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,
563 * The zero page is always resident, we don't need to pin it
564 * and it falls into our invalid/reserved test so we don't
565 * unpin in put_pfn(). Unpin all zero pages in the batch here.
567 for (i = 0 ; i < ret; i++) {
568 if (unlikely(is_zero_pfn(page_to_pfn(pages[i]))))
569 unpin_user_page(pages[i]);
572 *pfn = page_to_pfn(pages[0]);
576 vaddr = untagged_addr(vaddr);
579 vma = vma_lookup(mm, vaddr);
581 if (vma && vma->vm_flags & VM_PFNMAP) {
582 ret = follow_fault_pfn(vma, mm, vaddr, pfn, prot & IOMMU_WRITE);
587 if (is_invalid_reserved_pfn(*pfn))
594 mmap_read_unlock(mm);
598 static int vfio_wait(struct vfio_iommu *iommu)
602 prepare_to_wait(&iommu->vaddr_wait, &wait, TASK_KILLABLE);
603 mutex_unlock(&iommu->lock);
605 mutex_lock(&iommu->lock);
606 finish_wait(&iommu->vaddr_wait, &wait);
607 if (kthread_should_stop() || !iommu->container_open ||
608 fatal_signal_pending(current)) {
615 * Find dma struct and wait for its vaddr to be valid. iommu lock is dropped
616 * if the task waits, but is re-locked on return. Return result in *dma_p.
617 * Return 0 on success with no waiting, WAITED on success if waited, and -errno
620 static int vfio_find_dma_valid(struct vfio_iommu *iommu, dma_addr_t start,
621 size_t size, struct vfio_dma **dma_p)
626 *dma_p = vfio_find_dma(iommu, start, size);
629 else if (!(*dma_p)->vaddr_invalid)
632 ret = vfio_wait(iommu);
633 } while (ret == WAITED);
639 * Wait for all vaddr in the dma_list to become valid. iommu lock is dropped
640 * if the task waits, but is re-locked on return. Return 0 on success with no
641 * waiting, WAITED on success if waited, and -errno on error.
643 static int vfio_wait_all_valid(struct vfio_iommu *iommu)
647 while (iommu->vaddr_invalid_count && ret >= 0)
648 ret = vfio_wait(iommu);
654 * Attempt to pin pages. We really don't want to track all the pfns and
655 * the iommu can only map chunks of consecutive pfns anyway, so get the
656 * first page and all consecutive pages with the same locking.
658 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
659 long npage, unsigned long *pfn_base,
660 unsigned long limit, struct vfio_batch *batch)
663 struct mm_struct *mm = current->mm;
664 long ret, pinned = 0, lock_acct = 0;
666 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
668 /* This code path is only user initiated */
673 /* Leftover pages in batch from an earlier call. */
674 *pfn_base = page_to_pfn(batch->pages[batch->offset]);
676 rsvd = is_invalid_reserved_pfn(*pfn_base);
683 /* Empty batch, so refill it. */
684 long req_pages = min_t(long, npage, batch->capacity);
686 ret = vaddr_get_pfns(mm, vaddr, req_pages, dma->prot,
696 rsvd = is_invalid_reserved_pfn(*pfn_base);
701 * pfn is preset for the first iteration of this inner loop and
702 * updated at the end to handle a VM_PFNMAP pfn. In that case,
703 * batch->pages isn't valid (there's no struct page), so allow
704 * batch->pages to be touched only when there's more than one
705 * pfn to check, which guarantees the pfns are from a
709 if (pfn != *pfn_base + pinned ||
710 rsvd != is_invalid_reserved_pfn(pfn))
714 * Reserved pages aren't counted against the user,
715 * externally pinned pages are already counted against
718 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
719 if (!dma->lock_cap &&
720 mm->locked_vm + lock_acct + 1 > limit) {
721 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
722 __func__, limit << PAGE_SHIFT);
739 pfn = page_to_pfn(batch->pages[batch->offset]);
742 if (unlikely(disable_hugepages))
747 ret = vfio_lock_acct(dma, lock_acct, false);
750 if (batch->size == 1 && !batch->offset) {
751 /* May be a VM_PFNMAP pfn, which the batch can't remember. */
752 put_pfn(pfn, dma->prot);
757 if (pinned && !rsvd) {
758 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
759 put_pfn(pfn, dma->prot);
761 vfio_batch_unpin(batch, dma);
769 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
770 unsigned long pfn, long npage,
773 long unlocked = 0, locked = 0;
776 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
777 if (put_pfn(pfn++, dma->prot)) {
779 if (vfio_find_vpfn(dma, iova))
785 vfio_lock_acct(dma, locked - unlocked, true);
790 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
791 unsigned long *pfn_base, bool do_accounting)
793 struct page *pages[1];
794 struct mm_struct *mm;
797 mm = get_task_mm(dma->task);
801 ret = vaddr_get_pfns(mm, vaddr, 1, dma->prot, pfn_base, pages);
807 if (do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
808 ret = vfio_lock_acct(dma, 1, true);
810 put_pfn(*pfn_base, dma->prot);
812 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
813 "(%ld) exceeded\n", __func__,
814 dma->task->comm, task_pid_nr(dma->task),
815 task_rlimit(dma->task, RLIMIT_MEMLOCK));
824 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
828 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
833 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
836 vfio_lock_acct(dma, -unlocked, true);
841 static int vfio_iommu_type1_pin_pages(void *iommu_data,
842 struct iommu_group *iommu_group,
843 unsigned long *user_pfn,
845 unsigned long *phys_pfn)
847 struct vfio_iommu *iommu = iommu_data;
848 struct vfio_iommu_group *group;
850 unsigned long remote_vaddr;
851 struct vfio_dma *dma;
855 if (!iommu || !user_pfn || !phys_pfn)
858 /* Supported for v2 version only */
862 mutex_lock(&iommu->lock);
865 * Wait for all necessary vaddr's to be valid so they can be used in
866 * the main loop without dropping the lock, to avoid racing vs unmap.
869 if (iommu->vaddr_invalid_count) {
870 for (i = 0; i < npage; i++) {
871 iova = user_pfn[i] << PAGE_SHIFT;
872 ret = vfio_find_dma_valid(iommu, iova, PAGE_SIZE, &dma);
880 /* Fail if notifier list is empty */
881 if (!iommu->notifier.head) {
887 * If iommu capable domain exist in the container then all pages are
888 * already pinned and accounted. Accounting should be done if there is no
889 * iommu capable domain in the container.
891 do_accounting = list_empty(&iommu->domain_list);
893 for (i = 0; i < npage; i++) {
894 struct vfio_pfn *vpfn;
896 iova = user_pfn[i] << PAGE_SHIFT;
897 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
903 if ((dma->prot & prot) != prot) {
908 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
910 phys_pfn[i] = vpfn->pfn;
914 remote_vaddr = dma->vaddr + (iova - dma->iova);
915 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
920 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
922 if (put_pfn(phys_pfn[i], dma->prot) && do_accounting)
923 vfio_lock_acct(dma, -1, true);
927 if (iommu->dirty_page_tracking) {
928 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
931 * Bitmap populated with the smallest supported page
934 bitmap_set(dma->bitmap,
935 (iova - dma->iova) >> pgshift, 1);
940 group = vfio_iommu_find_iommu_group(iommu, iommu_group);
941 if (!group->pinned_page_dirty_scope) {
942 group->pinned_page_dirty_scope = true;
943 iommu->num_non_pinned_groups--;
950 for (j = 0; j < i; j++) {
953 iova = user_pfn[j] << PAGE_SHIFT;
954 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
955 vfio_unpin_page_external(dma, iova, do_accounting);
959 mutex_unlock(&iommu->lock);
963 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
964 unsigned long *user_pfn,
967 struct vfio_iommu *iommu = iommu_data;
971 if (!iommu || !user_pfn || npage <= 0)
974 /* Supported for v2 version only */
978 mutex_lock(&iommu->lock);
980 do_accounting = list_empty(&iommu->domain_list);
981 for (i = 0; i < npage; i++) {
982 struct vfio_dma *dma;
985 iova = user_pfn[i] << PAGE_SHIFT;
986 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
990 vfio_unpin_page_external(dma, iova, do_accounting);
993 mutex_unlock(&iommu->lock);
994 return i > 0 ? i : -EINVAL;
997 static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
998 struct list_head *regions,
999 struct iommu_iotlb_gather *iotlb_gather)
1002 struct vfio_regions *entry, *next;
1004 iommu_iotlb_sync(domain->domain, iotlb_gather);
1006 list_for_each_entry_safe(entry, next, regions, list) {
1007 unlocked += vfio_unpin_pages_remote(dma,
1009 entry->phys >> PAGE_SHIFT,
1010 entry->len >> PAGE_SHIFT,
1012 list_del(&entry->list);
1022 * Generally, VFIO needs to unpin remote pages after each IOTLB flush.
1023 * Therefore, when using IOTLB flush sync interface, VFIO need to keep track
1024 * of these regions (currently using a list).
1026 * This value specifies maximum number of regions for each IOTLB flush sync.
1028 #define VFIO_IOMMU_TLB_SYNC_MAX 512
1030 static size_t unmap_unpin_fast(struct vfio_domain *domain,
1031 struct vfio_dma *dma, dma_addr_t *iova,
1032 size_t len, phys_addr_t phys, long *unlocked,
1033 struct list_head *unmapped_list,
1035 struct iommu_iotlb_gather *iotlb_gather)
1037 size_t unmapped = 0;
1038 struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
1041 unmapped = iommu_unmap_fast(domain->domain, *iova, len,
1047 entry->iova = *iova;
1049 entry->len = unmapped;
1050 list_add_tail(&entry->list, unmapped_list);
1058 * Sync if the number of fast-unmap regions hits the limit
1059 * or in case of errors.
1061 if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
1062 *unlocked += vfio_sync_unpin(dma, domain, unmapped_list,
1070 static size_t unmap_unpin_slow(struct vfio_domain *domain,
1071 struct vfio_dma *dma, dma_addr_t *iova,
1072 size_t len, phys_addr_t phys,
1075 size_t unmapped = iommu_unmap(domain->domain, *iova, len);
1078 *unlocked += vfio_unpin_pages_remote(dma, *iova,
1080 unmapped >> PAGE_SHIFT,
1088 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
1091 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
1092 struct vfio_domain *domain, *d;
1093 LIST_HEAD(unmapped_region_list);
1094 struct iommu_iotlb_gather iotlb_gather;
1095 int unmapped_region_cnt = 0;
1101 if (list_empty(&iommu->domain_list))
1105 * We use the IOMMU to track the physical addresses, otherwise we'd
1106 * need a much more complicated tracking system. Unfortunately that
1107 * means we need to use one of the iommu domains to figure out the
1108 * pfns to unpin. The rest need to be unmapped in advance so we have
1109 * no iommu translations remaining when the pages are unpinned.
1111 domain = d = list_first_entry(&iommu->domain_list,
1112 struct vfio_domain, next);
1114 list_for_each_entry_continue(d, &iommu->domain_list, next) {
1115 iommu_unmap(d->domain, dma->iova, dma->size);
1119 iommu_iotlb_gather_init(&iotlb_gather);
1120 while (iova < end) {
1121 size_t unmapped, len;
1122 phys_addr_t phys, next;
1124 phys = iommu_iova_to_phys(domain->domain, iova);
1125 if (WARN_ON(!phys)) {
1131 * To optimize for fewer iommu_unmap() calls, each of which
1132 * may require hardware cache flushing, try to find the
1133 * largest contiguous physical memory chunk to unmap.
1135 for (len = PAGE_SIZE;
1136 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
1137 next = iommu_iova_to_phys(domain->domain, iova + len);
1138 if (next != phys + len)
1143 * First, try to use fast unmap/unpin. In case of failure,
1144 * switch to slow unmap/unpin path.
1146 unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys,
1147 &unlocked, &unmapped_region_list,
1148 &unmapped_region_cnt,
1151 unmapped = unmap_unpin_slow(domain, dma, &iova, len,
1153 if (WARN_ON(!unmapped))
1158 dma->iommu_mapped = false;
1160 if (unmapped_region_cnt) {
1161 unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list,
1165 if (do_accounting) {
1166 vfio_lock_acct(dma, -unlocked, true);
1172 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
1174 WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list));
1175 vfio_unmap_unpin(iommu, dma, true);
1176 vfio_unlink_dma(iommu, dma);
1177 put_task_struct(dma->task);
1178 vfio_dma_bitmap_free(dma);
1179 if (dma->vaddr_invalid) {
1180 iommu->vaddr_invalid_count--;
1181 wake_up_all(&iommu->vaddr_wait);
1187 static void vfio_update_pgsize_bitmap(struct vfio_iommu *iommu)
1189 struct vfio_domain *domain;
1191 iommu->pgsize_bitmap = ULONG_MAX;
1193 list_for_each_entry(domain, &iommu->domain_list, next)
1194 iommu->pgsize_bitmap &= domain->domain->pgsize_bitmap;
1197 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
1198 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
1199 * That way the user will be able to map/unmap buffers whose size/
1200 * start address is aligned with PAGE_SIZE. Pinning code uses that
1201 * granularity while iommu driver can use the sub-PAGE_SIZE size
1202 * to map the buffer.
1204 if (iommu->pgsize_bitmap & ~PAGE_MASK) {
1205 iommu->pgsize_bitmap &= PAGE_MASK;
1206 iommu->pgsize_bitmap |= PAGE_SIZE;
1210 static int update_user_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1211 struct vfio_dma *dma, dma_addr_t base_iova,
1214 unsigned long pgshift = __ffs(pgsize);
1215 unsigned long nbits = dma->size >> pgshift;
1216 unsigned long bit_offset = (dma->iova - base_iova) >> pgshift;
1217 unsigned long copy_offset = bit_offset / BITS_PER_LONG;
1218 unsigned long shift = bit_offset % BITS_PER_LONG;
1219 unsigned long leftover;
1222 * mark all pages dirty if any IOMMU capable device is not able
1223 * to report dirty pages and all pages are pinned and mapped.
1225 if (iommu->num_non_pinned_groups && dma->iommu_mapped)
1226 bitmap_set(dma->bitmap, 0, nbits);
1229 bitmap_shift_left(dma->bitmap, dma->bitmap, shift,
1232 if (copy_from_user(&leftover,
1233 (void __user *)(bitmap + copy_offset),
1237 bitmap_or(dma->bitmap, dma->bitmap, &leftover, shift);
1240 if (copy_to_user((void __user *)(bitmap + copy_offset), dma->bitmap,
1241 DIRTY_BITMAP_BYTES(nbits + shift)))
1247 static int vfio_iova_dirty_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1248 dma_addr_t iova, size_t size, size_t pgsize)
1250 struct vfio_dma *dma;
1252 unsigned long pgshift = __ffs(pgsize);
1256 * GET_BITMAP request must fully cover vfio_dma mappings. Multiple
1257 * vfio_dma mappings may be clubbed by specifying large ranges, but
1258 * there must not be any previous mappings bisected by the range.
1259 * An error will be returned if these conditions are not met.
1261 dma = vfio_find_dma(iommu, iova, 1);
1262 if (dma && dma->iova != iova)
1265 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1266 if (dma && dma->iova + dma->size != iova + size)
1269 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1270 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1272 if (dma->iova < iova)
1275 if (dma->iova > iova + size - 1)
1278 ret = update_user_bitmap(bitmap, iommu, dma, iova, pgsize);
1283 * Re-populate bitmap to include all pinned pages which are
1284 * considered as dirty but exclude pages which are unpinned and
1285 * pages which are marked dirty by vfio_dma_rw()
1287 bitmap_clear(dma->bitmap, 0, dma->size >> pgshift);
1288 vfio_dma_populate_bitmap(dma, pgsize);
1293 static int verify_bitmap_size(uint64_t npages, uint64_t bitmap_size)
1295 if (!npages || !bitmap_size || (bitmap_size > DIRTY_BITMAP_SIZE_MAX) ||
1296 (bitmap_size < DIRTY_BITMAP_BYTES(npages)))
1302 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
1303 struct vfio_iommu_type1_dma_unmap *unmap,
1304 struct vfio_bitmap *bitmap)
1306 struct vfio_dma *dma, *dma_last = NULL;
1307 size_t unmapped = 0, pgsize;
1308 int ret = -EINVAL, retries = 0;
1309 unsigned long pgshift;
1310 dma_addr_t iova = unmap->iova;
1311 u64 size = unmap->size;
1312 bool unmap_all = unmap->flags & VFIO_DMA_UNMAP_FLAG_ALL;
1313 bool invalidate_vaddr = unmap->flags & VFIO_DMA_UNMAP_FLAG_VADDR;
1314 struct rb_node *n, *first_n;
1316 mutex_lock(&iommu->lock);
1318 pgshift = __ffs(iommu->pgsize_bitmap);
1319 pgsize = (size_t)1 << pgshift;
1321 if (iova & (pgsize - 1))
1328 } else if (!size || size & (pgsize - 1) ||
1329 iova + size - 1 < iova || size > SIZE_MAX) {
1333 /* When dirty tracking is enabled, allow only min supported pgsize */
1334 if ((unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
1335 (!iommu->dirty_page_tracking || (bitmap->pgsize != pgsize))) {
1339 WARN_ON((pgsize - 1) & PAGE_MASK);
1342 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
1343 * avoid tracking individual mappings. This means that the granularity
1344 * of the original mapping was lost and the user was allowed to attempt
1345 * to unmap any range. Depending on the contiguousness of physical
1346 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
1347 * or may not have worked. We only guaranteed unmap granularity
1348 * matching the original mapping; even though it was untracked here,
1349 * the original mappings are reflected in IOMMU mappings. This
1350 * resulted in a couple unusual behaviors. First, if a range is not
1351 * able to be unmapped, ex. a set of 4k pages that was mapped as a
1352 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
1353 * a zero sized unmap. Also, if an unmap request overlaps the first
1354 * address of a hugepage, the IOMMU will unmap the entire hugepage.
1355 * This also returns success and the returned unmap size reflects the
1356 * actual size unmapped.
1358 * We attempt to maintain compatibility with this "v1" interface, but
1359 * we take control out of the hands of the IOMMU. Therefore, an unmap
1360 * request offset from the beginning of the original mapping will
1361 * return success with zero sized unmap. And an unmap request covering
1362 * the first iova of mapping will unmap the entire range.
1364 * The v2 version of this interface intends to be more deterministic.
1365 * Unmap requests must fully cover previous mappings. Multiple
1366 * mappings may still be unmaped by specifying large ranges, but there
1367 * must not be any previous mappings bisected by the range. An error
1368 * will be returned if these conditions are not met. The v2 interface
1369 * will only return success and a size of zero if there were no
1370 * mappings within the range.
1372 if (iommu->v2 && !unmap_all) {
1373 dma = vfio_find_dma(iommu, iova, 1);
1374 if (dma && dma->iova != iova)
1377 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1378 if (dma && dma->iova + dma->size != iova + size)
1383 n = first_n = vfio_find_dma_first_node(iommu, iova, size);
1386 dma = rb_entry(n, struct vfio_dma, node);
1387 if (dma->iova >= iova + size)
1390 if (!iommu->v2 && iova > dma->iova)
1393 * Task with same address space who mapped this iova range is
1394 * allowed to unmap the iova range.
1396 if (dma->task->mm != current->mm)
1399 if (invalidate_vaddr) {
1400 if (dma->vaddr_invalid) {
1401 struct rb_node *last_n = n;
1403 for (n = first_n; n != last_n; n = rb_next(n)) {
1405 struct vfio_dma, node);
1406 dma->vaddr_invalid = false;
1407 iommu->vaddr_invalid_count--;
1413 dma->vaddr_invalid = true;
1414 iommu->vaddr_invalid_count++;
1415 unmapped += dma->size;
1420 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
1421 struct vfio_iommu_type1_dma_unmap nb_unmap;
1423 if (dma_last == dma) {
1424 BUG_ON(++retries > 10);
1430 nb_unmap.iova = dma->iova;
1431 nb_unmap.size = dma->size;
1434 * Notify anyone (mdev vendor drivers) to invalidate and
1435 * unmap iovas within the range we're about to unmap.
1436 * Vendor drivers MUST unpin pages in response to an
1439 mutex_unlock(&iommu->lock);
1440 blocking_notifier_call_chain(&iommu->notifier,
1441 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
1443 mutex_lock(&iommu->lock);
1447 if (unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
1448 ret = update_user_bitmap(bitmap->data, iommu, dma,
1454 unmapped += dma->size;
1456 vfio_remove_dma(iommu, dma);
1460 mutex_unlock(&iommu->lock);
1462 /* Report how much was unmapped */
1463 unmap->size = unmapped;
1468 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
1469 unsigned long pfn, long npage, int prot)
1471 struct vfio_domain *d;
1474 list_for_each_entry(d, &iommu->domain_list, next) {
1475 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
1476 npage << PAGE_SHIFT, prot | IOMMU_CACHE);
1486 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) {
1487 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
1494 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
1497 dma_addr_t iova = dma->iova;
1498 unsigned long vaddr = dma->vaddr;
1499 struct vfio_batch batch;
1500 size_t size = map_size;
1502 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1505 vfio_batch_init(&batch);
1508 /* Pin a contiguous chunk of memory */
1509 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
1510 size >> PAGE_SHIFT, &pfn, limit,
1519 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
1522 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
1524 vfio_batch_unpin(&batch, dma);
1528 size -= npage << PAGE_SHIFT;
1529 dma->size += npage << PAGE_SHIFT;
1532 vfio_batch_fini(&batch);
1533 dma->iommu_mapped = true;
1536 vfio_remove_dma(iommu, dma);
1542 * Check dma map request is within a valid iova range
1544 static bool vfio_iommu_iova_dma_valid(struct vfio_iommu *iommu,
1545 dma_addr_t start, dma_addr_t end)
1547 struct list_head *iova = &iommu->iova_list;
1548 struct vfio_iova *node;
1550 list_for_each_entry(node, iova, list) {
1551 if (start >= node->start && end <= node->end)
1556 * Check for list_empty() as well since a container with
1557 * a single mdev device will have an empty list.
1559 return list_empty(iova);
1562 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1563 struct vfio_iommu_type1_dma_map *map)
1565 bool set_vaddr = map->flags & VFIO_DMA_MAP_FLAG_VADDR;
1566 dma_addr_t iova = map->iova;
1567 unsigned long vaddr = map->vaddr;
1568 size_t size = map->size;
1569 int ret = 0, prot = 0;
1571 struct vfio_dma *dma;
1573 /* Verify that none of our __u64 fields overflow */
1574 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1577 /* READ/WRITE from device perspective */
1578 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1579 prot |= IOMMU_WRITE;
1580 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1583 if ((prot && set_vaddr) || (!prot && !set_vaddr))
1586 mutex_lock(&iommu->lock);
1588 pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
1590 WARN_ON((pgsize - 1) & PAGE_MASK);
1592 if (!size || (size | iova | vaddr) & (pgsize - 1)) {
1597 /* Don't allow IOVA or virtual address wrap */
1598 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) {
1603 dma = vfio_find_dma(iommu, iova, size);
1607 } else if (!dma->vaddr_invalid || dma->iova != iova ||
1608 dma->size != size) {
1612 dma->vaddr_invalid = false;
1613 iommu->vaddr_invalid_count--;
1614 wake_up_all(&iommu->vaddr_wait);
1622 if (!iommu->dma_avail) {
1627 if (!vfio_iommu_iova_dma_valid(iommu, iova, iova + size - 1)) {
1632 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1644 * We need to be able to both add to a task's locked memory and test
1645 * against the locked memory limit and we need to be able to do both
1646 * outside of this call path as pinning can be asynchronous via the
1647 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1648 * task_struct and VM locked pages requires an mm_struct, however
1649 * holding an indefinite mm reference is not recommended, therefore we
1650 * only hold a reference to a task. We could hold a reference to
1651 * current, however QEMU uses this call path through vCPU threads,
1652 * which can be killed resulting in a NULL mm and failure in the unmap
1653 * path when called via a different thread. Avoid this problem by
1654 * using the group_leader as threads within the same group require
1655 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1658 * Previously we also used the task for testing CAP_IPC_LOCK at the
1659 * time of pinning and accounting, however has_capability() makes use
1660 * of real_cred, a copy-on-write field, so we can't guarantee that it
1661 * matches group_leader, or in fact that it might not change by the
1662 * time it's evaluated. If a process were to call MAP_DMA with
1663 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1664 * possibly see different results for an iommu_mapped vfio_dma vs
1665 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1666 * time of calling MAP_DMA.
1668 get_task_struct(current->group_leader);
1669 dma->task = current->group_leader;
1670 dma->lock_cap = capable(CAP_IPC_LOCK);
1672 dma->pfn_list = RB_ROOT;
1674 /* Insert zero-sized and grow as we map chunks of it */
1675 vfio_link_dma(iommu, dma);
1677 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1678 if (list_empty(&iommu->domain_list))
1681 ret = vfio_pin_map_dma(iommu, dma, size);
1683 if (!ret && iommu->dirty_page_tracking) {
1684 ret = vfio_dma_bitmap_alloc(dma, pgsize);
1686 vfio_remove_dma(iommu, dma);
1690 mutex_unlock(&iommu->lock);
1694 static int vfio_bus_type(struct device *dev, void *data)
1696 struct bus_type **bus = data;
1698 if (*bus && *bus != dev->bus)
1706 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1707 struct vfio_domain *domain)
1709 struct vfio_batch batch;
1710 struct vfio_domain *d = NULL;
1712 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1715 ret = vfio_wait_all_valid(iommu);
1719 /* Arbitrarily pick the first domain in the list for lookups */
1720 if (!list_empty(&iommu->domain_list))
1721 d = list_first_entry(&iommu->domain_list,
1722 struct vfio_domain, next);
1724 vfio_batch_init(&batch);
1726 n = rb_first(&iommu->dma_list);
1728 for (; n; n = rb_next(n)) {
1729 struct vfio_dma *dma;
1732 dma = rb_entry(n, struct vfio_dma, node);
1735 while (iova < dma->iova + dma->size) {
1739 if (dma->iommu_mapped) {
1743 if (WARN_ON(!d)) { /* mapped w/o a domain?! */
1748 phys = iommu_iova_to_phys(d->domain, iova);
1750 if (WARN_ON(!phys)) {
1758 while (i < dma->iova + dma->size &&
1759 p == iommu_iova_to_phys(d->domain, i)) {
1766 unsigned long vaddr = dma->vaddr +
1768 size_t n = dma->iova + dma->size - iova;
1771 npage = vfio_pin_pages_remote(dma, vaddr,
1781 phys = pfn << PAGE_SHIFT;
1782 size = npage << PAGE_SHIFT;
1785 ret = iommu_map(domain->domain, iova, phys,
1786 size, dma->prot | IOMMU_CACHE);
1788 if (!dma->iommu_mapped) {
1789 vfio_unpin_pages_remote(dma, iova,
1793 vfio_batch_unpin(&batch, dma);
1802 /* All dmas are now mapped, defer to second tree walk for unwind */
1803 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1804 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1806 dma->iommu_mapped = true;
1809 vfio_batch_fini(&batch);
1813 for (; n; n = rb_prev(n)) {
1814 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1817 if (dma->iommu_mapped) {
1818 iommu_unmap(domain->domain, dma->iova, dma->size);
1823 while (iova < dma->iova + dma->size) {
1824 phys_addr_t phys, p;
1828 phys = iommu_iova_to_phys(domain->domain, iova);
1837 while (i < dma->iova + dma->size &&
1838 p == iommu_iova_to_phys(domain->domain, i)) {
1844 iommu_unmap(domain->domain, iova, size);
1845 vfio_unpin_pages_remote(dma, iova, phys >> PAGE_SHIFT,
1846 size >> PAGE_SHIFT, true);
1850 vfio_batch_fini(&batch);
1855 * We change our unmap behavior slightly depending on whether the IOMMU
1856 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1857 * for practically any contiguous power-of-two mapping we give it. This means
1858 * we don't need to look for contiguous chunks ourselves to make unmapping
1859 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1860 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1861 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1862 * hugetlbfs is in use.
1864 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1867 int ret, order = get_order(PAGE_SIZE * 2);
1869 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1873 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1874 IOMMU_READ | IOMMU_WRITE | IOMMU_CACHE);
1876 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1878 if (unmapped == PAGE_SIZE)
1879 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1881 domain->fgsp = true;
1884 __free_pages(pages, order);
1887 static struct vfio_iommu_group *find_iommu_group(struct vfio_domain *domain,
1888 struct iommu_group *iommu_group)
1890 struct vfio_iommu_group *g;
1892 list_for_each_entry(g, &domain->group_list, next) {
1893 if (g->iommu_group == iommu_group)
1900 static struct vfio_iommu_group*
1901 vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
1902 struct iommu_group *iommu_group)
1904 struct vfio_iommu_group *group;
1905 struct vfio_domain *domain;
1907 list_for_each_entry(domain, &iommu->domain_list, next) {
1908 group = find_iommu_group(domain, iommu_group);
1913 list_for_each_entry(group, &iommu->emulated_iommu_groups, next)
1914 if (group->iommu_group == iommu_group)
1919 static bool vfio_iommu_has_sw_msi(struct list_head *group_resv_regions,
1922 struct iommu_resv_region *region;
1925 list_for_each_entry(region, group_resv_regions, list) {
1927 * The presence of any 'real' MSI regions should take
1928 * precedence over the software-managed one if the
1929 * IOMMU driver happens to advertise both types.
1931 if (region->type == IOMMU_RESV_MSI) {
1936 if (region->type == IOMMU_RESV_SW_MSI) {
1937 *base = region->start;
1946 * This is a helper function to insert an address range to iova list.
1947 * The list is initially created with a single entry corresponding to
1948 * the IOMMU domain geometry to which the device group is attached.
1949 * The list aperture gets modified when a new domain is added to the
1950 * container if the new aperture doesn't conflict with the current one
1951 * or with any existing dma mappings. The list is also modified to
1952 * exclude any reserved regions associated with the device group.
1954 static int vfio_iommu_iova_insert(struct list_head *head,
1955 dma_addr_t start, dma_addr_t end)
1957 struct vfio_iova *region;
1959 region = kmalloc(sizeof(*region), GFP_KERNEL);
1963 INIT_LIST_HEAD(®ion->list);
1964 region->start = start;
1967 list_add_tail(®ion->list, head);
1972 * Check the new iommu aperture conflicts with existing aper or with any
1973 * existing dma mappings.
1975 static bool vfio_iommu_aper_conflict(struct vfio_iommu *iommu,
1976 dma_addr_t start, dma_addr_t end)
1978 struct vfio_iova *first, *last;
1979 struct list_head *iova = &iommu->iova_list;
1981 if (list_empty(iova))
1984 /* Disjoint sets, return conflict */
1985 first = list_first_entry(iova, struct vfio_iova, list);
1986 last = list_last_entry(iova, struct vfio_iova, list);
1987 if (start > last->end || end < first->start)
1990 /* Check for any existing dma mappings below the new start */
1991 if (start > first->start) {
1992 if (vfio_find_dma(iommu, first->start, start - first->start))
1996 /* Check for any existing dma mappings beyond the new end */
1997 if (end < last->end) {
1998 if (vfio_find_dma(iommu, end + 1, last->end - end))
2006 * Resize iommu iova aperture window. This is called only if the new
2007 * aperture has no conflict with existing aperture and dma mappings.
2009 static int vfio_iommu_aper_resize(struct list_head *iova,
2010 dma_addr_t start, dma_addr_t end)
2012 struct vfio_iova *node, *next;
2014 if (list_empty(iova))
2015 return vfio_iommu_iova_insert(iova, start, end);
2017 /* Adjust iova list start */
2018 list_for_each_entry_safe(node, next, iova, list) {
2019 if (start < node->start)
2021 if (start >= node->start && start < node->end) {
2022 node->start = start;
2025 /* Delete nodes before new start */
2026 list_del(&node->list);
2030 /* Adjust iova list end */
2031 list_for_each_entry_safe(node, next, iova, list) {
2032 if (end > node->end)
2034 if (end > node->start && end <= node->end) {
2038 /* Delete nodes after new end */
2039 list_del(&node->list);
2047 * Check reserved region conflicts with existing dma mappings
2049 static bool vfio_iommu_resv_conflict(struct vfio_iommu *iommu,
2050 struct list_head *resv_regions)
2052 struct iommu_resv_region *region;
2054 /* Check for conflict with existing dma mappings */
2055 list_for_each_entry(region, resv_regions, list) {
2056 if (region->type == IOMMU_RESV_DIRECT_RELAXABLE)
2059 if (vfio_find_dma(iommu, region->start, region->length))
2067 * Check iova region overlap with reserved regions and
2068 * exclude them from the iommu iova range
2070 static int vfio_iommu_resv_exclude(struct list_head *iova,
2071 struct list_head *resv_regions)
2073 struct iommu_resv_region *resv;
2074 struct vfio_iova *n, *next;
2076 list_for_each_entry(resv, resv_regions, list) {
2077 phys_addr_t start, end;
2079 if (resv->type == IOMMU_RESV_DIRECT_RELAXABLE)
2082 start = resv->start;
2083 end = resv->start + resv->length - 1;
2085 list_for_each_entry_safe(n, next, iova, list) {
2089 if (start > n->end || end < n->start)
2092 * Insert a new node if current node overlaps with the
2093 * reserve region to exclude that from valid iova range.
2094 * Note that, new node is inserted before the current
2095 * node and finally the current node is deleted keeping
2096 * the list updated and sorted.
2098 if (start > n->start)
2099 ret = vfio_iommu_iova_insert(&n->list, n->start,
2101 if (!ret && end < n->end)
2102 ret = vfio_iommu_iova_insert(&n->list, end + 1,
2112 if (list_empty(iova))
2118 static void vfio_iommu_resv_free(struct list_head *resv_regions)
2120 struct iommu_resv_region *n, *next;
2122 list_for_each_entry_safe(n, next, resv_regions, list) {
2128 static void vfio_iommu_iova_free(struct list_head *iova)
2130 struct vfio_iova *n, *next;
2132 list_for_each_entry_safe(n, next, iova, list) {
2138 static int vfio_iommu_iova_get_copy(struct vfio_iommu *iommu,
2139 struct list_head *iova_copy)
2141 struct list_head *iova = &iommu->iova_list;
2142 struct vfio_iova *n;
2145 list_for_each_entry(n, iova, list) {
2146 ret = vfio_iommu_iova_insert(iova_copy, n->start, n->end);
2154 vfio_iommu_iova_free(iova_copy);
2158 static void vfio_iommu_iova_insert_copy(struct vfio_iommu *iommu,
2159 struct list_head *iova_copy)
2161 struct list_head *iova = &iommu->iova_list;
2163 vfio_iommu_iova_free(iova);
2165 list_splice_tail(iova_copy, iova);
2168 static int vfio_iommu_type1_attach_group(void *iommu_data,
2169 struct iommu_group *iommu_group, enum vfio_group_type type)
2171 struct vfio_iommu *iommu = iommu_data;
2172 struct vfio_iommu_group *group;
2173 struct vfio_domain *domain, *d;
2174 struct bus_type *bus = NULL;
2175 bool resv_msi, msi_remap;
2176 phys_addr_t resv_msi_base = 0;
2177 struct iommu_domain_geometry *geo;
2178 LIST_HEAD(iova_copy);
2179 LIST_HEAD(group_resv_regions);
2182 mutex_lock(&iommu->lock);
2184 /* Check for duplicates */
2185 if (vfio_iommu_find_iommu_group(iommu, iommu_group))
2189 group = kzalloc(sizeof(*group), GFP_KERNEL);
2192 group->iommu_group = iommu_group;
2194 if (type == VFIO_EMULATED_IOMMU) {
2195 list_add(&group->next, &iommu->emulated_iommu_groups);
2197 * An emulated IOMMU group cannot dirty memory directly, it can
2198 * only use interfaces that provide dirty tracking.
2199 * The iommu scope can only be promoted with the addition of a
2200 * dirty tracking group.
2202 group->pinned_page_dirty_scope = true;
2207 /* Determine bus_type in order to allocate a domain */
2208 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
2210 goto out_free_group;
2213 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2215 goto out_free_group;
2218 domain->domain = iommu_domain_alloc(bus);
2219 if (!domain->domain)
2220 goto out_free_domain;
2222 if (iommu->nesting) {
2223 ret = iommu_enable_nesting(domain->domain);
2228 ret = iommu_attach_group(domain->domain, group->iommu_group);
2232 /* Get aperture info */
2233 geo = &domain->domain->geometry;
2234 if (vfio_iommu_aper_conflict(iommu, geo->aperture_start,
2235 geo->aperture_end)) {
2240 ret = iommu_get_group_resv_regions(iommu_group, &group_resv_regions);
2244 if (vfio_iommu_resv_conflict(iommu, &group_resv_regions)) {
2250 * We don't want to work on the original iova list as the list
2251 * gets modified and in case of failure we have to retain the
2252 * original list. Get a copy here.
2254 ret = vfio_iommu_iova_get_copy(iommu, &iova_copy);
2258 ret = vfio_iommu_aper_resize(&iova_copy, geo->aperture_start,
2263 ret = vfio_iommu_resv_exclude(&iova_copy, &group_resv_regions);
2267 resv_msi = vfio_iommu_has_sw_msi(&group_resv_regions, &resv_msi_base);
2269 INIT_LIST_HEAD(&domain->group_list);
2270 list_add(&group->next, &domain->group_list);
2272 msi_remap = irq_domain_check_msi_remap() ||
2273 iommu_capable(bus, IOMMU_CAP_INTR_REMAP);
2275 if (!allow_unsafe_interrupts && !msi_remap) {
2276 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
2283 * If the IOMMU can block non-coherent operations (ie PCIe TLPs with
2284 * no-snoop set) then VFIO always turns this feature on because on Intel
2285 * platforms it optimizes KVM to disable wbinvd emulation.
2287 if (domain->domain->ops->enforce_cache_coherency)
2288 domain->enforce_cache_coherency =
2289 domain->domain->ops->enforce_cache_coherency(
2293 * Try to match an existing compatible domain. We don't want to
2294 * preclude an IOMMU driver supporting multiple bus_types and being
2295 * able to include different bus_types in the same IOMMU domain, so
2296 * we test whether the domains use the same iommu_ops rather than
2297 * testing if they're on the same bus_type.
2299 list_for_each_entry(d, &iommu->domain_list, next) {
2300 if (d->domain->ops == domain->domain->ops &&
2301 d->enforce_cache_coherency ==
2302 domain->enforce_cache_coherency) {
2303 iommu_detach_group(domain->domain, group->iommu_group);
2304 if (!iommu_attach_group(d->domain,
2305 group->iommu_group)) {
2306 list_add(&group->next, &d->group_list);
2307 iommu_domain_free(domain->domain);
2312 ret = iommu_attach_group(domain->domain,
2313 group->iommu_group);
2319 vfio_test_domain_fgsp(domain);
2321 /* replay mappings on new domains */
2322 ret = vfio_iommu_replay(iommu, domain);
2327 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
2328 if (ret && ret != -ENODEV)
2332 list_add(&domain->next, &iommu->domain_list);
2333 vfio_update_pgsize_bitmap(iommu);
2335 /* Delete the old one and insert new iova list */
2336 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2339 * An iommu backed group can dirty memory directly and therefore
2340 * demotes the iommu scope until it declares itself dirty tracking
2341 * capable via the page pinning interface.
2343 iommu->num_non_pinned_groups++;
2344 mutex_unlock(&iommu->lock);
2345 vfio_iommu_resv_free(&group_resv_regions);
2350 iommu_detach_group(domain->domain, group->iommu_group);
2352 iommu_domain_free(domain->domain);
2353 vfio_iommu_iova_free(&iova_copy);
2354 vfio_iommu_resv_free(&group_resv_regions);
2360 mutex_unlock(&iommu->lock);
2364 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
2366 struct rb_node *node;
2368 while ((node = rb_first(&iommu->dma_list)))
2369 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
2372 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
2374 struct rb_node *n, *p;
2376 n = rb_first(&iommu->dma_list);
2377 for (; n; n = rb_next(n)) {
2378 struct vfio_dma *dma;
2379 long locked = 0, unlocked = 0;
2381 dma = rb_entry(n, struct vfio_dma, node);
2382 unlocked += vfio_unmap_unpin(iommu, dma, false);
2383 p = rb_first(&dma->pfn_list);
2384 for (; p; p = rb_next(p)) {
2385 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
2388 if (!is_invalid_reserved_pfn(vpfn->pfn))
2391 vfio_lock_acct(dma, locked - unlocked, true);
2396 * Called when a domain is removed in detach. It is possible that
2397 * the removed domain decided the iova aperture window. Modify the
2398 * iova aperture with the smallest window among existing domains.
2400 static void vfio_iommu_aper_expand(struct vfio_iommu *iommu,
2401 struct list_head *iova_copy)
2403 struct vfio_domain *domain;
2404 struct vfio_iova *node;
2405 dma_addr_t start = 0;
2406 dma_addr_t end = (dma_addr_t)~0;
2408 if (list_empty(iova_copy))
2411 list_for_each_entry(domain, &iommu->domain_list, next) {
2412 struct iommu_domain_geometry *geo = &domain->domain->geometry;
2414 if (geo->aperture_start > start)
2415 start = geo->aperture_start;
2416 if (geo->aperture_end < end)
2417 end = geo->aperture_end;
2420 /* Modify aperture limits. The new aper is either same or bigger */
2421 node = list_first_entry(iova_copy, struct vfio_iova, list);
2422 node->start = start;
2423 node = list_last_entry(iova_copy, struct vfio_iova, list);
2428 * Called when a group is detached. The reserved regions for that
2429 * group can be part of valid iova now. But since reserved regions
2430 * may be duplicated among groups, populate the iova valid regions
2433 static int vfio_iommu_resv_refresh(struct vfio_iommu *iommu,
2434 struct list_head *iova_copy)
2436 struct vfio_domain *d;
2437 struct vfio_iommu_group *g;
2438 struct vfio_iova *node;
2439 dma_addr_t start, end;
2440 LIST_HEAD(resv_regions);
2443 if (list_empty(iova_copy))
2446 list_for_each_entry(d, &iommu->domain_list, next) {
2447 list_for_each_entry(g, &d->group_list, next) {
2448 ret = iommu_get_group_resv_regions(g->iommu_group,
2455 node = list_first_entry(iova_copy, struct vfio_iova, list);
2456 start = node->start;
2457 node = list_last_entry(iova_copy, struct vfio_iova, list);
2460 /* purge the iova list and create new one */
2461 vfio_iommu_iova_free(iova_copy);
2463 ret = vfio_iommu_aper_resize(iova_copy, start, end);
2467 /* Exclude current reserved regions from iova ranges */
2468 ret = vfio_iommu_resv_exclude(iova_copy, &resv_regions);
2470 vfio_iommu_resv_free(&resv_regions);
2474 static void vfio_iommu_type1_detach_group(void *iommu_data,
2475 struct iommu_group *iommu_group)
2477 struct vfio_iommu *iommu = iommu_data;
2478 struct vfio_domain *domain;
2479 struct vfio_iommu_group *group;
2480 bool update_dirty_scope = false;
2481 LIST_HEAD(iova_copy);
2483 mutex_lock(&iommu->lock);
2484 list_for_each_entry(group, &iommu->emulated_iommu_groups, next) {
2485 if (group->iommu_group != iommu_group)
2487 update_dirty_scope = !group->pinned_page_dirty_scope;
2488 list_del(&group->next);
2491 if (list_empty(&iommu->emulated_iommu_groups) &&
2492 list_empty(&iommu->domain_list)) {
2493 WARN_ON(iommu->notifier.head);
2494 vfio_iommu_unmap_unpin_all(iommu);
2496 goto detach_group_done;
2500 * Get a copy of iova list. This will be used to update
2501 * and to replace the current one later. Please note that
2502 * we will leave the original list as it is if update fails.
2504 vfio_iommu_iova_get_copy(iommu, &iova_copy);
2506 list_for_each_entry(domain, &iommu->domain_list, next) {
2507 group = find_iommu_group(domain, iommu_group);
2511 iommu_detach_group(domain->domain, group->iommu_group);
2512 update_dirty_scope = !group->pinned_page_dirty_scope;
2513 list_del(&group->next);
2516 * Group ownership provides privilege, if the group list is
2517 * empty, the domain goes away. If it's the last domain with
2518 * iommu and external domain doesn't exist, then all the
2519 * mappings go away too. If it's the last domain with iommu and
2520 * external domain exist, update accounting
2522 if (list_empty(&domain->group_list)) {
2523 if (list_is_singular(&iommu->domain_list)) {
2524 if (list_empty(&iommu->emulated_iommu_groups)) {
2525 WARN_ON(iommu->notifier.head);
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 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
2587 init_waitqueue_head(&iommu->vaddr_wait);
2588 iommu->pgsize_bitmap = PAGE_MASK;
2589 INIT_LIST_HEAD(&iommu->emulated_iommu_groups);
2594 static void vfio_release_domain(struct vfio_domain *domain)
2596 struct vfio_iommu_group *group, *group_tmp;
2598 list_for_each_entry_safe(group, group_tmp,
2599 &domain->group_list, next) {
2600 iommu_detach_group(domain->domain, group->iommu_group);
2601 list_del(&group->next);
2605 iommu_domain_free(domain->domain);
2608 static void vfio_iommu_type1_release(void *iommu_data)
2610 struct vfio_iommu *iommu = iommu_data;
2611 struct vfio_domain *domain, *domain_tmp;
2612 struct vfio_iommu_group *group, *next_group;
2614 list_for_each_entry_safe(group, next_group,
2615 &iommu->emulated_iommu_groups, next) {
2616 list_del(&group->next);
2620 vfio_iommu_unmap_unpin_all(iommu);
2622 list_for_each_entry_safe(domain, domain_tmp,
2623 &iommu->domain_list, next) {
2624 vfio_release_domain(domain);
2625 list_del(&domain->next);
2629 vfio_iommu_iova_free(&iommu->iova_list);
2634 static int vfio_domains_have_enforce_cache_coherency(struct vfio_iommu *iommu)
2636 struct vfio_domain *domain;
2639 mutex_lock(&iommu->lock);
2640 list_for_each_entry(domain, &iommu->domain_list, next) {
2641 if (!(domain->enforce_cache_coherency)) {
2646 mutex_unlock(&iommu->lock);
2651 static int vfio_iommu_type1_check_extension(struct vfio_iommu *iommu,
2655 case VFIO_TYPE1_IOMMU:
2656 case VFIO_TYPE1v2_IOMMU:
2657 case VFIO_TYPE1_NESTING_IOMMU:
2658 case VFIO_UNMAP_ALL:
2659 case VFIO_UPDATE_VADDR:
2661 case VFIO_DMA_CC_IOMMU:
2664 return vfio_domains_have_enforce_cache_coherency(iommu);
2670 static int vfio_iommu_iova_add_cap(struct vfio_info_cap *caps,
2671 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas,
2674 struct vfio_info_cap_header *header;
2675 struct vfio_iommu_type1_info_cap_iova_range *iova_cap;
2677 header = vfio_info_cap_add(caps, size,
2678 VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE, 1);
2680 return PTR_ERR(header);
2682 iova_cap = container_of(header,
2683 struct vfio_iommu_type1_info_cap_iova_range,
2685 iova_cap->nr_iovas = cap_iovas->nr_iovas;
2686 memcpy(iova_cap->iova_ranges, cap_iovas->iova_ranges,
2687 cap_iovas->nr_iovas * sizeof(*cap_iovas->iova_ranges));
2691 static int vfio_iommu_iova_build_caps(struct vfio_iommu *iommu,
2692 struct vfio_info_cap *caps)
2694 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas;
2695 struct vfio_iova *iova;
2697 int iovas = 0, i = 0, ret;
2699 list_for_each_entry(iova, &iommu->iova_list, list)
2704 * Return 0 as a container with a single mdev device
2705 * will have an empty list
2710 size = struct_size(cap_iovas, iova_ranges, iovas);
2712 cap_iovas = kzalloc(size, GFP_KERNEL);
2716 cap_iovas->nr_iovas = iovas;
2718 list_for_each_entry(iova, &iommu->iova_list, list) {
2719 cap_iovas->iova_ranges[i].start = iova->start;
2720 cap_iovas->iova_ranges[i].end = iova->end;
2724 ret = vfio_iommu_iova_add_cap(caps, cap_iovas, size);
2730 static int vfio_iommu_migration_build_caps(struct vfio_iommu *iommu,
2731 struct vfio_info_cap *caps)
2733 struct vfio_iommu_type1_info_cap_migration cap_mig;
2735 cap_mig.header.id = VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION;
2736 cap_mig.header.version = 1;
2739 /* support minimum pgsize */
2740 cap_mig.pgsize_bitmap = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2741 cap_mig.max_dirty_bitmap_size = DIRTY_BITMAP_SIZE_MAX;
2743 return vfio_info_add_capability(caps, &cap_mig.header, sizeof(cap_mig));
2746 static int vfio_iommu_dma_avail_build_caps(struct vfio_iommu *iommu,
2747 struct vfio_info_cap *caps)
2749 struct vfio_iommu_type1_info_dma_avail cap_dma_avail;
2751 cap_dma_avail.header.id = VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL;
2752 cap_dma_avail.header.version = 1;
2754 cap_dma_avail.avail = iommu->dma_avail;
2756 return vfio_info_add_capability(caps, &cap_dma_avail.header,
2757 sizeof(cap_dma_avail));
2760 static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu,
2763 struct vfio_iommu_type1_info info;
2764 unsigned long minsz;
2765 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
2766 unsigned long capsz;
2769 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
2771 /* For backward compatibility, cannot require this */
2772 capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
2774 if (copy_from_user(&info, (void __user *)arg, minsz))
2777 if (info.argsz < minsz)
2780 if (info.argsz >= capsz) {
2782 info.cap_offset = 0; /* output, no-recopy necessary */
2785 mutex_lock(&iommu->lock);
2786 info.flags = VFIO_IOMMU_INFO_PGSIZES;
2788 info.iova_pgsizes = iommu->pgsize_bitmap;
2790 ret = vfio_iommu_migration_build_caps(iommu, &caps);
2793 ret = vfio_iommu_dma_avail_build_caps(iommu, &caps);
2796 ret = vfio_iommu_iova_build_caps(iommu, &caps);
2798 mutex_unlock(&iommu->lock);
2804 info.flags |= VFIO_IOMMU_INFO_CAPS;
2806 if (info.argsz < sizeof(info) + caps.size) {
2807 info.argsz = sizeof(info) + caps.size;
2809 vfio_info_cap_shift(&caps, sizeof(info));
2810 if (copy_to_user((void __user *)arg +
2811 sizeof(info), caps.buf,
2816 info.cap_offset = sizeof(info);
2822 return copy_to_user((void __user *)arg, &info, minsz) ?
2826 static int vfio_iommu_type1_map_dma(struct vfio_iommu *iommu,
2829 struct vfio_iommu_type1_dma_map map;
2830 unsigned long minsz;
2831 uint32_t mask = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE |
2832 VFIO_DMA_MAP_FLAG_VADDR;
2834 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
2836 if (copy_from_user(&map, (void __user *)arg, minsz))
2839 if (map.argsz < minsz || map.flags & ~mask)
2842 return vfio_dma_do_map(iommu, &map);
2845 static int vfio_iommu_type1_unmap_dma(struct vfio_iommu *iommu,
2848 struct vfio_iommu_type1_dma_unmap unmap;
2849 struct vfio_bitmap bitmap = { 0 };
2850 uint32_t mask = VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP |
2851 VFIO_DMA_UNMAP_FLAG_VADDR |
2852 VFIO_DMA_UNMAP_FLAG_ALL;
2853 unsigned long minsz;
2856 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
2858 if (copy_from_user(&unmap, (void __user *)arg, minsz))
2861 if (unmap.argsz < minsz || unmap.flags & ~mask)
2864 if ((unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
2865 (unmap.flags & (VFIO_DMA_UNMAP_FLAG_ALL |
2866 VFIO_DMA_UNMAP_FLAG_VADDR)))
2869 if (unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
2870 unsigned long pgshift;
2872 if (unmap.argsz < (minsz + sizeof(bitmap)))
2875 if (copy_from_user(&bitmap,
2876 (void __user *)(arg + minsz),
2880 if (!access_ok((void __user *)bitmap.data, bitmap.size))
2883 pgshift = __ffs(bitmap.pgsize);
2884 ret = verify_bitmap_size(unmap.size >> pgshift,
2890 ret = vfio_dma_do_unmap(iommu, &unmap, &bitmap);
2894 return copy_to_user((void __user *)arg, &unmap, minsz) ?
2898 static int vfio_iommu_type1_dirty_pages(struct vfio_iommu *iommu,
2901 struct vfio_iommu_type1_dirty_bitmap dirty;
2902 uint32_t mask = VFIO_IOMMU_DIRTY_PAGES_FLAG_START |
2903 VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP |
2904 VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP;
2905 unsigned long minsz;
2911 minsz = offsetofend(struct vfio_iommu_type1_dirty_bitmap, flags);
2913 if (copy_from_user(&dirty, (void __user *)arg, minsz))
2916 if (dirty.argsz < minsz || dirty.flags & ~mask)
2919 /* only one flag should be set at a time */
2920 if (__ffs(dirty.flags) != __fls(dirty.flags))
2923 if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_START) {
2926 mutex_lock(&iommu->lock);
2927 pgsize = 1 << __ffs(iommu->pgsize_bitmap);
2928 if (!iommu->dirty_page_tracking) {
2929 ret = vfio_dma_bitmap_alloc_all(iommu, pgsize);
2931 iommu->dirty_page_tracking = true;
2933 mutex_unlock(&iommu->lock);
2935 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP) {
2936 mutex_lock(&iommu->lock);
2937 if (iommu->dirty_page_tracking) {
2938 iommu->dirty_page_tracking = false;
2939 vfio_dma_bitmap_free_all(iommu);
2941 mutex_unlock(&iommu->lock);
2943 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP) {
2944 struct vfio_iommu_type1_dirty_bitmap_get range;
2945 unsigned long pgshift;
2946 size_t data_size = dirty.argsz - minsz;
2947 size_t iommu_pgsize;
2949 if (!data_size || data_size < sizeof(range))
2952 if (copy_from_user(&range, (void __user *)(arg + minsz),
2956 if (range.iova + range.size < range.iova)
2958 if (!access_ok((void __user *)range.bitmap.data,
2962 pgshift = __ffs(range.bitmap.pgsize);
2963 ret = verify_bitmap_size(range.size >> pgshift,
2968 mutex_lock(&iommu->lock);
2970 iommu_pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2972 /* allow only smallest supported pgsize */
2973 if (range.bitmap.pgsize != iommu_pgsize) {
2977 if (range.iova & (iommu_pgsize - 1)) {
2981 if (!range.size || range.size & (iommu_pgsize - 1)) {
2986 if (iommu->dirty_page_tracking)
2987 ret = vfio_iova_dirty_bitmap(range.bitmap.data,
2990 range.bitmap.pgsize);
2994 mutex_unlock(&iommu->lock);
3002 static long vfio_iommu_type1_ioctl(void *iommu_data,
3003 unsigned int cmd, unsigned long arg)
3005 struct vfio_iommu *iommu = iommu_data;
3008 case VFIO_CHECK_EXTENSION:
3009 return vfio_iommu_type1_check_extension(iommu, arg);
3010 case VFIO_IOMMU_GET_INFO:
3011 return vfio_iommu_type1_get_info(iommu, arg);
3012 case VFIO_IOMMU_MAP_DMA:
3013 return vfio_iommu_type1_map_dma(iommu, arg);
3014 case VFIO_IOMMU_UNMAP_DMA:
3015 return vfio_iommu_type1_unmap_dma(iommu, arg);
3016 case VFIO_IOMMU_DIRTY_PAGES:
3017 return vfio_iommu_type1_dirty_pages(iommu, arg);
3023 static int vfio_iommu_type1_register_notifier(void *iommu_data,
3024 unsigned long *events,
3025 struct notifier_block *nb)
3027 struct vfio_iommu *iommu = iommu_data;
3029 /* clear known events */
3030 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
3032 /* refuse to register if still events remaining */
3036 return blocking_notifier_chain_register(&iommu->notifier, nb);
3039 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
3040 struct notifier_block *nb)
3042 struct vfio_iommu *iommu = iommu_data;
3044 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
3047 static int vfio_iommu_type1_dma_rw_chunk(struct vfio_iommu *iommu,
3048 dma_addr_t user_iova, void *data,
3049 size_t count, bool write,
3052 struct mm_struct *mm;
3053 unsigned long vaddr;
3054 struct vfio_dma *dma;
3055 bool kthread = current->mm == NULL;
3061 ret = vfio_find_dma_valid(iommu, user_iova, 1, &dma);
3065 if ((write && !(dma->prot & IOMMU_WRITE)) ||
3066 !(dma->prot & IOMMU_READ))
3069 mm = get_task_mm(dma->task);
3076 else if (current->mm != mm)
3079 offset = user_iova - dma->iova;
3081 if (count > dma->size - offset)
3082 count = dma->size - offset;
3084 vaddr = dma->vaddr + offset;
3087 *copied = copy_to_user((void __user *)vaddr, data,
3089 if (*copied && iommu->dirty_page_tracking) {
3090 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
3092 * Bitmap populated with the smallest supported page
3095 bitmap_set(dma->bitmap, offset >> pgshift,
3096 ((offset + *copied - 1) >> pgshift) -
3097 (offset >> pgshift) + 1);
3100 *copied = copy_from_user(data, (void __user *)vaddr,
3103 kthread_unuse_mm(mm);
3106 return *copied ? 0 : -EFAULT;
3109 static int vfio_iommu_type1_dma_rw(void *iommu_data, dma_addr_t user_iova,
3110 void *data, size_t count, bool write)
3112 struct vfio_iommu *iommu = iommu_data;
3116 mutex_lock(&iommu->lock);
3118 ret = vfio_iommu_type1_dma_rw_chunk(iommu, user_iova, data,
3119 count, write, &done);
3128 mutex_unlock(&iommu->lock);
3132 static struct iommu_domain *
3133 vfio_iommu_type1_group_iommu_domain(void *iommu_data,
3134 struct iommu_group *iommu_group)
3136 struct iommu_domain *domain = ERR_PTR(-ENODEV);
3137 struct vfio_iommu *iommu = iommu_data;
3138 struct vfio_domain *d;
3140 if (!iommu || !iommu_group)
3141 return ERR_PTR(-EINVAL);
3143 mutex_lock(&iommu->lock);
3144 list_for_each_entry(d, &iommu->domain_list, next) {
3145 if (find_iommu_group(d, iommu_group)) {
3150 mutex_unlock(&iommu->lock);
3155 static void vfio_iommu_type1_notify(void *iommu_data,
3156 enum vfio_iommu_notify_type event)
3158 struct vfio_iommu *iommu = iommu_data;
3160 if (event != VFIO_IOMMU_CONTAINER_CLOSE)
3162 mutex_lock(&iommu->lock);
3163 iommu->container_open = false;
3164 mutex_unlock(&iommu->lock);
3165 wake_up_all(&iommu->vaddr_wait);
3168 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
3169 .name = "vfio-iommu-type1",
3170 .owner = THIS_MODULE,
3171 .open = vfio_iommu_type1_open,
3172 .release = vfio_iommu_type1_release,
3173 .ioctl = vfio_iommu_type1_ioctl,
3174 .attach_group = vfio_iommu_type1_attach_group,
3175 .detach_group = vfio_iommu_type1_detach_group,
3176 .pin_pages = vfio_iommu_type1_pin_pages,
3177 .unpin_pages = vfio_iommu_type1_unpin_pages,
3178 .register_notifier = vfio_iommu_type1_register_notifier,
3179 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
3180 .dma_rw = vfio_iommu_type1_dma_rw,
3181 .group_iommu_domain = vfio_iommu_type1_group_iommu_domain,
3182 .notify = vfio_iommu_type1_notify,
3185 static int __init vfio_iommu_type1_init(void)
3187 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
3190 static void __exit vfio_iommu_type1_cleanup(void)
3192 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
3195 module_init(vfio_iommu_type1_init);
3196 module_exit(vfio_iommu_type1_cleanup);
3198 MODULE_VERSION(DRIVER_VERSION);
3199 MODULE_LICENSE("GPL v2");
3200 MODULE_AUTHOR(DRIVER_AUTHOR);
3201 MODULE_DESCRIPTION(DRIVER_DESC);