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>
42 #include <linux/dma-iommu.h>
43 #include <linux/irqdomain.h>
45 #define DRIVER_VERSION "0.2"
46 #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
47 #define DRIVER_DESC "Type1 IOMMU driver for VFIO"
49 static bool allow_unsafe_interrupts;
50 module_param_named(allow_unsafe_interrupts,
51 allow_unsafe_interrupts, bool, S_IRUGO | S_IWUSR);
52 MODULE_PARM_DESC(allow_unsafe_interrupts,
53 "Enable VFIO IOMMU support for on platforms without interrupt remapping support.");
55 static bool disable_hugepages;
56 module_param_named(disable_hugepages,
57 disable_hugepages, bool, S_IRUGO | S_IWUSR);
58 MODULE_PARM_DESC(disable_hugepages,
59 "Disable VFIO IOMMU support for IOMMU hugepages.");
62 struct list_head domain_list;
63 struct vfio_domain *external_domain; /* domain for external user */
65 struct rb_root dma_list;
66 struct blocking_notifier_head notifier;
72 struct iommu_domain *domain;
73 struct list_head next;
74 struct list_head group_list;
75 int prot; /* IOMMU_CACHE */
76 bool fgsp; /* Fine-grained super pages */
81 dma_addr_t iova; /* Device address */
82 unsigned long vaddr; /* Process virtual addr */
83 size_t size; /* Map size (bytes) */
84 int prot; /* IOMMU_READ/WRITE */
86 struct task_struct *task;
87 struct rb_root pfn_list; /* Ex-user pinned pfn list */
91 struct iommu_group *iommu_group;
92 struct list_head next;
96 * Guest RAM pinning working set or DMA target
100 dma_addr_t iova; /* Device address */
101 unsigned long pfn; /* Host pfn */
105 #define IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu) \
106 (!list_empty(&iommu->domain_list))
108 static int put_pfn(unsigned long pfn, int prot);
111 * This code handles mapping and unmapping of user data buffers
112 * into DMA'ble space using the IOMMU
115 static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
116 dma_addr_t start, size_t size)
118 struct rb_node *node = iommu->dma_list.rb_node;
121 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
123 if (start + size <= dma->iova)
124 node = node->rb_left;
125 else if (start >= dma->iova + dma->size)
126 node = node->rb_right;
134 static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new)
136 struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL;
137 struct vfio_dma *dma;
141 dma = rb_entry(parent, struct vfio_dma, node);
143 if (new->iova + new->size <= dma->iova)
144 link = &(*link)->rb_left;
146 link = &(*link)->rb_right;
149 rb_link_node(&new->node, parent, link);
150 rb_insert_color(&new->node, &iommu->dma_list);
153 static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
155 rb_erase(&old->node, &iommu->dma_list);
159 * Helper Functions for host iova-pfn list
161 static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
163 struct vfio_pfn *vpfn;
164 struct rb_node *node = dma->pfn_list.rb_node;
167 vpfn = rb_entry(node, struct vfio_pfn, node);
169 if (iova < vpfn->iova)
170 node = node->rb_left;
171 else if (iova > vpfn->iova)
172 node = node->rb_right;
179 static void vfio_link_pfn(struct vfio_dma *dma,
180 struct vfio_pfn *new)
182 struct rb_node **link, *parent = NULL;
183 struct vfio_pfn *vpfn;
185 link = &dma->pfn_list.rb_node;
188 vpfn = rb_entry(parent, struct vfio_pfn, node);
190 if (new->iova < vpfn->iova)
191 link = &(*link)->rb_left;
193 link = &(*link)->rb_right;
196 rb_link_node(&new->node, parent, link);
197 rb_insert_color(&new->node, &dma->pfn_list);
200 static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old)
202 rb_erase(&old->node, &dma->pfn_list);
205 static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova,
208 struct vfio_pfn *vpfn;
210 vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL);
216 atomic_set(&vpfn->ref_count, 1);
217 vfio_link_pfn(dma, vpfn);
221 static void vfio_remove_from_pfn_list(struct vfio_dma *dma,
222 struct vfio_pfn *vpfn)
224 vfio_unlink_pfn(dma, vpfn);
228 static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma,
231 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
234 atomic_inc(&vpfn->ref_count);
238 static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn)
242 if (atomic_dec_and_test(&vpfn->ref_count)) {
243 ret = put_pfn(vpfn->pfn, dma->prot);
244 vfio_remove_from_pfn_list(dma, vpfn);
250 struct mm_struct *mm;
252 struct work_struct work;
255 /* delayed decrement/increment for locked_vm */
256 static void vfio_lock_acct_bg(struct work_struct *work)
258 struct vwork *vwork = container_of(work, struct vwork, work);
259 struct mm_struct *mm;
262 down_write(&mm->mmap_sem);
263 mm->locked_vm += vwork->npage;
264 up_write(&mm->mmap_sem);
269 static void vfio_lock_acct(struct task_struct *task, long npage)
272 struct mm_struct *mm;
278 is_current = (task->mm == current->mm);
280 mm = is_current ? task->mm : get_task_mm(task);
282 return; /* process exited */
284 if (down_write_trylock(&mm->mmap_sem)) {
285 mm->locked_vm += npage;
286 up_write(&mm->mmap_sem);
293 mm = get_task_mm(task);
299 * Couldn't get mmap_sem lock, so must setup to update
300 * mm->locked_vm later. If locked_vm were atomic, we
301 * wouldn't need this silliness
303 vwork = kmalloc(sizeof(struct vwork), GFP_KERNEL);
304 if (WARN_ON(!vwork)) {
308 INIT_WORK(&vwork->work, vfio_lock_acct_bg);
310 vwork->npage = npage;
311 schedule_work(&vwork->work);
315 * Some mappings aren't backed by a struct page, for example an mmap'd
316 * MMIO range for our own or another device. These use a different
317 * pfn conversion and shouldn't be tracked as locked pages.
319 static bool is_invalid_reserved_pfn(unsigned long pfn)
321 if (pfn_valid(pfn)) {
323 struct page *tail = pfn_to_page(pfn);
324 struct page *head = compound_head(tail);
325 reserved = !!(PageReserved(head));
328 * "head" is not a dangling pointer
329 * (compound_head takes care of that)
330 * but the hugepage may have been split
331 * from under us (and we may not hold a
332 * reference count on the head page so it can
333 * be reused before we run PageReferenced), so
334 * we've to check PageTail before returning
341 return PageReserved(tail);
347 static int put_pfn(unsigned long pfn, int prot)
349 if (!is_invalid_reserved_pfn(pfn)) {
350 struct page *page = pfn_to_page(pfn);
351 if (prot & IOMMU_WRITE)
359 static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,
360 int prot, unsigned long *pfn)
362 struct page *page[1];
363 struct vm_area_struct *vma;
366 if (mm == current->mm) {
367 ret = get_user_pages_fast(vaddr, 1, !!(prot & IOMMU_WRITE),
370 unsigned int flags = 0;
372 if (prot & IOMMU_WRITE)
375 down_read(&mm->mmap_sem);
376 ret = get_user_pages_remote(NULL, mm, vaddr, 1, flags, page,
378 up_read(&mm->mmap_sem);
382 *pfn = page_to_pfn(page[0]);
386 down_read(&mm->mmap_sem);
388 vma = find_vma_intersection(mm, vaddr, vaddr + 1);
390 if (vma && vma->vm_flags & VM_PFNMAP) {
391 *pfn = ((vaddr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
392 if (is_invalid_reserved_pfn(*pfn))
396 up_read(&mm->mmap_sem);
401 * Attempt to pin pages. We really don't want to track all the pfns and
402 * the iommu can only map chunks of consecutive pfns anyway, so get the
403 * first page and all consecutive pages with the same locking.
405 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
406 long npage, unsigned long *pfn_base)
408 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
409 bool lock_cap = capable(CAP_IPC_LOCK);
410 long ret, pinned = 0, lock_acct = 0;
412 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
414 /* This code path is only user initiated */
418 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base);
423 rsvd = is_invalid_reserved_pfn(*pfn_base);
426 * Reserved pages aren't counted against the user, externally pinned
427 * pages are already counted against the user.
429 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
430 if (!lock_cap && current->mm->locked_vm + 1 > limit) {
431 put_pfn(*pfn_base, dma->prot);
432 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__,
433 limit << PAGE_SHIFT);
439 if (unlikely(disable_hugepages))
442 /* Lock all the consecutive pages from pfn_base */
443 for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage;
444 pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) {
445 unsigned long pfn = 0;
447 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn);
451 if (pfn != *pfn_base + pinned ||
452 rsvd != is_invalid_reserved_pfn(pfn)) {
453 put_pfn(pfn, dma->prot);
457 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
459 current->mm->locked_vm + lock_acct + 1 > limit) {
460 put_pfn(pfn, dma->prot);
461 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
462 __func__, limit << PAGE_SHIFT);
470 vfio_lock_acct(current, lock_acct);
475 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
476 unsigned long pfn, long npage,
479 long unlocked = 0, locked = 0;
482 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
483 if (put_pfn(pfn++, dma->prot)) {
485 if (vfio_find_vpfn(dma, iova))
491 vfio_lock_acct(dma->task, locked - unlocked);
496 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
497 unsigned long *pfn_base, bool do_accounting)
500 bool lock_cap = ns_capable(task_active_pid_ns(dma->task)->user_ns,
502 struct mm_struct *mm;
506 mm = get_task_mm(dma->task);
510 ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base);
514 rsvd = is_invalid_reserved_pfn(*pfn_base);
515 limit = task_rlimit(dma->task, RLIMIT_MEMLOCK) >> PAGE_SHIFT;
517 if (!rsvd && !lock_cap && mm->locked_vm + 1 > limit) {
518 put_pfn(*pfn_base, dma->prot);
519 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK (%ld) exceeded\n",
520 __func__, dma->task->comm, task_pid_nr(dma->task),
521 limit << PAGE_SHIFT);
526 if (!rsvd && do_accounting)
527 vfio_lock_acct(dma->task, 1);
535 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
539 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
544 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
547 vfio_lock_acct(dma->task, -unlocked);
552 static int vfio_iommu_type1_pin_pages(void *iommu_data,
553 unsigned long *user_pfn,
555 unsigned long *phys_pfn)
557 struct vfio_iommu *iommu = iommu_data;
559 unsigned long remote_vaddr;
560 struct vfio_dma *dma;
563 if (!iommu || !user_pfn || !phys_pfn)
566 /* Supported for v2 version only */
570 mutex_lock(&iommu->lock);
572 /* Fail if notifier list is empty */
573 if ((!iommu->external_domain) || (!iommu->notifier.head)) {
579 * If iommu capable domain exist in the container then all pages are
580 * already pinned and accounted. Accouting should be done if there is no
581 * iommu capable domain in the container.
583 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
585 for (i = 0; i < npage; i++) {
587 struct vfio_pfn *vpfn;
589 iova = user_pfn[i] << PAGE_SHIFT;
590 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
596 if ((dma->prot & prot) != prot) {
601 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
603 phys_pfn[i] = vpfn->pfn;
607 remote_vaddr = dma->vaddr + iova - dma->iova;
608 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
615 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
617 vfio_unpin_page_external(dma, iova, do_accounting);
627 for (j = 0; j < i; j++) {
630 iova = user_pfn[j] << PAGE_SHIFT;
631 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
632 vfio_unpin_page_external(dma, iova, do_accounting);
636 mutex_unlock(&iommu->lock);
640 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
641 unsigned long *user_pfn,
644 struct vfio_iommu *iommu = iommu_data;
648 if (!iommu || !user_pfn)
651 /* Supported for v2 version only */
655 mutex_lock(&iommu->lock);
657 if (!iommu->external_domain) {
658 mutex_unlock(&iommu->lock);
662 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
663 for (i = 0; i < npage; i++) {
664 struct vfio_dma *dma;
667 iova = user_pfn[i] << PAGE_SHIFT;
668 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
671 vfio_unpin_page_external(dma, iova, do_accounting);
675 mutex_unlock(&iommu->lock);
676 return i > npage ? npage : (i > 0 ? i : -EINVAL);
679 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
682 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
683 struct vfio_domain *domain, *d;
689 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
693 * We use the IOMMU to track the physical addresses, otherwise we'd
694 * need a much more complicated tracking system. Unfortunately that
695 * means we need to use one of the iommu domains to figure out the
696 * pfns to unpin. The rest need to be unmapped in advance so we have
697 * no iommu translations remaining when the pages are unpinned.
699 domain = d = list_first_entry(&iommu->domain_list,
700 struct vfio_domain, next);
702 list_for_each_entry_continue(d, &iommu->domain_list, next) {
703 iommu_unmap(d->domain, dma->iova, dma->size);
708 size_t unmapped, len;
709 phys_addr_t phys, next;
711 phys = iommu_iova_to_phys(domain->domain, iova);
712 if (WARN_ON(!phys)) {
718 * To optimize for fewer iommu_unmap() calls, each of which
719 * may require hardware cache flushing, try to find the
720 * largest contiguous physical memory chunk to unmap.
722 for (len = PAGE_SIZE;
723 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
724 next = iommu_iova_to_phys(domain->domain, iova + len);
725 if (next != phys + len)
729 unmapped = iommu_unmap(domain->domain, iova, len);
730 if (WARN_ON(!unmapped))
733 unlocked += vfio_unpin_pages_remote(dma, iova,
735 unmapped >> PAGE_SHIFT,
742 dma->iommu_mapped = false;
744 vfio_lock_acct(dma->task, -unlocked);
750 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
752 vfio_unmap_unpin(iommu, dma, true);
753 vfio_unlink_dma(iommu, dma);
754 put_task_struct(dma->task);
758 static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu)
760 struct vfio_domain *domain;
761 unsigned long bitmap = ULONG_MAX;
763 mutex_lock(&iommu->lock);
764 list_for_each_entry(domain, &iommu->domain_list, next)
765 bitmap &= domain->domain->pgsize_bitmap;
766 mutex_unlock(&iommu->lock);
769 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
770 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
771 * That way the user will be able to map/unmap buffers whose size/
772 * start address is aligned with PAGE_SIZE. Pinning code uses that
773 * granularity while iommu driver can use the sub-PAGE_SIZE size
776 if (bitmap & ~PAGE_MASK) {
784 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
785 struct vfio_iommu_type1_dma_unmap *unmap)
788 struct vfio_dma *dma, *dma_last = NULL;
790 int ret = 0, retries = 0;
792 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
794 if (unmap->iova & mask)
796 if (!unmap->size || unmap->size & mask)
799 WARN_ON(mask & PAGE_MASK);
801 mutex_lock(&iommu->lock);
804 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
805 * avoid tracking individual mappings. This means that the granularity
806 * of the original mapping was lost and the user was allowed to attempt
807 * to unmap any range. Depending on the contiguousness of physical
808 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
809 * or may not have worked. We only guaranteed unmap granularity
810 * matching the original mapping; even though it was untracked here,
811 * the original mappings are reflected in IOMMU mappings. This
812 * resulted in a couple unusual behaviors. First, if a range is not
813 * able to be unmapped, ex. a set of 4k pages that was mapped as a
814 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
815 * a zero sized unmap. Also, if an unmap request overlaps the first
816 * address of a hugepage, the IOMMU will unmap the entire hugepage.
817 * This also returns success and the returned unmap size reflects the
818 * actual size unmapped.
820 * We attempt to maintain compatibility with this "v1" interface, but
821 * we take control out of the hands of the IOMMU. Therefore, an unmap
822 * request offset from the beginning of the original mapping will
823 * return success with zero sized unmap. And an unmap request covering
824 * the first iova of mapping will unmap the entire range.
826 * The v2 version of this interface intends to be more deterministic.
827 * Unmap requests must fully cover previous mappings. Multiple
828 * mappings may still be unmaped by specifying large ranges, but there
829 * must not be any previous mappings bisected by the range. An error
830 * will be returned if these conditions are not met. The v2 interface
831 * will only return success and a size of zero if there were no
832 * mappings within the range.
835 dma = vfio_find_dma(iommu, unmap->iova, 1);
836 if (dma && dma->iova != unmap->iova) {
840 dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0);
841 if (dma && dma->iova + dma->size != unmap->iova + unmap->size) {
847 while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
848 if (!iommu->v2 && unmap->iova > dma->iova)
851 * Task with same address space who mapped this iova range is
852 * allowed to unmap the iova range.
854 if (dma->task->mm != current->mm)
857 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
858 struct vfio_iommu_type1_dma_unmap nb_unmap;
860 if (dma_last == dma) {
861 BUG_ON(++retries > 10);
867 nb_unmap.iova = dma->iova;
868 nb_unmap.size = dma->size;
871 * Notify anyone (mdev vendor drivers) to invalidate and
872 * unmap iovas within the range we're about to unmap.
873 * Vendor drivers MUST unpin pages in response to an
876 mutex_unlock(&iommu->lock);
877 blocking_notifier_call_chain(&iommu->notifier,
878 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
882 unmapped += dma->size;
883 vfio_remove_dma(iommu, dma);
887 mutex_unlock(&iommu->lock);
889 /* Report how much was unmapped */
890 unmap->size = unmapped;
896 * Turns out AMD IOMMU has a page table bug where it won't map large pages
897 * to a region that previously mapped smaller pages. This should be fixed
898 * soon, so this is just a temporary workaround to break mappings down into
899 * PAGE_SIZE. Better to map smaller pages than nothing.
901 static int map_try_harder(struct vfio_domain *domain, dma_addr_t iova,
902 unsigned long pfn, long npage, int prot)
907 for (i = 0; i < npage; i++, pfn++, iova += PAGE_SIZE) {
908 ret = iommu_map(domain->domain, iova,
909 (phys_addr_t)pfn << PAGE_SHIFT,
910 PAGE_SIZE, prot | domain->prot);
915 for (; i < npage && i > 0; i--, iova -= PAGE_SIZE)
916 iommu_unmap(domain->domain, iova, PAGE_SIZE);
921 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
922 unsigned long pfn, long npage, int prot)
924 struct vfio_domain *d;
927 list_for_each_entry(d, &iommu->domain_list, next) {
928 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
929 npage << PAGE_SHIFT, prot | d->prot);
932 map_try_harder(d, iova, pfn, npage, prot))
942 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next)
943 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
948 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
951 dma_addr_t iova = dma->iova;
952 unsigned long vaddr = dma->vaddr;
953 size_t size = map_size;
959 /* Pin a contiguous chunk of memory */
960 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
961 size >> PAGE_SHIFT, &pfn);
969 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
972 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
977 size -= npage << PAGE_SHIFT;
978 dma->size += npage << PAGE_SHIFT;
981 dma->iommu_mapped = true;
984 vfio_remove_dma(iommu, dma);
989 static int vfio_dma_do_map(struct vfio_iommu *iommu,
990 struct vfio_iommu_type1_dma_map *map)
992 dma_addr_t iova = map->iova;
993 unsigned long vaddr = map->vaddr;
994 size_t size = map->size;
995 int ret = 0, prot = 0;
997 struct vfio_dma *dma;
999 /* Verify that none of our __u64 fields overflow */
1000 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1003 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
1005 WARN_ON(mask & PAGE_MASK);
1007 /* READ/WRITE from device perspective */
1008 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1009 prot |= IOMMU_WRITE;
1010 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1013 if (!prot || !size || (size | iova | vaddr) & mask)
1016 /* Don't allow IOVA or virtual address wrap */
1017 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr)
1020 mutex_lock(&iommu->lock);
1022 if (vfio_find_dma(iommu, iova, size)) {
1027 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1036 get_task_struct(current);
1037 dma->task = current;
1038 dma->pfn_list = RB_ROOT;
1040 /* Insert zero-sized and grow as we map chunks of it */
1041 vfio_link_dma(iommu, dma);
1043 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1044 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1047 ret = vfio_pin_map_dma(iommu, dma, size);
1050 mutex_unlock(&iommu->lock);
1054 static int vfio_bus_type(struct device *dev, void *data)
1056 struct bus_type **bus = data;
1058 if (*bus && *bus != dev->bus)
1066 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1067 struct vfio_domain *domain)
1069 struct vfio_domain *d;
1073 /* Arbitrarily pick the first domain in the list for lookups */
1074 d = list_first_entry(&iommu->domain_list, struct vfio_domain, next);
1075 n = rb_first(&iommu->dma_list);
1077 for (; n; n = rb_next(n)) {
1078 struct vfio_dma *dma;
1081 dma = rb_entry(n, struct vfio_dma, node);
1084 while (iova < dma->iova + dma->size) {
1088 if (dma->iommu_mapped) {
1092 phys = iommu_iova_to_phys(d->domain, iova);
1094 if (WARN_ON(!phys)) {
1102 while (i < dma->iova + dma->size &&
1103 p == iommu_iova_to_phys(d->domain, i)) {
1110 unsigned long vaddr = dma->vaddr +
1112 size_t n = dma->iova + dma->size - iova;
1115 npage = vfio_pin_pages_remote(dma, vaddr,
1124 phys = pfn << PAGE_SHIFT;
1125 size = npage << PAGE_SHIFT;
1128 ret = iommu_map(domain->domain, iova, phys,
1129 size, dma->prot | domain->prot);
1135 dma->iommu_mapped = true;
1141 * We change our unmap behavior slightly depending on whether the IOMMU
1142 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1143 * for practically any contiguous power-of-two mapping we give it. This means
1144 * we don't need to look for contiguous chunks ourselves to make unmapping
1145 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1146 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1147 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1148 * hugetlbfs is in use.
1150 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1153 int ret, order = get_order(PAGE_SIZE * 2);
1155 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1159 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1160 IOMMU_READ | IOMMU_WRITE | domain->prot);
1162 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1164 if (unmapped == PAGE_SIZE)
1165 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1167 domain->fgsp = true;
1170 __free_pages(pages, order);
1173 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1174 struct iommu_group *iommu_group)
1176 struct vfio_group *g;
1178 list_for_each_entry(g, &domain->group_list, next) {
1179 if (g->iommu_group == iommu_group)
1186 static bool vfio_iommu_has_resv_msi(struct iommu_group *group,
1189 struct list_head group_resv_regions;
1190 struct iommu_resv_region *region, *next;
1193 INIT_LIST_HEAD(&group_resv_regions);
1194 iommu_get_group_resv_regions(group, &group_resv_regions);
1195 list_for_each_entry(region, &group_resv_regions, list) {
1196 if (region->type & IOMMU_RESV_MSI) {
1197 *base = region->start;
1203 list_for_each_entry_safe(region, next, &group_resv_regions, list)
1208 static int vfio_iommu_type1_attach_group(void *iommu_data,
1209 struct iommu_group *iommu_group)
1211 struct vfio_iommu *iommu = iommu_data;
1212 struct vfio_group *group;
1213 struct vfio_domain *domain, *d;
1214 struct bus_type *bus = NULL, *mdev_bus;
1216 bool resv_msi, msi_remap;
1217 phys_addr_t resv_msi_base;
1219 mutex_lock(&iommu->lock);
1221 list_for_each_entry(d, &iommu->domain_list, next) {
1222 if (find_iommu_group(d, iommu_group)) {
1223 mutex_unlock(&iommu->lock);
1228 if (iommu->external_domain) {
1229 if (find_iommu_group(iommu->external_domain, iommu_group)) {
1230 mutex_unlock(&iommu->lock);
1235 group = kzalloc(sizeof(*group), GFP_KERNEL);
1236 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
1237 if (!group || !domain) {
1242 group->iommu_group = iommu_group;
1244 /* Determine bus_type in order to allocate a domain */
1245 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
1249 mdev_bus = symbol_get(mdev_bus_type);
1252 if ((bus == mdev_bus) && !iommu_present(bus)) {
1253 symbol_put(mdev_bus_type);
1254 if (!iommu->external_domain) {
1255 INIT_LIST_HEAD(&domain->group_list);
1256 iommu->external_domain = domain;
1260 list_add(&group->next,
1261 &iommu->external_domain->group_list);
1262 mutex_unlock(&iommu->lock);
1265 symbol_put(mdev_bus_type);
1268 domain->domain = iommu_domain_alloc(bus);
1269 if (!domain->domain) {
1274 if (iommu->nesting) {
1277 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
1283 ret = iommu_attach_group(domain->domain, iommu_group);
1287 resv_msi = vfio_iommu_has_resv_msi(iommu_group, &resv_msi_base);
1289 INIT_LIST_HEAD(&domain->group_list);
1290 list_add(&group->next, &domain->group_list);
1292 msi_remap = resv_msi ? irq_domain_check_msi_remap() :
1293 iommu_capable(bus, IOMMU_CAP_INTR_REMAP);
1295 if (!allow_unsafe_interrupts && !msi_remap) {
1296 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
1302 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
1303 domain->prot |= IOMMU_CACHE;
1306 * Try to match an existing compatible domain. We don't want to
1307 * preclude an IOMMU driver supporting multiple bus_types and being
1308 * able to include different bus_types in the same IOMMU domain, so
1309 * we test whether the domains use the same iommu_ops rather than
1310 * testing if they're on the same bus_type.
1312 list_for_each_entry(d, &iommu->domain_list, next) {
1313 if (d->domain->ops == domain->domain->ops &&
1314 d->prot == domain->prot) {
1315 iommu_detach_group(domain->domain, iommu_group);
1316 if (!iommu_attach_group(d->domain, iommu_group)) {
1317 list_add(&group->next, &d->group_list);
1318 iommu_domain_free(domain->domain);
1320 mutex_unlock(&iommu->lock);
1324 ret = iommu_attach_group(domain->domain, iommu_group);
1330 vfio_test_domain_fgsp(domain);
1332 /* replay mappings on new domains */
1333 ret = vfio_iommu_replay(iommu, domain);
1337 if (resv_msi && iommu_get_msi_cookie(domain->domain, resv_msi_base))
1340 list_add(&domain->next, &iommu->domain_list);
1342 mutex_unlock(&iommu->lock);
1347 iommu_detach_group(domain->domain, iommu_group);
1349 iommu_domain_free(domain->domain);
1353 mutex_unlock(&iommu->lock);
1357 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
1359 struct rb_node *node;
1361 while ((node = rb_first(&iommu->dma_list)))
1362 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
1365 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
1367 struct rb_node *n, *p;
1369 n = rb_first(&iommu->dma_list);
1370 for (; n; n = rb_next(n)) {
1371 struct vfio_dma *dma;
1372 long locked = 0, unlocked = 0;
1374 dma = rb_entry(n, struct vfio_dma, node);
1375 unlocked += vfio_unmap_unpin(iommu, dma, false);
1376 p = rb_first(&dma->pfn_list);
1377 for (; p; p = rb_next(p)) {
1378 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
1381 if (!is_invalid_reserved_pfn(vpfn->pfn))
1384 vfio_lock_acct(dma->task, locked - unlocked);
1388 static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu)
1392 n = rb_first(&iommu->dma_list);
1393 for (; n; n = rb_next(n)) {
1394 struct vfio_dma *dma;
1396 dma = rb_entry(n, struct vfio_dma, node);
1398 if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list)))
1401 /* mdev vendor driver must unregister notifier */
1402 WARN_ON(iommu->notifier.head);
1405 static void vfio_iommu_type1_detach_group(void *iommu_data,
1406 struct iommu_group *iommu_group)
1408 struct vfio_iommu *iommu = iommu_data;
1409 struct vfio_domain *domain;
1410 struct vfio_group *group;
1412 mutex_lock(&iommu->lock);
1414 if (iommu->external_domain) {
1415 group = find_iommu_group(iommu->external_domain, iommu_group);
1417 list_del(&group->next);
1420 if (list_empty(&iommu->external_domain->group_list)) {
1421 vfio_sanity_check_pfn_list(iommu);
1423 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1424 vfio_iommu_unmap_unpin_all(iommu);
1426 kfree(iommu->external_domain);
1427 iommu->external_domain = NULL;
1429 goto detach_group_done;
1433 list_for_each_entry(domain, &iommu->domain_list, next) {
1434 group = find_iommu_group(domain, iommu_group);
1438 iommu_detach_group(domain->domain, iommu_group);
1439 list_del(&group->next);
1442 * Group ownership provides privilege, if the group list is
1443 * empty, the domain goes away. If it's the last domain with
1444 * iommu and external domain doesn't exist, then all the
1445 * mappings go away too. If it's the last domain with iommu and
1446 * external domain exist, update accounting
1448 if (list_empty(&domain->group_list)) {
1449 if (list_is_singular(&iommu->domain_list)) {
1450 if (!iommu->external_domain)
1451 vfio_iommu_unmap_unpin_all(iommu);
1453 vfio_iommu_unmap_unpin_reaccount(iommu);
1455 iommu_domain_free(domain->domain);
1456 list_del(&domain->next);
1463 mutex_unlock(&iommu->lock);
1466 static void *vfio_iommu_type1_open(unsigned long arg)
1468 struct vfio_iommu *iommu;
1470 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
1472 return ERR_PTR(-ENOMEM);
1475 case VFIO_TYPE1_IOMMU:
1477 case VFIO_TYPE1_NESTING_IOMMU:
1478 iommu->nesting = true;
1479 case VFIO_TYPE1v2_IOMMU:
1484 return ERR_PTR(-EINVAL);
1487 INIT_LIST_HEAD(&iommu->domain_list);
1488 iommu->dma_list = RB_ROOT;
1489 mutex_init(&iommu->lock);
1490 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
1495 static void vfio_release_domain(struct vfio_domain *domain, bool external)
1497 struct vfio_group *group, *group_tmp;
1499 list_for_each_entry_safe(group, group_tmp,
1500 &domain->group_list, next) {
1502 iommu_detach_group(domain->domain, group->iommu_group);
1503 list_del(&group->next);
1508 iommu_domain_free(domain->domain);
1511 static void vfio_iommu_type1_release(void *iommu_data)
1513 struct vfio_iommu *iommu = iommu_data;
1514 struct vfio_domain *domain, *domain_tmp;
1516 if (iommu->external_domain) {
1517 vfio_release_domain(iommu->external_domain, true);
1518 vfio_sanity_check_pfn_list(iommu);
1519 kfree(iommu->external_domain);
1522 vfio_iommu_unmap_unpin_all(iommu);
1524 list_for_each_entry_safe(domain, domain_tmp,
1525 &iommu->domain_list, next) {
1526 vfio_release_domain(domain, false);
1527 list_del(&domain->next);
1533 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
1535 struct vfio_domain *domain;
1538 mutex_lock(&iommu->lock);
1539 list_for_each_entry(domain, &iommu->domain_list, next) {
1540 if (!(domain->prot & IOMMU_CACHE)) {
1545 mutex_unlock(&iommu->lock);
1550 static long vfio_iommu_type1_ioctl(void *iommu_data,
1551 unsigned int cmd, unsigned long arg)
1553 struct vfio_iommu *iommu = iommu_data;
1554 unsigned long minsz;
1556 if (cmd == VFIO_CHECK_EXTENSION) {
1558 case VFIO_TYPE1_IOMMU:
1559 case VFIO_TYPE1v2_IOMMU:
1560 case VFIO_TYPE1_NESTING_IOMMU:
1562 case VFIO_DMA_CC_IOMMU:
1565 return vfio_domains_have_iommu_cache(iommu);
1569 } else if (cmd == VFIO_IOMMU_GET_INFO) {
1570 struct vfio_iommu_type1_info info;
1572 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
1574 if (copy_from_user(&info, (void __user *)arg, minsz))
1577 if (info.argsz < minsz)
1580 info.flags = VFIO_IOMMU_INFO_PGSIZES;
1582 info.iova_pgsizes = vfio_pgsize_bitmap(iommu);
1584 return copy_to_user((void __user *)arg, &info, minsz) ?
1587 } else if (cmd == VFIO_IOMMU_MAP_DMA) {
1588 struct vfio_iommu_type1_dma_map map;
1589 uint32_t mask = VFIO_DMA_MAP_FLAG_READ |
1590 VFIO_DMA_MAP_FLAG_WRITE;
1592 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
1594 if (copy_from_user(&map, (void __user *)arg, minsz))
1597 if (map.argsz < minsz || map.flags & ~mask)
1600 return vfio_dma_do_map(iommu, &map);
1602 } else if (cmd == VFIO_IOMMU_UNMAP_DMA) {
1603 struct vfio_iommu_type1_dma_unmap unmap;
1606 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
1608 if (copy_from_user(&unmap, (void __user *)arg, minsz))
1611 if (unmap.argsz < minsz || unmap.flags)
1614 ret = vfio_dma_do_unmap(iommu, &unmap);
1618 return copy_to_user((void __user *)arg, &unmap, minsz) ?
1625 static int vfio_iommu_type1_register_notifier(void *iommu_data,
1626 unsigned long *events,
1627 struct notifier_block *nb)
1629 struct vfio_iommu *iommu = iommu_data;
1631 /* clear known events */
1632 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
1634 /* refuse to register if still events remaining */
1638 return blocking_notifier_chain_register(&iommu->notifier, nb);
1641 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
1642 struct notifier_block *nb)
1644 struct vfio_iommu *iommu = iommu_data;
1646 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
1649 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
1650 .name = "vfio-iommu-type1",
1651 .owner = THIS_MODULE,
1652 .open = vfio_iommu_type1_open,
1653 .release = vfio_iommu_type1_release,
1654 .ioctl = vfio_iommu_type1_ioctl,
1655 .attach_group = vfio_iommu_type1_attach_group,
1656 .detach_group = vfio_iommu_type1_detach_group,
1657 .pin_pages = vfio_iommu_type1_pin_pages,
1658 .unpin_pages = vfio_iommu_type1_unpin_pages,
1659 .register_notifier = vfio_iommu_type1_register_notifier,
1660 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
1663 static int __init vfio_iommu_type1_init(void)
1665 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
1668 static void __exit vfio_iommu_type1_cleanup(void)
1670 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
1673 module_init(vfio_iommu_type1_init);
1674 module_exit(vfio_iommu_type1_cleanup);
1676 MODULE_VERSION(DRIVER_VERSION);
1677 MODULE_LICENSE("GPL v2");
1678 MODULE_AUTHOR(DRIVER_AUTHOR);
1679 MODULE_DESCRIPTION(DRIVER_DESC);