empty(iova))
2102 static void vfio_iommu_resv_free(struct list_head *resv_regions)
2104 struct iommu_resv_region *n, *next;
2106 list_for_each_entry_safe(n, next, resv_regions, list) {
2112 static void vfio_iommu_iova_free(struct list_head *iova)
2114 struct vfio_iova *n, *next;
2116 list_for_each_entry_safe(n, next, iova, list) {
2122 static int vfio_iommu_iova_get_copy(struct vfio_iommu *iommu,
2123 struct list_head *iova_copy)
2125 struct list_head *iova = &iommu->iova_list;
2126 struct vfio_iova *n;
2129 list_for_each_entry(n, iova, list) {
2130 ret = vfio_iommu_iova_insert(iova_copy, n->start, n->end);
2138 vfio_iommu_iova_free(iova_copy);
2142 static void vfio_iommu_iova_insert_copy(struct vfio_iommu *iommu,
2143 struct list_head *iova_copy)
2145 struct list_head *iova = &iommu->iova_list;
2147 vfio_iommu_iova_free(iova);
2149 list_splice_tail(iova_copy, iova);
2152 /* Redundantly walks non-present capabilities to simplify caller */
2153 static int vfio_iommu_device_capable(struct device *dev, void *data)
2155 return device_iommu_capable(dev, (enum iommu_cap)data);
2158 static int vfio_iommu_domain_alloc(struct device *dev, void *data)
2160 struct iommu_domain **domain = data;
2162 *domain = iommu_domain_alloc(dev->bus);
2163 return 1; /* Don't iterate */
2166 static int vfio_iommu_type1_attach_group(void *iommu_data,
2167 struct iommu_group *iommu_group, enum vfio_group_type type)
2169 struct vfio_iommu *iommu = iommu_data;
2170 struct vfio_iommu_group *group;
2171 struct vfio_domain *domain, *d;
2172 bool resv_msi, msi_remap;
2173 phys_addr_t resv_msi_base = 0;
2174 struct iommu_domain_geometry *geo;
2175 LIST_HEAD(iova_copy);
2176 LIST_HEAD(group_resv_regions);
2179 mutex_lock(&iommu->lock);
2181 /* Check for duplicates */
2182 if (vfio_iommu_find_iommu_group(iommu, iommu_group))
2186 group = kzalloc(sizeof(*group), GFP_KERNEL);
2189 group->iommu_group = iommu_group;
2191 if (type == VFIO_EMULATED_IOMMU) {
2192 list_add(&group->next, &iommu->emulated_iommu_groups);
2194 * An emulated IOMMU group cannot dirty memory directly, it can
2195 * only use interfaces that provide dirty tracking.
2196 * The iommu scope can only be promoted with the addition of a
2197 * dirty tracking group.
2199 group->pinned_page_dirty_scope = true;
2205 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2207 goto out_free_group;
2210 * Going via the iommu_group iterator avoids races, and trivially gives
2211 * us a representative device for the IOMMU API call. We don't actually
2212 * want to iterate beyond the first device (if any).
2215 iommu_group_for_each_dev(iommu_group, &domain->domain,
2216 vfio_iommu_domain_alloc);
2217 if (!domain->domain)
2218 goto out_free_domain;
2220 if (iommu->nesting) {
2221 ret = iommu_enable_nesting(domain->domain);
2226 ret = iommu_attach_group(domain->domain, group->iommu_group);
2230 /* Get aperture info */
2231 geo = &domain->domain->geometry;
2232 if (vfio_iommu_aper_conflict(iommu, geo->aperture_start,
2233 geo->aperture_end)) {
2238 ret = iommu_get_group_resv_regions(iommu_group, &group_resv_regions);
2242 if (vfio_iommu_resv_conflict(iommu, &group_resv_regions)) {
2248 * We don't want to work on the original iova list as the list
2249 * gets modified and in case of failure we have to retain the
2250 * original list. Get a copy here.
2252 ret = vfio_iommu_iova_get_copy(iommu, &iova_copy);
2256 ret = vfio_iommu_aper_resize(&iova_copy, geo->aperture_start,
2261 ret = vfio_iommu_resv_exclude(&iova_copy, &group_resv_regions);
2265 resv_msi = vfio_iommu_has_sw_msi(&group_resv_regions, &resv_msi_base);
2267 INIT_LIST_HEAD(&domain->group_list);
2268 list_add(&group->next, &domain->group_list);
2270 msi_remap = irq_domain_check_msi_remap() ||
2271 iommu_group_for_each_dev(iommu_group, (void *)IOMMU_CAP_INTR_REMAP,
2272 vfio_iommu_device_capable);
2274 if (!allow_unsafe_interrupts && !msi_remap) {
2275 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
2282 * If the IOMMU can block non-coherent operations (ie PCIe TLPs with
2283 * no-snoop set) then VFIO always turns this feature on because on Intel
2284 * platforms it optimizes KVM to disable wbinvd emulation.
2286 if (domain->domain->ops->enforce_cache_coherency)
2287 domain->enforce_cache_coherency =
2288 domain->domain->ops->enforce_cache_coherency(
2292 * Try to match an existing compatible domain. We don't want to
2293 * preclude an IOMMU driver supporting multiple bus_types and being
2294 * able to include different bus_types in the same IOMMU domain, so
2295 * we test whether the domains use the same iommu_ops rather than
2296 * testing if they're on the same bus_type.
2298 list_for_each_entry(d, &iommu->domain_list, next) {
2299 if (d->domain->ops == domain->domain->ops &&
2300 d->enforce_cache_coherency ==
2301 domain->enforce_cache_coherency) {
2302 iommu_detach_group(domain->domain, group->iommu_group);
2303 if (!iommu_attach_group(d->domain,
2304 group->iommu_group)) {
2305 list_add(&group->next, &d->group_list);
2306 iommu_domain_free(domain->domain);
2311 ret = iommu_attach_group(domain->domain,
2312 group->iommu_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 iommu_detach_group(domain->domain, group->iommu_group);
2351 iommu_domain_free(domain->domain);
2352 vfio_iommu_iova_free(&iova_copy);
2353 vfio_iommu_resv_free(&group_resv_regions);
2359 mutex_unlock(&iommu->lock);
2363 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
2365 struct rb_node *node;
2367 while ((node = rb_first(&iommu->dma_list)))
2368 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
2371 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
2373 struct rb_node *n, *p;
2375 n = rb_first(&iommu->dma_list);
2376 for (; n; n = rb_next(n)) {
2377 struct vfio_dma *dma;
2378 long locked = 0, unlocked = 0;
2380 dma = rb_entry(n, struct vfio_dma, node);
2381 unlocked += vfio_unmap_unpin(iommu, dma, false);
2382 p = rb_first(&dma->pfn_list);
2383 for (; p; p = rb_next(p)) {
2384 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
2387 if (!is_invalid_reserved_pfn(vpfn->pfn))
2390 vfio_lock_acct(dma, locked - unlocked, true);
2395 * Called when a domain is removed in detach. It is possible that
2396 * the removed domain decided the iova aperture window. Modify the
2397 * iova aperture with the smallest window among existing domains.
2399 static void vfio_iommu_aper_expand(struct vfio_iommu *iommu,
2400 struct list_head *iova_copy)
2402 struct vfio_domain *domain;
2403 struct vfio_iova *node;
2404 dma_addr_t start = 0;
2405 dma_addr_t end = (dma_addr_t)~0;
2407 if (list_empty(iova_copy))
2410 list_for_each_entry(domain, &iommu->domain_list, next) {
2411 struct iommu_domain_geometry *geo = &domain->domain->geometry;
2413 if (geo->aperture_start > start)
2414 start = geo->aperture_start;
2415 if (geo->aperture_end < end)
2416 end = geo->aperture_end;
2419 /* Modify aperture limits. The new aper is either same or bigger */
2420 node = list_first_entry(iova_copy, struct vfio_iova, list);
2421 node->start = start;
2422 node = list_last_entry(iova_copy, struct vfio_iova, list);
2427 * Called when a group is detached. The reserved regions for that
2428 * group can be part of valid iova now. But since reserved regions
2429 * may be duplicated among groups, populate the iova valid regions
2432 static int vfio_iommu_resv_refresh(struct vfio_iommu *iommu,
2433 struct list_head *iova_copy)
2435 struct vfio_domain *d;
2436 struct vfio_iommu_group *g;
2437 struct vfio_iova *node;
2438 dma_addr_t start, end;
2439 LIST_HEAD(resv_regions);
2442 if (list_empty(iova_copy))
2445 list_for_each_entry(d, &iommu->domain_list, next) {
2446 list_for_each_entry(g, &d->group_list, next) {
2447 ret = iommu_get_group_resv_regions(g->iommu_group,
2454 node = list_first_entry(iova_copy, struct vfio_iova, list);
2455 start = node->start;
2456 node = list_last_entry(iova_copy, struct vfio_iova, list);
2459 /* purge the iova list and create new one */
2460 vfio_iommu_iova_free(iova_copy);
2462 ret = vfio_iommu_aper_resize(iova_copy, start, end);
2466 /* Exclude current reserved regions from iova ranges */
2467 ret = vfio_iommu_resv_exclude(iova_copy, &resv_regions);
2469 vfio_iommu_resv_free(&resv_regions);
2473 static void vfio_iommu_type1_detach_group(void *iommu_data,
2474 struct iommu_group *iommu_group)
2476 struct vfio_iommu *iommu = iommu_data;
2477 struct vfio_domain *domain;
2478 struct vfio_iommu_group *group;
2479 bool update_dirty_scope = false;
2480 LIST_HEAD(iova_copy);
2482 mutex_lock(&iommu->lock);
2483 list_for_each_entry(group, &iommu->emulated_iommu_groups, next) {
2484 if (group->iommu_group != iommu_group)
2486 update_dirty_scope = !group->pinned_page_dirty_scope;
2487 list_del(&group->next);
2490 if (list_empty(&iommu->emulated_iommu_groups) &&
2491 list_empty(&iommu->domain_list)) {
2492 WARN_ON(!list_empty(&iommu->device_list));
2493 vfio_iommu_unmap_unpin_all(iommu);
2495 goto detach_group_done;
2499 * Get a copy of iova list. This will be used to update
2500 * and to replace the current one later. Please note that
2501 * we will leave the original list as it is if update fails.
2503 vfio_iommu_iova_get_copy(iommu, &iova_copy);
2505 list_for_each_entry(domain, &iommu->domain_list, next) {
2506 group = find_iommu_group(domain, iommu_group);
2510 iommu_detach_group(domain->domain, group->iommu_group);
2511 update_dirty_scope = !group->pinned_page_dirty_scope;
2512 list_del(&group->next);
2515 * Group ownership provides privilege, if the group list is
2516 * empty, the domain goes away. If it's the last domain with
2517 * iommu and external domain doesn't exist, then all the
2518 * mappings go away too. If it's the last domain with iommu and
2519 * external domain exist, update accounting
2521 if (list_empty(&domain->group_list)) {
2522 if (list_is_singular(&iommu->domain_list)) {
2523 if (list_empty(&iommu->emulated_iommu_groups)) {
2524 WARN_ON(!list_empty(
2525 &iommu->device_list));
2526 vfio_iommu_unmap_unpin_all(iommu);
2528 vfio_iommu_unmap_unpin_reaccount(iommu);
2531 iommu_domain_free(domain->domain);
2532 list_del(&domain->next);
2534 vfio_iommu_aper_expand(iommu, &iova_copy);
2535 vfio_update_pgsize_bitmap(iommu);
2540 if (!vfio_iommu_resv_refresh(iommu, &iova_copy))
2541 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2543 vfio_iommu_iova_free(&iova_copy);
2547 * Removal of a group without dirty tracking may allow the iommu scope
2550 if (update_dirty_scope) {
2551 iommu->num_non_pinned_groups--;
2552 if (iommu->dirty_page_tracking)
2553 vfio_iommu_populate_bitmap_full(iommu);
2555 mutex_unlock(&iommu->lock);
2558 static void *vfio_iommu_type1_open(unsigned long arg)
2560 struct vfio_iommu *iommu;
2562 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
2564 return ERR_PTR(-ENOMEM);
2567 case VFIO_TYPE1_IOMMU:
2569 case VFIO_TYPE1_NESTING_IOMMU:
2570 iommu->nesting = true;
2572 case VFIO_TYPE1v2_IOMMU:
2577 return ERR_PTR(-EINVAL);
2580 INIT_LIST_HEAD(&iommu->domain_list);
2581 INIT_LIST_HEAD(&iommu->iova_list);
2582 iommu->dma_list = RB_ROOT;
2583 iommu->dma_avail = dma_entry_limit;
2584 iommu->container_open = true;
2585 mutex_init(&iommu->lock);
2586 mutex_init(&iommu->device_list_lock);
2587 INIT_LIST_HEAD(&iommu->device_list);
2588 init_waitqueue_head(&iommu->vaddr_wait);
2589 iommu->pgsize_bitmap = PAGE_MASK;
2590 INIT_LIST_HEAD(&iommu->emulated_iommu_groups);
2595 static void vfio_release_domain(struct vfio_domain *domain)
2597 struct vfio_iommu_group *group, *group_tmp;
2599 list_for_each_entry_safe(group, group_tmp,
2600 &domain->group_list, next) {
2601 iommu_detach_group(domain->domain, group->iommu_group);
2602 list_del(&group->next);
2606 iommu_domain_free(domain->domain);
2609 static void vfio_iommu_type1_release(void *iommu_data)
2611 struct vfio_iommu *iommu = iommu_data;
2612 struct vfio_domain *domain, *domain_tmp;
2613 struct vfio_iommu_group *group, *next_group;
2615 list_for_each_entry_safe(group, next_group,
2616 &iommu->emulated_iommu_groups, next) {
2617 list_del(&group->next);
2621 vfio_iommu_unmap_unpin_all(iommu);
2623 list_for_each_entry_safe(domain, domain_tmp,
2624 &iommu->domain_list, next) {
2625 vfio_release_domain(domain);
2626 list_del(&domain->next);
2630 vfio_iommu_iova_free(&iommu->iova_list);
2635 static int vfio_domains_have_enforce_cache_coherency(struct vfio_iommu *iommu)
2637 struct vfio_domain *domain;
2640 mutex_lock(&iommu->lock);
2641 list_for_each_entry(domain, &iommu->domain_list, next) {
2642 if (!(domain->enforce_cache_coherency)) {
2647 mutex_unlock(&iommu->lock);
2652 static int vfio_iommu_type1_check_extension(struct vfio_iommu *iommu,
2656 case VFIO_TYPE1_IOMMU:
2657 case VFIO_TYPE1v2_IOMMU:
2658 case VFIO_TYPE1_NESTING_IOMMU:
2659 case VFIO_UNMAP_ALL:
2660 case VFIO_UPDATE_VADDR:
2662 case VFIO_DMA_CC_IOMMU:
2665 return vfio_domains_have_enforce_cache_coherency(iommu);
2671 static int vfio_iommu_iova_add_cap(struct vfio_info_cap *caps,
2672 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas,
2675 struct vfio_info_cap_header *header;
2676 struct vfio_iommu_type1_info_cap_iova_range *iova_cap;
2678 header = vfio_info_cap_add(caps, size,
2679 VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE, 1);
2681 return PTR_ERR(header);
2683 iova_cap = container_of(header,
2684 struct vfio_iommu_type1_info_cap_iova_range,
2686 iova_cap->nr_iovas = cap_iovas->nr_iovas;
2687 memcpy(iova_cap->iova_ranges, cap_iovas->iova_ranges,
2688 cap_iovas->nr_iovas * sizeof(*cap_iovas->iova_ranges));
2692 static int vfio_iommu_iova_build_caps(struct vfio_iommu *iommu,
2693 struct vfio_info_cap *caps)
2695 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas;
2696 struct vfio_iova *iova;
2698 int iovas = 0, i = 0, ret;
2700 list_for_each_entry(iova, &iommu->iova_list, list)
2705 * Return 0 as a container with a single mdev device
2706 * will have an empty list
2711 size = struct_size(cap_iovas, iova_ranges, iovas);
2713 cap_iovas = kzalloc(size, GFP_KERNEL);
2717 cap_iovas->nr_iovas = iovas;
2719 list_for_each_entry(iova, &iommu->iova_list, list) {
2720 cap_iovas->iova_ranges[i].start = iova->start;
2721 cap_iovas->iova_ranges[i].end = iova->end;
2725 ret = vfio_iommu_iova_add_cap(caps, cap_iovas, size);
2731 static int vfio_iommu_migration_build_caps(struct vfio_iommu *iommu,
2732 struct vfio_info_cap *caps)
2734 struct vfio_iommu_type1_info_cap_migration cap_mig;
2736 cap_mig.header.id = VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION;
2737 cap_mig.header.version = 1;
2740 /* support minimum pgsize */
2741 cap_mig.pgsize_bitmap = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2742 cap_mig.max_dirty_bitmap_size = DIRTY_BITMAP_SIZE_MAX;
2744 return vfio_info_add_capability(caps, &cap_mig.header, sizeof(cap_mig));
2747 static int vfio_iommu_dma_avail_build_caps(struct vfio_iommu *iommu,
2748 struct vfio_info_cap *caps)
2750 struct vfio_iommu_type1_info_dma_avail cap_dma_avail;
2752 cap_dma_avail.header.id = VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL;
2753 cap_dma_avail.header.version = 1;
2755 cap_dma_avail.avail = iommu->dma_avail;
2757 return vfio_info_add_capability(caps, &cap_dma_avail.header,
2758 sizeof(cap_dma_avail));
2761 static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu,
2764 struct vfio_iommu_type1_info info;
2765 unsigned long minsz;
2766 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
2767 unsigned long capsz;
2770 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
2772 /* For backward compatibility, cannot require this */
2773 capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
2775 if (copy_from_user(&info, (void __user *)arg, minsz))
2778 if (info.argsz < minsz)
2781 if (info.argsz >= capsz) {
2783 info.cap_offset = 0; /* output, no-recopy necessary */
2786 mutex_lock(&iommu->lock);
2787 info.flags = VFIO_IOMMU_INFO_PGSIZES;
2789 info.iova_pgsizes = iommu->pgsize_bitmap;
2791 ret = vfio_iommu_migration_build_caps(iommu, &caps);
2794 ret = vfio_iommu_dma_avail_build_caps(iommu, &caps);
2797 ret = vfio_iommu_iova_build_caps(iommu, &caps);
2799 mutex_unlock(&iommu->lock);
2805 info.flags |= VFIO_IOMMU_INFO_CAPS;
2807 if (info.argsz < sizeof(info) + caps.size) {
2808 info.argsz = sizeof(info) + caps.size;
2810 vfio_info_cap_shift(&caps, sizeof(info));
2811 if (copy_to_user((void __user *)arg +
2812 sizeof(info), caps.buf,
2817 info.cap_offset = sizeof(info);
2823 return copy_to_user((void __user *)arg, &info, minsz) ?
2827 static int vfio_iommu_type1_map_dma(struct vfio_iommu *iommu,
2830 struct vfio_iommu_type1_dma_map map;
2831 unsigned long minsz;
2832 uint32_t mask = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE |
2833 VFIO_DMA_MAP_FLAG_VADDR;
2835 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
2837 if (copy_from_user(&map, (void __user *)arg, minsz))
2840 if (map.argsz < minsz || map.flags & ~mask)
2843 return vfio_dma_do_map(iommu, &map);
2846 static int vfio_iommu_type1_unmap_dma(struct vfio_iommu *iommu,
2849 struct vfio_iommu_type1_dma_unmap unmap;
2850 struct vfio_bitmap bitmap = { 0 };
2851 uint32_t mask = VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP |
2852 VFIO_DMA_UNMAP_FLAG_VADDR |
2853 VFIO_DMA_UNMAP_FLAG_ALL;
2854 unsigned long minsz;
2857 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
2859 if (copy_from_user(&unmap, (void __user *)arg, minsz))
2862 if (unmap.argsz < minsz || unmap.flags & ~mask)
2865 if ((unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
2866 (unmap.flags & (VFIO_DMA_UNMAP_FLAG_ALL |
2867 VFIO_DMA_UNMAP_FLAG_VADDR)))
2870 if (unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
2871 unsigned long pgshift;
2873 if (unmap.argsz < (minsz + sizeof(bitmap)))
2876 if (copy_from_user(&bitmap,
2877 (void __user *)(arg + minsz),
2881 if (!access_ok((void __user *)bitmap.data, bitmap.size))
2884 pgshift = __ffs(bitmap.pgsize);
2885 ret = verify_bitmap_size(unmap.size >> pgshift,
2891 ret = vfio_dma_do_unmap(iommu, &unmap, &bitmap);
2895 return copy_to_user((void __user *)arg, &unmap, minsz) ?
2899 static int vfio_iommu_type1_dirty_pages(struct vfio_iommu *iommu,
2902 struct vfio_iommu_type1_dirty_bitmap dirty;
2903 uint32_t mask = VFIO_IOMMU_DIRTY_PAGES_FLAG_START |
2904 VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP |
2905 VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP;
2906 unsigned long minsz;
2912 minsz = offsetofend(struct vfio_iommu_type1_dirty_bitmap, flags);
2914 if (copy_from_user(&dirty, (void __user *)arg, minsz))
2917 if (dirty.argsz < minsz || dirty.flags & ~mask)
2920 /* only one flag should be set at a time */
2921 if (__ffs(dirty.flags) != __fls(dirty.flags))
2924 if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_START) {
2927 mutex_lock(&iommu->lock);
2928 pgsize = 1 << __ffs(iommu->pgsize_bitmap);
2929 if (!iommu->dirty_page_tracking) {
2930 ret = vfio_dma_bitmap_alloc_all(iommu, pgsize);
2932 iommu->dirty_page_tracking = true;
2934 mutex_unlock(&iommu->lock);
2936 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP) {
2937 mutex_lock(&iommu->lock);
2938 if (iommu->dirty_page_tracking) {
2939 iommu->dirty_page_tracking = false;
2940 vfio_dma_bitmap_free_all(iommu);
2942 mutex_unlock(&iommu->lock);
2944 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP) {
2945 struct vfio_iommu_type1_dirty_bitmap_get range;
2946 unsigned long pgshift;
2947 size_t data_size = dirty.argsz - minsz;
2948 size_t iommu_pgsize;
2950 if (!data_size || data_size < sizeof(range))
2953 if (copy_from_user(&range, (void __user *)(arg + minsz),
2957 if (range.iova + range.size < range.iova)
2959 if (!access_ok((void __user *)range.bitmap.data,
2963 pgshift = __ffs(range.bitmap.pgsize);
2964 ret = verify_bitmap_size(range.size >> pgshift,
2969 mutex_lock(&iommu->lock);
2971 iommu_pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2973 /* allow only smallest supported pgsize */
2974 if (range.bitmap.pgsize != iommu_pgsize) {
2978 if (range.iova & (iommu_pgsize - 1)) {
2982 if (!range.size || range.size & (iommu_pgsize - 1)) {
2987 if (iommu->dirty_page_tracking)
2988 ret = vfio_iova_dirty_bitmap(range.bitmap.data,
2991 range.bitmap.pgsize);
2995 mutex_unlock(&iommu->lock);
3003 static long vfio_iommu_type1_ioctl(void *iommu_data,
3004 unsigned int cmd, unsigned long arg)
3006 struct vfio_iommu *iommu = iommu_data;
3009 case VFIO_CHECK_EXTENSION:
3010 return vfio_iommu_type1_check_extension(iommu, arg);
3011 case VFIO_IOMMU_GET_INFO:
3012 return vfio_iommu_type1_get_info(iommu, arg);
3013 case VFIO_IOMMU_MAP_DMA:
3014 return vfio_iommu_type1_map_dma(iommu, arg);
3015 case VFIO_IOMMU_UNMAP_DMA:
3016 return vfio_iommu_type1_unmap_dma(iommu, arg);
3017 case VFIO_IOMMU_DIRTY_PAGES:
3018 return vfio_iommu_type1_dirty_pages(iommu, arg);
3024 static void vfio_iommu_type1_register_device(void *iommu_data,
3025 struct vfio_device *vdev)
3027 struct vfio_iommu *iommu = iommu_data;
3029 if (!vdev->ops->dma_unmap)
3033 * list_empty(&iommu->device_list) is tested under the iommu->lock while
3034 * iteration for dma_unmap must be done under the device_list_lock.
3035 * Holding both locks here allows avoiding the device_list_lock in
3036 * several fast paths. See vfio_notify_dma_unmap()
3038 mutex_lock(&iommu->lock);
3039 mutex_lock(&iommu->device_list_lock);
3040 list_add(&vdev->iommu_entry, &iommu->device_list);
3041 mutex_unlock(&iommu->device_list_lock);
3042 mutex_unlock(&iommu->lock);
3045 static void vfio_iommu_type1_unregister_device(void *iommu_data,
3046 struct vfio_device *vdev)
3048 struct vfio_iommu *iommu = iommu_data;
3050 if (!vdev->ops->dma_unmap)
3053 mutex_lock(&iommu->lock);
3054 mutex_lock(&iommu->device_list_lock);
3055 list_del(&vdev->iommu_entry);
3056 mutex_unlock(&iommu->device_list_lock);
3057 mutex_unlock(&iommu->lock);
3060 static int vfio_iommu_type1_dma_rw_chunk(struct vfio_iommu *iommu,
3061 dma_addr_t user_iova, void *data,
3062 size_t count, bool write,
3065 struct mm_struct *mm;
3066 unsigned long vaddr;
3067 struct vfio_dma *dma;
3068 bool kthread = current->mm == NULL;
3074 ret = vfio_find_dma_valid(iommu, user_iova, 1, &dma);
3078 if ((write && !(dma->prot & IOMMU_WRITE)) ||
3079 !(dma->prot & IOMMU_READ))
3082 mm = get_task_mm(dma->task);
3089 else if (current->mm != mm)
3092 offset = user_iova - dma->iova;
3094 if (count > dma->size - offset)
3095 count = dma->size - offset;
3097 vaddr = dma->vaddr + offset;
3100 *copied = copy_to_user((void __user *)vaddr, data,
3102 if (*copied && iommu->dirty_page_tracking) {
3103 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
3105 * Bitmap populated with the smallest supported page
3108 bitmap_set(dma->bitmap, offset >> pgshift,
3109 ((offset + *copied - 1) >> pgshift) -
3110 (offset >> pgshift) + 1);
3113 *copied = copy_from_user(data, (void __user *)vaddr,
3116 kthread_unuse_mm(mm);
3119 return *copied ? 0 : -EFAULT;
3122 static int vfio_iommu_type1_dma_rw(void *iommu_data, dma_addr_t user_iova,
3123 void *data, size_t count, bool write)
3125 struct vfio_iommu *iommu = iommu_data;
3129 mutex_lock(&iommu->lock);
3131 ret = vfio_iommu_type1_dma_rw_chunk(iommu, user_iova, data,
3132 count, write, &done);
3141 mutex_unlock(&iommu->lock);
3145 static struct iommu_domain *
3146 vfio_iommu_type1_group_iommu_domain(void *iommu_data,
3147 struct iommu_group *iommu_group)
3149 struct iommu_domain *domain = ERR_PTR(-ENODEV);
3150 struct vfio_iommu *iommu = iommu_data;
3151 struct vfio_domain *d;
3153 if (!iommu || !iommu_group)
3154 return ERR_PTR(-EINVAL);
3156 mutex_lock(&iommu->lock);
3157 list_for_each_entry(d, &iommu->domain_list, next) {
3158 if (find_iommu_group(d, iommu_group)) {
3163 mutex_unlock(&iommu->lock);
3168 static void vfio_iommu_type1_notify(void *iommu_data,
3169 enum vfio_iommu_notify_type event)
3171 struct vfio_iommu *iommu = iommu_data;
3173 if (event != VFIO_IOMMU_CONTAINER_CLOSE)
3175 mutex_lock(&iommu->lock);
3176 iommu->container_open = false;
3177 mutex_unlock(&iommu->lock);
3178 wake_up_all(&iommu->vaddr_wait);
3181 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
3182 .name = "vfio-iommu-type1",
3183 .owner = THIS_MODULE,
3184 .open = vfio_iommu_type1_open,
3185 .release = vfio_iommu_type1_release,
3186 .ioctl = vfio_iommu_type1_ioctl,
3187 .attach_group = vfio_iommu_type1_attach_group,
3188 .detach_group = vfio_iommu_type1_detach_group,
3189 .pin_pages = vfio_iommu_type1_pin_pages,
3190 .unpin_pages = vfio_iommu_type1_unpin_pages,
3191 .register_device = vfio_iommu_type1_register_device,
3192 .unregister_device = vfio_iommu_type1_unregister_device,
3193 .dma_rw = vfio_iommu_type1_dma_rw,
3194 .group_iommu_domain = vfio_iommu_type1_group_iommu_domain,
3195 .notify = vfio_iommu_type1_notify,
3198 static int __init vfio_iommu_type1_init(void)
3200 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
3203 static void __exit vfio_iommu_type1_cleanup(void)
3205 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
3208 module_init(vfio_iommu_type1_init);
3209 module_exit(vfio_iommu_type1_cleanup);
3211 MODULE_VERSION(DRIVER_VERSION);
3212 MODULE_LICENSE("GPL v2");
3213 MODULE_AUTHOR(DRIVER_AUTHOR);
3214 MODULE_DESCRIPTION(DRIVER_DESC);