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, size_t 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);
743 if (pinned && !rsvd) {
744 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
745 put_pfn(pfn, dma->prot);
747 vfio_batch_unpin(batch, dma);
755 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
756 unsigned long pfn, long npage,
759 long unlocked = 0, locked = 0;
762 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
763 if (put_pfn(pfn++, dma->prot)) {
765 if (vfio_find_vpfn(dma, iova))
771 vfio_lock_acct(dma, locked - unlocked, true);
776 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
777 unsigned long *pfn_base, bool do_accounting)
779 struct page *pages[1];
780 struct mm_struct *mm;
783 mm = get_task_mm(dma->task);
787 ret = vaddr_get_pfns(mm, vaddr, 1, dma->prot, pfn_base, pages);
788 if (ret == 1 && do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
789 ret = vfio_lock_acct(dma, 1, true);
791 put_pfn(*pfn_base, dma->prot);
793 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
794 "(%ld) exceeded\n", __func__,
795 dma->task->comm, task_pid_nr(dma->task),
796 task_rlimit(dma->task, RLIMIT_MEMLOCK));
804 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
808 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
813 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
816 vfio_lock_acct(dma, -unlocked, true);
821 static int vfio_iommu_type1_pin_pages(void *iommu_data,
822 struct iommu_group *iommu_group,
823 unsigned long *user_pfn,
825 unsigned long *phys_pfn)
827 struct vfio_iommu *iommu = iommu_data;
828 struct vfio_group *group;
830 unsigned long remote_vaddr;
831 struct vfio_dma *dma;
835 if (!iommu || !user_pfn || !phys_pfn)
838 /* Supported for v2 version only */
842 mutex_lock(&iommu->lock);
845 * Wait for all necessary vaddr's to be valid so they can be used in
846 * the main loop without dropping the lock, to avoid racing vs unmap.
849 if (iommu->vaddr_invalid_count) {
850 for (i = 0; i < npage; i++) {
851 iova = user_pfn[i] << PAGE_SHIFT;
852 ret = vfio_find_dma_valid(iommu, iova, PAGE_SIZE, &dma);
860 /* Fail if notifier list is empty */
861 if (!iommu->notifier.head) {
867 * If iommu capable domain exist in the container then all pages are
868 * already pinned and accounted. Accouting should be done if there is no
869 * iommu capable domain in the container.
871 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
873 for (i = 0; i < npage; i++) {
874 struct vfio_pfn *vpfn;
876 iova = user_pfn[i] << PAGE_SHIFT;
877 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
883 if ((dma->prot & prot) != prot) {
888 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
890 phys_pfn[i] = vpfn->pfn;
894 remote_vaddr = dma->vaddr + (iova - dma->iova);
895 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
900 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
902 if (put_pfn(phys_pfn[i], dma->prot) && do_accounting)
903 vfio_lock_acct(dma, -1, true);
907 if (iommu->dirty_page_tracking) {
908 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
911 * Bitmap populated with the smallest supported page
914 bitmap_set(dma->bitmap,
915 (iova - dma->iova) >> pgshift, 1);
920 group = vfio_iommu_find_iommu_group(iommu, iommu_group);
921 if (!group->pinned_page_dirty_scope) {
922 group->pinned_page_dirty_scope = true;
923 iommu->num_non_pinned_groups--;
930 for (j = 0; j < i; j++) {
933 iova = user_pfn[j] << PAGE_SHIFT;
934 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
935 vfio_unpin_page_external(dma, iova, do_accounting);
939 mutex_unlock(&iommu->lock);
943 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
944 unsigned long *user_pfn,
947 struct vfio_iommu *iommu = iommu_data;
951 if (!iommu || !user_pfn)
954 /* Supported for v2 version only */
958 mutex_lock(&iommu->lock);
960 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
961 for (i = 0; i < npage; i++) {
962 struct vfio_dma *dma;
965 iova = user_pfn[i] << PAGE_SHIFT;
966 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
969 vfio_unpin_page_external(dma, iova, do_accounting);
973 mutex_unlock(&iommu->lock);
974 return i > npage ? npage : (i > 0 ? i : -EINVAL);
977 static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
978 struct list_head *regions,
979 struct iommu_iotlb_gather *iotlb_gather)
982 struct vfio_regions *entry, *next;
984 iommu_iotlb_sync(domain->domain, iotlb_gather);
986 list_for_each_entry_safe(entry, next, regions, list) {
987 unlocked += vfio_unpin_pages_remote(dma,
989 entry->phys >> PAGE_SHIFT,
990 entry->len >> PAGE_SHIFT,
992 list_del(&entry->list);
1002 * Generally, VFIO needs to unpin remote pages after each IOTLB flush.
1003 * Therefore, when using IOTLB flush sync interface, VFIO need to keep track
1004 * of these regions (currently using a list).
1006 * This value specifies maximum number of regions for each IOTLB flush sync.
1008 #define VFIO_IOMMU_TLB_SYNC_MAX 512
1010 static size_t unmap_unpin_fast(struct vfio_domain *domain,
1011 struct vfio_dma *dma, dma_addr_t *iova,
1012 size_t len, phys_addr_t phys, long *unlocked,
1013 struct list_head *unmapped_list,
1015 struct iommu_iotlb_gather *iotlb_gather)
1017 size_t unmapped = 0;
1018 struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
1021 unmapped = iommu_unmap_fast(domain->domain, *iova, len,
1027 entry->iova = *iova;
1029 entry->len = unmapped;
1030 list_add_tail(&entry->list, unmapped_list);
1038 * Sync if the number of fast-unmap regions hits the limit
1039 * or in case of errors.
1041 if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
1042 *unlocked += vfio_sync_unpin(dma, domain, unmapped_list,
1050 static size_t unmap_unpin_slow(struct vfio_domain *domain,
1051 struct vfio_dma *dma, dma_addr_t *iova,
1052 size_t len, phys_addr_t phys,
1055 size_t unmapped = iommu_unmap(domain->domain, *iova, len);
1058 *unlocked += vfio_unpin_pages_remote(dma, *iova,
1060 unmapped >> PAGE_SHIFT,
1068 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
1071 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
1072 struct vfio_domain *domain, *d;
1073 LIST_HEAD(unmapped_region_list);
1074 struct iommu_iotlb_gather iotlb_gather;
1075 int unmapped_region_cnt = 0;
1081 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1085 * We use the IOMMU to track the physical addresses, otherwise we'd
1086 * need a much more complicated tracking system. Unfortunately that
1087 * means we need to use one of the iommu domains to figure out the
1088 * pfns to unpin. The rest need to be unmapped in advance so we have
1089 * no iommu translations remaining when the pages are unpinned.
1091 domain = d = list_first_entry(&iommu->domain_list,
1092 struct vfio_domain, next);
1094 list_for_each_entry_continue(d, &iommu->domain_list, next) {
1095 iommu_unmap(d->domain, dma->iova, dma->size);
1099 iommu_iotlb_gather_init(&iotlb_gather);
1100 while (iova < end) {
1101 size_t unmapped, len;
1102 phys_addr_t phys, next;
1104 phys = iommu_iova_to_phys(domain->domain, iova);
1105 if (WARN_ON(!phys)) {
1111 * To optimize for fewer iommu_unmap() calls, each of which
1112 * may require hardware cache flushing, try to find the
1113 * largest contiguous physical memory chunk to unmap.
1115 for (len = PAGE_SIZE;
1116 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
1117 next = iommu_iova_to_phys(domain->domain, iova + len);
1118 if (next != phys + len)
1123 * First, try to use fast unmap/unpin. In case of failure,
1124 * switch to slow unmap/unpin path.
1126 unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys,
1127 &unlocked, &unmapped_region_list,
1128 &unmapped_region_cnt,
1131 unmapped = unmap_unpin_slow(domain, dma, &iova, len,
1133 if (WARN_ON(!unmapped))
1138 dma->iommu_mapped = false;
1140 if (unmapped_region_cnt) {
1141 unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list,
1145 if (do_accounting) {
1146 vfio_lock_acct(dma, -unlocked, true);
1152 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
1154 WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list));
1155 vfio_unmap_unpin(iommu, dma, true);
1156 vfio_unlink_dma(iommu, dma);
1157 put_task_struct(dma->task);
1158 vfio_dma_bitmap_free(dma);
1159 if (dma->vaddr_invalid) {
1160 iommu->vaddr_invalid_count--;
1161 wake_up_all(&iommu->vaddr_wait);
1167 static void vfio_update_pgsize_bitmap(struct vfio_iommu *iommu)
1169 struct vfio_domain *domain;
1171 iommu->pgsize_bitmap = ULONG_MAX;
1173 list_for_each_entry(domain, &iommu->domain_list, next)
1174 iommu->pgsize_bitmap &= domain->domain->pgsize_bitmap;
1177 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
1178 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
1179 * That way the user will be able to map/unmap buffers whose size/
1180 * start address is aligned with PAGE_SIZE. Pinning code uses that
1181 * granularity while iommu driver can use the sub-PAGE_SIZE size
1182 * to map the buffer.
1184 if (iommu->pgsize_bitmap & ~PAGE_MASK) {
1185 iommu->pgsize_bitmap &= PAGE_MASK;
1186 iommu->pgsize_bitmap |= PAGE_SIZE;
1190 static int update_user_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1191 struct vfio_dma *dma, dma_addr_t base_iova,
1194 unsigned long pgshift = __ffs(pgsize);
1195 unsigned long nbits = dma->size >> pgshift;
1196 unsigned long bit_offset = (dma->iova - base_iova) >> pgshift;
1197 unsigned long copy_offset = bit_offset / BITS_PER_LONG;
1198 unsigned long shift = bit_offset % BITS_PER_LONG;
1199 unsigned long leftover;
1202 * mark all pages dirty if any IOMMU capable device is not able
1203 * to report dirty pages and all pages are pinned and mapped.
1205 if (iommu->num_non_pinned_groups && dma->iommu_mapped)
1206 bitmap_set(dma->bitmap, 0, nbits);
1209 bitmap_shift_left(dma->bitmap, dma->bitmap, shift,
1212 if (copy_from_user(&leftover,
1213 (void __user *)(bitmap + copy_offset),
1217 bitmap_or(dma->bitmap, dma->bitmap, &leftover, shift);
1220 if (copy_to_user((void __user *)(bitmap + copy_offset), dma->bitmap,
1221 DIRTY_BITMAP_BYTES(nbits + shift)))
1227 static int vfio_iova_dirty_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1228 dma_addr_t iova, size_t size, size_t pgsize)
1230 struct vfio_dma *dma;
1232 unsigned long pgshift = __ffs(pgsize);
1236 * GET_BITMAP request must fully cover vfio_dma mappings. Multiple
1237 * vfio_dma mappings may be clubbed by specifying large ranges, but
1238 * there must not be any previous mappings bisected by the range.
1239 * An error will be returned if these conditions are not met.
1241 dma = vfio_find_dma(iommu, iova, 1);
1242 if (dma && dma->iova != iova)
1245 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1246 if (dma && dma->iova + dma->size != iova + size)
1249 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1250 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1252 if (dma->iova < iova)
1255 if (dma->iova > iova + size - 1)
1258 ret = update_user_bitmap(bitmap, iommu, dma, iova, pgsize);
1263 * Re-populate bitmap to include all pinned pages which are
1264 * considered as dirty but exclude pages which are unpinned and
1265 * pages which are marked dirty by vfio_dma_rw()
1267 bitmap_clear(dma->bitmap, 0, dma->size >> pgshift);
1268 vfio_dma_populate_bitmap(dma, pgsize);
1273 static int verify_bitmap_size(uint64_t npages, uint64_t bitmap_size)
1275 if (!npages || !bitmap_size || (bitmap_size > DIRTY_BITMAP_SIZE_MAX) ||
1276 (bitmap_size < DIRTY_BITMAP_BYTES(npages)))
1282 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
1283 struct vfio_iommu_type1_dma_unmap *unmap,
1284 struct vfio_bitmap *bitmap)
1286 struct vfio_dma *dma, *dma_last = NULL;
1287 size_t unmapped = 0, pgsize;
1288 int ret = -EINVAL, retries = 0;
1289 unsigned long pgshift;
1290 dma_addr_t iova = unmap->iova;
1291 unsigned long size = unmap->size;
1292 bool unmap_all = unmap->flags & VFIO_DMA_UNMAP_FLAG_ALL;
1293 bool invalidate_vaddr = unmap->flags & VFIO_DMA_UNMAP_FLAG_VADDR;
1294 struct rb_node *n, *first_n;
1296 mutex_lock(&iommu->lock);
1298 pgshift = __ffs(iommu->pgsize_bitmap);
1299 pgsize = (size_t)1 << pgshift;
1301 if (iova & (pgsize - 1))
1308 } else if (!size || size & (pgsize - 1)) {
1312 if (iova + size - 1 < iova || size > SIZE_MAX)
1315 /* When dirty tracking is enabled, allow only min supported pgsize */
1316 if ((unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
1317 (!iommu->dirty_page_tracking || (bitmap->pgsize != pgsize))) {
1321 WARN_ON((pgsize - 1) & PAGE_MASK);
1324 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
1325 * avoid tracking individual mappings. This means that the granularity
1326 * of the original mapping was lost and the user was allowed to attempt
1327 * to unmap any range. Depending on the contiguousness of physical
1328 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
1329 * or may not have worked. We only guaranteed unmap granularity
1330 * matching the original mapping; even though it was untracked here,
1331 * the original mappings are reflected in IOMMU mappings. This
1332 * resulted in a couple unusual behaviors. First, if a range is not
1333 * able to be unmapped, ex. a set of 4k pages that was mapped as a
1334 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
1335 * a zero sized unmap. Also, if an unmap request overlaps the first
1336 * address of a hugepage, the IOMMU will unmap the entire hugepage.
1337 * This also returns success and the returned unmap size reflects the
1338 * actual size unmapped.
1340 * We attempt to maintain compatibility with this "v1" interface, but
1341 * we take control out of the hands of the IOMMU. Therefore, an unmap
1342 * request offset from the beginning of the original mapping will
1343 * return success with zero sized unmap. And an unmap request covering
1344 * the first iova of mapping will unmap the entire range.
1346 * The v2 version of this interface intends to be more deterministic.
1347 * Unmap requests must fully cover previous mappings. Multiple
1348 * mappings may still be unmaped by specifying large ranges, but there
1349 * must not be any previous mappings bisected by the range. An error
1350 * will be returned if these conditions are not met. The v2 interface
1351 * will only return success and a size of zero if there were no
1352 * mappings within the range.
1354 if (iommu->v2 && !unmap_all) {
1355 dma = vfio_find_dma(iommu, iova, 1);
1356 if (dma && dma->iova != iova)
1359 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1360 if (dma && dma->iova + dma->size != iova + size)
1365 n = first_n = vfio_find_dma_first_node(iommu, iova, size);
1368 dma = rb_entry(n, struct vfio_dma, node);
1369 if (dma->iova >= iova + size)
1372 if (!iommu->v2 && iova > dma->iova)
1375 * Task with same address space who mapped this iova range is
1376 * allowed to unmap the iova range.
1378 if (dma->task->mm != current->mm)
1381 if (invalidate_vaddr) {
1382 if (dma->vaddr_invalid) {
1383 struct rb_node *last_n = n;
1385 for (n = first_n; n != last_n; n = rb_next(n)) {
1387 struct vfio_dma, node);
1388 dma->vaddr_invalid = false;
1389 iommu->vaddr_invalid_count--;
1395 dma->vaddr_invalid = true;
1396 iommu->vaddr_invalid_count++;
1397 unmapped += dma->size;
1402 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
1403 struct vfio_iommu_type1_dma_unmap nb_unmap;
1405 if (dma_last == dma) {
1406 BUG_ON(++retries > 10);
1412 nb_unmap.iova = dma->iova;
1413 nb_unmap.size = dma->size;
1416 * Notify anyone (mdev vendor drivers) to invalidate and
1417 * unmap iovas within the range we're about to unmap.
1418 * Vendor drivers MUST unpin pages in response to an
1421 mutex_unlock(&iommu->lock);
1422 blocking_notifier_call_chain(&iommu->notifier,
1423 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
1425 mutex_lock(&iommu->lock);
1429 if (unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
1430 ret = update_user_bitmap(bitmap->data, iommu, dma,
1436 unmapped += dma->size;
1438 vfio_remove_dma(iommu, dma);
1442 mutex_unlock(&iommu->lock);
1444 /* Report how much was unmapped */
1445 unmap->size = unmapped;
1450 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
1451 unsigned long pfn, long npage, int prot)
1453 struct vfio_domain *d;
1456 list_for_each_entry(d, &iommu->domain_list, next) {
1457 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
1458 npage << PAGE_SHIFT, prot | d->prot);
1468 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) {
1469 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
1476 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
1479 dma_addr_t iova = dma->iova;
1480 unsigned long vaddr = dma->vaddr;
1481 struct vfio_batch batch;
1482 size_t size = map_size;
1484 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1487 vfio_batch_init(&batch);
1490 /* Pin a contiguous chunk of memory */
1491 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
1492 size >> PAGE_SHIFT, &pfn, limit,
1501 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
1504 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
1506 vfio_batch_unpin(&batch, dma);
1510 size -= npage << PAGE_SHIFT;
1511 dma->size += npage << PAGE_SHIFT;
1514 vfio_batch_fini(&batch);
1515 dma->iommu_mapped = true;
1518 vfio_remove_dma(iommu, dma);
1524 * Check dma map request is within a valid iova range
1526 static bool vfio_iommu_iova_dma_valid(struct vfio_iommu *iommu,
1527 dma_addr_t start, dma_addr_t end)
1529 struct list_head *iova = &iommu->iova_list;
1530 struct vfio_iova *node;
1532 list_for_each_entry(node, iova, list) {
1533 if (start >= node->start && end <= node->end)
1538 * Check for list_empty() as well since a container with
1539 * a single mdev device will have an empty list.
1541 return list_empty(iova);
1544 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1545 struct vfio_iommu_type1_dma_map *map)
1547 bool set_vaddr = map->flags & VFIO_DMA_MAP_FLAG_VADDR;
1548 dma_addr_t iova = map->iova;
1549 unsigned long vaddr = map->vaddr;
1550 size_t size = map->size;
1551 int ret = 0, prot = 0;
1553 struct vfio_dma *dma;
1555 /* Verify that none of our __u64 fields overflow */
1556 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1559 /* READ/WRITE from device perspective */
1560 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1561 prot |= IOMMU_WRITE;
1562 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1565 if ((prot && set_vaddr) || (!prot && !set_vaddr))
1568 mutex_lock(&iommu->lock);
1570 pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
1572 WARN_ON((pgsize - 1) & PAGE_MASK);
1574 if (!size || (size | iova | vaddr) & (pgsize - 1)) {
1579 /* Don't allow IOVA or virtual address wrap */
1580 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) {
1585 dma = vfio_find_dma(iommu, iova, size);
1589 } else if (!dma->vaddr_invalid || dma->iova != iova ||
1590 dma->size != size) {
1594 dma->vaddr_invalid = false;
1595 iommu->vaddr_invalid_count--;
1596 wake_up_all(&iommu->vaddr_wait);
1604 if (!iommu->dma_avail) {
1609 if (!vfio_iommu_iova_dma_valid(iommu, iova, iova + size - 1)) {
1614 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1626 * We need to be able to both add to a task's locked memory and test
1627 * against the locked memory limit and we need to be able to do both
1628 * outside of this call path as pinning can be asynchronous via the
1629 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1630 * task_struct and VM locked pages requires an mm_struct, however
1631 * holding an indefinite mm reference is not recommended, therefore we
1632 * only hold a reference to a task. We could hold a reference to
1633 * current, however QEMU uses this call path through vCPU threads,
1634 * which can be killed resulting in a NULL mm and failure in the unmap
1635 * path when called via a different thread. Avoid this problem by
1636 * using the group_leader as threads within the same group require
1637 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1640 * Previously we also used the task for testing CAP_IPC_LOCK at the
1641 * time of pinning and accounting, however has_capability() makes use
1642 * of real_cred, a copy-on-write field, so we can't guarantee that it
1643 * matches group_leader, or in fact that it might not change by the
1644 * time it's evaluated. If a process were to call MAP_DMA with
1645 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1646 * possibly see different results for an iommu_mapped vfio_dma vs
1647 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1648 * time of calling MAP_DMA.
1650 get_task_struct(current->group_leader);
1651 dma->task = current->group_leader;
1652 dma->lock_cap = capable(CAP_IPC_LOCK);
1654 dma->pfn_list = RB_ROOT;
1656 /* Insert zero-sized and grow as we map chunks of it */
1657 vfio_link_dma(iommu, dma);
1659 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1660 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1663 ret = vfio_pin_map_dma(iommu, dma, size);
1665 if (!ret && iommu->dirty_page_tracking) {
1666 ret = vfio_dma_bitmap_alloc(dma, pgsize);
1668 vfio_remove_dma(iommu, dma);
1672 mutex_unlock(&iommu->lock);
1676 static int vfio_bus_type(struct device *dev, void *data)
1678 struct bus_type **bus = data;
1680 if (*bus && *bus != dev->bus)
1688 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1689 struct vfio_domain *domain)
1691 struct vfio_batch batch;
1692 struct vfio_domain *d = NULL;
1694 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1697 ret = vfio_wait_all_valid(iommu);
1701 /* Arbitrarily pick the first domain in the list for lookups */
1702 if (!list_empty(&iommu->domain_list))
1703 d = list_first_entry(&iommu->domain_list,
1704 struct vfio_domain, next);
1706 vfio_batch_init(&batch);
1708 n = rb_first(&iommu->dma_list);
1710 for (; n; n = rb_next(n)) {
1711 struct vfio_dma *dma;
1714 dma = rb_entry(n, struct vfio_dma, node);
1717 while (iova < dma->iova + dma->size) {
1721 if (dma->iommu_mapped) {
1725 if (WARN_ON(!d)) { /* mapped w/o a domain?! */
1730 phys = iommu_iova_to_phys(d->domain, iova);
1732 if (WARN_ON(!phys)) {
1740 while (i < dma->iova + dma->size &&
1741 p == iommu_iova_to_phys(d->domain, i)) {
1748 unsigned long vaddr = dma->vaddr +
1750 size_t n = dma->iova + dma->size - iova;
1753 npage = vfio_pin_pages_remote(dma, vaddr,
1763 phys = pfn << PAGE_SHIFT;
1764 size = npage << PAGE_SHIFT;
1767 ret = iommu_map(domain->domain, iova, phys,
1768 size, dma->prot | domain->prot);
1770 if (!dma->iommu_mapped) {
1771 vfio_unpin_pages_remote(dma, iova,
1775 vfio_batch_unpin(&batch, dma);
1784 /* All dmas are now mapped, defer to second tree walk for unwind */
1785 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1786 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1788 dma->iommu_mapped = true;
1791 vfio_batch_fini(&batch);
1795 for (; n; n = rb_prev(n)) {
1796 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1799 if (dma->iommu_mapped) {
1800 iommu_unmap(domain->domain, dma->iova, dma->size);
1805 while (iova < dma->iova + dma->size) {
1806 phys_addr_t phys, p;
1810 phys = iommu_iova_to_phys(domain->domain, iova);
1819 while (i < dma->iova + dma->size &&
1820 p == iommu_iova_to_phys(domain->domain, i)) {
1826 iommu_unmap(domain->domain, iova, size);
1827 vfio_unpin_pages_remote(dma, iova, phys >> PAGE_SHIFT,
1828 size >> PAGE_SHIFT, true);
1832 vfio_batch_fini(&batch);
1837 * We change our unmap behavior slightly depending on whether the IOMMU
1838 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1839 * for practically any contiguous power-of-two mapping we give it. This means
1840 * we don't need to look for contiguous chunks ourselves to make unmapping
1841 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1842 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1843 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1844 * hugetlbfs is in use.
1846 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1849 int ret, order = get_order(PAGE_SIZE * 2);
1851 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1855 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1856 IOMMU_READ | IOMMU_WRITE | domain->prot);
1858 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1860 if (unmapped == PAGE_SIZE)
1861 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1863 domain->fgsp = true;
1866 __free_pages(pages, order);
1869 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1870 struct iommu_group *iommu_group)
1872 struct vfio_group *g;
1874 list_for_each_entry(g, &domain->group_list, next) {
1875 if (g->iommu_group == iommu_group)
1882 static struct vfio_group *vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
1883 struct iommu_group *iommu_group)
1885 struct vfio_domain *domain;
1886 struct vfio_group *group = NULL;
1888 list_for_each_entry(domain, &iommu->domain_list, next) {
1889 group = find_iommu_group(domain, iommu_group);
1894 if (iommu->external_domain)
1895 group = find_iommu_group(iommu->external_domain, iommu_group);
1900 static bool vfio_iommu_has_sw_msi(struct list_head *group_resv_regions,
1903 struct iommu_resv_region *region;
1906 list_for_each_entry(region, group_resv_regions, list) {
1908 * The presence of any 'real' MSI regions should take
1909 * precedence over the software-managed one if the
1910 * IOMMU driver happens to advertise both types.
1912 if (region->type == IOMMU_RESV_MSI) {
1917 if (region->type == IOMMU_RESV_SW_MSI) {
1918 *base = region->start;
1926 static struct device *vfio_mdev_get_iommu_device(struct device *dev)
1928 struct device *(*fn)(struct device *dev);
1929 struct device *iommu_device;
1931 fn = symbol_get(mdev_get_iommu_device);
1933 iommu_device = fn(dev);
1934 symbol_put(mdev_get_iommu_device);
1936 return iommu_device;
1942 static int vfio_mdev_attach_domain(struct device *dev, void *data)
1944 struct iommu_domain *domain = data;
1945 struct device *iommu_device;
1947 iommu_device = vfio_mdev_get_iommu_device(dev);
1949 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1950 return iommu_aux_attach_device(domain, iommu_device);
1952 return iommu_attach_device(domain, iommu_device);
1958 static int vfio_mdev_detach_domain(struct device *dev, void *data)
1960 struct iommu_domain *domain = data;
1961 struct device *iommu_device;
1963 iommu_device = vfio_mdev_get_iommu_device(dev);
1965 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1966 iommu_aux_detach_device(domain, iommu_device);
1968 iommu_detach_device(domain, iommu_device);
1974 static int vfio_iommu_attach_group(struct vfio_domain *domain,
1975 struct vfio_group *group)
1977 if (group->mdev_group)
1978 return iommu_group_for_each_dev(group->iommu_group,
1980 vfio_mdev_attach_domain);
1982 return iommu_attach_group(domain->domain, group->iommu_group);
1985 static void vfio_iommu_detach_group(struct vfio_domain *domain,
1986 struct vfio_group *group)
1988 if (group->mdev_group)
1989 iommu_group_for_each_dev(group->iommu_group, domain->domain,
1990 vfio_mdev_detach_domain);
1992 iommu_detach_group(domain->domain, group->iommu_group);
1995 static bool vfio_bus_is_mdev(struct bus_type *bus)
1997 struct bus_type *mdev_bus;
2000 mdev_bus = symbol_get(mdev_bus_type);
2002 ret = (bus == mdev_bus);
2003 symbol_put(mdev_bus_type);
2009 static int vfio_mdev_iommu_device(struct device *dev, void *data)
2011 struct device **old = data, *new;
2013 new = vfio_mdev_get_iommu_device(dev);
2014 if (!new || (*old && *old != new))
2023 * This is a helper function to insert an address range to iova list.
2024 * The list is initially created with a single entry corresponding to
2025 * the IOMMU domain geometry to which the device group is attached.
2026 * The list aperture gets modified when a new domain is added to the
2027 * container if the new aperture doesn't conflict with the current one
2028 * or with any existing dma mappings. The list is also modified to
2029 * exclude any reserved regions associated with the device group.
2031 static int vfio_iommu_iova_insert(struct list_head *head,
2032 dma_addr_t start, dma_addr_t end)
2034 struct vfio_iova *region;
2036 region = kmalloc(sizeof(*region), GFP_KERNEL);
2040 INIT_LIST_HEAD(®ion->list);
2041 region->start = start;
2044 list_add_tail(®ion->list, head);
2049 * Check the new iommu aperture conflicts with existing aper or with any
2050 * existing dma mappings.
2052 static bool vfio_iommu_aper_conflict(struct vfio_iommu *iommu,
2053 dma_addr_t start, dma_addr_t end)
2055 struct vfio_iova *first, *last;
2056 struct list_head *iova = &iommu->iova_list;
2058 if (list_empty(iova))
2061 /* Disjoint sets, return conflict */
2062 first = list_first_entry(iova, struct vfio_iova, list);
2063 last = list_last_entry(iova, struct vfio_iova, list);
2064 if (start > last->end || end < first->start)
2067 /* Check for any existing dma mappings below the new start */
2068 if (start > first->start) {
2069 if (vfio_find_dma(iommu, first->start, start - first->start))
2073 /* Check for any existing dma mappings beyond the new end */
2074 if (end < last->end) {
2075 if (vfio_find_dma(iommu, end + 1, last->end - end))
2083 * Resize iommu iova aperture window. This is called only if the new
2084 * aperture has no conflict with existing aperture and dma mappings.
2086 static int vfio_iommu_aper_resize(struct list_head *iova,
2087 dma_addr_t start, dma_addr_t end)
2089 struct vfio_iova *node, *next;
2091 if (list_empty(iova))
2092 return vfio_iommu_iova_insert(iova, start, end);
2094 /* Adjust iova list start */
2095 list_for_each_entry_safe(node, next, iova, list) {
2096 if (start < node->start)
2098 if (start >= node->start && start < node->end) {
2099 node->start = start;
2102 /* Delete nodes before new start */
2103 list_del(&node->list);
2107 /* Adjust iova list end */
2108 list_for_each_entry_safe(node, next, iova, list) {
2109 if (end > node->end)
2111 if (end > node->start && end <= node->end) {
2115 /* Delete nodes after new end */
2116 list_del(&node->list);
2124 * Check reserved region conflicts with existing dma mappings
2126 static bool vfio_iommu_resv_conflict(struct vfio_iommu *iommu,
2127 struct list_head *resv_regions)
2129 struct iommu_resv_region *region;
2131 /* Check for conflict with existing dma mappings */
2132 list_for_each_entry(region, resv_regions, list) {
2133 if (region->type == IOMMU_RESV_DIRECT_RELAXABLE)
2136 if (vfio_find_dma(iommu, region->start, region->length))
2144 * Check iova region overlap with reserved regions and
2145 * exclude them from the iommu iova range
2147 static int vfio_iommu_resv_exclude(struct list_head *iova,
2148 struct list_head *resv_regions)
2150 struct iommu_resv_region *resv;
2151 struct vfio_iova *n, *next;
2153 list_for_each_entry(resv, resv_regions, list) {
2154 phys_addr_t start, end;
2156 if (resv->type == IOMMU_RESV_DIRECT_RELAXABLE)
2159 start = resv->start;
2160 end = resv->start + resv->length - 1;
2162 list_for_each_entry_safe(n, next, iova, list) {
2166 if (start > n->end || end < n->start)
2169 * Insert a new node if current node overlaps with the
2170 * reserve region to exlude that from valid iova range.
2171 * Note that, new node is inserted before the current
2172 * node and finally the current node is deleted keeping
2173 * the list updated and sorted.
2175 if (start > n->start)
2176 ret = vfio_iommu_iova_insert(&n->list, n->start,
2178 if (!ret && end < n->end)
2179 ret = vfio_iommu_iova_insert(&n->list, end + 1,
2189 if (list_empty(iova))
2195 static void vfio_iommu_resv_free(struct list_head *resv_regions)
2197 struct iommu_resv_region *n, *next;
2199 list_for_each_entry_safe(n, next, resv_regions, list) {
2205 static void vfio_iommu_iova_free(struct list_head *iova)
2207 struct vfio_iova *n, *next;
2209 list_for_each_entry_safe(n, next, iova, list) {
2215 static int vfio_iommu_iova_get_copy(struct vfio_iommu *iommu,
2216 struct list_head *iova_copy)
2218 struct list_head *iova = &iommu->iova_list;
2219 struct vfio_iova *n;
2222 list_for_each_entry(n, iova, list) {
2223 ret = vfio_iommu_iova_insert(iova_copy, n->start, n->end);
2231 vfio_iommu_iova_free(iova_copy);
2235 static void vfio_iommu_iova_insert_copy(struct vfio_iommu *iommu,
2236 struct list_head *iova_copy)
2238 struct list_head *iova = &iommu->iova_list;
2240 vfio_iommu_iova_free(iova);
2242 list_splice_tail(iova_copy, iova);
2245 static int vfio_iommu_type1_attach_group(void *iommu_data,
2246 struct iommu_group *iommu_group)
2248 struct vfio_iommu *iommu = iommu_data;
2249 struct vfio_group *group;
2250 struct vfio_domain *domain, *d;
2251 struct bus_type *bus = NULL;
2253 bool resv_msi, msi_remap;
2254 phys_addr_t resv_msi_base = 0;
2255 struct iommu_domain_geometry geo;
2256 LIST_HEAD(iova_copy);
2257 LIST_HEAD(group_resv_regions);
2259 mutex_lock(&iommu->lock);
2261 /* Check for duplicates */
2262 if (vfio_iommu_find_iommu_group(iommu, iommu_group)) {
2263 mutex_unlock(&iommu->lock);
2267 group = kzalloc(sizeof(*group), GFP_KERNEL);
2268 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2269 if (!group || !domain) {
2274 group->iommu_group = iommu_group;
2276 /* Determine bus_type in order to allocate a domain */
2277 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
2281 if (vfio_bus_is_mdev(bus)) {
2282 struct device *iommu_device = NULL;
2284 group->mdev_group = true;
2286 /* Determine the isolation type */
2287 ret = iommu_group_for_each_dev(iommu_group, &iommu_device,
2288 vfio_mdev_iommu_device);
2289 if (ret || !iommu_device) {
2290 if (!iommu->external_domain) {
2291 INIT_LIST_HEAD(&domain->group_list);
2292 iommu->external_domain = domain;
2293 vfio_update_pgsize_bitmap(iommu);
2298 list_add(&group->next,
2299 &iommu->external_domain->group_list);
2301 * Non-iommu backed group cannot dirty memory directly,
2302 * it can only use interfaces that provide dirty
2304 * The iommu scope can only be promoted with the
2305 * addition of a dirty tracking group.
2307 group->pinned_page_dirty_scope = true;
2308 mutex_unlock(&iommu->lock);
2313 bus = iommu_device->bus;
2316 domain->domain = iommu_domain_alloc(bus);
2317 if (!domain->domain) {
2322 if (iommu->nesting) {
2325 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
2331 ret = vfio_iommu_attach_group(domain, group);
2335 /* Get aperture info */
2336 iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY, &geo);
2338 if (vfio_iommu_aper_conflict(iommu, geo.aperture_start,
2339 geo.aperture_end)) {
2344 ret = iommu_get_group_resv_regions(iommu_group, &group_resv_regions);
2348 if (vfio_iommu_resv_conflict(iommu, &group_resv_regions)) {
2354 * We don't want to work on the original iova list as the list
2355 * gets modified and in case of failure we have to retain the
2356 * original list. Get a copy here.
2358 ret = vfio_iommu_iova_get_copy(iommu, &iova_copy);
2362 ret = vfio_iommu_aper_resize(&iova_copy, geo.aperture_start,
2367 ret = vfio_iommu_resv_exclude(&iova_copy, &group_resv_regions);
2371 resv_msi = vfio_iommu_has_sw_msi(&group_resv_regions, &resv_msi_base);
2373 INIT_LIST_HEAD(&domain->group_list);
2374 list_add(&group->next, &domain->group_list);
2376 msi_remap = irq_domain_check_msi_remap() ||
2377 iommu_capable(bus, IOMMU_CAP_INTR_REMAP);
2379 if (!allow_unsafe_interrupts && !msi_remap) {
2380 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
2386 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
2387 domain->prot |= IOMMU_CACHE;
2390 * Try to match an existing compatible domain. We don't want to
2391 * preclude an IOMMU driver supporting multiple bus_types and being
2392 * able to include different bus_types in the same IOMMU domain, so
2393 * we test whether the domains use the same iommu_ops rather than
2394 * testing if they're on the same bus_type.
2396 list_for_each_entry(d, &iommu->domain_list, next) {
2397 if (d->domain->ops == domain->domain->ops &&
2398 d->prot == domain->prot) {
2399 vfio_iommu_detach_group(domain, group);
2400 if (!vfio_iommu_attach_group(d, group)) {
2401 list_add(&group->next, &d->group_list);
2402 iommu_domain_free(domain->domain);
2407 ret = vfio_iommu_attach_group(domain, group);
2413 vfio_test_domain_fgsp(domain);
2415 /* replay mappings on new domains */
2416 ret = vfio_iommu_replay(iommu, domain);
2421 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
2422 if (ret && ret != -ENODEV)
2426 list_add(&domain->next, &iommu->domain_list);
2427 vfio_update_pgsize_bitmap(iommu);
2429 /* Delete the old one and insert new iova list */
2430 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2433 * An iommu backed group can dirty memory directly and therefore
2434 * demotes the iommu scope until it declares itself dirty tracking
2435 * capable via the page pinning interface.
2437 iommu->num_non_pinned_groups++;
2438 mutex_unlock(&iommu->lock);
2439 vfio_iommu_resv_free(&group_resv_regions);
2444 vfio_iommu_detach_group(domain, group);
2446 iommu_domain_free(domain->domain);
2447 vfio_iommu_iova_free(&iova_copy);
2448 vfio_iommu_resv_free(&group_resv_regions);
2452 mutex_unlock(&iommu->lock);
2456 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
2458 struct rb_node *node;
2460 while ((node = rb_first(&iommu->dma_list)))
2461 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
2464 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
2466 struct rb_node *n, *p;
2468 n = rb_first(&iommu->dma_list);
2469 for (; n; n = rb_next(n)) {
2470 struct vfio_dma *dma;
2471 long locked = 0, unlocked = 0;
2473 dma = rb_entry(n, struct vfio_dma, node);
2474 unlocked += vfio_unmap_unpin(iommu, dma, false);
2475 p = rb_first(&dma->pfn_list);
2476 for (; p; p = rb_next(p)) {
2477 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
2480 if (!is_invalid_reserved_pfn(vpfn->pfn))
2483 vfio_lock_acct(dma, locked - unlocked, true);
2488 * Called when a domain is removed in detach. It is possible that
2489 * the removed domain decided the iova aperture window. Modify the
2490 * iova aperture with the smallest window among existing domains.
2492 static void vfio_iommu_aper_expand(struct vfio_iommu *iommu,
2493 struct list_head *iova_copy)
2495 struct vfio_domain *domain;
2496 struct iommu_domain_geometry geo;
2497 struct vfio_iova *node;
2498 dma_addr_t start = 0;
2499 dma_addr_t end = (dma_addr_t)~0;
2501 if (list_empty(iova_copy))
2504 list_for_each_entry(domain, &iommu->domain_list, next) {
2505 iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY,
2507 if (geo.aperture_start > start)
2508 start = geo.aperture_start;
2509 if (geo.aperture_end < end)
2510 end = geo.aperture_end;
2513 /* Modify aperture limits. The new aper is either same or bigger */
2514 node = list_first_entry(iova_copy, struct vfio_iova, list);
2515 node->start = start;
2516 node = list_last_entry(iova_copy, struct vfio_iova, list);
2521 * Called when a group is detached. The reserved regions for that
2522 * group can be part of valid iova now. But since reserved regions
2523 * may be duplicated among groups, populate the iova valid regions
2526 static int vfio_iommu_resv_refresh(struct vfio_iommu *iommu,
2527 struct list_head *iova_copy)
2529 struct vfio_domain *d;
2530 struct vfio_group *g;
2531 struct vfio_iova *node;
2532 dma_addr_t start, end;
2533 LIST_HEAD(resv_regions);
2536 if (list_empty(iova_copy))
2539 list_for_each_entry(d, &iommu->domain_list, next) {
2540 list_for_each_entry(g, &d->group_list, next) {
2541 ret = iommu_get_group_resv_regions(g->iommu_group,
2548 node = list_first_entry(iova_copy, struct vfio_iova, list);
2549 start = node->start;
2550 node = list_last_entry(iova_copy, struct vfio_iova, list);
2553 /* purge the iova list and create new one */
2554 vfio_iommu_iova_free(iova_copy);
2556 ret = vfio_iommu_aper_resize(iova_copy, start, end);
2560 /* Exclude current reserved regions from iova ranges */
2561 ret = vfio_iommu_resv_exclude(iova_copy, &resv_regions);
2563 vfio_iommu_resv_free(&resv_regions);
2567 static void vfio_iommu_type1_detach_group(void *iommu_data,
2568 struct iommu_group *iommu_group)
2570 struct vfio_iommu *iommu = iommu_data;
2571 struct vfio_domain *domain;
2572 struct vfio_group *group;
2573 bool update_dirty_scope = false;
2574 LIST_HEAD(iova_copy);
2576 mutex_lock(&iommu->lock);
2578 if (iommu->external_domain) {
2579 group = find_iommu_group(iommu->external_domain, iommu_group);
2581 update_dirty_scope = !group->pinned_page_dirty_scope;
2582 list_del(&group->next);
2585 if (list_empty(&iommu->external_domain->group_list)) {
2586 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) {
2587 WARN_ON(iommu->notifier.head);
2588 vfio_iommu_unmap_unpin_all(iommu);
2591 kfree(iommu->external_domain);
2592 iommu->external_domain = NULL;
2594 goto detach_group_done;
2599 * Get a copy of iova list. This will be used to update
2600 * and to replace the current one later. Please note that
2601 * we will leave the original list as it is if update fails.
2603 vfio_iommu_iova_get_copy(iommu, &iova_copy);
2605 list_for_each_entry(domain, &iommu->domain_list, next) {
2606 group = find_iommu_group(domain, iommu_group);
2610 vfio_iommu_detach_group(domain, group);
2611 update_dirty_scope = !group->pinned_page_dirty_scope;
2612 list_del(&group->next);
2615 * Group ownership provides privilege, if the group list is
2616 * empty, the domain goes away. If it's the last domain with
2617 * iommu and external domain doesn't exist, then all the
2618 * mappings go away too. If it's the last domain with iommu and
2619 * external domain exist, update accounting
2621 if (list_empty(&domain->group_list)) {
2622 if (list_is_singular(&iommu->domain_list)) {
2623 if (!iommu->external_domain) {
2624 WARN_ON(iommu->notifier.head);
2625 vfio_iommu_unmap_unpin_all(iommu);
2627 vfio_iommu_unmap_unpin_reaccount(iommu);
2630 iommu_domain_free(domain->domain);
2631 list_del(&domain->next);
2633 vfio_iommu_aper_expand(iommu, &iova_copy);
2634 vfio_update_pgsize_bitmap(iommu);
2639 if (!vfio_iommu_resv_refresh(iommu, &iova_copy))
2640 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2642 vfio_iommu_iova_free(&iova_copy);
2646 * Removal of a group without dirty tracking may allow the iommu scope
2649 if (update_dirty_scope) {
2650 iommu->num_non_pinned_groups--;
2651 if (iommu->dirty_page_tracking)
2652 vfio_iommu_populate_bitmap_full(iommu);
2654 mutex_unlock(&iommu->lock);
2657 static void *vfio_iommu_type1_open(unsigned long arg)
2659 struct vfio_iommu *iommu;
2661 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
2663 return ERR_PTR(-ENOMEM);
2666 case VFIO_TYPE1_IOMMU:
2668 case VFIO_TYPE1_NESTING_IOMMU:
2669 iommu->nesting = true;
2671 case VFIO_TYPE1v2_IOMMU:
2676 return ERR_PTR(-EINVAL);
2679 INIT_LIST_HEAD(&iommu->domain_list);
2680 INIT_LIST_HEAD(&iommu->iova_list);
2681 iommu->dma_list = RB_ROOT;
2682 iommu->dma_avail = dma_entry_limit;
2683 iommu->container_open = true;
2684 mutex_init(&iommu->lock);
2685 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
2686 init_waitqueue_head(&iommu->vaddr_wait);
2691 static void vfio_release_domain(struct vfio_domain *domain, bool external)
2693 struct vfio_group *group, *group_tmp;
2695 list_for_each_entry_safe(group, group_tmp,
2696 &domain->group_list, next) {
2698 vfio_iommu_detach_group(domain, group);
2699 list_del(&group->next);
2704 iommu_domain_free(domain->domain);
2707 static void vfio_iommu_type1_release(void *iommu_data)
2709 struct vfio_iommu *iommu = iommu_data;
2710 struct vfio_domain *domain, *domain_tmp;
2712 if (iommu->external_domain) {
2713 vfio_release_domain(iommu->external_domain, true);
2714 kfree(iommu->external_domain);
2717 vfio_iommu_unmap_unpin_all(iommu);
2719 list_for_each_entry_safe(domain, domain_tmp,
2720 &iommu->domain_list, next) {
2721 vfio_release_domain(domain, false);
2722 list_del(&domain->next);
2726 vfio_iommu_iova_free(&iommu->iova_list);
2731 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
2733 struct vfio_domain *domain;
2736 mutex_lock(&iommu->lock);
2737 list_for_each_entry(domain, &iommu->domain_list, next) {
2738 if (!(domain->prot & IOMMU_CACHE)) {
2743 mutex_unlock(&iommu->lock);
2748 static int vfio_iommu_type1_check_extension(struct vfio_iommu *iommu,
2752 case VFIO_TYPE1_IOMMU:
2753 case VFIO_TYPE1v2_IOMMU:
2754 case VFIO_TYPE1_NESTING_IOMMU:
2755 case VFIO_UNMAP_ALL:
2756 case VFIO_UPDATE_VADDR:
2758 case VFIO_DMA_CC_IOMMU:
2761 return vfio_domains_have_iommu_cache(iommu);
2767 static int vfio_iommu_iova_add_cap(struct vfio_info_cap *caps,
2768 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas,
2771 struct vfio_info_cap_header *header;
2772 struct vfio_iommu_type1_info_cap_iova_range *iova_cap;
2774 header = vfio_info_cap_add(caps, size,
2775 VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE, 1);
2777 return PTR_ERR(header);
2779 iova_cap = container_of(header,
2780 struct vfio_iommu_type1_info_cap_iova_range,
2782 iova_cap->nr_iovas = cap_iovas->nr_iovas;
2783 memcpy(iova_cap->iova_ranges, cap_iovas->iova_ranges,
2784 cap_iovas->nr_iovas * sizeof(*cap_iovas->iova_ranges));
2788 static int vfio_iommu_iova_build_caps(struct vfio_iommu *iommu,
2789 struct vfio_info_cap *caps)
2791 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas;
2792 struct vfio_iova *iova;
2794 int iovas = 0, i = 0, ret;
2796 list_for_each_entry(iova, &iommu->iova_list, list)
2801 * Return 0 as a container with a single mdev device
2802 * will have an empty list
2807 size = sizeof(*cap_iovas) + (iovas * sizeof(*cap_iovas->iova_ranges));
2809 cap_iovas = kzalloc(size, GFP_KERNEL);
2813 cap_iovas->nr_iovas = iovas;
2815 list_for_each_entry(iova, &iommu->iova_list, list) {
2816 cap_iovas->iova_ranges[i].start = iova->start;
2817 cap_iovas->iova_ranges[i].end = iova->end;
2821 ret = vfio_iommu_iova_add_cap(caps, cap_iovas, size);
2827 static int vfio_iommu_migration_build_caps(struct vfio_iommu *iommu,
2828 struct vfio_info_cap *caps)
2830 struct vfio_iommu_type1_info_cap_migration cap_mig;
2832 cap_mig.header.id = VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION;
2833 cap_mig.header.version = 1;
2836 /* support minimum pgsize */
2837 cap_mig.pgsize_bitmap = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2838 cap_mig.max_dirty_bitmap_size = DIRTY_BITMAP_SIZE_MAX;
2840 return vfio_info_add_capability(caps, &cap_mig.header, sizeof(cap_mig));
2843 static int vfio_iommu_dma_avail_build_caps(struct vfio_iommu *iommu,
2844 struct vfio_info_cap *caps)
2846 struct vfio_iommu_type1_info_dma_avail cap_dma_avail;
2848 cap_dma_avail.header.id = VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL;
2849 cap_dma_avail.header.version = 1;
2851 cap_dma_avail.avail = iommu->dma_avail;
2853 return vfio_info_add_capability(caps, &cap_dma_avail.header,
2854 sizeof(cap_dma_avail));
2857 static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu,
2860 struct vfio_iommu_type1_info info;
2861 unsigned long minsz;
2862 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
2863 unsigned long capsz;
2866 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
2868 /* For backward compatibility, cannot require this */
2869 capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
2871 if (copy_from_user(&info, (void __user *)arg, minsz))
2874 if (info.argsz < minsz)
2877 if (info.argsz >= capsz) {
2879 info.cap_offset = 0; /* output, no-recopy necessary */
2882 mutex_lock(&iommu->lock);
2883 info.flags = VFIO_IOMMU_INFO_PGSIZES;
2885 info.iova_pgsizes = iommu->pgsize_bitmap;
2887 ret = vfio_iommu_migration_build_caps(iommu, &caps);
2890 ret = vfio_iommu_dma_avail_build_caps(iommu, &caps);
2893 ret = vfio_iommu_iova_build_caps(iommu, &caps);
2895 mutex_unlock(&iommu->lock);
2901 info.flags |= VFIO_IOMMU_INFO_CAPS;
2903 if (info.argsz < sizeof(info) + caps.size) {
2904 info.argsz = sizeof(info) + caps.size;
2906 vfio_info_cap_shift(&caps, sizeof(info));
2907 if (copy_to_user((void __user *)arg +
2908 sizeof(info), caps.buf,
2913 info.cap_offset = sizeof(info);
2919 return copy_to_user((void __user *)arg, &info, minsz) ?
2923 static int vfio_iommu_type1_map_dma(struct vfio_iommu *iommu,
2926 struct vfio_iommu_type1_dma_map map;
2927 unsigned long minsz;
2928 uint32_t mask = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE |
2929 VFIO_DMA_MAP_FLAG_VADDR;
2931 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
2933 if (copy_from_user(&map, (void __user *)arg, minsz))
2936 if (map.argsz < minsz || map.flags & ~mask)
2939 return vfio_dma_do_map(iommu, &map);
2942 static int vfio_iommu_type1_unmap_dma(struct vfio_iommu *iommu,
2945 struct vfio_iommu_type1_dma_unmap unmap;
2946 struct vfio_bitmap bitmap = { 0 };
2947 uint32_t mask = VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP |
2948 VFIO_DMA_UNMAP_FLAG_VADDR |
2949 VFIO_DMA_UNMAP_FLAG_ALL;
2950 unsigned long minsz;
2953 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
2955 if (copy_from_user(&unmap, (void __user *)arg, minsz))
2958 if (unmap.argsz < minsz || unmap.flags & ~mask)
2961 if ((unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
2962 (unmap.flags & (VFIO_DMA_UNMAP_FLAG_ALL |
2963 VFIO_DMA_UNMAP_FLAG_VADDR)))
2966 if (unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
2967 unsigned long pgshift;
2969 if (unmap.argsz < (minsz + sizeof(bitmap)))
2972 if (copy_from_user(&bitmap,
2973 (void __user *)(arg + minsz),
2977 if (!access_ok((void __user *)bitmap.data, bitmap.size))
2980 pgshift = __ffs(bitmap.pgsize);
2981 ret = verify_bitmap_size(unmap.size >> pgshift,
2987 ret = vfio_dma_do_unmap(iommu, &unmap, &bitmap);
2991 return copy_to_user((void __user *)arg, &unmap, minsz) ?
2995 static int vfio_iommu_type1_dirty_pages(struct vfio_iommu *iommu,
2998 struct vfio_iommu_type1_dirty_bitmap dirty;
2999 uint32_t mask = VFIO_IOMMU_DIRTY_PAGES_FLAG_START |
3000 VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP |
3001 VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP;
3002 unsigned long minsz;
3008 minsz = offsetofend(struct vfio_iommu_type1_dirty_bitmap, flags);
3010 if (copy_from_user(&dirty, (void __user *)arg, minsz))
3013 if (dirty.argsz < minsz || dirty.flags & ~mask)
3016 /* only one flag should be set at a time */
3017 if (__ffs(dirty.flags) != __fls(dirty.flags))
3020 if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_START) {
3023 mutex_lock(&iommu->lock);
3024 pgsize = 1 << __ffs(iommu->pgsize_bitmap);
3025 if (!iommu->dirty_page_tracking) {
3026 ret = vfio_dma_bitmap_alloc_all(iommu, pgsize);
3028 iommu->dirty_page_tracking = true;
3030 mutex_unlock(&iommu->lock);
3032 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP) {
3033 mutex_lock(&iommu->lock);
3034 if (iommu->dirty_page_tracking) {
3035 iommu->dirty_page_tracking = false;
3036 vfio_dma_bitmap_free_all(iommu);
3038 mutex_unlock(&iommu->lock);
3040 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP) {
3041 struct vfio_iommu_type1_dirty_bitmap_get range;
3042 unsigned long pgshift;
3043 size_t data_size = dirty.argsz - minsz;
3044 size_t iommu_pgsize;
3046 if (!data_size || data_size < sizeof(range))
3049 if (copy_from_user(&range, (void __user *)(arg + minsz),
3053 if (range.iova + range.size < range.iova)
3055 if (!access_ok((void __user *)range.bitmap.data,
3059 pgshift = __ffs(range.bitmap.pgsize);
3060 ret = verify_bitmap_size(range.size >> pgshift,
3065 mutex_lock(&iommu->lock);
3067 iommu_pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
3069 /* allow only smallest supported pgsize */
3070 if (range.bitmap.pgsize != iommu_pgsize) {
3074 if (range.iova & (iommu_pgsize - 1)) {
3078 if (!range.size || range.size & (iommu_pgsize - 1)) {
3083 if (iommu->dirty_page_tracking)
3084 ret = vfio_iova_dirty_bitmap(range.bitmap.data,
3087 range.bitmap.pgsize);
3091 mutex_unlock(&iommu->lock);
3099 static long vfio_iommu_type1_ioctl(void *iommu_data,
3100 unsigned int cmd, unsigned long arg)
3102 struct vfio_iommu *iommu = iommu_data;
3105 case VFIO_CHECK_EXTENSION:
3106 return vfio_iommu_type1_check_extension(iommu, arg);
3107 case VFIO_IOMMU_GET_INFO:
3108 return vfio_iommu_type1_get_info(iommu, arg);
3109 case VFIO_IOMMU_MAP_DMA:
3110 return vfio_iommu_type1_map_dma(iommu, arg);
3111 case VFIO_IOMMU_UNMAP_DMA:
3112 return vfio_iommu_type1_unmap_dma(iommu, arg);
3113 case VFIO_IOMMU_DIRTY_PAGES:
3114 return vfio_iommu_type1_dirty_pages(iommu, arg);
3120 static int vfio_iommu_type1_register_notifier(void *iommu_data,
3121 unsigned long *events,
3122 struct notifier_block *nb)
3124 struct vfio_iommu *iommu = iommu_data;
3126 /* clear known events */
3127 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
3129 /* refuse to register if still events remaining */
3133 return blocking_notifier_chain_register(&iommu->notifier, nb);
3136 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
3137 struct notifier_block *nb)
3139 struct vfio_iommu *iommu = iommu_data;
3141 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
3144 static int vfio_iommu_type1_dma_rw_chunk(struct vfio_iommu *iommu,
3145 dma_addr_t user_iova, void *data,
3146 size_t count, bool write,
3149 struct mm_struct *mm;
3150 unsigned long vaddr;
3151 struct vfio_dma *dma;
3152 bool kthread = current->mm == NULL;
3158 ret = vfio_find_dma_valid(iommu, user_iova, 1, &dma);
3162 if ((write && !(dma->prot & IOMMU_WRITE)) ||
3163 !(dma->prot & IOMMU_READ))
3166 mm = get_task_mm(dma->task);
3173 else if (current->mm != mm)
3176 offset = user_iova - dma->iova;
3178 if (count > dma->size - offset)
3179 count = dma->size - offset;
3181 vaddr = dma->vaddr + offset;
3184 *copied = copy_to_user((void __user *)vaddr, data,
3186 if (*copied && iommu->dirty_page_tracking) {
3187 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
3189 * Bitmap populated with the smallest supported page
3192 bitmap_set(dma->bitmap, offset >> pgshift,
3193 ((offset + *copied - 1) >> pgshift) -
3194 (offset >> pgshift) + 1);
3197 *copied = copy_from_user(data, (void __user *)vaddr,
3200 kthread_unuse_mm(mm);
3203 return *copied ? 0 : -EFAULT;
3206 static int vfio_iommu_type1_dma_rw(void *iommu_data, dma_addr_t user_iova,
3207 void *data, size_t count, bool write)
3209 struct vfio_iommu *iommu = iommu_data;
3213 mutex_lock(&iommu->lock);
3215 ret = vfio_iommu_type1_dma_rw_chunk(iommu, user_iova, data,
3216 count, write, &done);
3225 mutex_unlock(&iommu->lock);
3229 static struct iommu_domain *
3230 vfio_iommu_type1_group_iommu_domain(void *iommu_data,
3231 struct iommu_group *iommu_group)
3233 struct iommu_domain *domain = ERR_PTR(-ENODEV);
3234 struct vfio_iommu *iommu = iommu_data;
3235 struct vfio_domain *d;
3237 if (!iommu || !iommu_group)
3238 return ERR_PTR(-EINVAL);
3240 mutex_lock(&iommu->lock);
3241 list_for_each_entry(d, &iommu->domain_list, next) {
3242 if (find_iommu_group(d, iommu_group)) {
3247 mutex_unlock(&iommu->lock);
3252 static void vfio_iommu_type1_notify(void *iommu_data,
3253 enum vfio_iommu_notify_type event)
3255 struct vfio_iommu *iommu = iommu_data;
3257 if (event != VFIO_IOMMU_CONTAINER_CLOSE)
3259 mutex_lock(&iommu->lock);
3260 iommu->container_open = false;
3261 mutex_unlock(&iommu->lock);
3262 wake_up_all(&iommu->vaddr_wait);
3265 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
3266 .name = "vfio-iommu-type1",
3267 .owner = THIS_MODULE,
3268 .open = vfio_iommu_type1_open,
3269 .release = vfio_iommu_type1_release,
3270 .ioctl = vfio_iommu_type1_ioctl,
3271 .attach_group = vfio_iommu_type1_attach_group,
3272 .detach_group = vfio_iommu_type1_detach_group,
3273 .pin_pages = vfio_iommu_type1_pin_pages,
3274 .unpin_pages = vfio_iommu_type1_unpin_pages,
3275 .register_notifier = vfio_iommu_type1_register_notifier,
3276 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
3277 .dma_rw = vfio_iommu_type1_dma_rw,
3278 .group_iommu_domain = vfio_iommu_type1_group_iommu_domain,
3279 .notify = vfio_iommu_type1_notify,
3282 static int __init vfio_iommu_type1_init(void)
3284 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
3287 static void __exit vfio_iommu_type1_cleanup(void)
3289 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
3292 module_init(vfio_iommu_type1_init);
3293 module_exit(vfio_iommu_type1_cleanup);
3295 MODULE_VERSION(DRIVER_VERSION);
3296 MODULE_LICENSE("GPL v2");
3297 MODULE_AUTHOR(DRIVER_AUTHOR);
3298 MODULE_DESCRIPTION(DRIVER_DESC);