2 * VFIO: IOMMU DMA mapping support for Type1 IOMMU
4 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
5 * Author: Alex Williamson <alex.williamson@redhat.com>
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
11 * Derived from original vfio:
12 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
13 * Author: Tom Lyon, pugs@cisco.com
15 * We arbitrarily define a Type1 IOMMU as one matching the below code.
16 * It could be called the x86 IOMMU as it's designed for AMD-Vi & Intel
17 * VT-d, but that makes it harder to re-use as theoretically anyone
18 * implementing a similar IOMMU could make use of this. We expect the
19 * IOMMU to support the IOMMU API and have few to no restrictions around
20 * the IOVA range that can be mapped. The Type1 IOMMU is currently
21 * optimized for relatively static mappings of a userspace process with
22 * userpsace pages pinned into memory. We also assume devices and IOMMU
23 * domains are PCI based as the IOMMU API is still centered around a
24 * device/bus interface rather than a group interface.
27 #include <linux/compat.h>
28 #include <linux/device.h>
30 #include <linux/iommu.h>
31 #include <linux/module.h>
33 #include <linux/rbtree.h>
34 #include <linux/sched.h>
35 #include <linux/slab.h>
36 #include <linux/uaccess.h>
37 #include <linux/vfio.h>
38 #include <linux/workqueue.h>
39 #include <linux/pid_namespace.h>
40 #include <linux/mdev.h>
41 #include <linux/notifier.h>
43 #define DRIVER_VERSION "0.2"
44 #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
45 #define DRIVER_DESC "Type1 IOMMU driver for VFIO"
47 static bool allow_unsafe_interrupts;
48 module_param_named(allow_unsafe_interrupts,
49 allow_unsafe_interrupts, bool, S_IRUGO | S_IWUSR);
50 MODULE_PARM_DESC(allow_unsafe_interrupts,
51 "Enable VFIO IOMMU support for on platforms without interrupt remapping support.");
53 static bool disable_hugepages;
54 module_param_named(disable_hugepages,
55 disable_hugepages, bool, S_IRUGO | S_IWUSR);
56 MODULE_PARM_DESC(disable_hugepages,
57 "Disable VFIO IOMMU support for IOMMU hugepages.");
60 struct list_head domain_list;
61 struct vfio_domain *external_domain; /* domain for external user */
63 struct rb_root dma_list;
64 struct blocking_notifier_head notifier;
70 struct iommu_domain *domain;
71 struct list_head next;
72 struct list_head group_list;
73 int prot; /* IOMMU_CACHE */
74 bool fgsp; /* Fine-grained super pages */
79 dma_addr_t iova; /* Device address */
80 unsigned long vaddr; /* Process virtual addr */
81 size_t size; /* Map size (bytes) */
82 int prot; /* IOMMU_READ/WRITE */
84 struct task_struct *task;
85 struct rb_root pfn_list; /* Ex-user pinned pfn list */
89 struct iommu_group *iommu_group;
90 struct list_head next;
94 * Guest RAM pinning working set or DMA target
98 dma_addr_t iova; /* Device address */
99 unsigned long pfn; /* Host pfn */
103 #define IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu) \
104 (!list_empty(&iommu->domain_list))
106 static int put_pfn(unsigned long pfn, int prot);
109 * This code handles mapping and unmapping of user data buffers
110 * into DMA'ble space using the IOMMU
113 static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
114 dma_addr_t start, size_t size)
116 struct rb_node *node = iommu->dma_list.rb_node;
119 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
121 if (start + size <= dma->iova)
122 node = node->rb_left;
123 else if (start >= dma->iova + dma->size)
124 node = node->rb_right;
132 static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new)
134 struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL;
135 struct vfio_dma *dma;
139 dma = rb_entry(parent, struct vfio_dma, node);
141 if (new->iova + new->size <= dma->iova)
142 link = &(*link)->rb_left;
144 link = &(*link)->rb_right;
147 rb_link_node(&new->node, parent, link);
148 rb_insert_color(&new->node, &iommu->dma_list);
151 static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
153 rb_erase(&old->node, &iommu->dma_list);
157 * Helper Functions for host iova-pfn list
159 static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
161 struct vfio_pfn *vpfn;
162 struct rb_node *node = dma->pfn_list.rb_node;
165 vpfn = rb_entry(node, struct vfio_pfn, node);
167 if (iova < vpfn->iova)
168 node = node->rb_left;
169 else if (iova > vpfn->iova)
170 node = node->rb_right;
177 static void vfio_link_pfn(struct vfio_dma *dma,
178 struct vfio_pfn *new)
180 struct rb_node **link, *parent = NULL;
181 struct vfio_pfn *vpfn;
183 link = &dma->pfn_list.rb_node;
186 vpfn = rb_entry(parent, struct vfio_pfn, node);
188 if (new->iova < vpfn->iova)
189 link = &(*link)->rb_left;
191 link = &(*link)->rb_right;
194 rb_link_node(&new->node, parent, link);
195 rb_insert_color(&new->node, &dma->pfn_list);
198 static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old)
200 rb_erase(&old->node, &dma->pfn_list);
203 static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova,
206 struct vfio_pfn *vpfn;
208 vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL);
214 atomic_set(&vpfn->ref_count, 1);
215 vfio_link_pfn(dma, vpfn);
219 static void vfio_remove_from_pfn_list(struct vfio_dma *dma,
220 struct vfio_pfn *vpfn)
222 vfio_unlink_pfn(dma, vpfn);
226 static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma,
229 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
232 atomic_inc(&vpfn->ref_count);
236 static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn)
240 if (atomic_dec_and_test(&vpfn->ref_count)) {
241 ret = put_pfn(vpfn->pfn, dma->prot);
242 vfio_remove_from_pfn_list(dma, vpfn);
248 struct mm_struct *mm;
250 struct work_struct work;
253 /* delayed decrement/increment for locked_vm */
254 static void vfio_lock_acct_bg(struct work_struct *work)
256 struct vwork *vwork = container_of(work, struct vwork, work);
257 struct mm_struct *mm;
260 down_write(&mm->mmap_sem);
261 mm->locked_vm += vwork->npage;
262 up_write(&mm->mmap_sem);
267 static void vfio_lock_acct(struct task_struct *task, long npage)
270 struct mm_struct *mm;
275 mm = get_task_mm(task);
277 return; /* process exited or nothing to do */
279 if (down_write_trylock(&mm->mmap_sem)) {
280 mm->locked_vm += npage;
281 up_write(&mm->mmap_sem);
287 * Couldn't get mmap_sem lock, so must setup to update
288 * mm->locked_vm later. If locked_vm were atomic, we
289 * wouldn't need this silliness
291 vwork = kmalloc(sizeof(struct vwork), GFP_KERNEL);
296 INIT_WORK(&vwork->work, vfio_lock_acct_bg);
298 vwork->npage = npage;
299 schedule_work(&vwork->work);
303 * Some mappings aren't backed by a struct page, for example an mmap'd
304 * MMIO range for our own or another device. These use a different
305 * pfn conversion and shouldn't be tracked as locked pages.
307 static bool is_invalid_reserved_pfn(unsigned long pfn)
309 if (pfn_valid(pfn)) {
311 struct page *tail = pfn_to_page(pfn);
312 struct page *head = compound_head(tail);
313 reserved = !!(PageReserved(head));
316 * "head" is not a dangling pointer
317 * (compound_head takes care of that)
318 * but the hugepage may have been split
319 * from under us (and we may not hold a
320 * reference count on the head page so it can
321 * be reused before we run PageReferenced), so
322 * we've to check PageTail before returning
329 return PageReserved(tail);
335 static int put_pfn(unsigned long pfn, int prot)
337 if (!is_invalid_reserved_pfn(pfn)) {
338 struct page *page = pfn_to_page(pfn);
339 if (prot & IOMMU_WRITE)
347 static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,
348 int prot, unsigned long *pfn)
350 struct page *page[1];
351 struct vm_area_struct *vma;
354 if (mm == current->mm) {
355 ret = get_user_pages_fast(vaddr, 1, !!(prot & IOMMU_WRITE),
358 unsigned int flags = 0;
360 if (prot & IOMMU_WRITE)
363 down_read(&mm->mmap_sem);
364 ret = get_user_pages_remote(NULL, mm, vaddr, 1, flags, page,
366 up_read(&mm->mmap_sem);
370 *pfn = page_to_pfn(page[0]);
374 down_read(&mm->mmap_sem);
376 vma = find_vma_intersection(mm, vaddr, vaddr + 1);
378 if (vma && vma->vm_flags & VM_PFNMAP) {
379 *pfn = ((vaddr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
380 if (is_invalid_reserved_pfn(*pfn))
384 up_read(&mm->mmap_sem);
389 * Attempt to pin pages. We really don't want to track all the pfns and
390 * the iommu can only map chunks of consecutive pfns anyway, so get the
391 * first page and all consecutive pages with the same locking.
393 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
394 long npage, unsigned long *pfn_base)
397 bool lock_cap = ns_capable(task_active_pid_ns(dma->task)->user_ns,
399 struct mm_struct *mm;
400 long ret, i = 0, lock_acct = 0;
402 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
404 mm = get_task_mm(dma->task);
408 ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base);
410 goto pin_pg_remote_exit;
412 rsvd = is_invalid_reserved_pfn(*pfn_base);
413 limit = task_rlimit(dma->task, RLIMIT_MEMLOCK) >> PAGE_SHIFT;
416 * Reserved pages aren't counted against the user, externally pinned
417 * pages are already counted against the user.
419 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
420 if (!lock_cap && mm->locked_vm + 1 > limit) {
421 put_pfn(*pfn_base, dma->prot);
422 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__,
423 limit << PAGE_SHIFT);
425 goto pin_pg_remote_exit;
431 if (likely(!disable_hugepages)) {
432 /* Lock all the consecutive pages from pfn_base */
433 for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; i < npage;
434 i++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) {
435 unsigned long pfn = 0;
437 ret = vaddr_get_pfn(mm, vaddr, dma->prot, &pfn);
441 if (pfn != *pfn_base + i ||
442 rsvd != is_invalid_reserved_pfn(pfn)) {
443 put_pfn(pfn, dma->prot);
447 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
449 mm->locked_vm + lock_acct + 1 > limit) {
450 put_pfn(pfn, dma->prot);
451 pr_warn("%s: RLIMIT_MEMLOCK (%ld) "
452 "exceeded\n", __func__,
453 limit << PAGE_SHIFT);
461 vfio_lock_acct(dma->task, lock_acct);
469 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
470 unsigned long pfn, long npage,
473 long unlocked = 0, locked = 0;
476 for (i = 0; i < npage; i++) {
477 if (put_pfn(pfn++, dma->prot)) {
479 if (vfio_find_vpfn(dma, iova + (i << PAGE_SHIFT)))
485 vfio_lock_acct(dma->task, locked - unlocked);
490 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
491 unsigned long *pfn_base, bool do_accounting)
494 bool lock_cap = ns_capable(task_active_pid_ns(dma->task)->user_ns,
496 struct mm_struct *mm;
500 mm = get_task_mm(dma->task);
504 ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base);
508 rsvd = is_invalid_reserved_pfn(*pfn_base);
509 limit = task_rlimit(dma->task, RLIMIT_MEMLOCK) >> PAGE_SHIFT;
511 if (!rsvd && !lock_cap && mm->locked_vm + 1 > limit) {
512 put_pfn(*pfn_base, dma->prot);
513 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK (%ld) exceeded\n",
514 __func__, dma->task->comm, task_pid_nr(dma->task),
515 limit << PAGE_SHIFT);
520 if (!rsvd && do_accounting)
521 vfio_lock_acct(dma->task, 1);
529 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
533 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
538 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
541 vfio_lock_acct(dma->task, -unlocked);
546 static int vfio_iommu_type1_pin_pages(void *iommu_data,
547 unsigned long *user_pfn,
549 unsigned long *phys_pfn)
551 struct vfio_iommu *iommu = iommu_data;
553 unsigned long remote_vaddr;
554 struct vfio_dma *dma;
557 if (!iommu || !user_pfn || !phys_pfn)
560 /* Supported for v2 version only */
564 mutex_lock(&iommu->lock);
566 /* Fail if notifier list is empty */
567 if ((!iommu->external_domain) || (!iommu->notifier.head)) {
573 * If iommu capable domain exist in the container then all pages are
574 * already pinned and accounted. Accouting should be done if there is no
575 * iommu capable domain in the container.
577 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
579 for (i = 0; i < npage; i++) {
581 struct vfio_pfn *vpfn;
583 iova = user_pfn[i] << PAGE_SHIFT;
584 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
590 if ((dma->prot & prot) != prot) {
595 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
597 phys_pfn[i] = vpfn->pfn;
601 remote_vaddr = dma->vaddr + iova - dma->iova;
602 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
609 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
611 vfio_unpin_page_external(dma, iova, do_accounting);
621 for (j = 0; j < i; j++) {
624 iova = user_pfn[j] << PAGE_SHIFT;
625 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
626 vfio_unpin_page_external(dma, iova, do_accounting);
630 mutex_unlock(&iommu->lock);
634 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
635 unsigned long *user_pfn,
638 struct vfio_iommu *iommu = iommu_data;
642 if (!iommu || !user_pfn)
645 /* Supported for v2 version only */
649 mutex_lock(&iommu->lock);
651 if (!iommu->external_domain) {
652 mutex_unlock(&iommu->lock);
656 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
657 for (i = 0; i < npage; i++) {
658 struct vfio_dma *dma;
661 iova = user_pfn[i] << PAGE_SHIFT;
662 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
665 vfio_unpin_page_external(dma, iova, do_accounting);
669 mutex_unlock(&iommu->lock);
670 return i > npage ? npage : (i > 0 ? i : -EINVAL);
673 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
676 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
677 struct vfio_domain *domain, *d;
683 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
687 * We use the IOMMU to track the physical addresses, otherwise we'd
688 * need a much more complicated tracking system. Unfortunately that
689 * means we need to use one of the iommu domains to figure out the
690 * pfns to unpin. The rest need to be unmapped in advance so we have
691 * no iommu translations remaining when the pages are unpinned.
693 domain = d = list_first_entry(&iommu->domain_list,
694 struct vfio_domain, next);
696 list_for_each_entry_continue(d, &iommu->domain_list, next) {
697 iommu_unmap(d->domain, dma->iova, dma->size);
702 size_t unmapped, len;
703 phys_addr_t phys, next;
705 phys = iommu_iova_to_phys(domain->domain, iova);
706 if (WARN_ON(!phys)) {
712 * To optimize for fewer iommu_unmap() calls, each of which
713 * may require hardware cache flushing, try to find the
714 * largest contiguous physical memory chunk to unmap.
716 for (len = PAGE_SIZE;
717 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
718 next = iommu_iova_to_phys(domain->domain, iova + len);
719 if (next != phys + len)
723 unmapped = iommu_unmap(domain->domain, iova, len);
724 if (WARN_ON(!unmapped))
727 unlocked += vfio_unpin_pages_remote(dma, iova,
729 unmapped >> PAGE_SHIFT,
736 dma->iommu_mapped = false;
738 vfio_lock_acct(dma->task, -unlocked);
744 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
746 vfio_unmap_unpin(iommu, dma, true);
747 vfio_unlink_dma(iommu, dma);
748 put_task_struct(dma->task);
752 static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu)
754 struct vfio_domain *domain;
755 unsigned long bitmap = ULONG_MAX;
757 mutex_lock(&iommu->lock);
758 list_for_each_entry(domain, &iommu->domain_list, next)
759 bitmap &= domain->domain->pgsize_bitmap;
760 mutex_unlock(&iommu->lock);
763 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
764 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
765 * That way the user will be able to map/unmap buffers whose size/
766 * start address is aligned with PAGE_SIZE. Pinning code uses that
767 * granularity while iommu driver can use the sub-PAGE_SIZE size
770 if (bitmap & ~PAGE_MASK) {
778 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
779 struct vfio_iommu_type1_dma_unmap *unmap)
782 struct vfio_dma *dma, *dma_last = NULL;
784 int ret = 0, retries = 0;
786 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
788 if (unmap->iova & mask)
790 if (!unmap->size || unmap->size & mask)
793 WARN_ON(mask & PAGE_MASK);
795 mutex_lock(&iommu->lock);
798 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
799 * avoid tracking individual mappings. This means that the granularity
800 * of the original mapping was lost and the user was allowed to attempt
801 * to unmap any range. Depending on the contiguousness of physical
802 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
803 * or may not have worked. We only guaranteed unmap granularity
804 * matching the original mapping; even though it was untracked here,
805 * the original mappings are reflected in IOMMU mappings. This
806 * resulted in a couple unusual behaviors. First, if a range is not
807 * able to be unmapped, ex. a set of 4k pages that was mapped as a
808 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
809 * a zero sized unmap. Also, if an unmap request overlaps the first
810 * address of a hugepage, the IOMMU will unmap the entire hugepage.
811 * This also returns success and the returned unmap size reflects the
812 * actual size unmapped.
814 * We attempt to maintain compatibility with this "v1" interface, but
815 * we take control out of the hands of the IOMMU. Therefore, an unmap
816 * request offset from the beginning of the original mapping will
817 * return success with zero sized unmap. And an unmap request covering
818 * the first iova of mapping will unmap the entire range.
820 * The v2 version of this interface intends to be more deterministic.
821 * Unmap requests must fully cover previous mappings. Multiple
822 * mappings may still be unmaped by specifying large ranges, but there
823 * must not be any previous mappings bisected by the range. An error
824 * will be returned if these conditions are not met. The v2 interface
825 * will only return success and a size of zero if there were no
826 * mappings within the range.
829 dma = vfio_find_dma(iommu, unmap->iova, 1);
830 if (dma && dma->iova != unmap->iova) {
834 dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0);
835 if (dma && dma->iova + dma->size != unmap->iova + unmap->size) {
841 while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
842 if (!iommu->v2 && unmap->iova > dma->iova)
845 * Task with same address space who mapped this iova range is
846 * allowed to unmap the iova range.
848 if (dma->task->mm != current->mm)
851 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
852 struct vfio_iommu_type1_dma_unmap nb_unmap;
854 if (dma_last == dma) {
855 BUG_ON(++retries > 10);
861 nb_unmap.iova = dma->iova;
862 nb_unmap.size = dma->size;
865 * Notify anyone (mdev vendor drivers) to invalidate and
866 * unmap iovas within the range we're about to unmap.
867 * Vendor drivers MUST unpin pages in response to an
870 mutex_unlock(&iommu->lock);
871 blocking_notifier_call_chain(&iommu->notifier,
872 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
876 unmapped += dma->size;
877 vfio_remove_dma(iommu, dma);
881 mutex_unlock(&iommu->lock);
883 /* Report how much was unmapped */
884 unmap->size = unmapped;
890 * Turns out AMD IOMMU has a page table bug where it won't map large pages
891 * to a region that previously mapped smaller pages. This should be fixed
892 * soon, so this is just a temporary workaround to break mappings down into
893 * PAGE_SIZE. Better to map smaller pages than nothing.
895 static int map_try_harder(struct vfio_domain *domain, dma_addr_t iova,
896 unsigned long pfn, long npage, int prot)
901 for (i = 0; i < npage; i++, pfn++, iova += PAGE_SIZE) {
902 ret = iommu_map(domain->domain, iova,
903 (phys_addr_t)pfn << PAGE_SHIFT,
904 PAGE_SIZE, prot | domain->prot);
909 for (; i < npage && i > 0; i--, iova -= PAGE_SIZE)
910 iommu_unmap(domain->domain, iova, PAGE_SIZE);
915 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
916 unsigned long pfn, long npage, int prot)
918 struct vfio_domain *d;
921 list_for_each_entry(d, &iommu->domain_list, next) {
922 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
923 npage << PAGE_SHIFT, prot | d->prot);
926 map_try_harder(d, iova, pfn, npage, prot))
936 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next)
937 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
942 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
945 dma_addr_t iova = dma->iova;
946 unsigned long vaddr = dma->vaddr;
947 size_t size = map_size;
953 /* Pin a contiguous chunk of memory */
954 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
955 size >> PAGE_SHIFT, &pfn);
963 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
966 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
971 size -= npage << PAGE_SHIFT;
972 dma->size += npage << PAGE_SHIFT;
975 dma->iommu_mapped = true;
978 vfio_remove_dma(iommu, dma);
983 static int vfio_dma_do_map(struct vfio_iommu *iommu,
984 struct vfio_iommu_type1_dma_map *map)
986 dma_addr_t iova = map->iova;
987 unsigned long vaddr = map->vaddr;
988 size_t size = map->size;
989 int ret = 0, prot = 0;
991 struct vfio_dma *dma;
993 /* Verify that none of our __u64 fields overflow */
994 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
997 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
999 WARN_ON(mask & PAGE_MASK);
1001 /* READ/WRITE from device perspective */
1002 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1003 prot |= IOMMU_WRITE;
1004 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1007 if (!prot || !size || (size | iova | vaddr) & mask)
1010 /* Don't allow IOVA or virtual address wrap */
1011 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr)
1014 mutex_lock(&iommu->lock);
1016 if (vfio_find_dma(iommu, iova, size)) {
1021 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1030 get_task_struct(current);
1031 dma->task = current;
1032 dma->pfn_list = RB_ROOT;
1034 /* Insert zero-sized and grow as we map chunks of it */
1035 vfio_link_dma(iommu, dma);
1037 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1038 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1041 ret = vfio_pin_map_dma(iommu, dma, size);
1044 mutex_unlock(&iommu->lock);
1048 static int vfio_bus_type(struct device *dev, void *data)
1050 struct bus_type **bus = data;
1052 if (*bus && *bus != dev->bus)
1060 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1061 struct vfio_domain *domain)
1063 struct vfio_domain *d;
1067 /* Arbitrarily pick the first domain in the list for lookups */
1068 d = list_first_entry(&iommu->domain_list, struct vfio_domain, next);
1069 n = rb_first(&iommu->dma_list);
1071 for (; n; n = rb_next(n)) {
1072 struct vfio_dma *dma;
1075 dma = rb_entry(n, struct vfio_dma, node);
1078 while (iova < dma->iova + dma->size) {
1082 if (dma->iommu_mapped) {
1086 phys = iommu_iova_to_phys(d->domain, iova);
1088 if (WARN_ON(!phys)) {
1096 while (i < dma->iova + dma->size &&
1097 p == iommu_iova_to_phys(d->domain, i)) {
1104 unsigned long vaddr = dma->vaddr +
1106 size_t n = dma->iova + dma->size - iova;
1109 npage = vfio_pin_pages_remote(dma, vaddr,
1118 phys = pfn << PAGE_SHIFT;
1119 size = npage << PAGE_SHIFT;
1122 ret = iommu_map(domain->domain, iova, phys,
1123 size, dma->prot | domain->prot);
1129 dma->iommu_mapped = true;
1135 * We change our unmap behavior slightly depending on whether the IOMMU
1136 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1137 * for practically any contiguous power-of-two mapping we give it. This means
1138 * we don't need to look for contiguous chunks ourselves to make unmapping
1139 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1140 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1141 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1142 * hugetlbfs is in use.
1144 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1147 int ret, order = get_order(PAGE_SIZE * 2);
1149 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1153 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1154 IOMMU_READ | IOMMU_WRITE | domain->prot);
1156 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1158 if (unmapped == PAGE_SIZE)
1159 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1161 domain->fgsp = true;
1164 __free_pages(pages, order);
1167 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1168 struct iommu_group *iommu_group)
1170 struct vfio_group *g;
1172 list_for_each_entry(g, &domain->group_list, next) {
1173 if (g->iommu_group == iommu_group)
1180 static int vfio_iommu_type1_attach_group(void *iommu_data,
1181 struct iommu_group *iommu_group)
1183 struct vfio_iommu *iommu = iommu_data;
1184 struct vfio_group *group;
1185 struct vfio_domain *domain, *d;
1186 struct bus_type *bus = NULL, *mdev_bus;
1189 mutex_lock(&iommu->lock);
1191 list_for_each_entry(d, &iommu->domain_list, next) {
1192 if (find_iommu_group(d, iommu_group)) {
1193 mutex_unlock(&iommu->lock);
1198 if (iommu->external_domain) {
1199 if (find_iommu_group(iommu->external_domain, iommu_group)) {
1200 mutex_unlock(&iommu->lock);
1205 group = kzalloc(sizeof(*group), GFP_KERNEL);
1206 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
1207 if (!group || !domain) {
1212 group->iommu_group = iommu_group;
1214 /* Determine bus_type in order to allocate a domain */
1215 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
1219 mdev_bus = symbol_get(mdev_bus_type);
1222 if ((bus == mdev_bus) && !iommu_present(bus)) {
1223 symbol_put(mdev_bus_type);
1224 if (!iommu->external_domain) {
1225 INIT_LIST_HEAD(&domain->group_list);
1226 iommu->external_domain = domain;
1230 list_add(&group->next,
1231 &iommu->external_domain->group_list);
1232 mutex_unlock(&iommu->lock);
1235 symbol_put(mdev_bus_type);
1238 domain->domain = iommu_domain_alloc(bus);
1239 if (!domain->domain) {
1244 if (iommu->nesting) {
1247 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
1253 ret = iommu_attach_group(domain->domain, iommu_group);
1257 INIT_LIST_HEAD(&domain->group_list);
1258 list_add(&group->next, &domain->group_list);
1260 if (!allow_unsafe_interrupts &&
1261 !iommu_capable(bus, IOMMU_CAP_INTR_REMAP)) {
1262 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
1268 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
1269 domain->prot |= IOMMU_CACHE;
1272 * Try to match an existing compatible domain. We don't want to
1273 * preclude an IOMMU driver supporting multiple bus_types and being
1274 * able to include different bus_types in the same IOMMU domain, so
1275 * we test whether the domains use the same iommu_ops rather than
1276 * testing if they're on the same bus_type.
1278 list_for_each_entry(d, &iommu->domain_list, next) {
1279 if (d->domain->ops == domain->domain->ops &&
1280 d->prot == domain->prot) {
1281 iommu_detach_group(domain->domain, iommu_group);
1282 if (!iommu_attach_group(d->domain, iommu_group)) {
1283 list_add(&group->next, &d->group_list);
1284 iommu_domain_free(domain->domain);
1286 mutex_unlock(&iommu->lock);
1290 ret = iommu_attach_group(domain->domain, iommu_group);
1296 vfio_test_domain_fgsp(domain);
1298 /* replay mappings on new domains */
1299 ret = vfio_iommu_replay(iommu, domain);
1303 list_add(&domain->next, &iommu->domain_list);
1305 mutex_unlock(&iommu->lock);
1310 iommu_detach_group(domain->domain, iommu_group);
1312 iommu_domain_free(domain->domain);
1316 mutex_unlock(&iommu->lock);
1320 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
1322 struct rb_node *node;
1324 while ((node = rb_first(&iommu->dma_list)))
1325 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
1328 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
1330 struct rb_node *n, *p;
1332 n = rb_first(&iommu->dma_list);
1333 for (; n; n = rb_next(n)) {
1334 struct vfio_dma *dma;
1335 long locked = 0, unlocked = 0;
1337 dma = rb_entry(n, struct vfio_dma, node);
1338 unlocked += vfio_unmap_unpin(iommu, dma, false);
1339 p = rb_first(&dma->pfn_list);
1340 for (; p; p = rb_next(p)) {
1341 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
1344 if (!is_invalid_reserved_pfn(vpfn->pfn))
1347 vfio_lock_acct(dma->task, locked - unlocked);
1351 static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu)
1355 n = rb_first(&iommu->dma_list);
1356 for (; n; n = rb_next(n)) {
1357 struct vfio_dma *dma;
1359 dma = rb_entry(n, struct vfio_dma, node);
1361 if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list)))
1364 /* mdev vendor driver must unregister notifier */
1365 WARN_ON(iommu->notifier.head);
1368 static void vfio_iommu_type1_detach_group(void *iommu_data,
1369 struct iommu_group *iommu_group)
1371 struct vfio_iommu *iommu = iommu_data;
1372 struct vfio_domain *domain;
1373 struct vfio_group *group;
1375 mutex_lock(&iommu->lock);
1377 if (iommu->external_domain) {
1378 group = find_iommu_group(iommu->external_domain, iommu_group);
1380 list_del(&group->next);
1383 if (list_empty(&iommu->external_domain->group_list)) {
1384 vfio_sanity_check_pfn_list(iommu);
1386 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1387 vfio_iommu_unmap_unpin_all(iommu);
1389 kfree(iommu->external_domain);
1390 iommu->external_domain = NULL;
1392 goto detach_group_done;
1396 list_for_each_entry(domain, &iommu->domain_list, next) {
1397 group = find_iommu_group(domain, iommu_group);
1401 iommu_detach_group(domain->domain, iommu_group);
1402 list_del(&group->next);
1405 * Group ownership provides privilege, if the group list is
1406 * empty, the domain goes away. If it's the last domain with
1407 * iommu and external domain doesn't exist, then all the
1408 * mappings go away too. If it's the last domain with iommu and
1409 * external domain exist, update accounting
1411 if (list_empty(&domain->group_list)) {
1412 if (list_is_singular(&iommu->domain_list)) {
1413 if (!iommu->external_domain)
1414 vfio_iommu_unmap_unpin_all(iommu);
1416 vfio_iommu_unmap_unpin_reaccount(iommu);
1418 iommu_domain_free(domain->domain);
1419 list_del(&domain->next);
1426 mutex_unlock(&iommu->lock);
1429 static void *vfio_iommu_type1_open(unsigned long arg)
1431 struct vfio_iommu *iommu;
1433 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
1435 return ERR_PTR(-ENOMEM);
1438 case VFIO_TYPE1_IOMMU:
1440 case VFIO_TYPE1_NESTING_IOMMU:
1441 iommu->nesting = true;
1442 case VFIO_TYPE1v2_IOMMU:
1447 return ERR_PTR(-EINVAL);
1450 INIT_LIST_HEAD(&iommu->domain_list);
1451 iommu->dma_list = RB_ROOT;
1452 mutex_init(&iommu->lock);
1453 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
1458 static void vfio_release_domain(struct vfio_domain *domain, bool external)
1460 struct vfio_group *group, *group_tmp;
1462 list_for_each_entry_safe(group, group_tmp,
1463 &domain->group_list, next) {
1465 iommu_detach_group(domain->domain, group->iommu_group);
1466 list_del(&group->next);
1471 iommu_domain_free(domain->domain);
1474 static void vfio_iommu_type1_release(void *iommu_data)
1476 struct vfio_iommu *iommu = iommu_data;
1477 struct vfio_domain *domain, *domain_tmp;
1479 if (iommu->external_domain) {
1480 vfio_release_domain(iommu->external_domain, true);
1481 vfio_sanity_check_pfn_list(iommu);
1482 kfree(iommu->external_domain);
1485 vfio_iommu_unmap_unpin_all(iommu);
1487 list_for_each_entry_safe(domain, domain_tmp,
1488 &iommu->domain_list, next) {
1489 vfio_release_domain(domain, false);
1490 list_del(&domain->next);
1496 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
1498 struct vfio_domain *domain;
1501 mutex_lock(&iommu->lock);
1502 list_for_each_entry(domain, &iommu->domain_list, next) {
1503 if (!(domain->prot & IOMMU_CACHE)) {
1508 mutex_unlock(&iommu->lock);
1513 static long vfio_iommu_type1_ioctl(void *iommu_data,
1514 unsigned int cmd, unsigned long arg)
1516 struct vfio_iommu *iommu = iommu_data;
1517 unsigned long minsz;
1519 if (cmd == VFIO_CHECK_EXTENSION) {
1521 case VFIO_TYPE1_IOMMU:
1522 case VFIO_TYPE1v2_IOMMU:
1523 case VFIO_TYPE1_NESTING_IOMMU:
1525 case VFIO_DMA_CC_IOMMU:
1528 return vfio_domains_have_iommu_cache(iommu);
1532 } else if (cmd == VFIO_IOMMU_GET_INFO) {
1533 struct vfio_iommu_type1_info info;
1535 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
1537 if (copy_from_user(&info, (void __user *)arg, minsz))
1540 if (info.argsz < minsz)
1543 info.flags = VFIO_IOMMU_INFO_PGSIZES;
1545 info.iova_pgsizes = vfio_pgsize_bitmap(iommu);
1547 return copy_to_user((void __user *)arg, &info, minsz) ?
1550 } else if (cmd == VFIO_IOMMU_MAP_DMA) {
1551 struct vfio_iommu_type1_dma_map map;
1552 uint32_t mask = VFIO_DMA_MAP_FLAG_READ |
1553 VFIO_DMA_MAP_FLAG_WRITE;
1555 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
1557 if (copy_from_user(&map, (void __user *)arg, minsz))
1560 if (map.argsz < minsz || map.flags & ~mask)
1563 return vfio_dma_do_map(iommu, &map);
1565 } else if (cmd == VFIO_IOMMU_UNMAP_DMA) {
1566 struct vfio_iommu_type1_dma_unmap unmap;
1569 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
1571 if (copy_from_user(&unmap, (void __user *)arg, minsz))
1574 if (unmap.argsz < minsz || unmap.flags)
1577 ret = vfio_dma_do_unmap(iommu, &unmap);
1581 return copy_to_user((void __user *)arg, &unmap, minsz) ?
1588 static int vfio_iommu_type1_register_notifier(void *iommu_data,
1589 unsigned long *events,
1590 struct notifier_block *nb)
1592 struct vfio_iommu *iommu = iommu_data;
1594 /* clear known events */
1595 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
1597 /* refuse to register if still events remaining */
1601 return blocking_notifier_chain_register(&iommu->notifier, nb);
1604 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
1605 struct notifier_block *nb)
1607 struct vfio_iommu *iommu = iommu_data;
1609 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
1612 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
1613 .name = "vfio-iommu-type1",
1614 .owner = THIS_MODULE,
1615 .open = vfio_iommu_type1_open,
1616 .release = vfio_iommu_type1_release,
1617 .ioctl = vfio_iommu_type1_ioctl,
1618 .attach_group = vfio_iommu_type1_attach_group,
1619 .detach_group = vfio_iommu_type1_detach_group,
1620 .pin_pages = vfio_iommu_type1_pin_pages,
1621 .unpin_pages = vfio_iommu_type1_unpin_pages,
1622 .register_notifier = vfio_iommu_type1_register_notifier,
1623 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
1626 static int __init vfio_iommu_type1_init(void)
1628 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
1631 static void __exit vfio_iommu_type1_cleanup(void)
1633 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
1636 module_init(vfio_iommu_type1_init);
1637 module_exit(vfio_iommu_type1_cleanup);
1639 MODULE_VERSION(DRIVER_VERSION);
1640 MODULE_LICENSE("GPL v2");
1641 MODULE_AUTHOR(DRIVER_AUTHOR);
1642 MODULE_DESCRIPTION(DRIVER_DESC);