1 // SPDX-License-Identifier: GPL-2.0-only
3 * VFIO: IOMMU DMA mapping support for Type1 IOMMU
5 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
6 * Author: Alex Williamson <alex.williamson@redhat.com>
8 * Derived from original vfio:
9 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
10 * Author: Tom Lyon, pugs@cisco.com
12 * We arbitrarily define a Type1 IOMMU as one matching the below code.
13 * It could be called the x86 IOMMU as it's designed for AMD-Vi & Intel
14 * VT-d, but that makes it harder to re-use as theoretically anyone
15 * implementing a similar IOMMU could make use of this. We expect the
16 * IOMMU to support the IOMMU API and have few to no restrictions around
17 * the IOVA range that can be mapped. The Type1 IOMMU is currently
18 * optimized for relatively static mappings of a userspace process with
19 * userspace pages pinned into memory. We also assume devices and IOMMU
20 * domains are PCI based as the IOMMU API is still centered around a
21 * device/bus interface rather than a group interface.
24 #include <linux/compat.h>
25 #include <linux/device.h>
27 #include <linux/highmem.h>
28 #include <linux/iommu.h>
29 #include <linux/module.h>
31 #include <linux/kthread.h>
32 #include <linux/rbtree.h>
33 #include <linux/sched/signal.h>
34 #include <linux/sched/mm.h>
35 #include <linux/slab.h>
36 #include <linux/uaccess.h>
37 #include <linux/vfio.h>
38 #include <linux/workqueue.h>
39 #include <linux/notifier.h>
40 #include <linux/dma-iommu.h>
41 #include <linux/irqdomain.h>
44 #define DRIVER_VERSION "0.2"
45 #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
46 #define DRIVER_DESC "Type1 IOMMU driver for VFIO"
48 static bool allow_unsafe_interrupts;
49 module_param_named(allow_unsafe_interrupts,
50 allow_unsafe_interrupts, bool, S_IRUGO | S_IWUSR);
51 MODULE_PARM_DESC(allow_unsafe_interrupts,
52 "Enable VFIO IOMMU support for on platforms without interrupt remapping support.");
54 static bool disable_hugepages;
55 module_param_named(disable_hugepages,
56 disable_hugepages, bool, S_IRUGO | S_IWUSR);
57 MODULE_PARM_DESC(disable_hugepages,
58 "Disable VFIO IOMMU support for IOMMU hugepages.");
60 static unsigned int dma_entry_limit __read_mostly = U16_MAX;
61 module_param_named(dma_entry_limit, dma_entry_limit, uint, 0644);
62 MODULE_PARM_DESC(dma_entry_limit,
63 "Maximum number of user DMA mappings per container (65535).");
66 struct list_head domain_list;
67 struct list_head iova_list;
69 struct rb_root dma_list;
70 struct list_head device_list;
71 struct mutex device_list_lock;
72 unsigned int dma_avail;
73 unsigned int vaddr_invalid_count;
74 uint64_t pgsize_bitmap;
75 uint64_t num_non_pinned_groups;
76 wait_queue_head_t vaddr_wait;
79 bool dirty_page_tracking;
81 struct list_head emulated_iommu_groups;
85 struct iommu_domain *domain;
86 struct list_head next;
87 struct list_head group_list;
88 bool fgsp : 1; /* Fine-grained super pages */
89 bool enforce_cache_coherency : 1;
94 dma_addr_t iova; /* Device address */
95 unsigned long vaddr; /* Process virtual addr */
96 size_t size; /* Map size (bytes) */
97 int prot; /* IOMMU_READ/WRITE */
99 bool lock_cap; /* capable(CAP_IPC_LOCK) */
101 struct task_struct *task;
102 struct rb_root pfn_list; /* Ex-user pinned pfn list */
103 unsigned long *bitmap;
107 struct page **pages; /* for pin_user_pages_remote */
108 struct page *fallback_page; /* if pages alloc fails */
109 int capacity; /* length of pages array */
110 int size; /* of batch currently */
111 int offset; /* of next entry in pages */
114 struct vfio_iommu_group {
115 struct iommu_group *iommu_group;
116 struct list_head next;
117 bool pinned_page_dirty_scope;
121 struct list_head list;
127 * Guest RAM pinning working set or DMA target
131 dma_addr_t iova; /* Device address */
132 unsigned long pfn; /* Host pfn */
133 unsigned int ref_count;
136 struct vfio_regions {
137 struct list_head list;
143 #define DIRTY_BITMAP_BYTES(n) (ALIGN(n, BITS_PER_TYPE(u64)) / BITS_PER_BYTE)
146 * Input argument of number of bits to bitmap_set() is unsigned integer, which
147 * further casts to signed integer for unaligned multi-bit operation,
149 * Then maximum bitmap size supported is 2^31 bits divided by 2^3 bits/byte,
150 * that is 2^28 (256 MB) which maps to 2^31 * 2^12 = 2^43 (8TB) on 4K page
153 #define DIRTY_BITMAP_PAGES_MAX ((u64)INT_MAX)
154 #define DIRTY_BITMAP_SIZE_MAX DIRTY_BITMAP_BYTES(DIRTY_BITMAP_PAGES_MAX)
158 static int put_pfn(unsigned long pfn, int prot);
160 static struct vfio_iommu_group*
161 vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
162 struct iommu_group *iommu_group);
165 * This code handles mapping and unmapping of user data buffers
166 * into DMA'ble space using the IOMMU
169 static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
170 dma_addr_t start, size_t size)
172 struct rb_node *node = iommu->dma_list.rb_node;
175 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
177 if (start + size <= dma->iova)
178 node = node->rb_left;
179 else if (start >= dma->iova + dma->size)
180 node = node->rb_right;
188 static struct rb_node *vfio_find_dma_first_node(struct vfio_iommu *iommu,
189 dma_addr_t start, u64 size)
191 struct rb_node *res = NULL;
192 struct rb_node *node = iommu->dma_list.rb_node;
193 struct vfio_dma *dma_res = NULL;
196 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
198 if (start < dma->iova + dma->size) {
201 if (start >= dma->iova)
203 node = node->rb_left;
205 node = node->rb_right;
208 if (res && size && dma_res->iova >= start + size)
213 static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new)
215 struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL;
216 struct vfio_dma *dma;
220 dma = rb_entry(parent, struct vfio_dma, node);
222 if (new->iova + new->size <= dma->iova)
223 link = &(*link)->rb_left;
225 link = &(*link)->rb_right;
228 rb_link_node(&new->node, parent, link);
229 rb_insert_color(&new->node, &iommu->dma_list);
232 static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
234 rb_erase(&old->node, &iommu->dma_list);
238 static int vfio_dma_bitmap_alloc(struct vfio_dma *dma, size_t pgsize)
240 uint64_t npages = dma->size / pgsize;
242 if (npages > DIRTY_BITMAP_PAGES_MAX)
246 * Allocate extra 64 bits that are used to calculate shift required for
247 * bitmap_shift_left() to manipulate and club unaligned number of pages
248 * in adjacent vfio_dma ranges.
250 dma->bitmap = kvzalloc(DIRTY_BITMAP_BYTES(npages) + sizeof(u64),
258 static void vfio_dma_bitmap_free(struct vfio_dma *dma)
264 static void vfio_dma_populate_bitmap(struct vfio_dma *dma, size_t pgsize)
267 unsigned long pgshift = __ffs(pgsize);
269 for (p = rb_first(&dma->pfn_list); p; p = rb_next(p)) {
270 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn, node);
272 bitmap_set(dma->bitmap, (vpfn->iova - dma->iova) >> pgshift, 1);
276 static void vfio_iommu_populate_bitmap_full(struct vfio_iommu *iommu)
279 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
281 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
282 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
284 bitmap_set(dma->bitmap, 0, dma->size >> pgshift);
288 static int vfio_dma_bitmap_alloc_all(struct vfio_iommu *iommu, size_t pgsize)
292 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
293 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
296 ret = vfio_dma_bitmap_alloc(dma, pgsize);
300 for (p = rb_prev(n); p; p = rb_prev(p)) {
301 struct vfio_dma *dma = rb_entry(n,
302 struct vfio_dma, node);
304 vfio_dma_bitmap_free(dma);
308 vfio_dma_populate_bitmap(dma, pgsize);
313 static void vfio_dma_bitmap_free_all(struct vfio_iommu *iommu)
317 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
318 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
320 vfio_dma_bitmap_free(dma);
325 * Helper Functions for host iova-pfn list
327 static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
329 struct vfio_pfn *vpfn;
330 struct rb_node *node = dma->pfn_list.rb_node;
333 vpfn = rb_entry(node, struct vfio_pfn, node);
335 if (iova < vpfn->iova)
336 node = node->rb_left;
337 else if (iova > vpfn->iova)
338 node = node->rb_right;
345 static void vfio_link_pfn(struct vfio_dma *dma,
346 struct vfio_pfn *new)
348 struct rb_node **link, *parent = NULL;
349 struct vfio_pfn *vpfn;
351 link = &dma->pfn_list.rb_node;
354 vpfn = rb_entry(parent, struct vfio_pfn, node);
356 if (new->iova < vpfn->iova)
357 link = &(*link)->rb_left;
359 link = &(*link)->rb_right;
362 rb_link_node(&new->node, parent, link);
363 rb_insert_color(&new->node, &dma->pfn_list);
366 static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old)
368 rb_erase(&old->node, &dma->pfn_list);
371 static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova,
374 struct vfio_pfn *vpfn;
376 vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL);
383 vfio_link_pfn(dma, vpfn);
387 static void vfio_remove_from_pfn_list(struct vfio_dma *dma,
388 struct vfio_pfn *vpfn)
390 vfio_unlink_pfn(dma, vpfn);
394 static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma,
397 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
404 static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn)
409 if (!vpfn->ref_count) {
410 ret = put_pfn(vpfn->pfn, dma->prot);
411 vfio_remove_from_pfn_list(dma, vpfn);
416 static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
418 struct mm_struct *mm;
424 mm = async ? get_task_mm(dma->task) : dma->task->mm;
426 return -ESRCH; /* process exited */
428 ret = mmap_write_lock_killable(mm);
430 ret = __account_locked_vm(mm, abs(npage), npage > 0, dma->task,
432 mmap_write_unlock(mm);
442 * Some mappings aren't backed by a struct page, for example an mmap'd
443 * MMIO range for our own or another device. These use a different
444 * pfn conversion and shouldn't be tracked as locked pages.
445 * For compound pages, any driver that sets the reserved bit in head
446 * page needs to set the reserved bit in all subpages to be safe.
448 static bool is_invalid_reserved_pfn(unsigned long pfn)
451 return PageReserved(pfn_to_page(pfn));
456 static int put_pfn(unsigned long pfn, int prot)
458 if (!is_invalid_reserved_pfn(pfn)) {
459 struct page *page = pfn_to_page(pfn);
461 unpin_user_pages_dirty_lock(&page, 1, prot & IOMMU_WRITE);
467 #define VFIO_BATCH_MAX_CAPACITY (PAGE_SIZE / sizeof(struct page *))
469 static void vfio_batch_init(struct vfio_batch *batch)
474 if (unlikely(disable_hugepages))
477 batch->pages = (struct page **) __get_free_page(GFP_KERNEL);
481 batch->capacity = VFIO_BATCH_MAX_CAPACITY;
485 batch->pages = &batch->fallback_page;
489 static void vfio_batch_unpin(struct vfio_batch *batch, struct vfio_dma *dma)
491 while (batch->size) {
492 unsigned long pfn = page_to_pfn(batch->pages[batch->offset]);
494 put_pfn(pfn, dma->prot);
500 static void vfio_batch_fini(struct vfio_batch *batch)
502 if (batch->capacity == VFIO_BATCH_MAX_CAPACITY)
503 free_page((unsigned long)batch->pages);
506 static int follow_fault_pfn(struct vm_area_struct *vma, struct mm_struct *mm,
507 unsigned long vaddr, unsigned long *pfn,
514 ret = follow_pte(vma->vm_mm, vaddr, &ptep, &ptl);
516 bool unlocked = false;
518 ret = fixup_user_fault(mm, vaddr,
520 (write_fault ? FAULT_FLAG_WRITE : 0),
528 ret = follow_pte(vma->vm_mm, vaddr, &ptep, &ptl);
533 if (write_fault && !pte_write(*ptep))
536 *pfn = pte_pfn(*ptep);
538 pte_unmap_unlock(ptep, ptl);
543 * Returns the positive number of pfns successfully obtained or a negative
546 static int vaddr_get_pfns(struct mm_struct *mm, unsigned long vaddr,
547 long npages, int prot, unsigned long *pfn,
550 struct vm_area_struct *vma;
551 unsigned int flags = 0;
554 if (prot & IOMMU_WRITE)
558 ret = pin_user_pages_remote(mm, vaddr, npages, flags | FOLL_LONGTERM,
564 * The zero page is always resident, we don't need to pin it
565 * and it falls into our invalid/reserved test so we don't
566 * unpin in put_pfn(). Unpin all zero pages in the batch here.
568 for (i = 0 ; i < ret; i++) {
569 if (unlikely(is_zero_pfn(page_to_pfn(pages[i]))))
570 unpin_user_page(pages[i]);
573 *pfn = page_to_pfn(pages[0]);
577 vaddr = untagged_addr(vaddr);
580 vma = vma_lookup(mm, vaddr);
582 if (vma && vma->vm_flags & VM_PFNMAP) {
583 ret = follow_fault_pfn(vma, mm, vaddr, pfn, prot & IOMMU_WRITE);
588 if (is_invalid_reserved_pfn(*pfn))
595 mmap_read_unlock(mm);
599 static int vfio_wait(struct vfio_iommu *iommu)
603 prepare_to_wait(&iommu->vaddr_wait, &wait, TASK_KILLABLE);
604 mutex_unlock(&iommu->lock);
606 mutex_lock(&iommu->lock);
607 finish_wait(&iommu->vaddr_wait, &wait);
608 if (kthread_should_stop() || !iommu->container_open ||
609 fatal_signal_pending(current)) {
616 * Find dma struct and wait for its vaddr to be valid. iommu lock is dropped
617 * if the task waits, but is re-locked on return. Return result in *dma_p.
618 * Return 0 on success with no waiting, WAITED on success if waited, and -errno
621 static int vfio_find_dma_valid(struct vfio_iommu *iommu, dma_addr_t start,
622 size_t size, struct vfio_dma **dma_p)
627 *dma_p = vfio_find_dma(iommu, start, size);
630 else if (!(*dma_p)->vaddr_invalid)
633 ret = vfio_wait(iommu);
634 } while (ret == WAITED);
640 * Wait for all vaddr in the dma_list to become valid. iommu lock is dropped
641 * if the task waits, but is re-locked on return. Return 0 on success with no
642 * waiting, WAITED on success if waited, and -errno on error.
644 static int vfio_wait_all_valid(struct vfio_iommu *iommu)
648 while (iommu->vaddr_invalid_count && ret >= 0)
649 ret = vfio_wait(iommu);
655 * Attempt to pin pages. We really don't want to track all the pfns and
656 * the iommu can only map chunks of consecutive pfns anyway, so get the
657 * first page and all consecutive pages with the same locking.
659 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
660 long npage, unsigned long *pfn_base,
661 unsigned long limit, struct vfio_batch *batch)
664 struct mm_struct *mm = current->mm;
665 long ret, pinned = 0, lock_acct = 0;
667 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
669 /* This code path is only user initiated */
674 /* Leftover pages in batch from an earlier call. */
675 *pfn_base = page_to_pfn(batch->pages[batch->offset]);
677 rsvd = is_invalid_reserved_pfn(*pfn_base);
684 /* Empty batch, so refill it. */
685 long req_pages = min_t(long, npage, batch->capacity);
687 ret = vaddr_get_pfns(mm, vaddr, req_pages, dma->prot,
697 rsvd = is_invalid_reserved_pfn(*pfn_base);
702 * pfn is preset for the first iteration of this inner loop and
703 * updated at the end to handle a VM_PFNMAP pfn. In that case,
704 * batch->pages isn't valid (there's no struct page), so allow
705 * batch->pages to be touched only when there's more than one
706 * pfn to check, which guarantees the pfns are from a
710 if (pfn != *pfn_base + pinned ||
711 rsvd != is_invalid_reserved_pfn(pfn))
715 * Reserved pages aren't counted against the user,
716 * externally pinned pages are already counted against
719 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
720 if (!dma->lock_cap &&
721 mm->locked_vm + lock_acct + 1 > limit) {
722 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
723 __func__, limit << PAGE_SHIFT);
740 pfn = page_to_pfn(batch->pages[batch->offset]);
743 if (unlikely(disable_hugepages))
748 ret = vfio_lock_acct(dma, lock_acct, false);
751 if (batch->size == 1 && !batch->offset) {
752 /* May be a VM_PFNMAP pfn, which the batch can't remember. */
753 put_pfn(pfn, dma->prot);
758 if (pinned && !rsvd) {
759 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
760 put_pfn(pfn, dma->prot);
762 vfio_batch_unpin(batch, dma);
770 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
771 unsigned long pfn, long npage,
774 long unlocked = 0, locked = 0;
777 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
778 if (put_pfn(pfn++, dma->prot)) {
780 if (vfio_find_vpfn(dma, iova))
786 vfio_lock_acct(dma, locked - unlocked, true);
791 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
792 unsigned long *pfn_base, bool do_accounting)
794 struct page *pages[1];
795 struct mm_struct *mm;
798 mm = get_task_mm(dma->task);
802 ret = vaddr_get_pfns(mm, vaddr, 1, dma->prot, pfn_base, pages);
808 if (do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
809 ret = vfio_lock_acct(dma, 1, true);
811 put_pfn(*pfn_base, dma->prot);
813 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
814 "(%ld) exceeded\n", __func__,
815 dma->task->comm, task_pid_nr(dma->task),
816 task_rlimit(dma->task, RLIMIT_MEMLOCK));
825 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
829 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
834 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
837 vfio_lock_acct(dma, -unlocked, true);
842 static int vfio_iommu_type1_pin_pages(void *iommu_data,
843 struct iommu_group *iommu_group,
844 dma_addr_t user_iova,
848 struct vfio_iommu *iommu = iommu_data;
849 struct vfio_iommu_group *group;
851 unsigned long remote_vaddr;
852 struct vfio_dma *dma;
856 if (!iommu || !pages)
859 /* Supported for v2 version only */
863 mutex_lock(&iommu->lock);
866 * Wait for all necessary vaddr's to be valid so they can be used in
867 * the main loop without dropping the lock, to avoid racing vs unmap.
870 if (iommu->vaddr_invalid_count) {
871 for (i = 0; i < npage; i++) {
872 iova = user_iova + PAGE_SIZE * i;
873 ret = vfio_find_dma_valid(iommu, iova, PAGE_SIZE, &dma);
881 /* Fail if no dma_umap notifier is registered */
882 if (list_empty(&iommu->device_list)) {
888 * If iommu capable domain exist in the container then all pages are
889 * already pinned and accounted. Accounting should be done if there is no
890 * iommu capable domain in the container.
892 do_accounting = list_empty(&iommu->domain_list);
894 for (i = 0; i < npage; i++) {
895 unsigned long phys_pfn;
896 struct vfio_pfn *vpfn;
898 iova = user_iova + PAGE_SIZE * i;
899 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
905 if ((dma->prot & prot) != prot) {
910 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
912 pages[i] = pfn_to_page(vpfn->pfn);
916 remote_vaddr = dma->vaddr + (iova - dma->iova);
917 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn,
922 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn);
924 if (put_pfn(phys_pfn, dma->prot) && do_accounting)
925 vfio_lock_acct(dma, -1, true);
929 pages[i] = pfn_to_page(phys_pfn);
931 if (iommu->dirty_page_tracking) {
932 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
935 * Bitmap populated with the smallest supported page
938 bitmap_set(dma->bitmap,
939 (iova - dma->iova) >> pgshift, 1);
944 group = vfio_iommu_find_iommu_group(iommu, iommu_group);
945 if (!group->pinned_page_dirty_scope) {
946 group->pinned_page_dirty_scope = true;
947 iommu->num_non_pinned_groups--;
954 for (j = 0; j < i; j++) {
957 iova = user_iova + PAGE_SIZE * j;
958 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
959 vfio_unpin_page_external(dma, iova, do_accounting);
963 mutex_unlock(&iommu->lock);
967 static void vfio_iommu_type1_unpin_pages(void *iommu_data,
968 dma_addr_t user_iova, int npage)
970 struct vfio_iommu *iommu = iommu_data;
974 /* Supported for v2 version only */
975 if (WARN_ON(!iommu->v2))
978 mutex_lock(&iommu->lock);
980 do_accounting = list_empty(&iommu->domain_list);
981 for (i = 0; i < npage; i++) {
982 dma_addr_t iova = user_iova + PAGE_SIZE * i;
983 struct vfio_dma *dma;
985 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
989 vfio_unpin_page_external(dma, iova, do_accounting);
992 mutex_unlock(&iommu->lock);
997 static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
998 struct list_head *regions,
999 struct iommu_iotlb_gather *iotlb_gather)
1002 struct vfio_regions *entry, *next;
1004 iommu_iotlb_sync(domain->domain, iotlb_gather);
1006 list_for_each_entry_safe(entry, next, regions, list) {
1007 unlocked += vfio_unpin_pages_remote(dma,
1009 entry->phys >> PAGE_SHIFT,
1010 entry->len >> PAGE_SHIFT,
1012 list_del(&entry->list);
1022 * Generally, VFIO needs to unpin remote pages after each IOTLB flush.
1023 * Therefore, when using IOTLB flush sync interface, VFIO need to keep track
1024 * of these regions (currently using a list).
1026 * This value specifies maximum number of regions for each IOTLB flush sync.
1028 #define VFIO_IOMMU_TLB_SYNC_MAX 512
1030 static size_t unmap_unpin_fast(struct vfio_domain *domain,
1031 struct vfio_dma *dma, dma_addr_t *iova,
1032 size_t len, phys_addr_t phys, long *unlocked,
1033 struct list_head *unmapped_list,
1035 struct iommu_iotlb_gather *iotlb_gather)
1037 size_t unmapped = 0;
1038 struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
1041 unmapped = iommu_unmap_fast(domain->domain, *iova, len,
1047 entry->iova = *iova;
1049 entry->len = unmapped;
1050 list_add_tail(&entry->list, unmapped_list);
1058 * Sync if the number of fast-unmap regions hits the limit
1059 * or in case of errors.
1061 if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
1062 *unlocked += vfio_sync_unpin(dma, domain, unmapped_list,
1070 static size_t unmap_unpin_slow(struct vfio_domain *domain,
1071 struct vfio_dma *dma, dma_addr_t *iova,
1072 size_t len, phys_addr_t phys,
1075 size_t unmapped = iommu_unmap(domain->domain, *iova, len);
1078 *unlocked += vfio_unpin_pages_remote(dma, *iova,
1080 unmapped >> PAGE_SHIFT,
1088 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
1091 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
1092 struct vfio_domain *domain, *d;
1093 LIST_HEAD(unmapped_region_list);
1094 struct iommu_iotlb_gather iotlb_gather;
1095 int unmapped_region_cnt = 0;
1101 if (list_empty(&iommu->domain_list))
1105 * We use the IOMMU to track the physical addresses, otherwise we'd
1106 * need a much more complicated tracking system. Unfortunately that
1107 * means we need to use one of the iommu domains to figure out the
1108 * pfns to unpin. The rest need to be unmapped in advance so we have
1109 * no iommu translations remaining when the pages are unpinned.
1111 domain = d = list_first_entry(&iommu->domain_list,
1112 struct vfio_domain, next);
1114 list_for_each_entry_continue(d, &iommu->domain_list, next) {
1115 iommu_unmap(d->domain, dma->iova, dma->size);
1119 iommu_iotlb_gather_init(&iotlb_gather);
1120 while (iova < end) {
1121 size_t unmapped, len;
1122 phys_addr_t phys, next;
1124 phys = iommu_iova_to_phys(domain->domain, iova);
1125 if (WARN_ON(!phys)) {
1131 * To optimize for fewer iommu_unmap() calls, each of which
1132 * may require hardware cache flushing, try to find the
1133 * largest contiguous physical memory chunk to unmap.
1135 for (len = PAGE_SIZE;
1136 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
1137 next = iommu_iova_to_phys(domain->domain, iova + len);
1138 if (next != phys + len)
1143 * First, try to use fast unmap/unpin. In case of failure,
1144 * switch to slow unmap/unpin path.
1146 unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys,
1147 &unlocked, &unmapped_region_list,
1148 &unmapped_region_cnt,
1151 unmapped = unmap_unpin_slow(domain, dma, &iova, len,
1153 if (WARN_ON(!unmapped))
1158 dma->iommu_mapped = false;
1160 if (unmapped_region_cnt) {
1161 unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list,
1165 if (do_accounting) {
1166 vfio_lock_acct(dma, -unlocked, true);
1172 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
1174 WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list));
1175 vfio_unmap_unpin(iommu, dma, true);
1176 vfio_unlink_dma(iommu, dma);
1177 put_task_struct(dma->task);
1178 vfio_dma_bitmap_free(dma);
1179 if (dma->vaddr_invalid) {
1180 iommu->vaddr_invalid_count--;
1181 wake_up_all(&iommu->vaddr_wait);
1187 static void vfio_update_pgsize_bitmap(struct vfio_iommu *iommu)
1189 struct vfio_domain *domain;
1191 iommu->pgsize_bitmap = ULONG_MAX;
1193 list_for_each_entry(domain, &iommu->domain_list, next)
1194 iommu->pgsize_bitmap &= domain->domain->pgsize_bitmap;
1197 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
1198 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
1199 * That way the user will be able to map/unmap buffers whose size/
1200 * start address is aligned with PAGE_SIZE. Pinning code uses that
1201 * granularity while iommu driver can use the sub-PAGE_SIZE size
1202 * to map the buffer.
1204 if (iommu->pgsize_bitmap & ~PAGE_MASK) {
1205 iommu->pgsize_bitmap &= PAGE_MASK;
1206 iommu->pgsize_bitmap |= PAGE_SIZE;
1210 static int update_user_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1211 struct vfio_dma *dma, dma_addr_t base_iova,
1214 unsigned long pgshift = __ffs(pgsize);
1215 unsigned long nbits = dma->size >> pgshift;
1216 unsigned long bit_offset = (dma->iova - base_iova) >> pgshift;
1217 unsigned long copy_offset = bit_offset / BITS_PER_LONG;
1218 unsigned long shift = bit_offset % BITS_PER_LONG;
1219 unsigned long leftover;
1222 * mark all pages dirty if any IOMMU capable device is not able
1223 * to report dirty pages and all pages are pinned and mapped.
1225 if (iommu->num_non_pinned_groups && dma->iommu_mapped)
1226 bitmap_set(dma->bitmap, 0, nbits);
1229 bitmap_shift_left(dma->bitmap, dma->bitmap, shift,
1232 if (copy_from_user(&leftover,
1233 (void __user *)(bitmap + copy_offset),
1237 bitmap_or(dma->bitmap, dma->bitmap, &leftover, shift);
1240 if (copy_to_user((void __user *)(bitmap + copy_offset), dma->bitmap,
1241 DIRTY_BITMAP_BYTES(nbits + shift)))
1247 static int vfio_iova_dirty_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1248 dma_addr_t iova, size_t size, size_t pgsize)
1250 struct vfio_dma *dma;
1252 unsigned long pgshift = __ffs(pgsize);
1256 * GET_BITMAP request must fully cover vfio_dma mappings. Multiple
1257 * vfio_dma mappings may be clubbed by specifying large ranges, but
1258 * there must not be any previous mappings bisected by the range.
1259 * An error will be returned if these conditions are not met.
1261 dma = vfio_find_dma(iommu, iova, 1);
1262 if (dma && dma->iova != iova)
1265 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1266 if (dma && dma->iova + dma->size != iova + size)
1269 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1270 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1272 if (dma->iova < iova)
1275 if (dma->iova > iova + size - 1)
1278 ret = update_user_bitmap(bitmap, iommu, dma, iova, pgsize);
1283 * Re-populate bitmap to include all pinned pages which are
1284 * considered as dirty but exclude pages which are unpinned and
1285 * pages which are marked dirty by vfio_dma_rw()
1287 bitmap_clear(dma->bitmap, 0, dma->size >> pgshift);
1288 vfio_dma_populate_bitmap(dma, pgsize);
1293 static int verify_bitmap_size(uint64_t npages, uint64_t bitmap_size)
1295 if (!npages || !bitmap_size || (bitmap_size > DIRTY_BITMAP_SIZE_MAX) ||
1296 (bitmap_size < DIRTY_BITMAP_BYTES(npages)))
1303 * Notify VFIO drivers using vfio_register_emulated_iommu_dev() to invalidate
1304 * and unmap iovas within the range we're about to unmap. Drivers MUST unpin
1305 * pages in response to an invalidation.
1307 static void vfio_notify_dma_unmap(struct vfio_iommu *iommu,
1308 struct vfio_dma *dma)
1310 struct vfio_device *device;
1312 if (list_empty(&iommu->device_list))
1316 * The device is expected to call vfio_unpin_pages() for any IOVA it has
1317 * pinned within the range. Since vfio_unpin_pages() will eventually
1318 * call back down to this code and try to obtain the iommu->lock we must
1321 mutex_lock(&iommu->device_list_lock);
1322 mutex_unlock(&iommu->lock);
1324 list_for_each_entry(device, &iommu->device_list, iommu_entry)
1325 device->ops->dma_unmap(device, dma->iova, dma->size);
1327 mutex_unlock(&iommu->device_list_lock);
1328 mutex_lock(&iommu->lock);
1331 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
1332 struct vfio_iommu_type1_dma_unmap *unmap,
1333 struct vfio_bitmap *bitmap)
1335 struct vfio_dma *dma, *dma_last = NULL;
1336 size_t unmapped = 0, pgsize;
1337 int ret = -EINVAL, retries = 0;
1338 unsigned long pgshift;
1339 dma_addr_t iova = unmap->iova;
1340 u64 size = unmap->size;
1341 bool unmap_all = unmap->flags & VFIO_DMA_UNMAP_FLAG_ALL;
1342 bool invalidate_vaddr = unmap->flags & VFIO_DMA_UNMAP_FLAG_VADDR;
1343 struct rb_node *n, *first_n;
1345 mutex_lock(&iommu->lock);
1347 pgshift = __ffs(iommu->pgsize_bitmap);
1348 pgsize = (size_t)1 << pgshift;
1350 if (iova & (pgsize - 1))
1357 } else if (!size || size & (pgsize - 1) ||
1358 iova + size - 1 < iova || size > SIZE_MAX) {
1362 /* When dirty tracking is enabled, allow only min supported pgsize */
1363 if ((unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
1364 (!iommu->dirty_page_tracking || (bitmap->pgsize != pgsize))) {
1368 WARN_ON((pgsize - 1) & PAGE_MASK);
1371 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
1372 * avoid tracking individual mappings. This means that the granularity
1373 * of the original mapping was lost and the user was allowed to attempt
1374 * to unmap any range. Depending on the contiguousness of physical
1375 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
1376 * or may not have worked. We only guaranteed unmap granularity
1377 * matching the original mapping; even though it was untracked here,
1378 * the original mappings are reflected in IOMMU mappings. This
1379 * resulted in a couple unusual behaviors. First, if a range is not
1380 * able to be unmapped, ex. a set of 4k pages that was mapped as a
1381 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
1382 * a zero sized unmap. Also, if an unmap request overlaps the first
1383 * address of a hugepage, the IOMMU will unmap the entire hugepage.
1384 * This also returns success and the returned unmap size reflects the
1385 * actual size unmapped.
1387 * We attempt to maintain compatibility with this "v1" interface, but
1388 * we take control out of the hands of the IOMMU. Therefore, an unmap
1389 * request offset from the beginning of the original mapping will
1390 * return success with zero sized unmap. And an unmap request covering
1391 * the first iova of mapping will unmap the entire range.
1393 * The v2 version of this interface intends to be more deterministic.
1394 * Unmap requests must fully cover previous mappings. Multiple
1395 * mappings may still be unmaped by specifying large ranges, but there
1396 * must not be any previous mappings bisected by the range. An error
1397 * will be returned if these conditions are not met. The v2 interface
1398 * will only return success and a size of zero if there were no
1399 * mappings within the range.
1401 if (iommu->v2 && !unmap_all) {
1402 dma = vfio_find_dma(iommu, iova, 1);
1403 if (dma && dma->iova != iova)
1406 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1407 if (dma && dma->iova + dma->size != iova + size)
1412 n = first_n = vfio_find_dma_first_node(iommu, iova, size);
1415 dma = rb_entry(n, struct vfio_dma, node);
1416 if (dma->iova >= iova + size)
1419 if (!iommu->v2 && iova > dma->iova)
1422 if (invalidate_vaddr) {
1423 if (dma->vaddr_invalid) {
1424 struct rb_node *last_n = n;
1426 for (n = first_n; n != last_n; n = rb_next(n)) {
1428 struct vfio_dma, node);
1429 dma->vaddr_invalid = false;
1430 iommu->vaddr_invalid_count--;
1436 dma->vaddr_invalid = true;
1437 iommu->vaddr_invalid_count++;
1438 unmapped += dma->size;
1443 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
1444 if (dma_last == dma) {
1445 BUG_ON(++retries > 10);
1451 vfio_notify_dma_unmap(iommu, dma);
1455 if (unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
1456 ret = update_user_bitmap(bitmap->data, iommu, dma,
1462 unmapped += dma->size;
1464 vfio_remove_dma(iommu, dma);
1468 mutex_unlock(&iommu->lock);
1470 /* Report how much was unmapped */
1471 unmap->size = unmapped;
1476 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
1477 unsigned long pfn, long npage, int prot)
1479 struct vfio_domain *d;
1482 list_for_each_entry(d, &iommu->domain_list, next) {
1483 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
1484 npage << PAGE_SHIFT, prot | IOMMU_CACHE);
1494 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) {
1495 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
1502 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
1505 dma_addr_t iova = dma->iova;
1506 unsigned long vaddr = dma->vaddr;
1507 struct vfio_batch batch;
1508 size_t size = map_size;
1510 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1513 vfio_batch_init(&batch);
1516 /* Pin a contiguous chunk of memory */
1517 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
1518 size >> PAGE_SHIFT, &pfn, limit,
1527 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
1530 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
1532 vfio_batch_unpin(&batch, dma);
1536 size -= npage << PAGE_SHIFT;
1537 dma->size += npage << PAGE_SHIFT;
1540 vfio_batch_fini(&batch);
1541 dma->iommu_mapped = true;
1544 vfio_remove_dma(iommu, dma);
1550 * Check dma map request is within a valid iova range
1552 static bool vfio_iommu_iova_dma_valid(struct vfio_iommu *iommu,
1553 dma_addr_t start, dma_addr_t end)
1555 struct list_head *iova = &iommu->iova_list;
1556 struct vfio_iova *node;
1558 list_for_each_entry(node, iova, list) {
1559 if (start >= node->start && end <= node->end)
1564 * Check for list_empty() as well since a container with
1565 * a single mdev device will have an empty list.
1567 return list_empty(iova);
1570 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1571 struct vfio_iommu_type1_dma_map *map)
1573 bool set_vaddr = map->flags & VFIO_DMA_MAP_FLAG_VADDR;
1574 dma_addr_t iova = map->iova;
1575 unsigned long vaddr = map->vaddr;
1576 size_t size = map->size;
1577 int ret = 0, prot = 0;
1579 struct vfio_dma *dma;
1581 /* Verify that none of our __u64 fields overflow */
1582 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1585 /* READ/WRITE from device perspective */
1586 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1587 prot |= IOMMU_WRITE;
1588 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1591 if ((prot && set_vaddr) || (!prot && !set_vaddr))
1594 mutex_lock(&iommu->lock);
1596 pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
1598 WARN_ON((pgsize - 1) & PAGE_MASK);
1600 if (!size || (size | iova | vaddr) & (pgsize - 1)) {
1605 /* Don't allow IOVA or virtual address wrap */
1606 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) {
1611 dma = vfio_find_dma(iommu, iova, size);
1615 } else if (!dma->vaddr_invalid || dma->iova != iova ||
1616 dma->size != size) {
1620 dma->vaddr_invalid = false;
1621 iommu->vaddr_invalid_count--;
1622 wake_up_all(&iommu->vaddr_wait);
1630 if (!iommu->dma_avail) {
1635 if (!vfio_iommu_iova_dma_valid(iommu, iova, iova + size - 1)) {
1640 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1652 * We need to be able to both add to a task's locked memory and test
1653 * against the locked memory limit and we need to be able to do both
1654 * outside of this call path as pinning can be asynchronous via the
1655 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1656 * task_struct and VM locked pages requires an mm_struct, however
1657 * holding an indefinite mm reference is not recommended, therefore we
1658 * only hold a reference to a task. We could hold a reference to
1659 * current, however QEMU uses this call path through vCPU threads,
1660 * which can be killed resulting in a NULL mm and failure in the unmap
1661 * path when called via a different thread. Avoid this problem by
1662 * using the group_leader as threads within the same group require
1663 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1666 * Previously we also used the task for testing CAP_IPC_LOCK at the
1667 * time of pinning and accounting, however has_capability() makes use
1668 * of real_cred, a copy-on-write field, so we can't guarantee that it
1669 * matches group_leader, or in fact that it might not change by the
1670 * time it's evaluated. If a process were to call MAP_DMA with
1671 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1672 * possibly see different results for an iommu_mapped vfio_dma vs
1673 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1674 * time of calling MAP_DMA.
1676 get_task_struct(current->group_leader);
1677 dma->task = current->group_leader;
1678 dma->lock_cap = capable(CAP_IPC_LOCK);
1680 dma->pfn_list = RB_ROOT;
1682 /* Insert zero-sized and grow as we map chunks of it */
1683 vfio_link_dma(iommu, dma);
1685 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1686 if (list_empty(&iommu->domain_list))
1689 ret = vfio_pin_map_dma(iommu, dma, size);
1691 if (!ret && iommu->dirty_page_tracking) {
1692 ret = vfio_dma_bitmap_alloc(dma, pgsize);
1694 vfio_remove_dma(iommu, dma);
1698 mutex_unlock(&iommu->lock);
1702 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1703 struct vfio_domain *domain)
1705 struct vfio_batch batch;
1706 struct vfio_domain *d = NULL;
1708 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1711 ret = vfio_wait_all_valid(iommu);
1715 /* Arbitrarily pick the first domain in the list for lookups */
1716 if (!list_empty(&iommu->domain_list))
1717 d = list_first_entry(&iommu->domain_list,
1718 struct vfio_domain, next);
1720 vfio_batch_init(&batch);
1722 n = rb_first(&iommu->dma_list);
1724 for (; n; n = rb_next(n)) {
1725 struct vfio_dma *dma;
1728 dma = rb_entry(n, struct vfio_dma, node);
1731 while (iova < dma->iova + dma->size) {
1735 if (dma->iommu_mapped) {
1739 if (WARN_ON(!d)) { /* mapped w/o a domain?! */
1744 phys = iommu_iova_to_phys(d->domain, iova);
1746 if (WARN_ON(!phys)) {
1754 while (i < dma->iova + dma->size &&
1755 p == iommu_iova_to_phys(d->domain, i)) {
1762 unsigned long vaddr = dma->vaddr +
1764 size_t n = dma->iova + dma->size - iova;
1767 npage = vfio_pin_pages_remote(dma, vaddr,
1777 phys = pfn << PAGE_SHIFT;
1778 size = npage << PAGE_SHIFT;
1781 ret = iommu_map(domain->domain, iova, phys,
1782 size, dma->prot | IOMMU_CACHE);
1784 if (!dma->iommu_mapped) {
1785 vfio_unpin_pages_remote(dma, iova,
1789 vfio_batch_unpin(&batch, dma);
1798 /* All dmas are now mapped, defer to second tree walk for unwind */
1799 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1800 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1802 dma->iommu_mapped = true;
1805 vfio_batch_fini(&batch);
1809 for (; n; n = rb_prev(n)) {
1810 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1813 if (dma->iommu_mapped) {
1814 iommu_unmap(domain->domain, dma->iova, dma->size);
1819 while (iova < dma->iova + dma->size) {
1820 phys_addr_t phys, p;
1824 phys = iommu_iova_to_phys(domain->domain, iova);
1833 while (i < dma->iova + dma->size &&
1834 p == iommu_iova_to_phys(domain->domain, i)) {
1840 iommu_unmap(domain->domain, iova, size);
1841 vfio_unpin_pages_remote(dma, iova, phys >> PAGE_SHIFT,
1842 size >> PAGE_SHIFT, true);
1846 vfio_batch_fini(&batch);
1851 * We change our unmap behavior slightly depending on whether the IOMMU
1852 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1853 * for practically any contiguous power-of-two mapping we give it. This means
1854 * we don't need to look for contiguous chunks ourselves to make unmapping
1855 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1856 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1857 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1858 * hugetlbfs is in use.
1860 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1863 int ret, order = get_order(PAGE_SIZE * 2);
1865 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1869 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1870 IOMMU_READ | IOMMU_WRITE | IOMMU_CACHE);
1872 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1874 if (unmapped == PAGE_SIZE)
1875 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1877 domain->fgsp = true;
1880 __free_pages(pages, order);
1883 static struct vfio_iommu_group *find_iommu_group(struct vfio_domain *domain,
1884 struct iommu_group *iommu_group)
1886 struct vfio_iommu_group *g;
1888 list_for_each_entry(g, &domain->group_list, next) {
1889 if (g->iommu_group == iommu_group)
1896 static struct vfio_iommu_group*
1897 vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
1898 struct iommu_group *iommu_group)
1900 struct vfio_iommu_group *group;
1901 struct vfio_domain *domain;
1903 list_for_each_entry(domain, &iommu->domain_list, next) {
1904 group = find_iommu_group(domain, iommu_group);
1909 list_for_each_entry(group, &iommu->emulated_iommu_groups, next)
1910 if (group->iommu_group == iommu_group)
1915 static bool vfio_iommu_has_sw_msi(struct list_head *group_resv_regions,
1918 struct iommu_resv_region *region;
1921 list_for_each_entry(region, group_resv_regions, list) {
1923 * The presence of any 'real' MSI regions should take
1924 * precedence over the software-managed one if the
1925 * IOMMU driver happens to advertise both types.
1927 if (region->type == IOMMU_RESV_MSI) {
1932 if (region->type == IOMMU_RESV_SW_MSI) {
1933 *base = region->start;
1942 * This is a helper function to insert an address range to iova list.
1943 * The list is initially created with a single entry corresponding to
1944 * the IOMMU domain geometry to which the device group is attached.
1945 * The list aperture gets modified when a new domain is added to the
1946 * container if the new aperture doesn't conflict with the current one
1947 * or with any existing dma mappings. The list is also modified to
1948 * exclude any reserved regions associated with the device group.
1950 static int vfio_iommu_iova_insert(struct list_head *head,
1951 dma_addr_t start, dma_addr_t end)
1953 struct vfio_iova *region;
1955 region = kmalloc(sizeof(*region), GFP_KERNEL);
1959 INIT_LIST_HEAD(®ion->list);
1960 region->start = start;
1963 list_add_tail(®ion->list, head);
1968 * Check the new iommu aperture conflicts with existing aper or with any
1969 * existing dma mappings.
1971 static bool vfio_iommu_aper_conflict(struct vfio_iommu *iommu,
1972 dma_addr_t start, dma_addr_t end)
1974 struct vfio_iova *first, *last;
1975 struct list_head *iova = &iommu->iova_list;
1977 if (list_empty(iova))
1980 /* Disjoint sets, return conflict */
1981 first = list_first_entry(iova, struct vfio_iova, list);
1982 last = list_last_entry(iova, struct vfio_iova, list);
1983 if (start > last->end || end < first->start)
1986 /* Check for any existing dma mappings below the new start */
1987 if (start > first->start) {
1988 if (vfio_find_dma(iommu, first->start, start - first->start))
1992 /* Check for any existing dma mappings beyond the new end */
1993 if (end < last->end) {
1994 if (vfio_find_dma(iommu, end + 1, last->end - end))
2002 * Resize iommu iova aperture window. This is called only if the new
2003 * aperture has no conflict with existing aperture and dma mappings.
2005 static int vfio_iommu_aper_resize(struct list_head *iova,
2006 dma_addr_t start, dma_addr_t end)
2008 struct vfio_iova *node, *next;
2010 if (list_empty(iova))
2011 return vfio_iommu_iova_insert(iova, start, end);
2013 /* Adjust iova list start */
2014 list_for_each_entry_safe(node, next, iova, list) {
2015 if (start < node->start)
2017 if (start >= node->start && start < node->end) {
2018 node->start = start;
2021 /* Delete nodes before new start */
2022 list_del(&node->list);
2026 /* Adjust iova list end */
2027 list_for_each_entry_safe(node, next, iova, list) {
2028 if (end > node->end)
2030 if (end > node->start && end <= node->end) {
2034 /* Delete nodes after new end */
2035 list_del(&node->list);
2043 * Check reserved region conflicts with existing dma mappings
2045 static bool vfio_iommu_resv_conflict(struct vfio_iommu *iommu,
2046 struct list_head *resv_regions)
2048 struct iommu_resv_region *region;
2050 /* Check for conflict with existing dma mappings */
2051 list_for_each_entry(region, resv_regions, list) {
2052 if (region->type == IOMMU_RESV_DIRECT_RELAXABLE)
2055 if (vfio_find_dma(iommu, region->start, region->length))
2063 * Check iova region overlap with reserved regions and
2064 * exclude them from the iommu iova range
2066 static int vfio_iommu_resv_exclude(struct list_head *iova,
2067 struct list_head *resv_regions)
2069 struct iommu_resv_region *resv;
2070 struct vfio_iova *n, *next;
2072 list_for_each_entry(resv, resv_regions, list) {
2073 phys_addr_t start, end;
2075 if (resv->type == IOMMU_RESV_DIRECT_RELAXABLE)
2078 start = resv->start;
2079 end = resv->start + resv->length - 1;
2081 list_for_each_entry_safe(n, next, iova, list) {
2085 if (start > n->end || end < n->start)
2088 * Insert a new node if current node overlaps with the
2089 * reserve region to exclude that from valid iova range.
2090 * Note that, new node is inserted before the current
2091 * node and finally the current node is deleted keeping
2092 * the list updated and sorted.
2094 if (start > n->start)
2095 ret = vfio_iommu_iova_insert(&n->list, n->start,
2097 if (!ret && end < n->end)
2098 ret = vfio_iommu_iova_insert(&n->list, end + 1,
2108 if (list_empty(iova))
2114 static void vfio_iommu_resv_free(struct list_head *resv_regions)
2116 struct iommu_resv_region *n, *next;
2118 list_for_each_entry_safe(n, next, resv_regions, list) {
2124 static void vfio_iommu_iova_free(struct list_head *iova)
2126 struct vfio_iova *n, *next;
2128 list_for_each_entry_safe(n, next, iova, list) {
2134 static int vfio_iommu_iova_get_copy(struct vfio_iommu *iommu,
2135 struct list_head *iova_copy)
2137 struct list_head *iova = &iommu->iova_list;
2138 struct vfio_iova *n;
2141 list_for_each_entry(n, iova, list) {
2142 ret = vfio_iommu_iova_insert(iova_copy, n->start, n->end);
2150 vfio_iommu_iova_free(iova_copy);
2154 static void vfio_iommu_iova_insert_copy(struct vfio_iommu *iommu,
2155 struct list_head *iova_copy)
2157 struct list_head *iova = &iommu->iova_list;
2159 vfio_iommu_iova_free(iova);
2161 list_splice_tail(iova_copy, iova);
2164 /* Redundantly walks non-present capabilities to simplify caller */
2165 static int vfio_iommu_device_capable(struct device *dev, void *data)
2167 return device_iommu_capable(dev, (enum iommu_cap)data);
2170 static int vfio_iommu_domain_alloc(struct device *dev, void *data)
2172 struct iommu_domain **domain = data;
2174 *domain = iommu_domain_alloc(dev->bus);
2175 return 1; /* Don't iterate */
2178 static int vfio_iommu_type1_attach_group(void *iommu_data,
2179 struct iommu_group *iommu_group, enum vfio_group_type type)
2181 struct vfio_iommu *iommu = iommu_data;
2182 struct vfio_iommu_group *group;
2183 struct vfio_domain *domain, *d;
2184 bool resv_msi, msi_remap;
2185 phys_addr_t resv_msi_base = 0;
2186 struct iommu_domain_geometry *geo;
2187 LIST_HEAD(iova_copy);
2188 LIST_HEAD(group_resv_regions);
2191 mutex_lock(&iommu->lock);
2193 /* Check for duplicates */
2194 if (vfio_iommu_find_iommu_group(iommu, iommu_group))
2198 group = kzalloc(sizeof(*group), GFP_KERNEL);
2201 group->iommu_group = iommu_group;
2203 if (type == VFIO_EMULATED_IOMMU) {
2204 list_add(&group->next, &iommu->emulated_iommu_groups);
2206 * An emulated IOMMU group cannot dirty memory directly, it can
2207 * only use interfaces that provide dirty tracking.
2208 * The iommu scope can only be promoted with the addition of a
2209 * dirty tracking group.
2211 group->pinned_page_dirty_scope = true;
2217 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2219 goto out_free_group;
2222 * Going via the iommu_group iterator avoids races, and trivially gives
2223 * us a representative device for the IOMMU API call. We don't actually
2224 * want to iterate beyond the first device (if any).
2227 iommu_group_for_each_dev(iommu_group, &domain->domain,
2228 vfio_iommu_domain_alloc);
2229 if (!domain->domain)
2230 goto out_free_domain;
2232 if (iommu->nesting) {
2233 ret = iommu_enable_nesting(domain->domain);
2238 ret = iommu_attach_group(domain->domain, group->iommu_group);
2242 /* Get aperture info */
2243 geo = &domain->domain->geometry;
2244 if (vfio_iommu_aper_conflict(iommu, geo->aperture_start,
2245 geo->aperture_end)) {
2250 ret = iommu_get_group_resv_regions(iommu_group, &group_resv_regions);
2254 if (vfio_iommu_resv_conflict(iommu, &group_resv_regions)) {
2260 * We don't want to work on the original iova list as the list
2261 * gets modified and in case of failure we have to retain the
2262 * original list. Get a copy here.
2264 ret = vfio_iommu_iova_get_copy(iommu, &iova_copy);
2268 ret = vfio_iommu_aper_resize(&iova_copy, geo->aperture_start,
2273 ret = vfio_iommu_resv_exclude(&iova_copy, &group_resv_regions);
2277 resv_msi = vfio_iommu_has_sw_msi(&group_resv_regions, &resv_msi_base);
2279 INIT_LIST_HEAD(&domain->group_list);
2280 list_add(&group->next, &domain->group_list);
2282 msi_remap = irq_domain_check_msi_remap() ||
2283 iommu_group_for_each_dev(iommu_group, (void *)IOMMU_CAP_INTR_REMAP,
2284 vfio_iommu_device_capable);
2286 if (!allow_unsafe_interrupts && !msi_remap) {
2287 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
2294 * If the IOMMU can block non-coherent operations (ie PCIe TLPs with
2295 * no-snoop set) then VFIO always turns this feature on because on Intel
2296 * platforms it optimizes KVM to disable wbinvd emulation.
2298 if (domain->domain->ops->enforce_cache_coherency)
2299 domain->enforce_cache_coherency =
2300 domain->domain->ops->enforce_cache_coherency(
2304 * Try to match an existing compatible domain. We don't want to
2305 * preclude an IOMMU driver supporting multiple bus_types and being
2306 * able to include different bus_types in the same IOMMU domain, so
2307 * we test whether the domains use the same iommu_ops rather than
2308 * testing if they're on the same bus_type.
2310 list_for_each_entry(d, &iommu->domain_list, next) {
2311 if (d->domain->ops == domain->domain->ops &&
2312 d->enforce_cache_coherency ==
2313 domain->enforce_cache_coherency) {
2314 iommu_detach_group(domain->domain, group->iommu_group);
2315 if (!iommu_attach_group(d->domain,
2316 group->iommu_group)) {
2317 list_add(&group->next, &d->group_list);
2318 iommu_domain_free(domain->domain);
2323 ret = iommu_attach_group(domain->domain,
2324 group->iommu_group);
2330 vfio_test_domain_fgsp(domain);
2332 /* replay mappings on new domains */
2333 ret = vfio_iommu_replay(iommu, domain);
2338 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
2339 if (ret && ret != -ENODEV)
2343 list_add(&domain->next, &iommu->domain_list);
2344 vfio_update_pgsize_bitmap(iommu);
2346 /* Delete the old one and insert new iova list */
2347 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2350 * An iommu backed group can dirty memory directly and therefore
2351 * demotes the iommu scope until it declares itself dirty tracking
2352 * capable via the page pinning interface.
2354 iommu->num_non_pinned_groups++;
2355 mutex_unlock(&iommu->lock);
2356 vfio_iommu_resv_free(&group_resv_regions);
2361 iommu_detach_group(domain->domain, group->iommu_group);
2363 iommu_domain_free(domain->domain);
2364 vfio_iommu_iova_free(&iova_copy);
2365 vfio_iommu_resv_free(&group_resv_regions);
2371 mutex_unlock(&iommu->lock);
2375 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
2377 struct rb_node *node;
2379 while ((node = rb_first(&iommu->dma_list)))
2380 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
2383 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
2385 struct rb_node *n, *p;
2387 n = rb_first(&iommu->dma_list);
2388 for (; n; n = rb_next(n)) {
2389 struct vfio_dma *dma;
2390 long locked = 0, unlocked = 0;
2392 dma = rb_entry(n, struct vfio_dma, node);
2393 unlocked += vfio_unmap_unpin(iommu, dma, false);
2394 p = rb_first(&dma->pfn_list);
2395 for (; p; p = rb_next(p)) {
2396 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
2399 if (!is_invalid_reserved_pfn(vpfn->pfn))
2402 vfio_lock_acct(dma, locked - unlocked, true);
2407 * Called when a domain is removed in detach. It is possible that
2408 * the removed domain decided the iova aperture window. Modify the
2409 * iova aperture with the smallest window among existing domains.
2411 static void vfio_iommu_aper_expand(struct vfio_iommu *iommu,
2412 struct list_head *iova_copy)
2414 struct vfio_domain *domain;
2415 struct vfio_iova *node;
2416 dma_addr_t start = 0;
2417 dma_addr_t end = (dma_addr_t)~0;
2419 if (list_empty(iova_copy))
2422 list_for_each_entry(domain, &iommu->domain_list, next) {
2423 struct iommu_domain_geometry *geo = &domain->domain->geometry;
2425 if (geo->aperture_start > start)
2426 start = geo->aperture_start;
2427 if (geo->aperture_end < end)
2428 end = geo->aperture_end;
2431 /* Modify aperture limits. The new aper is either same or bigger */
2432 node = list_first_entry(iova_copy, struct vfio_iova, list);
2433 node->start = start;
2434 node = list_last_entry(iova_copy, struct vfio_iova, list);
2439 * Called when a group is detached. The reserved regions for that
2440 * group can be part of valid iova now. But since reserved regions
2441 * may be duplicated among groups, populate the iova valid regions
2444 static int vfio_iommu_resv_refresh(struct vfio_iommu *iommu,
2445 struct list_head *iova_copy)
2447 struct vfio_domain *d;
2448 struct vfio_iommu_group *g;
2449 struct vfio_iova *node;
2450 dma_addr_t start, end;
2451 LIST_HEAD(resv_regions);
2454 if (list_empty(iova_copy))
2457 list_for_each_entry(d, &iommu->domain_list, next) {
2458 list_for_each_entry(g, &d->group_list, next) {
2459 ret = iommu_get_group_resv_regions(g->iommu_group,
2466 node = list_first_entry(iova_copy, struct vfio_iova, list);
2467 start = node->start;
2468 node = list_last_entry(iova_copy, struct vfio_iova, list);
2471 /* purge the iova list and create new one */
2472 vfio_iommu_iova_free(iova_copy);
2474 ret = vfio_iommu_aper_resize(iova_copy, start, end);
2478 /* Exclude current reserved regions from iova ranges */
2479 ret = vfio_iommu_resv_exclude(iova_copy, &resv_regions);
2481 vfio_iommu_resv_free(&resv_regions);
2485 static void vfio_iommu_type1_detach_group(void *iommu_data,
2486 struct iommu_group *iommu_group)
2488 struct vfio_iommu *iommu = iommu_data;
2489 struct vfio_domain *domain;
2490 struct vfio_iommu_group *group;
2491 bool update_dirty_scope = false;
2492 LIST_HEAD(iova_copy);
2494 mutex_lock(&iommu->lock);
2495 list_for_each_entry(group, &iommu->emulated_iommu_groups, next) {
2496 if (group->iommu_group != iommu_group)
2498 update_dirty_scope = !group->pinned_page_dirty_scope;
2499 list_del(&group->next);
2502 if (list_empty(&iommu->emulated_iommu_groups) &&
2503 list_empty(&iommu->domain_list)) {
2504 WARN_ON(!list_empty(&iommu->device_list));
2505 vfio_iommu_unmap_unpin_all(iommu);
2507 goto detach_group_done;
2511 * Get a copy of iova list. This will be used to update
2512 * and to replace the current one later. Please note that
2513 * we will leave the original list as it is if update fails.
2515 vfio_iommu_iova_get_copy(iommu, &iova_copy);
2517 list_for_each_entry(domain, &iommu->domain_list, next) {
2518 group = find_iommu_group(domain, iommu_group);
2522 iommu_detach_group(domain->domain, group->iommu_group);
2523 update_dirty_scope = !group->pinned_page_dirty_scope;
2524 list_del(&group->next);
2527 * Group ownership provides privilege, if the group list is
2528 * empty, the domain goes away. If it's the last domain with
2529 * iommu and external domain doesn't exist, then all the
2530 * mappings go away too. If it's the last domain with iommu and
2531 * external domain exist, update accounting
2533 if (list_empty(&domain->group_list)) {
2534 if (list_is_singular(&iommu->domain_list)) {
2535 if (list_empty(&iommu->emulated_iommu_groups)) {
2536 WARN_ON(!list_empty(
2537 &iommu->device_list));
2538 vfio_iommu_unmap_unpin_all(iommu);
2540 vfio_iommu_unmap_unpin_reaccount(iommu);
2543 iommu_domain_free(domain->domain);
2544 list_del(&domain->next);
2546 vfio_iommu_aper_expand(iommu, &iova_copy);
2547 vfio_update_pgsize_bitmap(iommu);
2552 if (!vfio_iommu_resv_refresh(iommu, &iova_copy))
2553 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2555 vfio_iommu_iova_free(&iova_copy);
2559 * Removal of a group without dirty tracking may allow the iommu scope
2562 if (update_dirty_scope) {
2563 iommu->num_non_pinned_groups--;
2564 if (iommu->dirty_page_tracking)
2565 vfio_iommu_populate_bitmap_full(iommu);
2567 mutex_unlock(&iommu->lock);
2570 static void *vfio_iommu_type1_open(unsigned long arg)
2572 struct vfio_iommu *iommu;
2574 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
2576 return ERR_PTR(-ENOMEM);
2579 case VFIO_TYPE1_IOMMU:
2581 case VFIO_TYPE1_NESTING_IOMMU:
2582 iommu->nesting = true;
2584 case VFIO_TYPE1v2_IOMMU:
2589 return ERR_PTR(-EINVAL);
2592 INIT_LIST_HEAD(&iommu->domain_list);
2593 INIT_LIST_HEAD(&iommu->iova_list);
2594 iommu->dma_list = RB_ROOT;
2595 iommu->dma_avail = dma_entry_limit;
2596 iommu->container_open = true;
2597 mutex_init(&iommu->lock);
2598 mutex_init(&iommu->device_list_lock);
2599 INIT_LIST_HEAD(&iommu->device_list);
2600 init_waitqueue_head(&iommu->vaddr_wait);
2601 iommu->pgsize_bitmap = PAGE_MASK;
2602 INIT_LIST_HEAD(&iommu->emulated_iommu_groups);
2607 static void vfio_release_domain(struct vfio_domain *domain)
2609 struct vfio_iommu_group *group, *group_tmp;
2611 list_for_each_entry_safe(group, group_tmp,
2612 &domain->group_list, next) {
2613 iommu_detach_group(domain->domain, group->iommu_group);
2614 list_del(&group->next);
2618 iommu_domain_free(domain->domain);
2621 static void vfio_iommu_type1_release(void *iommu_data)
2623 struct vfio_iommu *iommu = iommu_data;
2624 struct vfio_domain *domain, *domain_tmp;
2625 struct vfio_iommu_group *group, *next_group;
2627 list_for_each_entry_safe(group, next_group,
2628 &iommu->emulated_iommu_groups, next) {
2629 list_del(&group->next);
2633 vfio_iommu_unmap_unpin_all(iommu);
2635 list_for_each_entry_safe(domain, domain_tmp,
2636 &iommu->domain_list, next) {
2637 vfio_release_domain(domain);
2638 list_del(&domain->next);
2642 vfio_iommu_iova_free(&iommu->iova_list);
2647 static int vfio_domains_have_enforce_cache_coherency(struct vfio_iommu *iommu)
2649 struct vfio_domain *domain;
2652 mutex_lock(&iommu->lock);
2653 list_for_each_entry(domain, &iommu->domain_list, next) {
2654 if (!(domain->enforce_cache_coherency)) {
2659 mutex_unlock(&iommu->lock);
2664 static int vfio_iommu_type1_check_extension(struct vfio_iommu *iommu,
2668 case VFIO_TYPE1_IOMMU:
2669 case VFIO_TYPE1v2_IOMMU:
2670 case VFIO_TYPE1_NESTING_IOMMU:
2671 case VFIO_UNMAP_ALL:
2672 case VFIO_UPDATE_VADDR:
2674 case VFIO_DMA_CC_IOMMU:
2677 return vfio_domains_have_enforce_cache_coherency(iommu);
2683 static int vfio_iommu_iova_add_cap(struct vfio_info_cap *caps,
2684 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas,
2687 struct vfio_info_cap_header *header;
2688 struct vfio_iommu_type1_info_cap_iova_range *iova_cap;
2690 header = vfio_info_cap_add(caps, size,
2691 VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE, 1);
2693 return PTR_ERR(header);
2695 iova_cap = container_of(header,
2696 struct vfio_iommu_type1_info_cap_iova_range,
2698 iova_cap->nr_iovas = cap_iovas->nr_iovas;
2699 memcpy(iova_cap->iova_ranges, cap_iovas->iova_ranges,
2700 cap_iovas->nr_iovas * sizeof(*cap_iovas->iova_ranges));
2704 static int vfio_iommu_iova_build_caps(struct vfio_iommu *iommu,
2705 struct vfio_info_cap *caps)
2707 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas;
2708 struct vfio_iova *iova;
2710 int iovas = 0, i = 0, ret;
2712 list_for_each_entry(iova, &iommu->iova_list, list)
2717 * Return 0 as a container with a single mdev device
2718 * will have an empty list
2723 size = struct_size(cap_iovas, iova_ranges, iovas);
2725 cap_iovas = kzalloc(size, GFP_KERNEL);
2729 cap_iovas->nr_iovas = iovas;
2731 list_for_each_entry(iova, &iommu->iova_list, list) {
2732 cap_iovas->iova_ranges[i].start = iova->start;
2733 cap_iovas->iova_ranges[i].end = iova->end;
2737 ret = vfio_iommu_iova_add_cap(caps, cap_iovas, size);
2743 static int vfio_iommu_migration_build_caps(struct vfio_iommu *iommu,
2744 struct vfio_info_cap *caps)
2746 struct vfio_iommu_type1_info_cap_migration cap_mig;
2748 cap_mig.header.id = VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION;
2749 cap_mig.header.version = 1;
2752 /* support minimum pgsize */
2753 cap_mig.pgsize_bitmap = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2754 cap_mig.max_dirty_bitmap_size = DIRTY_BITMAP_SIZE_MAX;
2756 return vfio_info_add_capability(caps, &cap_mig.header, sizeof(cap_mig));
2759 static int vfio_iommu_dma_avail_build_caps(struct vfio_iommu *iommu,
2760 struct vfio_info_cap *caps)
2762 struct vfio_iommu_type1_info_dma_avail cap_dma_avail;
2764 cap_dma_avail.header.id = VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL;
2765 cap_dma_avail.header.version = 1;
2767 cap_dma_avail.avail = iommu->dma_avail;
2769 return vfio_info_add_capability(caps, &cap_dma_avail.header,
2770 sizeof(cap_dma_avail));
2773 static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu,
2776 struct vfio_iommu_type1_info info;
2777 unsigned long minsz;
2778 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
2779 unsigned long capsz;
2782 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
2784 /* For backward compatibility, cannot require this */
2785 capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
2787 if (copy_from_user(&info, (void __user *)arg, minsz))
2790 if (info.argsz < minsz)
2793 if (info.argsz >= capsz) {
2795 info.cap_offset = 0; /* output, no-recopy necessary */
2798 mutex_lock(&iommu->lock);
2799 info.flags = VFIO_IOMMU_INFO_PGSIZES;
2801 info.iova_pgsizes = iommu->pgsize_bitmap;
2803 ret = vfio_iommu_migration_build_caps(iommu, &caps);
2806 ret = vfio_iommu_dma_avail_build_caps(iommu, &caps);
2809 ret = vfio_iommu_iova_build_caps(iommu, &caps);
2811 mutex_unlock(&iommu->lock);
2817 info.flags |= VFIO_IOMMU_INFO_CAPS;
2819 if (info.argsz < sizeof(info) + caps.size) {
2820 info.argsz = sizeof(info) + caps.size;
2822 vfio_info_cap_shift(&caps, sizeof(info));
2823 if (copy_to_user((void __user *)arg +
2824 sizeof(info), caps.buf,
2829 info.cap_offset = sizeof(info);
2835 return copy_to_user((void __user *)arg, &info, minsz) ?
2839 static int vfio_iommu_type1_map_dma(struct vfio_iommu *iommu,
2842 struct vfio_iommu_type1_dma_map map;
2843 unsigned long minsz;
2844 uint32_t mask = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE |
2845 VFIO_DMA_MAP_FLAG_VADDR;
2847 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
2849 if (copy_from_user(&map, (void __user *)arg, minsz))
2852 if (map.argsz < minsz || map.flags & ~mask)
2855 return vfio_dma_do_map(iommu, &map);
2858 static int vfio_iommu_type1_unmap_dma(struct vfio_iommu *iommu,
2861 struct vfio_iommu_type1_dma_unmap unmap;
2862 struct vfio_bitmap bitmap = { 0 };
2863 uint32_t mask = VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP |
2864 VFIO_DMA_UNMAP_FLAG_VADDR |
2865 VFIO_DMA_UNMAP_FLAG_ALL;
2866 unsigned long minsz;
2869 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
2871 if (copy_from_user(&unmap, (void __user *)arg, minsz))
2874 if (unmap.argsz < minsz || unmap.flags & ~mask)
2877 if ((unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
2878 (unmap.flags & (VFIO_DMA_UNMAP_FLAG_ALL |
2879 VFIO_DMA_UNMAP_FLAG_VADDR)))
2882 if (unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
2883 unsigned long pgshift;
2885 if (unmap.argsz < (minsz + sizeof(bitmap)))
2888 if (copy_from_user(&bitmap,
2889 (void __user *)(arg + minsz),
2893 if (!access_ok((void __user *)bitmap.data, bitmap.size))
2896 pgshift = __ffs(bitmap.pgsize);
2897 ret = verify_bitmap_size(unmap.size >> pgshift,
2903 ret = vfio_dma_do_unmap(iommu, &unmap, &bitmap);
2907 return copy_to_user((void __user *)arg, &unmap, minsz) ?
2911 static int vfio_iommu_type1_dirty_pages(struct vfio_iommu *iommu,
2914 struct vfio_iommu_type1_dirty_bitmap dirty;
2915 uint32_t mask = VFIO_IOMMU_DIRTY_PAGES_FLAG_START |
2916 VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP |
2917 VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP;
2918 unsigned long minsz;
2924 minsz = offsetofend(struct vfio_iommu_type1_dirty_bitmap, flags);
2926 if (copy_from_user(&dirty, (void __user *)arg, minsz))
2929 if (dirty.argsz < minsz || dirty.flags & ~mask)
2932 /* only one flag should be set at a time */
2933 if (__ffs(dirty.flags) != __fls(dirty.flags))
2936 if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_START) {
2939 mutex_lock(&iommu->lock);
2940 pgsize = 1 << __ffs(iommu->pgsize_bitmap);
2941 if (!iommu->dirty_page_tracking) {
2942 ret = vfio_dma_bitmap_alloc_all(iommu, pgsize);
2944 iommu->dirty_page_tracking = true;
2946 mutex_unlock(&iommu->lock);
2948 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP) {
2949 mutex_lock(&iommu->lock);
2950 if (iommu->dirty_page_tracking) {
2951 iommu->dirty_page_tracking = false;
2952 vfio_dma_bitmap_free_all(iommu);
2954 mutex_unlock(&iommu->lock);
2956 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP) {
2957 struct vfio_iommu_type1_dirty_bitmap_get range;
2958 unsigned long pgshift;
2959 size_t data_size = dirty.argsz - minsz;
2960 size_t iommu_pgsize;
2962 if (!data_size || data_size < sizeof(range))
2965 if (copy_from_user(&range, (void __user *)(arg + minsz),
2969 if (range.iova + range.size < range.iova)
2971 if (!access_ok((void __user *)range.bitmap.data,
2975 pgshift = __ffs(range.bitmap.pgsize);
2976 ret = verify_bitmap_size(range.size >> pgshift,
2981 mutex_lock(&iommu->lock);
2983 iommu_pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2985 /* allow only smallest supported pgsize */
2986 if (range.bitmap.pgsize != iommu_pgsize) {
2990 if (range.iova & (iommu_pgsize - 1)) {
2994 if (!range.size || range.size & (iommu_pgsize - 1)) {
2999 if (iommu->dirty_page_tracking)
3000 ret = vfio_iova_dirty_bitmap(range.bitmap.data,
3003 range.bitmap.pgsize);
3007 mutex_unlock(&iommu->lock);
3015 static long vfio_iommu_type1_ioctl(void *iommu_data,
3016 unsigned int cmd, unsigned long arg)
3018 struct vfio_iommu *iommu = iommu_data;
3021 case VFIO_CHECK_EXTENSION:
3022 return vfio_iommu_type1_check_extension(iommu, arg);
3023 case VFIO_IOMMU_GET_INFO:
3024 return vfio_iommu_type1_get_info(iommu, arg);
3025 case VFIO_IOMMU_MAP_DMA:
3026 return vfio_iommu_type1_map_dma(iommu, arg);
3027 case VFIO_IOMMU_UNMAP_DMA:
3028 return vfio_iommu_type1_unmap_dma(iommu, arg);
3029 case VFIO_IOMMU_DIRTY_PAGES:
3030 return vfio_iommu_type1_dirty_pages(iommu, arg);
3036 static void vfio_iommu_type1_register_device(void *iommu_data,
3037 struct vfio_device *vdev)
3039 struct vfio_iommu *iommu = iommu_data;
3041 if (!vdev->ops->dma_unmap)
3045 * list_empty(&iommu->device_list) is tested under the iommu->lock while
3046 * iteration for dma_unmap must be done under the device_list_lock.
3047 * Holding both locks here allows avoiding the device_list_lock in
3048 * several fast paths. See vfio_notify_dma_unmap()
3050 mutex_lock(&iommu->lock);
3051 mutex_lock(&iommu->device_list_lock);
3052 list_add(&vdev->iommu_entry, &iommu->device_list);
3053 mutex_unlock(&iommu->device_list_lock);
3054 mutex_unlock(&iommu->lock);
3057 static void vfio_iommu_type1_unregister_device(void *iommu_data,
3058 struct vfio_device *vdev)
3060 struct vfio_iommu *iommu = iommu_data;
3062 if (!vdev->ops->dma_unmap)
3065 mutex_lock(&iommu->lock);
3066 mutex_lock(&iommu->device_list_lock);
3067 list_del(&vdev->iommu_entry);
3068 mutex_unlock(&iommu->device_list_lock);
3069 mutex_unlock(&iommu->lock);
3072 static int vfio_iommu_type1_dma_rw_chunk(struct vfio_iommu *iommu,
3073 dma_addr_t user_iova, void *data,
3074 size_t count, bool write,
3077 struct mm_struct *mm;
3078 unsigned long vaddr;
3079 struct vfio_dma *dma;
3080 bool kthread = current->mm == NULL;
3086 ret = vfio_find_dma_valid(iommu, user_iova, 1, &dma);
3090 if ((write && !(dma->prot & IOMMU_WRITE)) ||
3091 !(dma->prot & IOMMU_READ))
3094 mm = get_task_mm(dma->task);
3101 else if (current->mm != mm)
3104 offset = user_iova - dma->iova;
3106 if (count > dma->size - offset)
3107 count = dma->size - offset;
3109 vaddr = dma->vaddr + offset;
3112 *copied = copy_to_user((void __user *)vaddr, data,
3114 if (*copied && iommu->dirty_page_tracking) {
3115 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
3117 * Bitmap populated with the smallest supported page
3120 bitmap_set(dma->bitmap, offset >> pgshift,
3121 ((offset + *copied - 1) >> pgshift) -
3122 (offset >> pgshift) + 1);
3125 *copied = copy_from_user(data, (void __user *)vaddr,
3128 kthread_unuse_mm(mm);
3131 return *copied ? 0 : -EFAULT;
3134 static int vfio_iommu_type1_dma_rw(void *iommu_data, dma_addr_t user_iova,
3135 void *data, size_t count, bool write)
3137 struct vfio_iommu *iommu = iommu_data;
3141 mutex_lock(&iommu->lock);
3143 ret = vfio_iommu_type1_dma_rw_chunk(iommu, user_iova, data,
3144 count, write, &done);
3153 mutex_unlock(&iommu->lock);
3157 static struct iommu_domain *
3158 vfio_iommu_type1_group_iommu_domain(void *iommu_data,
3159 struct iommu_group *iommu_group)
3161 struct iommu_domain *domain = ERR_PTR(-ENODEV);
3162 struct vfio_iommu *iommu = iommu_data;
3163 struct vfio_domain *d;
3165 if (!iommu || !iommu_group)
3166 return ERR_PTR(-EINVAL);
3168 mutex_lock(&iommu->lock);
3169 list_for_each_entry(d, &iommu->domain_list, next) {
3170 if (find_iommu_group(d, iommu_group)) {
3175 mutex_unlock(&iommu->lock);
3180 static void vfio_iommu_type1_notify(void *iommu_data,
3181 enum vfio_iommu_notify_type event)
3183 struct vfio_iommu *iommu = iommu_data;
3185 if (event != VFIO_IOMMU_CONTAINER_CLOSE)
3187 mutex_lock(&iommu->lock);
3188 iommu->container_open = false;
3189 mutex_unlock(&iommu->lock);
3190 wake_up_all(&iommu->vaddr_wait);
3193 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
3194 .name = "vfio-iommu-type1",
3195 .owner = THIS_MODULE,
3196 .open = vfio_iommu_type1_open,
3197 .release = vfio_iommu_type1_release,
3198 .ioctl = vfio_iommu_type1_ioctl,
3199 .attach_group = vfio_iommu_type1_attach_group,
3200 .detach_group = vfio_iommu_type1_detach_group,
3201 .pin_pages = vfio_iommu_type1_pin_pages,
3202 .unpin_pages = vfio_iommu_type1_unpin_pages,
3203 .register_device = vfio_iommu_type1_register_device,
3204 .unregister_device = vfio_iommu_type1_unregister_device,
3205 .dma_rw = vfio_iommu_type1_dma_rw,
3206 .group_iommu_domain = vfio_iommu_type1_group_iommu_domain,
3207 .notify = vfio_iommu_type1_notify,
3210 static int __init vfio_iommu_type1_init(void)
3212 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
3215 static void __exit vfio_iommu_type1_cleanup(void)
3217 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
3220 module_init(vfio_iommu_type1_init);
3221 module_exit(vfio_iommu_type1_cleanup);
3223 MODULE_VERSION(DRIVER_VERSION);
3224 MODULE_LICENSE("GPL v2");
3225 MODULE_AUTHOR(DRIVER_AUTHOR);
3226 MODULE_DESCRIPTION(DRIVER_DESC);