1 // SPDX-License-Identifier: GPL-2.0-only
3 * VFIO: IOMMU DMA mapping support for Type1 IOMMU
5 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
6 * Author: Alex Williamson <alex.williamson@redhat.com>
8 * Derived from original vfio:
9 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
10 * Author: Tom Lyon, pugs@cisco.com
12 * We arbitrarily define a Type1 IOMMU as one matching the below code.
13 * It could be called the x86 IOMMU as it's designed for AMD-Vi & Intel
14 * VT-d, but that makes it harder to re-use as theoretically anyone
15 * implementing a similar IOMMU could make use of this. We expect the
16 * IOMMU to support the IOMMU API and have few to no restrictions around
17 * the IOVA range that can be mapped. The Type1 IOMMU is currently
18 * optimized for relatively static mappings of a userspace process with
19 * userpsace pages pinned into memory. We also assume devices and IOMMU
20 * domains are PCI based as the IOMMU API is still centered around a
21 * device/bus interface rather than a group interface.
24 #include <linux/compat.h>
25 #include <linux/device.h>
27 #include <linux/highmem.h>
28 #include <linux/iommu.h>
29 #include <linux/module.h>
31 #include <linux/kthread.h>
32 #include <linux/rbtree.h>
33 #include <linux/sched/signal.h>
34 #include <linux/sched/mm.h>
35 #include <linux/slab.h>
36 #include <linux/uaccess.h>
37 #include <linux/vfio.h>
38 #include <linux/workqueue.h>
39 #include <linux/mdev.h>
40 #include <linux/notifier.h>
41 #include <linux/dma-iommu.h>
42 #include <linux/irqdomain.h>
44 #define DRIVER_VERSION "0.2"
45 #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
46 #define DRIVER_DESC "Type1 IOMMU driver for VFIO"
48 static bool allow_unsafe_interrupts;
49 module_param_named(allow_unsafe_interrupts,
50 allow_unsafe_interrupts, bool, S_IRUGO | S_IWUSR);
51 MODULE_PARM_DESC(allow_unsafe_interrupts,
52 "Enable VFIO IOMMU support for on platforms without interrupt remapping support.");
54 static bool disable_hugepages;
55 module_param_named(disable_hugepages,
56 disable_hugepages, bool, S_IRUGO | S_IWUSR);
57 MODULE_PARM_DESC(disable_hugepages,
58 "Disable VFIO IOMMU support for IOMMU hugepages.");
60 static unsigned int dma_entry_limit __read_mostly = U16_MAX;
61 module_param_named(dma_entry_limit, dma_entry_limit, uint, 0644);
62 MODULE_PARM_DESC(dma_entry_limit,
63 "Maximum number of user DMA mappings per container (65535).");
66 struct list_head domain_list;
67 struct list_head iova_list;
68 struct vfio_domain *external_domain; /* domain for external user */
70 struct rb_root dma_list;
71 struct blocking_notifier_head notifier;
72 unsigned int dma_avail;
73 unsigned int vaddr_invalid_count;
74 uint64_t pgsize_bitmap;
75 uint64_t num_non_pinned_groups;
76 wait_queue_head_t vaddr_wait;
79 bool dirty_page_tracking;
80 bool pinned_page_dirty_scope;
85 struct iommu_domain *domain;
86 struct list_head next;
87 struct list_head group_list;
88 int prot; /* IOMMU_CACHE */
89 bool fgsp; /* Fine-grained super pages */
94 dma_addr_t iova; /* Device address */
95 unsigned long vaddr; /* Process virtual addr */
96 size_t size; /* Map size (bytes) */
97 int prot; /* IOMMU_READ/WRITE */
99 bool lock_cap; /* capable(CAP_IPC_LOCK) */
101 struct task_struct *task;
102 struct rb_root pfn_list; /* Ex-user pinned pfn list */
103 unsigned long *bitmap;
107 struct 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 */
115 struct iommu_group *iommu_group;
116 struct list_head next;
117 bool mdev_group; /* An mdev group */
118 bool pinned_page_dirty_scope;
122 struct list_head list;
128 * Guest RAM pinning working set or DMA target
132 dma_addr_t iova; /* Device address */
133 unsigned long pfn; /* Host pfn */
134 unsigned int ref_count;
137 struct vfio_regions {
138 struct list_head list;
144 #define IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu) \
145 (!list_empty(&iommu->domain_list))
147 #define DIRTY_BITMAP_BYTES(n) (ALIGN(n, BITS_PER_TYPE(u64)) / BITS_PER_BYTE)
150 * Input argument of number of bits to bitmap_set() is unsigned integer, which
151 * further casts to signed integer for unaligned multi-bit operation,
153 * Then maximum bitmap size supported is 2^31 bits divided by 2^3 bits/byte,
154 * that is 2^28 (256 MB) which maps to 2^31 * 2^12 = 2^43 (8TB) on 4K page
157 #define DIRTY_BITMAP_PAGES_MAX ((u64)INT_MAX)
158 #define DIRTY_BITMAP_SIZE_MAX DIRTY_BITMAP_BYTES(DIRTY_BITMAP_PAGES_MAX)
162 static int put_pfn(unsigned long pfn, int prot);
164 static struct vfio_group *vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
165 struct iommu_group *iommu_group);
168 * This code handles mapping and unmapping of user data buffers
169 * into DMA'ble space using the IOMMU
172 static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
173 dma_addr_t start, size_t size)
175 struct rb_node *node = iommu->dma_list.rb_node;
178 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
180 if (start + size <= dma->iova)
181 node = node->rb_left;
182 else if (start >= dma->iova + dma->size)
183 node = node->rb_right;
191 static struct rb_node *vfio_find_dma_first_node(struct vfio_iommu *iommu,
192 dma_addr_t start, u64 size)
194 struct rb_node *res = NULL;
195 struct rb_node *node = iommu->dma_list.rb_node;
196 struct vfio_dma *dma_res = NULL;
199 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
201 if (start < dma->iova + dma->size) {
204 if (start >= dma->iova)
206 node = node->rb_left;
208 node = node->rb_right;
211 if (res && size && dma_res->iova >= start + size)
216 static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new)
218 struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL;
219 struct vfio_dma *dma;
223 dma = rb_entry(parent, struct vfio_dma, node);
225 if (new->iova + new->size <= dma->iova)
226 link = &(*link)->rb_left;
228 link = &(*link)->rb_right;
231 rb_link_node(&new->node, parent, link);
232 rb_insert_color(&new->node, &iommu->dma_list);
235 static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
237 rb_erase(&old->node, &iommu->dma_list);
241 static int vfio_dma_bitmap_alloc(struct vfio_dma *dma, size_t pgsize)
243 uint64_t npages = dma->size / pgsize;
245 if (npages > DIRTY_BITMAP_PAGES_MAX)
249 * Allocate extra 64 bits that are used to calculate shift required for
250 * bitmap_shift_left() to manipulate and club unaligned number of pages
251 * in adjacent vfio_dma ranges.
253 dma->bitmap = kvzalloc(DIRTY_BITMAP_BYTES(npages) + sizeof(u64),
261 static void vfio_dma_bitmap_free(struct vfio_dma *dma)
267 static void vfio_dma_populate_bitmap(struct vfio_dma *dma, size_t pgsize)
270 unsigned long pgshift = __ffs(pgsize);
272 for (p = rb_first(&dma->pfn_list); p; p = rb_next(p)) {
273 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn, node);
275 bitmap_set(dma->bitmap, (vpfn->iova - dma->iova) >> pgshift, 1);
279 static void vfio_iommu_populate_bitmap_full(struct vfio_iommu *iommu)
282 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
284 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
285 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
287 bitmap_set(dma->bitmap, 0, dma->size >> pgshift);
291 static int vfio_dma_bitmap_alloc_all(struct vfio_iommu *iommu, size_t pgsize)
295 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
296 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
299 ret = vfio_dma_bitmap_alloc(dma, pgsize);
303 for (p = rb_prev(n); p; p = rb_prev(p)) {
304 struct vfio_dma *dma = rb_entry(n,
305 struct vfio_dma, node);
307 vfio_dma_bitmap_free(dma);
311 vfio_dma_populate_bitmap(dma, pgsize);
316 static void vfio_dma_bitmap_free_all(struct vfio_iommu *iommu)
320 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
321 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
323 vfio_dma_bitmap_free(dma);
328 * Helper Functions for host iova-pfn list
330 static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
332 struct vfio_pfn *vpfn;
333 struct rb_node *node = dma->pfn_list.rb_node;
336 vpfn = rb_entry(node, struct vfio_pfn, node);
338 if (iova < vpfn->iova)
339 node = node->rb_left;
340 else if (iova > vpfn->iova)
341 node = node->rb_right;
348 static void vfio_link_pfn(struct vfio_dma *dma,
349 struct vfio_pfn *new)
351 struct rb_node **link, *parent = NULL;
352 struct vfio_pfn *vpfn;
354 link = &dma->pfn_list.rb_node;
357 vpfn = rb_entry(parent, struct vfio_pfn, node);
359 if (new->iova < vpfn->iova)
360 link = &(*link)->rb_left;
362 link = &(*link)->rb_right;
365 rb_link_node(&new->node, parent, link);
366 rb_insert_color(&new->node, &dma->pfn_list);
369 static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old)
371 rb_erase(&old->node, &dma->pfn_list);
374 static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova,
377 struct vfio_pfn *vpfn;
379 vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL);
386 vfio_link_pfn(dma, vpfn);
390 static void vfio_remove_from_pfn_list(struct vfio_dma *dma,
391 struct vfio_pfn *vpfn)
393 vfio_unlink_pfn(dma, vpfn);
397 static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma,
400 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
407 static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn)
412 if (!vpfn->ref_count) {
413 ret = put_pfn(vpfn->pfn, dma->prot);
414 vfio_remove_from_pfn_list(dma, vpfn);
419 static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
421 struct mm_struct *mm;
427 mm = async ? get_task_mm(dma->task) : dma->task->mm;
429 return -ESRCH; /* process exited */
431 ret = mmap_write_lock_killable(mm);
433 ret = __account_locked_vm(mm, abs(npage), npage > 0, dma->task,
435 mmap_write_unlock(mm);
445 * Some mappings aren't backed by a struct page, for example an mmap'd
446 * MMIO range for our own or another device. These use a different
447 * pfn conversion and shouldn't be tracked as locked pages.
448 * For compound pages, any driver that sets the reserved bit in head
449 * page needs to set the reserved bit in all subpages to be safe.
451 static bool is_invalid_reserved_pfn(unsigned long pfn)
454 return PageReserved(pfn_to_page(pfn));
459 static int put_pfn(unsigned long pfn, int prot)
461 if (!is_invalid_reserved_pfn(pfn)) {
462 struct page *page = pfn_to_page(pfn);
464 unpin_user_pages_dirty_lock(&page, 1, prot & IOMMU_WRITE);
470 #define VFIO_BATCH_MAX_CAPACITY (PAGE_SIZE / sizeof(struct page *))
472 static void vfio_batch_init(struct vfio_batch *batch)
477 if (unlikely(disable_hugepages))
480 batch->pages = (struct page **) __get_free_page(GFP_KERNEL);
484 batch->capacity = VFIO_BATCH_MAX_CAPACITY;
488 batch->pages = &batch->fallback_page;
492 static void vfio_batch_unpin(struct vfio_batch *batch, struct vfio_dma *dma)
494 while (batch->size) {
495 unsigned long pfn = page_to_pfn(batch->pages[batch->offset]);
497 put_pfn(pfn, dma->prot);
503 static void vfio_batch_fini(struct vfio_batch *batch)
505 if (batch->capacity == VFIO_BATCH_MAX_CAPACITY)
506 free_page((unsigned long)batch->pages);
509 static int follow_fault_pfn(struct vm_area_struct *vma, struct mm_struct *mm,
510 unsigned long vaddr, unsigned long *pfn,
517 ret = follow_pte(vma->vm_mm, vaddr, &ptep, &ptl);
519 bool unlocked = false;
521 ret = fixup_user_fault(mm, vaddr,
523 (write_fault ? FAULT_FLAG_WRITE : 0),
531 ret = follow_pte(vma->vm_mm, vaddr, &ptep, &ptl);
536 if (write_fault && !pte_write(*ptep))
539 *pfn = pte_pfn(*ptep);
541 pte_unmap_unlock(ptep, ptl);
546 * Returns the positive number of pfns successfully obtained or a negative
549 static int vaddr_get_pfns(struct mm_struct *mm, unsigned long vaddr,
550 long npages, int prot, unsigned long *pfn,
553 struct vm_area_struct *vma;
554 unsigned int flags = 0;
557 if (prot & IOMMU_WRITE)
561 ret = pin_user_pages_remote(mm, vaddr, npages, flags | FOLL_LONGTERM,
564 *pfn = page_to_pfn(pages[0]);
568 vaddr = untagged_addr(vaddr);
571 vma = find_vma_intersection(mm, vaddr, vaddr + 1);
573 if (vma && vma->vm_flags & VM_PFNMAP) {
574 ret = follow_fault_pfn(vma, mm, vaddr, pfn, prot & IOMMU_WRITE);
579 if (is_invalid_reserved_pfn(*pfn))
586 mmap_read_unlock(mm);
590 static int vfio_wait(struct vfio_iommu *iommu)
594 prepare_to_wait(&iommu->vaddr_wait, &wait, TASK_KILLABLE);
595 mutex_unlock(&iommu->lock);
597 mutex_lock(&iommu->lock);
598 finish_wait(&iommu->vaddr_wait, &wait);
599 if (kthread_should_stop() || !iommu->container_open ||
600 fatal_signal_pending(current)) {
607 * Find dma struct and wait for its vaddr to be valid. iommu lock is dropped
608 * if the task waits, but is re-locked on return. Return result in *dma_p.
609 * Return 0 on success with no waiting, WAITED on success if waited, and -errno
612 static int vfio_find_dma_valid(struct vfio_iommu *iommu, dma_addr_t start,
613 size_t size, struct vfio_dma **dma_p)
618 *dma_p = vfio_find_dma(iommu, start, size);
621 else if (!(*dma_p)->vaddr_invalid)
624 ret = vfio_wait(iommu);
631 * Wait for all vaddr in the dma_list to become valid. iommu lock is dropped
632 * if the task waits, but is re-locked on return. Return 0 on success with no
633 * waiting, WAITED on success if waited, and -errno on error.
635 static int vfio_wait_all_valid(struct vfio_iommu *iommu)
639 while (iommu->vaddr_invalid_count && ret >= 0)
640 ret = vfio_wait(iommu);
646 * Attempt to pin pages. We really don't want to track all the pfns and
647 * the iommu can only map chunks of consecutive pfns anyway, so get the
648 * first page and all consecutive pages with the same locking.
650 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
651 long npage, unsigned long *pfn_base,
652 unsigned long limit, struct vfio_batch *batch)
655 struct mm_struct *mm = current->mm;
656 long ret, pinned = 0, lock_acct = 0;
658 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
660 /* This code path is only user initiated */
665 /* Leftover pages in batch from an earlier call. */
666 *pfn_base = page_to_pfn(batch->pages[batch->offset]);
668 rsvd = is_invalid_reserved_pfn(*pfn_base);
675 /* Empty batch, so refill it. */
676 long req_pages = min_t(long, npage, batch->capacity);
678 ret = vaddr_get_pfns(mm, vaddr, req_pages, dma->prot,
688 rsvd = is_invalid_reserved_pfn(*pfn_base);
693 * pfn is preset for the first iteration of this inner loop and
694 * updated at the end to handle a VM_PFNMAP pfn. In that case,
695 * batch->pages isn't valid (there's no struct page), so allow
696 * batch->pages to be touched only when there's more than one
697 * pfn to check, which guarantees the pfns are from a
701 if (pfn != *pfn_base + pinned ||
702 rsvd != is_invalid_reserved_pfn(pfn))
706 * Reserved pages aren't counted against the user,
707 * externally pinned pages are already counted against
710 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
711 if (!dma->lock_cap &&
712 mm->locked_vm + lock_acct + 1 > limit) {
713 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
714 __func__, limit << PAGE_SHIFT);
731 pfn = page_to_pfn(batch->pages[batch->offset]);
734 if (unlikely(disable_hugepages))
739 ret = vfio_lock_acct(dma, lock_acct, false);
742 if (batch->size == 1 && !batch->offset) {
743 /* May be a VM_PFNMAP pfn, which the batch can't remember. */
744 put_pfn(pfn, dma->prot);
749 if (pinned && !rsvd) {
750 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
751 put_pfn(pfn, dma->prot);
753 vfio_batch_unpin(batch, dma);
761 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
762 unsigned long pfn, long npage,
765 long unlocked = 0, locked = 0;
768 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
769 if (put_pfn(pfn++, dma->prot)) {
771 if (vfio_find_vpfn(dma, iova))
777 vfio_lock_acct(dma, locked - unlocked, true);
782 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
783 unsigned long *pfn_base, bool do_accounting)
785 struct page *pages[1];
786 struct mm_struct *mm;
789 mm = get_task_mm(dma->task);
793 ret = vaddr_get_pfns(mm, vaddr, 1, dma->prot, pfn_base, pages);
799 if (do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
800 ret = vfio_lock_acct(dma, 1, true);
802 put_pfn(*pfn_base, dma->prot);
804 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
805 "(%ld) exceeded\n", __func__,
806 dma->task->comm, task_pid_nr(dma->task),
807 task_rlimit(dma->task, RLIMIT_MEMLOCK));
816 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
820 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
825 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
828 vfio_lock_acct(dma, -unlocked, true);
833 static int vfio_iommu_type1_pin_pages(void *iommu_data,
834 struct iommu_group *iommu_group,
835 unsigned long *user_pfn,
837 unsigned long *phys_pfn)
839 struct vfio_iommu *iommu = iommu_data;
840 struct vfio_group *group;
842 unsigned long remote_vaddr;
843 struct vfio_dma *dma;
847 if (!iommu || !user_pfn || !phys_pfn)
850 /* Supported for v2 version only */
854 mutex_lock(&iommu->lock);
857 * Wait for all necessary vaddr's to be valid so they can be used in
858 * the main loop without dropping the lock, to avoid racing vs unmap.
861 if (iommu->vaddr_invalid_count) {
862 for (i = 0; i < npage; i++) {
863 iova = user_pfn[i] << PAGE_SHIFT;
864 ret = vfio_find_dma_valid(iommu, iova, PAGE_SIZE, &dma);
872 /* Fail if notifier list is empty */
873 if (!iommu->notifier.head) {
879 * If iommu capable domain exist in the container then all pages are
880 * already pinned and accounted. Accouting should be done if there is no
881 * iommu capable domain in the container.
883 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
885 for (i = 0; i < npage; i++) {
886 struct vfio_pfn *vpfn;
888 iova = user_pfn[i] << PAGE_SHIFT;
889 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
895 if ((dma->prot & prot) != prot) {
900 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
902 phys_pfn[i] = vpfn->pfn;
906 remote_vaddr = dma->vaddr + (iova - dma->iova);
907 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
912 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
914 if (put_pfn(phys_pfn[i], dma->prot) && do_accounting)
915 vfio_lock_acct(dma, -1, true);
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_pfn[j] << PAGE_SHIFT;
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 int vfio_iommu_type1_unpin_pages(void *iommu_data,
956 unsigned long *user_pfn,
959 struct vfio_iommu *iommu = iommu_data;
963 if (!iommu || !user_pfn)
966 /* Supported for v2 version only */
970 mutex_lock(&iommu->lock);
972 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
973 for (i = 0; i < npage; i++) {
974 struct vfio_dma *dma;
977 iova = user_pfn[i] << PAGE_SHIFT;
978 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
981 vfio_unpin_page_external(dma, iova, do_accounting);
985 mutex_unlock(&iommu->lock);
986 return i > npage ? npage : (i > 0 ? i : -EINVAL);
989 static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
990 struct list_head *regions,
991 struct iommu_iotlb_gather *iotlb_gather)
994 struct vfio_regions *entry, *next;
996 iommu_iotlb_sync(domain->domain, iotlb_gather);
998 list_for_each_entry_safe(entry, next, regions, list) {
999 unlocked += vfio_unpin_pages_remote(dma,
1001 entry->phys >> PAGE_SHIFT,
1002 entry->len >> PAGE_SHIFT,
1004 list_del(&entry->list);
1014 * Generally, VFIO needs to unpin remote pages after each IOTLB flush.
1015 * Therefore, when using IOTLB flush sync interface, VFIO need to keep track
1016 * of these regions (currently using a list).
1018 * This value specifies maximum number of regions for each IOTLB flush sync.
1020 #define VFIO_IOMMU_TLB_SYNC_MAX 512
1022 static size_t unmap_unpin_fast(struct vfio_domain *domain,
1023 struct vfio_dma *dma, dma_addr_t *iova,
1024 size_t len, phys_addr_t phys, long *unlocked,
1025 struct list_head *unmapped_list,
1027 struct iommu_iotlb_gather *iotlb_gather)
1029 size_t unmapped = 0;
1030 struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
1033 unmapped = iommu_unmap_fast(domain->domain, *iova, len,
1039 entry->iova = *iova;
1041 entry->len = unmapped;
1042 list_add_tail(&entry->list, unmapped_list);
1050 * Sync if the number of fast-unmap regions hits the limit
1051 * or in case of errors.
1053 if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
1054 *unlocked += vfio_sync_unpin(dma, domain, unmapped_list,
1062 static size_t unmap_unpin_slow(struct vfio_domain *domain,
1063 struct vfio_dma *dma, dma_addr_t *iova,
1064 size_t len, phys_addr_t phys,
1067 size_t unmapped = iommu_unmap(domain->domain, *iova, len);
1070 *unlocked += vfio_unpin_pages_remote(dma, *iova,
1072 unmapped >> PAGE_SHIFT,
1080 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
1083 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
1084 struct vfio_domain *domain, *d;
1085 LIST_HEAD(unmapped_region_list);
1086 struct iommu_iotlb_gather iotlb_gather;
1087 int unmapped_region_cnt = 0;
1093 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1097 * We use the IOMMU to track the physical addresses, otherwise we'd
1098 * need a much more complicated tracking system. Unfortunately that
1099 * means we need to use one of the iommu domains to figure out the
1100 * pfns to unpin. The rest need to be unmapped in advance so we have
1101 * no iommu translations remaining when the pages are unpinned.
1103 domain = d = list_first_entry(&iommu->domain_list,
1104 struct vfio_domain, next);
1106 list_for_each_entry_continue(d, &iommu->domain_list, next) {
1107 iommu_unmap(d->domain, dma->iova, dma->size);
1111 iommu_iotlb_gather_init(&iotlb_gather);
1112 while (iova < end) {
1113 size_t unmapped, len;
1114 phys_addr_t phys, next;
1116 phys = iommu_iova_to_phys(domain->domain, iova);
1117 if (WARN_ON(!phys)) {
1123 * To optimize for fewer iommu_unmap() calls, each of which
1124 * may require hardware cache flushing, try to find the
1125 * largest contiguous physical memory chunk to unmap.
1127 for (len = PAGE_SIZE;
1128 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
1129 next = iommu_iova_to_phys(domain->domain, iova + len);
1130 if (next != phys + len)
1135 * First, try to use fast unmap/unpin. In case of failure,
1136 * switch to slow unmap/unpin path.
1138 unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys,
1139 &unlocked, &unmapped_region_list,
1140 &unmapped_region_cnt,
1143 unmapped = unmap_unpin_slow(domain, dma, &iova, len,
1145 if (WARN_ON(!unmapped))
1150 dma->iommu_mapped = false;
1152 if (unmapped_region_cnt) {
1153 unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list,
1157 if (do_accounting) {
1158 vfio_lock_acct(dma, -unlocked, true);
1164 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
1166 WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list));
1167 vfio_unmap_unpin(iommu, dma, true);
1168 vfio_unlink_dma(iommu, dma);
1169 put_task_struct(dma->task);
1170 vfio_dma_bitmap_free(dma);
1171 if (dma->vaddr_invalid) {
1172 iommu->vaddr_invalid_count--;
1173 wake_up_all(&iommu->vaddr_wait);
1179 static void vfio_update_pgsize_bitmap(struct vfio_iommu *iommu)
1181 struct vfio_domain *domain;
1183 iommu->pgsize_bitmap = ULONG_MAX;
1185 list_for_each_entry(domain, &iommu->domain_list, next)
1186 iommu->pgsize_bitmap &= domain->domain->pgsize_bitmap;
1189 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
1190 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
1191 * That way the user will be able to map/unmap buffers whose size/
1192 * start address is aligned with PAGE_SIZE. Pinning code uses that
1193 * granularity while iommu driver can use the sub-PAGE_SIZE size
1194 * to map the buffer.
1196 if (iommu->pgsize_bitmap & ~PAGE_MASK) {
1197 iommu->pgsize_bitmap &= PAGE_MASK;
1198 iommu->pgsize_bitmap |= PAGE_SIZE;
1202 static int update_user_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1203 struct vfio_dma *dma, dma_addr_t base_iova,
1206 unsigned long pgshift = __ffs(pgsize);
1207 unsigned long nbits = dma->size >> pgshift;
1208 unsigned long bit_offset = (dma->iova - base_iova) >> pgshift;
1209 unsigned long copy_offset = bit_offset / BITS_PER_LONG;
1210 unsigned long shift = bit_offset % BITS_PER_LONG;
1211 unsigned long leftover;
1214 * mark all pages dirty if any IOMMU capable device is not able
1215 * to report dirty pages and all pages are pinned and mapped.
1217 if (iommu->num_non_pinned_groups && dma->iommu_mapped)
1218 bitmap_set(dma->bitmap, 0, nbits);
1221 bitmap_shift_left(dma->bitmap, dma->bitmap, shift,
1224 if (copy_from_user(&leftover,
1225 (void __user *)(bitmap + copy_offset),
1229 bitmap_or(dma->bitmap, dma->bitmap, &leftover, shift);
1232 if (copy_to_user((void __user *)(bitmap + copy_offset), dma->bitmap,
1233 DIRTY_BITMAP_BYTES(nbits + shift)))
1239 static int vfio_iova_dirty_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1240 dma_addr_t iova, size_t size, size_t pgsize)
1242 struct vfio_dma *dma;
1244 unsigned long pgshift = __ffs(pgsize);
1248 * GET_BITMAP request must fully cover vfio_dma mappings. Multiple
1249 * vfio_dma mappings may be clubbed by specifying large ranges, but
1250 * there must not be any previous mappings bisected by the range.
1251 * An error will be returned if these conditions are not met.
1253 dma = vfio_find_dma(iommu, iova, 1);
1254 if (dma && dma->iova != iova)
1257 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1258 if (dma && dma->iova + dma->size != iova + size)
1261 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1262 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1264 if (dma->iova < iova)
1267 if (dma->iova > iova + size - 1)
1270 ret = update_user_bitmap(bitmap, iommu, dma, iova, pgsize);
1275 * Re-populate bitmap to include all pinned pages which are
1276 * considered as dirty but exclude pages which are unpinned and
1277 * pages which are marked dirty by vfio_dma_rw()
1279 bitmap_clear(dma->bitmap, 0, dma->size >> pgshift);
1280 vfio_dma_populate_bitmap(dma, pgsize);
1285 static int verify_bitmap_size(uint64_t npages, uint64_t bitmap_size)
1287 if (!npages || !bitmap_size || (bitmap_size > DIRTY_BITMAP_SIZE_MAX) ||
1288 (bitmap_size < DIRTY_BITMAP_BYTES(npages)))
1294 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
1295 struct vfio_iommu_type1_dma_unmap *unmap,
1296 struct vfio_bitmap *bitmap)
1298 struct vfio_dma *dma, *dma_last = NULL;
1299 size_t unmapped = 0, pgsize;
1300 int ret = -EINVAL, retries = 0;
1301 unsigned long pgshift;
1302 dma_addr_t iova = unmap->iova;
1303 u64 size = unmap->size;
1304 bool unmap_all = unmap->flags & VFIO_DMA_UNMAP_FLAG_ALL;
1305 bool invalidate_vaddr = unmap->flags & VFIO_DMA_UNMAP_FLAG_VADDR;
1306 struct rb_node *n, *first_n;
1308 mutex_lock(&iommu->lock);
1310 pgshift = __ffs(iommu->pgsize_bitmap);
1311 pgsize = (size_t)1 << pgshift;
1313 if (iova & (pgsize - 1))
1320 } else if (!size || size & (pgsize - 1) ||
1321 iova + size - 1 < iova || size > SIZE_MAX) {
1325 /* When dirty tracking is enabled, allow only min supported pgsize */
1326 if ((unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
1327 (!iommu->dirty_page_tracking || (bitmap->pgsize != pgsize))) {
1331 WARN_ON((pgsize - 1) & PAGE_MASK);
1334 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
1335 * avoid tracking individual mappings. This means that the granularity
1336 * of the original mapping was lost and the user was allowed to attempt
1337 * to unmap any range. Depending on the contiguousness of physical
1338 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
1339 * or may not have worked. We only guaranteed unmap granularity
1340 * matching the original mapping; even though it was untracked here,
1341 * the original mappings are reflected in IOMMU mappings. This
1342 * resulted in a couple unusual behaviors. First, if a range is not
1343 * able to be unmapped, ex. a set of 4k pages that was mapped as a
1344 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
1345 * a zero sized unmap. Also, if an unmap request overlaps the first
1346 * address of a hugepage, the IOMMU will unmap the entire hugepage.
1347 * This also returns success and the returned unmap size reflects the
1348 * actual size unmapped.
1350 * We attempt to maintain compatibility with this "v1" interface, but
1351 * we take control out of the hands of the IOMMU. Therefore, an unmap
1352 * request offset from the beginning of the original mapping will
1353 * return success with zero sized unmap. And an unmap request covering
1354 * the first iova of mapping will unmap the entire range.
1356 * The v2 version of this interface intends to be more deterministic.
1357 * Unmap requests must fully cover previous mappings. Multiple
1358 * mappings may still be unmaped by specifying large ranges, but there
1359 * must not be any previous mappings bisected by the range. An error
1360 * will be returned if these conditions are not met. The v2 interface
1361 * will only return success and a size of zero if there were no
1362 * mappings within the range.
1364 if (iommu->v2 && !unmap_all) {
1365 dma = vfio_find_dma(iommu, iova, 1);
1366 if (dma && dma->iova != iova)
1369 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1370 if (dma && dma->iova + dma->size != iova + size)
1375 n = first_n = vfio_find_dma_first_node(iommu, iova, size);
1378 dma = rb_entry(n, struct vfio_dma, node);
1379 if (dma->iova >= iova + size)
1382 if (!iommu->v2 && iova > dma->iova)
1385 * Task with same address space who mapped this iova range is
1386 * allowed to unmap the iova range.
1388 if (dma->task->mm != current->mm)
1391 if (invalidate_vaddr) {
1392 if (dma->vaddr_invalid) {
1393 struct rb_node *last_n = n;
1395 for (n = first_n; n != last_n; n = rb_next(n)) {
1397 struct vfio_dma, node);
1398 dma->vaddr_invalid = false;
1399 iommu->vaddr_invalid_count--;
1405 dma->vaddr_invalid = true;
1406 iommu->vaddr_invalid_count++;
1407 unmapped += dma->size;
1412 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
1413 struct vfio_iommu_type1_dma_unmap nb_unmap;
1415 if (dma_last == dma) {
1416 BUG_ON(++retries > 10);
1422 nb_unmap.iova = dma->iova;
1423 nb_unmap.size = dma->size;
1426 * Notify anyone (mdev vendor drivers) to invalidate and
1427 * unmap iovas within the range we're about to unmap.
1428 * Vendor drivers MUST unpin pages in response to an
1431 mutex_unlock(&iommu->lock);
1432 blocking_notifier_call_chain(&iommu->notifier,
1433 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
1435 mutex_lock(&iommu->lock);
1439 if (unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
1440 ret = update_user_bitmap(bitmap->data, iommu, dma,
1446 unmapped += dma->size;
1448 vfio_remove_dma(iommu, dma);
1452 mutex_unlock(&iommu->lock);
1454 /* Report how much was unmapped */
1455 unmap->size = unmapped;
1460 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
1461 unsigned long pfn, long npage, int prot)
1463 struct vfio_domain *d;
1466 list_for_each_entry(d, &iommu->domain_list, next) {
1467 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
1468 npage << PAGE_SHIFT, prot | d->prot);
1478 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) {
1479 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
1486 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
1489 dma_addr_t iova = dma->iova;
1490 unsigned long vaddr = dma->vaddr;
1491 struct vfio_batch batch;
1492 size_t size = map_size;
1494 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1497 vfio_batch_init(&batch);
1500 /* Pin a contiguous chunk of memory */
1501 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
1502 size >> PAGE_SHIFT, &pfn, limit,
1511 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
1514 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
1516 vfio_batch_unpin(&batch, dma);
1520 size -= npage << PAGE_SHIFT;
1521 dma->size += npage << PAGE_SHIFT;
1524 vfio_batch_fini(&batch);
1525 dma->iommu_mapped = true;
1528 vfio_remove_dma(iommu, dma);
1534 * Check dma map request is within a valid iova range
1536 static bool vfio_iommu_iova_dma_valid(struct vfio_iommu *iommu,
1537 dma_addr_t start, dma_addr_t end)
1539 struct list_head *iova = &iommu->iova_list;
1540 struct vfio_iova *node;
1542 list_for_each_entry(node, iova, list) {
1543 if (start >= node->start && end <= node->end)
1548 * Check for list_empty() as well since a container with
1549 * a single mdev device will have an empty list.
1551 return list_empty(iova);
1554 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1555 struct vfio_iommu_type1_dma_map *map)
1557 bool set_vaddr = map->flags & VFIO_DMA_MAP_FLAG_VADDR;
1558 dma_addr_t iova = map->iova;
1559 unsigned long vaddr = map->vaddr;
1560 size_t size = map->size;
1561 int ret = 0, prot = 0;
1563 struct vfio_dma *dma;
1565 /* Verify that none of our __u64 fields overflow */
1566 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1569 /* READ/WRITE from device perspective */
1570 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1571 prot |= IOMMU_WRITE;
1572 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1575 if ((prot && set_vaddr) || (!prot && !set_vaddr))
1578 mutex_lock(&iommu->lock);
1580 pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
1582 WARN_ON((pgsize - 1) & PAGE_MASK);
1584 if (!size || (size | iova | vaddr) & (pgsize - 1)) {
1589 /* Don't allow IOVA or virtual address wrap */
1590 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) {
1595 dma = vfio_find_dma(iommu, iova, size);
1599 } else if (!dma->vaddr_invalid || dma->iova != iova ||
1600 dma->size != size) {
1604 dma->vaddr_invalid = false;
1605 iommu->vaddr_invalid_count--;
1606 wake_up_all(&iommu->vaddr_wait);
1614 if (!iommu->dma_avail) {
1619 if (!vfio_iommu_iova_dma_valid(iommu, iova, iova + size - 1)) {
1624 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1636 * We need to be able to both add to a task's locked memory and test
1637 * against the locked memory limit and we need to be able to do both
1638 * outside of this call path as pinning can be asynchronous via the
1639 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1640 * task_struct and VM locked pages requires an mm_struct, however
1641 * holding an indefinite mm reference is not recommended, therefore we
1642 * only hold a reference to a task. We could hold a reference to
1643 * current, however QEMU uses this call path through vCPU threads,
1644 * which can be killed resulting in a NULL mm and failure in the unmap
1645 * path when called via a different thread. Avoid this problem by
1646 * using the group_leader as threads within the same group require
1647 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1650 * Previously we also used the task for testing CAP_IPC_LOCK at the
1651 * time of pinning and accounting, however has_capability() makes use
1652 * of real_cred, a copy-on-write field, so we can't guarantee that it
1653 * matches group_leader, or in fact that it might not change by the
1654 * time it's evaluated. If a process were to call MAP_DMA with
1655 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1656 * possibly see different results for an iommu_mapped vfio_dma vs
1657 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1658 * time of calling MAP_DMA.
1660 get_task_struct(current->group_leader);
1661 dma->task = current->group_leader;
1662 dma->lock_cap = capable(CAP_IPC_LOCK);
1664 dma->pfn_list = RB_ROOT;
1666 /* Insert zero-sized and grow as we map chunks of it */
1667 vfio_link_dma(iommu, dma);
1669 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1670 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1673 ret = vfio_pin_map_dma(iommu, dma, size);
1675 if (!ret && iommu->dirty_page_tracking) {
1676 ret = vfio_dma_bitmap_alloc(dma, pgsize);
1678 vfio_remove_dma(iommu, dma);
1682 mutex_unlock(&iommu->lock);
1686 static int vfio_bus_type(struct device *dev, void *data)
1688 struct bus_type **bus = data;
1690 if (*bus && *bus != dev->bus)
1698 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1699 struct vfio_domain *domain)
1701 struct vfio_batch batch;
1702 struct vfio_domain *d = NULL;
1704 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1707 ret = vfio_wait_all_valid(iommu);
1711 /* Arbitrarily pick the first domain in the list for lookups */
1712 if (!list_empty(&iommu->domain_list))
1713 d = list_first_entry(&iommu->domain_list,
1714 struct vfio_domain, next);
1716 vfio_batch_init(&batch);
1718 n = rb_first(&iommu->dma_list);
1720 for (; n; n = rb_next(n)) {
1721 struct vfio_dma *dma;
1724 dma = rb_entry(n, struct vfio_dma, node);
1727 while (iova < dma->iova + dma->size) {
1731 if (dma->iommu_mapped) {
1735 if (WARN_ON(!d)) { /* mapped w/o a domain?! */
1740 phys = iommu_iova_to_phys(d->domain, iova);
1742 if (WARN_ON(!phys)) {
1750 while (i < dma->iova + dma->size &&
1751 p == iommu_iova_to_phys(d->domain, i)) {
1758 unsigned long vaddr = dma->vaddr +
1760 size_t n = dma->iova + dma->size - iova;
1763 npage = vfio_pin_pages_remote(dma, vaddr,
1773 phys = pfn << PAGE_SHIFT;
1774 size = npage << PAGE_SHIFT;
1777 ret = iommu_map(domain->domain, iova, phys,
1778 size, dma->prot | domain->prot);
1780 if (!dma->iommu_mapped) {
1781 vfio_unpin_pages_remote(dma, iova,
1785 vfio_batch_unpin(&batch, dma);
1794 /* All dmas are now mapped, defer to second tree walk for unwind */
1795 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1796 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1798 dma->iommu_mapped = true;
1801 vfio_batch_fini(&batch);
1805 for (; n; n = rb_prev(n)) {
1806 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1809 if (dma->iommu_mapped) {
1810 iommu_unmap(domain->domain, dma->iova, dma->size);
1815 while (iova < dma->iova + dma->size) {
1816 phys_addr_t phys, p;
1820 phys = iommu_iova_to_phys(domain->domain, iova);
1829 while (i < dma->iova + dma->size &&
1830 p == iommu_iova_to_phys(domain->domain, i)) {
1836 iommu_unmap(domain->domain, iova, size);
1837 vfio_unpin_pages_remote(dma, iova, phys >> PAGE_SHIFT,
1838 size >> PAGE_SHIFT, true);
1842 vfio_batch_fini(&batch);
1847 * We change our unmap behavior slightly depending on whether the IOMMU
1848 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1849 * for practically any contiguous power-of-two mapping we give it. This means
1850 * we don't need to look for contiguous chunks ourselves to make unmapping
1851 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1852 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1853 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1854 * hugetlbfs is in use.
1856 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1859 int ret, order = get_order(PAGE_SIZE * 2);
1861 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1865 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1866 IOMMU_READ | IOMMU_WRITE | domain->prot);
1868 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1870 if (unmapped == PAGE_SIZE)
1871 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1873 domain->fgsp = true;
1876 __free_pages(pages, order);
1879 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1880 struct iommu_group *iommu_group)
1882 struct vfio_group *g;
1884 list_for_each_entry(g, &domain->group_list, next) {
1885 if (g->iommu_group == iommu_group)
1892 static struct vfio_group *vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
1893 struct iommu_group *iommu_group)
1895 struct vfio_domain *domain;
1896 struct vfio_group *group = NULL;
1898 list_for_each_entry(domain, &iommu->domain_list, next) {
1899 group = find_iommu_group(domain, iommu_group);
1904 if (iommu->external_domain)
1905 group = find_iommu_group(iommu->external_domain, iommu_group);
1910 static bool vfio_iommu_has_sw_msi(struct list_head *group_resv_regions,
1913 struct iommu_resv_region *region;
1916 list_for_each_entry(region, group_resv_regions, list) {
1918 * The presence of any 'real' MSI regions should take
1919 * precedence over the software-managed one if the
1920 * IOMMU driver happens to advertise both types.
1922 if (region->type == IOMMU_RESV_MSI) {
1927 if (region->type == IOMMU_RESV_SW_MSI) {
1928 *base = region->start;
1936 static struct device *vfio_mdev_get_iommu_device(struct device *dev)
1938 struct device *(*fn)(struct device *dev);
1939 struct device *iommu_device;
1941 fn = symbol_get(mdev_get_iommu_device);
1943 iommu_device = fn(dev);
1944 symbol_put(mdev_get_iommu_device);
1946 return iommu_device;
1952 static int vfio_mdev_attach_domain(struct device *dev, void *data)
1954 struct iommu_domain *domain = data;
1955 struct device *iommu_device;
1957 iommu_device = vfio_mdev_get_iommu_device(dev);
1959 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1960 return iommu_aux_attach_device(domain, iommu_device);
1962 return iommu_attach_device(domain, iommu_device);
1968 static int vfio_mdev_detach_domain(struct device *dev, void *data)
1970 struct iommu_domain *domain = data;
1971 struct device *iommu_device;
1973 iommu_device = vfio_mdev_get_iommu_device(dev);
1975 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1976 iommu_aux_detach_device(domain, iommu_device);
1978 iommu_detach_device(domain, iommu_device);
1984 static int vfio_iommu_attach_group(struct vfio_domain *domain,
1985 struct vfio_group *group)
1987 if (group->mdev_group)
1988 return iommu_group_for_each_dev(group->iommu_group,
1990 vfio_mdev_attach_domain);
1992 return iommu_attach_group(domain->domain, group->iommu_group);
1995 static void vfio_iommu_detach_group(struct vfio_domain *domain,
1996 struct vfio_group *group)
1998 if (group->mdev_group)
1999 iommu_group_for_each_dev(group->iommu_group, domain->domain,
2000 vfio_mdev_detach_domain);
2002 iommu_detach_group(domain->domain, group->iommu_group);
2005 static bool vfio_bus_is_mdev(struct bus_type *bus)
2007 struct bus_type *mdev_bus;
2010 mdev_bus = symbol_get(mdev_bus_type);
2012 ret = (bus == mdev_bus);
2013 symbol_put(mdev_bus_type);
2019 static int vfio_mdev_iommu_device(struct device *dev, void *data)
2021 struct device **old = data, *new;
2023 new = vfio_mdev_get_iommu_device(dev);
2024 if (!new || (*old && *old != new))
2033 * This is a helper function to insert an address range to iova list.
2034 * The list is initially created with a single entry corresponding to
2035 * the IOMMU domain geometry to which the device group is attached.
2036 * The list aperture gets modified when a new domain is added to the
2037 * container if the new aperture doesn't conflict with the current one
2038 * or with any existing dma mappings. The list is also modified to
2039 * exclude any reserved regions associated with the device group.
2041 static int vfio_iommu_iova_insert(struct list_head *head,
2042 dma_addr_t start, dma_addr_t end)
2044 struct vfio_iova *region;
2046 region = kmalloc(sizeof(*region), GFP_KERNEL);
2050 INIT_LIST_HEAD(®ion->list);
2051 region->start = start;
2054 list_add_tail(®ion->list, head);
2059 * Check the new iommu aperture conflicts with existing aper or with any
2060 * existing dma mappings.
2062 static bool vfio_iommu_aper_conflict(struct vfio_iommu *iommu,
2063 dma_addr_t start, dma_addr_t end)
2065 struct vfio_iova *first, *last;
2066 struct list_head *iova = &iommu->iova_list;
2068 if (list_empty(iova))
2071 /* Disjoint sets, return conflict */
2072 first = list_first_entry(iova, struct vfio_iova, list);
2073 last = list_last_entry(iova, struct vfio_iova, list);
2074 if (start > last->end || end < first->start)
2077 /* Check for any existing dma mappings below the new start */
2078 if (start > first->start) {
2079 if (vfio_find_dma(iommu, first->start, start - first->start))
2083 /* Check for any existing dma mappings beyond the new end */
2084 if (end < last->end) {
2085 if (vfio_find_dma(iommu, end + 1, last->end - end))
2093 * Resize iommu iova aperture window. This is called only if the new
2094 * aperture has no conflict with existing aperture and dma mappings.
2096 static int vfio_iommu_aper_resize(struct list_head *iova,
2097 dma_addr_t start, dma_addr_t end)
2099 struct vfio_iova *node, *next;
2101 if (list_empty(iova))
2102 return vfio_iommu_iova_insert(iova, start, end);
2104 /* Adjust iova list start */
2105 list_for_each_entry_safe(node, next, iova, list) {
2106 if (start < node->start)
2108 if (start >= node->start && start < node->end) {
2109 node->start = start;
2112 /* Delete nodes before new start */
2113 list_del(&node->list);
2117 /* Adjust iova list end */
2118 list_for_each_entry_safe(node, next, iova, list) {
2119 if (end > node->end)
2121 if (end > node->start && end <= node->end) {
2125 /* Delete nodes after new end */
2126 list_del(&node->list);
2134 * Check reserved region conflicts with existing dma mappings
2136 static bool vfio_iommu_resv_conflict(struct vfio_iommu *iommu,
2137 struct list_head *resv_regions)
2139 struct iommu_resv_region *region;
2141 /* Check for conflict with existing dma mappings */
2142 list_for_each_entry(region, resv_regions, list) {
2143 if (region->type == IOMMU_RESV_DIRECT_RELAXABLE)
2146 if (vfio_find_dma(iommu, region->start, region->length))
2154 * Check iova region overlap with reserved regions and
2155 * exclude them from the iommu iova range
2157 static int vfio_iommu_resv_exclude(struct list_head *iova,
2158 struct list_head *resv_regions)
2160 struct iommu_resv_region *resv;
2161 struct vfio_iova *n, *next;
2163 list_for_each_entry(resv, resv_regions, list) {
2164 phys_addr_t start, end;
2166 if (resv->type == IOMMU_RESV_DIRECT_RELAXABLE)
2169 start = resv->start;
2170 end = resv->start + resv->length - 1;
2172 list_for_each_entry_safe(n, next, iova, list) {
2176 if (start > n->end || end < n->start)
2179 * Insert a new node if current node overlaps with the
2180 * reserve region to exlude that from valid iova range.
2181 * Note that, new node is inserted before the current
2182 * node and finally the current node is deleted keeping
2183 * the list updated and sorted.
2185 if (start > n->start)
2186 ret = vfio_iommu_iova_insert(&n->list, n->start,
2188 if (!ret && end < n->end)
2189 ret = vfio_iommu_iova_insert(&n->list, end + 1,
2199 if (list_empty(iova))
2205 static void vfio_iommu_resv_free(struct list_head *resv_regions)
2207 struct iommu_resv_region *n, *next;
2209 list_for_each_entry_safe(n, next, resv_regions, list) {
2215 static void vfio_iommu_iova_free(struct list_head *iova)
2217 struct vfio_iova *n, *next;
2219 list_for_each_entry_safe(n, next, iova, list) {
2225 static int vfio_iommu_iova_get_copy(struct vfio_iommu *iommu,
2226 struct list_head *iova_copy)
2228 struct list_head *iova = &iommu->iova_list;
2229 struct vfio_iova *n;
2232 list_for_each_entry(n, iova, list) {
2233 ret = vfio_iommu_iova_insert(iova_copy, n->start, n->end);
2241 vfio_iommu_iova_free(iova_copy);
2245 static void vfio_iommu_iova_insert_copy(struct vfio_iommu *iommu,
2246 struct list_head *iova_copy)
2248 struct list_head *iova = &iommu->iova_list;
2250 vfio_iommu_iova_free(iova);
2252 list_splice_tail(iova_copy, iova);
2255 static int vfio_iommu_type1_attach_group(void *iommu_data,
2256 struct iommu_group *iommu_group)
2258 struct vfio_iommu *iommu = iommu_data;
2259 struct vfio_group *group;
2260 struct vfio_domain *domain, *d;
2261 struct bus_type *bus = NULL;
2263 bool resv_msi, msi_remap;
2264 phys_addr_t resv_msi_base = 0;
2265 struct iommu_domain_geometry geo;
2266 LIST_HEAD(iova_copy);
2267 LIST_HEAD(group_resv_regions);
2269 mutex_lock(&iommu->lock);
2271 /* Check for duplicates */
2272 if (vfio_iommu_find_iommu_group(iommu, iommu_group)) {
2273 mutex_unlock(&iommu->lock);
2277 group = kzalloc(sizeof(*group), GFP_KERNEL);
2278 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2279 if (!group || !domain) {
2284 group->iommu_group = iommu_group;
2286 /* Determine bus_type in order to allocate a domain */
2287 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
2291 if (vfio_bus_is_mdev(bus)) {
2292 struct device *iommu_device = NULL;
2294 group->mdev_group = true;
2296 /* Determine the isolation type */
2297 ret = iommu_group_for_each_dev(iommu_group, &iommu_device,
2298 vfio_mdev_iommu_device);
2299 if (ret || !iommu_device) {
2300 if (!iommu->external_domain) {
2301 INIT_LIST_HEAD(&domain->group_list);
2302 iommu->external_domain = domain;
2303 vfio_update_pgsize_bitmap(iommu);
2308 list_add(&group->next,
2309 &iommu->external_domain->group_list);
2311 * Non-iommu backed group cannot dirty memory directly,
2312 * it can only use interfaces that provide dirty
2314 * The iommu scope can only be promoted with the
2315 * addition of a dirty tracking group.
2317 group->pinned_page_dirty_scope = true;
2318 mutex_unlock(&iommu->lock);
2323 bus = iommu_device->bus;
2326 domain->domain = iommu_domain_alloc(bus);
2327 if (!domain->domain) {
2332 if (iommu->nesting) {
2335 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
2341 ret = vfio_iommu_attach_group(domain, group);
2345 /* Get aperture info */
2346 iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY, &geo);
2348 if (vfio_iommu_aper_conflict(iommu, geo.aperture_start,
2349 geo.aperture_end)) {
2354 ret = iommu_get_group_resv_regions(iommu_group, &group_resv_regions);
2358 if (vfio_iommu_resv_conflict(iommu, &group_resv_regions)) {
2364 * We don't want to work on the original iova list as the list
2365 * gets modified and in case of failure we have to retain the
2366 * original list. Get a copy here.
2368 ret = vfio_iommu_iova_get_copy(iommu, &iova_copy);
2372 ret = vfio_iommu_aper_resize(&iova_copy, geo.aperture_start,
2377 ret = vfio_iommu_resv_exclude(&iova_copy, &group_resv_regions);
2381 resv_msi = vfio_iommu_has_sw_msi(&group_resv_regions, &resv_msi_base);
2383 INIT_LIST_HEAD(&domain->group_list);
2384 list_add(&group->next, &domain->group_list);
2386 msi_remap = irq_domain_check_msi_remap() ||
2387 iommu_capable(bus, IOMMU_CAP_INTR_REMAP);
2389 if (!allow_unsafe_interrupts && !msi_remap) {
2390 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
2396 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
2397 domain->prot |= IOMMU_CACHE;
2400 * Try to match an existing compatible domain. We don't want to
2401 * preclude an IOMMU driver supporting multiple bus_types and being
2402 * able to include different bus_types in the same IOMMU domain, so
2403 * we test whether the domains use the same iommu_ops rather than
2404 * testing if they're on the same bus_type.
2406 list_for_each_entry(d, &iommu->domain_list, next) {
2407 if (d->domain->ops == domain->domain->ops &&
2408 d->prot == domain->prot) {
2409 vfio_iommu_detach_group(domain, group);
2410 if (!vfio_iommu_attach_group(d, group)) {
2411 list_add(&group->next, &d->group_list);
2412 iommu_domain_free(domain->domain);
2417 ret = vfio_iommu_attach_group(domain, group);
2423 vfio_test_domain_fgsp(domain);
2425 /* replay mappings on new domains */
2426 ret = vfio_iommu_replay(iommu, domain);
2431 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
2432 if (ret && ret != -ENODEV)
2436 list_add(&domain->next, &iommu->domain_list);
2437 vfio_update_pgsize_bitmap(iommu);
2439 /* Delete the old one and insert new iova list */
2440 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2443 * An iommu backed group can dirty memory directly and therefore
2444 * demotes the iommu scope until it declares itself dirty tracking
2445 * capable via the page pinning interface.
2447 iommu->num_non_pinned_groups++;
2448 mutex_unlock(&iommu->lock);
2449 vfio_iommu_resv_free(&group_resv_regions);
2454 vfio_iommu_detach_group(domain, group);
2456 iommu_domain_free(domain->domain);
2457 vfio_iommu_iova_free(&iova_copy);
2458 vfio_iommu_resv_free(&group_resv_regions);
2462 mutex_unlock(&iommu->lock);
2466 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
2468 struct rb_node *node;
2470 while ((node = rb_first(&iommu->dma_list)))
2471 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
2474 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
2476 struct rb_node *n, *p;
2478 n = rb_first(&iommu->dma_list);
2479 for (; n; n = rb_next(n)) {
2480 struct vfio_dma *dma;
2481 long locked = 0, unlocked = 0;
2483 dma = rb_entry(n, struct vfio_dma, node);
2484 unlocked += vfio_unmap_unpin(iommu, dma, false);
2485 p = rb_first(&dma->pfn_list);
2486 for (; p; p = rb_next(p)) {
2487 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
2490 if (!is_invalid_reserved_pfn(vpfn->pfn))
2493 vfio_lock_acct(dma, locked - unlocked, true);
2498 * Called when a domain is removed in detach. It is possible that
2499 * the removed domain decided the iova aperture window. Modify the
2500 * iova aperture with the smallest window among existing domains.
2502 static void vfio_iommu_aper_expand(struct vfio_iommu *iommu,
2503 struct list_head *iova_copy)
2505 struct vfio_domain *domain;
2506 struct iommu_domain_geometry geo;
2507 struct vfio_iova *node;
2508 dma_addr_t start = 0;
2509 dma_addr_t end = (dma_addr_t)~0;
2511 if (list_empty(iova_copy))
2514 list_for_each_entry(domain, &iommu->domain_list, next) {
2515 iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY,
2517 if (geo.aperture_start > start)
2518 start = geo.aperture_start;
2519 if (geo.aperture_end < end)
2520 end = geo.aperture_end;
2523 /* Modify aperture limits. The new aper is either same or bigger */
2524 node = list_first_entry(iova_copy, struct vfio_iova, list);
2525 node->start = start;
2526 node = list_last_entry(iova_copy, struct vfio_iova, list);
2531 * Called when a group is detached. The reserved regions for that
2532 * group can be part of valid iova now. But since reserved regions
2533 * may be duplicated among groups, populate the iova valid regions
2536 static int vfio_iommu_resv_refresh(struct vfio_iommu *iommu,
2537 struct list_head *iova_copy)
2539 struct vfio_domain *d;
2540 struct vfio_group *g;
2541 struct vfio_iova *node;
2542 dma_addr_t start, end;
2543 LIST_HEAD(resv_regions);
2546 if (list_empty(iova_copy))
2549 list_for_each_entry(d, &iommu->domain_list, next) {
2550 list_for_each_entry(g, &d->group_list, next) {
2551 ret = iommu_get_group_resv_regions(g->iommu_group,
2558 node = list_first_entry(iova_copy, struct vfio_iova, list);
2559 start = node->start;
2560 node = list_last_entry(iova_copy, struct vfio_iova, list);
2563 /* purge the iova list and create new one */
2564 vfio_iommu_iova_free(iova_copy);
2566 ret = vfio_iommu_aper_resize(iova_copy, start, end);
2570 /* Exclude current reserved regions from iova ranges */
2571 ret = vfio_iommu_resv_exclude(iova_copy, &resv_regions);
2573 vfio_iommu_resv_free(&resv_regions);
2577 static void vfio_iommu_type1_detach_group(void *iommu_data,
2578 struct iommu_group *iommu_group)
2580 struct vfio_iommu *iommu = iommu_data;
2581 struct vfio_domain *domain;
2582 struct vfio_group *group;
2583 bool update_dirty_scope = false;
2584 LIST_HEAD(iova_copy);
2586 mutex_lock(&iommu->lock);
2588 if (iommu->external_domain) {
2589 group = find_iommu_group(iommu->external_domain, iommu_group);
2591 update_dirty_scope = !group->pinned_page_dirty_scope;
2592 list_del(&group->next);
2595 if (list_empty(&iommu->external_domain->group_list)) {
2596 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) {
2597 WARN_ON(iommu->notifier.head);
2598 vfio_iommu_unmap_unpin_all(iommu);
2601 kfree(iommu->external_domain);
2602 iommu->external_domain = NULL;
2604 goto detach_group_done;
2609 * Get a copy of iova list. This will be used to update
2610 * and to replace the current one later. Please note that
2611 * we will leave the original list as it is if update fails.
2613 vfio_iommu_iova_get_copy(iommu, &iova_copy);
2615 list_for_each_entry(domain, &iommu->domain_list, next) {
2616 group = find_iommu_group(domain, iommu_group);
2620 vfio_iommu_detach_group(domain, group);
2621 update_dirty_scope = !group->pinned_page_dirty_scope;
2622 list_del(&group->next);
2625 * Group ownership provides privilege, if the group list is
2626 * empty, the domain goes away. If it's the last domain with
2627 * iommu and external domain doesn't exist, then all the
2628 * mappings go away too. If it's the last domain with iommu and
2629 * external domain exist, update accounting
2631 if (list_empty(&domain->group_list)) {
2632 if (list_is_singular(&iommu->domain_list)) {
2633 if (!iommu->external_domain) {
2634 WARN_ON(iommu->notifier.head);
2635 vfio_iommu_unmap_unpin_all(iommu);
2637 vfio_iommu_unmap_unpin_reaccount(iommu);
2640 iommu_domain_free(domain->domain);
2641 list_del(&domain->next);
2643 vfio_iommu_aper_expand(iommu, &iova_copy);
2644 vfio_update_pgsize_bitmap(iommu);
2649 if (!vfio_iommu_resv_refresh(iommu, &iova_copy))
2650 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2652 vfio_iommu_iova_free(&iova_copy);
2656 * Removal of a group without dirty tracking may allow the iommu scope
2659 if (update_dirty_scope) {
2660 iommu->num_non_pinned_groups--;
2661 if (iommu->dirty_page_tracking)
2662 vfio_iommu_populate_bitmap_full(iommu);
2664 mutex_unlock(&iommu->lock);
2667 static void *vfio_iommu_type1_open(unsigned long arg)
2669 struct vfio_iommu *iommu;
2671 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
2673 return ERR_PTR(-ENOMEM);
2676 case VFIO_TYPE1_IOMMU:
2678 case VFIO_TYPE1_NESTING_IOMMU:
2679 iommu->nesting = true;
2681 case VFIO_TYPE1v2_IOMMU:
2686 return ERR_PTR(-EINVAL);
2689 INIT_LIST_HEAD(&iommu->domain_list);
2690 INIT_LIST_HEAD(&iommu->iova_list);
2691 iommu->dma_list = RB_ROOT;
2692 iommu->dma_avail = dma_entry_limit;
2693 iommu->container_open = true;
2694 mutex_init(&iommu->lock);
2695 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
2696 init_waitqueue_head(&iommu->vaddr_wait);
2701 static void vfio_release_domain(struct vfio_domain *domain, bool external)
2703 struct vfio_group *group, *group_tmp;
2705 list_for_each_entry_safe(group, group_tmp,
2706 &domain->group_list, next) {
2708 vfio_iommu_detach_group(domain, group);
2709 list_del(&group->next);
2714 iommu_domain_free(domain->domain);
2717 static void vfio_iommu_type1_release(void *iommu_data)
2719 struct vfio_iommu *iommu = iommu_data;
2720 struct vfio_domain *domain, *domain_tmp;
2722 if (iommu->external_domain) {
2723 vfio_release_domain(iommu->external_domain, true);
2724 kfree(iommu->external_domain);
2727 vfio_iommu_unmap_unpin_all(iommu);
2729 list_for_each_entry_safe(domain, domain_tmp,
2730 &iommu->domain_list, next) {
2731 vfio_release_domain(domain, false);
2732 list_del(&domain->next);
2736 vfio_iommu_iova_free(&iommu->iova_list);
2741 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
2743 struct vfio_domain *domain;
2746 mutex_lock(&iommu->lock);
2747 list_for_each_entry(domain, &iommu->domain_list, next) {
2748 if (!(domain->prot & IOMMU_CACHE)) {
2753 mutex_unlock(&iommu->lock);
2758 static int vfio_iommu_type1_check_extension(struct vfio_iommu *iommu,
2762 case VFIO_TYPE1_IOMMU:
2763 case VFIO_TYPE1v2_IOMMU:
2764 case VFIO_TYPE1_NESTING_IOMMU:
2765 case VFIO_UNMAP_ALL:
2766 case VFIO_UPDATE_VADDR:
2768 case VFIO_DMA_CC_IOMMU:
2771 return vfio_domains_have_iommu_cache(iommu);
2777 static int vfio_iommu_iova_add_cap(struct vfio_info_cap *caps,
2778 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas,
2781 struct vfio_info_cap_header *header;
2782 struct vfio_iommu_type1_info_cap_iova_range *iova_cap;
2784 header = vfio_info_cap_add(caps, size,
2785 VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE, 1);
2787 return PTR_ERR(header);
2789 iova_cap = container_of(header,
2790 struct vfio_iommu_type1_info_cap_iova_range,
2792 iova_cap->nr_iovas = cap_iovas->nr_iovas;
2793 memcpy(iova_cap->iova_ranges, cap_iovas->iova_ranges,
2794 cap_iovas->nr_iovas * sizeof(*cap_iovas->iova_ranges));
2798 static int vfio_iommu_iova_build_caps(struct vfio_iommu *iommu,
2799 struct vfio_info_cap *caps)
2801 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas;
2802 struct vfio_iova *iova;
2804 int iovas = 0, i = 0, ret;
2806 list_for_each_entry(iova, &iommu->iova_list, list)
2811 * Return 0 as a container with a single mdev device
2812 * will have an empty list
2817 size = sizeof(*cap_iovas) + (iovas * sizeof(*cap_iovas->iova_ranges));
2819 cap_iovas = kzalloc(size, GFP_KERNEL);
2823 cap_iovas->nr_iovas = iovas;
2825 list_for_each_entry(iova, &iommu->iova_list, list) {
2826 cap_iovas->iova_ranges[i].start = iova->start;
2827 cap_iovas->iova_ranges[i].end = iova->end;
2831 ret = vfio_iommu_iova_add_cap(caps, cap_iovas, size);
2837 static int vfio_iommu_migration_build_caps(struct vfio_iommu *iommu,
2838 struct vfio_info_cap *caps)
2840 struct vfio_iommu_type1_info_cap_migration cap_mig;
2842 cap_mig.header.id = VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION;
2843 cap_mig.header.version = 1;
2846 /* support minimum pgsize */
2847 cap_mig.pgsize_bitmap = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2848 cap_mig.max_dirty_bitmap_size = DIRTY_BITMAP_SIZE_MAX;
2850 return vfio_info_add_capability(caps, &cap_mig.header, sizeof(cap_mig));
2853 static int vfio_iommu_dma_avail_build_caps(struct vfio_iommu *iommu,
2854 struct vfio_info_cap *caps)
2856 struct vfio_iommu_type1_info_dma_avail cap_dma_avail;
2858 cap_dma_avail.header.id = VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL;
2859 cap_dma_avail.header.version = 1;
2861 cap_dma_avail.avail = iommu->dma_avail;
2863 return vfio_info_add_capability(caps, &cap_dma_avail.header,
2864 sizeof(cap_dma_avail));
2867 static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu,
2870 struct vfio_iommu_type1_info info;
2871 unsigned long minsz;
2872 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
2873 unsigned long capsz;
2876 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
2878 /* For backward compatibility, cannot require this */
2879 capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
2881 if (copy_from_user(&info, (void __user *)arg, minsz))
2884 if (info.argsz < minsz)
2887 if (info.argsz >= capsz) {
2889 info.cap_offset = 0; /* output, no-recopy necessary */
2892 mutex_lock(&iommu->lock);
2893 info.flags = VFIO_IOMMU_INFO_PGSIZES;
2895 info.iova_pgsizes = iommu->pgsize_bitmap;
2897 ret = vfio_iommu_migration_build_caps(iommu, &caps);
2900 ret = vfio_iommu_dma_avail_build_caps(iommu, &caps);
2903 ret = vfio_iommu_iova_build_caps(iommu, &caps);
2905 mutex_unlock(&iommu->lock);
2911 info.flags |= VFIO_IOMMU_INFO_CAPS;
2913 if (info.argsz < sizeof(info) + caps.size) {
2914 info.argsz = sizeof(info) + caps.size;
2916 vfio_info_cap_shift(&caps, sizeof(info));
2917 if (copy_to_user((void __user *)arg +
2918 sizeof(info), caps.buf,
2923 info.cap_offset = sizeof(info);
2929 return copy_to_user((void __user *)arg, &info, minsz) ?
2933 static int vfio_iommu_type1_map_dma(struct vfio_iommu *iommu,
2936 struct vfio_iommu_type1_dma_map map;
2937 unsigned long minsz;
2938 uint32_t mask = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE |
2939 VFIO_DMA_MAP_FLAG_VADDR;
2941 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
2943 if (copy_from_user(&map, (void __user *)arg, minsz))
2946 if (map.argsz < minsz || map.flags & ~mask)
2949 return vfio_dma_do_map(iommu, &map);
2952 static int vfio_iommu_type1_unmap_dma(struct vfio_iommu *iommu,
2955 struct vfio_iommu_type1_dma_unmap unmap;
2956 struct vfio_bitmap bitmap = { 0 };
2957 uint32_t mask = VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP |
2958 VFIO_DMA_UNMAP_FLAG_VADDR |
2959 VFIO_DMA_UNMAP_FLAG_ALL;
2960 unsigned long minsz;
2963 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
2965 if (copy_from_user(&unmap, (void __user *)arg, minsz))
2968 if (unmap.argsz < minsz || unmap.flags & ~mask)
2971 if ((unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
2972 (unmap.flags & (VFIO_DMA_UNMAP_FLAG_ALL |
2973 VFIO_DMA_UNMAP_FLAG_VADDR)))
2976 if (unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
2977 unsigned long pgshift;
2979 if (unmap.argsz < (minsz + sizeof(bitmap)))
2982 if (copy_from_user(&bitmap,
2983 (void __user *)(arg + minsz),
2987 if (!access_ok((void __user *)bitmap.data, bitmap.size))
2990 pgshift = __ffs(bitmap.pgsize);
2991 ret = verify_bitmap_size(unmap.size >> pgshift,
2997 ret = vfio_dma_do_unmap(iommu, &unmap, &bitmap);
3001 return copy_to_user((void __user *)arg, &unmap, minsz) ?
3005 static int vfio_iommu_type1_dirty_pages(struct vfio_iommu *iommu,
3008 struct vfio_iommu_type1_dirty_bitmap dirty;
3009 uint32_t mask = VFIO_IOMMU_DIRTY_PAGES_FLAG_START |
3010 VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP |
3011 VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP;
3012 unsigned long minsz;
3018 minsz = offsetofend(struct vfio_iommu_type1_dirty_bitmap, flags);
3020 if (copy_from_user(&dirty, (void __user *)arg, minsz))
3023 if (dirty.argsz < minsz || dirty.flags & ~mask)
3026 /* only one flag should be set at a time */
3027 if (__ffs(dirty.flags) != __fls(dirty.flags))
3030 if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_START) {
3033 mutex_lock(&iommu->lock);
3034 pgsize = 1 << __ffs(iommu->pgsize_bitmap);
3035 if (!iommu->dirty_page_tracking) {
3036 ret = vfio_dma_bitmap_alloc_all(iommu, pgsize);
3038 iommu->dirty_page_tracking = true;
3040 mutex_unlock(&iommu->lock);
3042 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP) {
3043 mutex_lock(&iommu->lock);
3044 if (iommu->dirty_page_tracking) {
3045 iommu->dirty_page_tracking = false;
3046 vfio_dma_bitmap_free_all(iommu);
3048 mutex_unlock(&iommu->lock);
3050 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP) {
3051 struct vfio_iommu_type1_dirty_bitmap_get range;
3052 unsigned long pgshift;
3053 size_t data_size = dirty.argsz - minsz;
3054 size_t iommu_pgsize;
3056 if (!data_size || data_size < sizeof(range))
3059 if (copy_from_user(&range, (void __user *)(arg + minsz),
3063 if (range.iova + range.size < range.iova)
3065 if (!access_ok((void __user *)range.bitmap.data,
3069 pgshift = __ffs(range.bitmap.pgsize);
3070 ret = verify_bitmap_size(range.size >> pgshift,
3075 mutex_lock(&iommu->lock);
3077 iommu_pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
3079 /* allow only smallest supported pgsize */
3080 if (range.bitmap.pgsize != iommu_pgsize) {
3084 if (range.iova & (iommu_pgsize - 1)) {
3088 if (!range.size || range.size & (iommu_pgsize - 1)) {
3093 if (iommu->dirty_page_tracking)
3094 ret = vfio_iova_dirty_bitmap(range.bitmap.data,
3097 range.bitmap.pgsize);
3101 mutex_unlock(&iommu->lock);
3109 static long vfio_iommu_type1_ioctl(void *iommu_data,
3110 unsigned int cmd, unsigned long arg)
3112 struct vfio_iommu *iommu = iommu_data;
3115 case VFIO_CHECK_EXTENSION:
3116 return vfio_iommu_type1_check_extension(iommu, arg);
3117 case VFIO_IOMMU_GET_INFO:
3118 return vfio_iommu_type1_get_info(iommu, arg);
3119 case VFIO_IOMMU_MAP_DMA:
3120 return vfio_iommu_type1_map_dma(iommu, arg);
3121 case VFIO_IOMMU_UNMAP_DMA:
3122 return vfio_iommu_type1_unmap_dma(iommu, arg);
3123 case VFIO_IOMMU_DIRTY_PAGES:
3124 return vfio_iommu_type1_dirty_pages(iommu, arg);
3130 static int vfio_iommu_type1_register_notifier(void *iommu_data,
3131 unsigned long *events,
3132 struct notifier_block *nb)
3134 struct vfio_iommu *iommu = iommu_data;
3136 /* clear known events */
3137 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
3139 /* refuse to register if still events remaining */
3143 return blocking_notifier_chain_register(&iommu->notifier, nb);
3146 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
3147 struct notifier_block *nb)
3149 struct vfio_iommu *iommu = iommu_data;
3151 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
3154 static int vfio_iommu_type1_dma_rw_chunk(struct vfio_iommu *iommu,
3155 dma_addr_t user_iova, void *data,
3156 size_t count, bool write,
3159 struct mm_struct *mm;
3160 unsigned long vaddr;
3161 struct vfio_dma *dma;
3162 bool kthread = current->mm == NULL;
3168 ret = vfio_find_dma_valid(iommu, user_iova, 1, &dma);
3172 if ((write && !(dma->prot & IOMMU_WRITE)) ||
3173 !(dma->prot & IOMMU_READ))
3176 mm = get_task_mm(dma->task);
3183 else if (current->mm != mm)
3186 offset = user_iova - dma->iova;
3188 if (count > dma->size - offset)
3189 count = dma->size - offset;
3191 vaddr = dma->vaddr + offset;
3194 *copied = copy_to_user((void __user *)vaddr, data,
3196 if (*copied && iommu->dirty_page_tracking) {
3197 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
3199 * Bitmap populated with the smallest supported page
3202 bitmap_set(dma->bitmap, offset >> pgshift,
3203 ((offset + *copied - 1) >> pgshift) -
3204 (offset >> pgshift) + 1);
3207 *copied = copy_from_user(data, (void __user *)vaddr,
3210 kthread_unuse_mm(mm);
3213 return *copied ? 0 : -EFAULT;
3216 static int vfio_iommu_type1_dma_rw(void *iommu_data, dma_addr_t user_iova,
3217 void *data, size_t count, bool write)
3219 struct vfio_iommu *iommu = iommu_data;
3223 mutex_lock(&iommu->lock);
3225 ret = vfio_iommu_type1_dma_rw_chunk(iommu, user_iova, data,
3226 count, write, &done);
3235 mutex_unlock(&iommu->lock);
3239 static struct iommu_domain *
3240 vfio_iommu_type1_group_iommu_domain(void *iommu_data,
3241 struct iommu_group *iommu_group)
3243 struct iommu_domain *domain = ERR_PTR(-ENODEV);
3244 struct vfio_iommu *iommu = iommu_data;
3245 struct vfio_domain *d;
3247 if (!iommu || !iommu_group)
3248 return ERR_PTR(-EINVAL);
3250 mutex_lock(&iommu->lock);
3251 list_for_each_entry(d, &iommu->domain_list, next) {
3252 if (find_iommu_group(d, iommu_group)) {
3257 mutex_unlock(&iommu->lock);
3262 static void vfio_iommu_type1_notify(void *iommu_data,
3263 enum vfio_iommu_notify_type event)
3265 struct vfio_iommu *iommu = iommu_data;
3267 if (event != VFIO_IOMMU_CONTAINER_CLOSE)
3269 mutex_lock(&iommu->lock);
3270 iommu->container_open = false;
3271 mutex_unlock(&iommu->lock);
3272 wake_up_all(&iommu->vaddr_wait);
3275 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
3276 .name = "vfio-iommu-type1",
3277 .owner = THIS_MODULE,
3278 .open = vfio_iommu_type1_open,
3279 .release = vfio_iommu_type1_release,
3280 .ioctl = vfio_iommu_type1_ioctl,
3281 .attach_group = vfio_iommu_type1_attach_group,
3282 .detach_group = vfio_iommu_type1_detach_group,
3283 .pin_pages = vfio_iommu_type1_pin_pages,
3284 .unpin_pages = vfio_iommu_type1_unpin_pages,
3285 .register_notifier = vfio_iommu_type1_register_notifier,
3286 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
3287 .dma_rw = vfio_iommu_type1_dma_rw,
3288 .group_iommu_domain = vfio_iommu_type1_group_iommu_domain,
3289 .notify = vfio_iommu_type1_notify,
3292 static int __init vfio_iommu_type1_init(void)
3294 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
3297 static void __exit vfio_iommu_type1_cleanup(void)
3299 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
3302 module_init(vfio_iommu_type1_init);
3303 module_exit(vfio_iommu_type1_cleanup);
3305 MODULE_VERSION(DRIVER_VERSION);
3306 MODULE_LICENSE("GPL v2");
3307 MODULE_AUTHOR(DRIVER_AUTHOR);
3308 MODULE_DESCRIPTION(DRIVER_DESC);