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/signal.h>
35 #include <linux/sched/mm.h>
36 #include <linux/slab.h>
37 #include <linux/uaccess.h>
38 #include <linux/vfio.h>
39 #include <linux/workqueue.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 = has_capability(dma->task, CAP_IPC_LOCK);
501 struct mm_struct *mm;
505 mm = get_task_mm(dma->task);
509 ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base);
513 rsvd = is_invalid_reserved_pfn(*pfn_base);
514 limit = task_rlimit(dma->task, RLIMIT_MEMLOCK) >> PAGE_SHIFT;
516 if (!rsvd && !lock_cap && mm->locked_vm + 1 > limit) {
517 put_pfn(*pfn_base, dma->prot);
518 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK (%ld) exceeded\n",
519 __func__, dma->task->comm, task_pid_nr(dma->task),
520 limit << PAGE_SHIFT);
525 if (!rsvd && do_accounting)
526 vfio_lock_acct(dma->task, 1);
534 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
538 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
543 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
546 vfio_lock_acct(dma->task, -unlocked);
551 static int vfio_iommu_type1_pin_pages(void *iommu_data,
552 unsigned long *user_pfn,
554 unsigned long *phys_pfn)
556 struct vfio_iommu *iommu = iommu_data;
558 unsigned long remote_vaddr;
559 struct vfio_dma *dma;
562 if (!iommu || !user_pfn || !phys_pfn)
565 /* Supported for v2 version only */
569 mutex_lock(&iommu->lock);
571 /* Fail if notifier list is empty */
572 if ((!iommu->external_domain) || (!iommu->notifier.head)) {
578 * If iommu capable domain exist in the container then all pages are
579 * already pinned and accounted. Accouting should be done if there is no
580 * iommu capable domain in the container.
582 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
584 for (i = 0; i < npage; i++) {
586 struct vfio_pfn *vpfn;
588 iova = user_pfn[i] << PAGE_SHIFT;
589 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
595 if ((dma->prot & prot) != prot) {
600 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
602 phys_pfn[i] = vpfn->pfn;
606 remote_vaddr = dma->vaddr + iova - dma->iova;
607 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
614 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
616 vfio_unpin_page_external(dma, iova, do_accounting);
626 for (j = 0; j < i; j++) {
629 iova = user_pfn[j] << PAGE_SHIFT;
630 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
631 vfio_unpin_page_external(dma, iova, do_accounting);
635 mutex_unlock(&iommu->lock);
639 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
640 unsigned long *user_pfn,
643 struct vfio_iommu *iommu = iommu_data;
647 if (!iommu || !user_pfn)
650 /* Supported for v2 version only */
654 mutex_lock(&iommu->lock);
656 if (!iommu->external_domain) {
657 mutex_unlock(&iommu->lock);
661 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
662 for (i = 0; i < npage; i++) {
663 struct vfio_dma *dma;
666 iova = user_pfn[i] << PAGE_SHIFT;
667 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
670 vfio_unpin_page_external(dma, iova, do_accounting);
674 mutex_unlock(&iommu->lock);
675 return i > npage ? npage : (i > 0 ? i : -EINVAL);
678 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
681 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
682 struct vfio_domain *domain, *d;
688 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
692 * We use the IOMMU to track the physical addresses, otherwise we'd
693 * need a much more complicated tracking system. Unfortunately that
694 * means we need to use one of the iommu domains to figure out the
695 * pfns to unpin. The rest need to be unmapped in advance so we have
696 * no iommu translations remaining when the pages are unpinned.
698 domain = d = list_first_entry(&iommu->domain_list,
699 struct vfio_domain, next);
701 list_for_each_entry_continue(d, &iommu->domain_list, next) {
702 iommu_unmap(d->domain, dma->iova, dma->size);
707 size_t unmapped, len;
708 phys_addr_t phys, next;
710 phys = iommu_iova_to_phys(domain->domain, iova);
711 if (WARN_ON(!phys)) {
717 * To optimize for fewer iommu_unmap() calls, each of which
718 * may require hardware cache flushing, try to find the
719 * largest contiguous physical memory chunk to unmap.
721 for (len = PAGE_SIZE;
722 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
723 next = iommu_iova_to_phys(domain->domain, iova + len);
724 if (next != phys + len)
728 unmapped = iommu_unmap(domain->domain, iova, len);
729 if (WARN_ON(!unmapped))
732 unlocked += vfio_unpin_pages_remote(dma, iova,
734 unmapped >> PAGE_SHIFT,
741 dma->iommu_mapped = false;
743 vfio_lock_acct(dma->task, -unlocked);
749 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
751 vfio_unmap_unpin(iommu, dma, true);
752 vfio_unlink_dma(iommu, dma);
753 put_task_struct(dma->task);
757 static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu)
759 struct vfio_domain *domain;
760 unsigned long bitmap = ULONG_MAX;
762 mutex_lock(&iommu->lock);
763 list_for_each_entry(domain, &iommu->domain_list, next)
764 bitmap &= domain->domain->pgsize_bitmap;
765 mutex_unlock(&iommu->lock);
768 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
769 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
770 * That way the user will be able to map/unmap buffers whose size/
771 * start address is aligned with PAGE_SIZE. Pinning code uses that
772 * granularity while iommu driver can use the sub-PAGE_SIZE size
775 if (bitmap & ~PAGE_MASK) {
783 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
784 struct vfio_iommu_type1_dma_unmap *unmap)
787 struct vfio_dma *dma, *dma_last = NULL;
789 int ret = 0, retries = 0;
791 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
793 if (unmap->iova & mask)
795 if (!unmap->size || unmap->size & mask)
798 WARN_ON(mask & PAGE_MASK);
800 mutex_lock(&iommu->lock);
803 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
804 * avoid tracking individual mappings. This means that the granularity
805 * of the original mapping was lost and the user was allowed to attempt
806 * to unmap any range. Depending on the contiguousness of physical
807 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
808 * or may not have worked. We only guaranteed unmap granularity
809 * matching the original mapping; even though it was untracked here,
810 * the original mappings are reflected in IOMMU mappings. This
811 * resulted in a couple unusual behaviors. First, if a range is not
812 * able to be unmapped, ex. a set of 4k pages that was mapped as a
813 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
814 * a zero sized unmap. Also, if an unmap request overlaps the first
815 * address of a hugepage, the IOMMU will unmap the entire hugepage.
816 * This also returns success and the returned unmap size reflects the
817 * actual size unmapped.
819 * We attempt to maintain compatibility with this "v1" interface, but
820 * we take control out of the hands of the IOMMU. Therefore, an unmap
821 * request offset from the beginning of the original mapping will
822 * return success with zero sized unmap. And an unmap request covering
823 * the first iova of mapping will unmap the entire range.
825 * The v2 version of this interface intends to be more deterministic.
826 * Unmap requests must fully cover previous mappings. Multiple
827 * mappings may still be unmaped by specifying large ranges, but there
828 * must not be any previous mappings bisected by the range. An error
829 * will be returned if these conditions are not met. The v2 interface
830 * will only return success and a size of zero if there were no
831 * mappings within the range.
834 dma = vfio_find_dma(iommu, unmap->iova, 1);
835 if (dma && dma->iova != unmap->iova) {
839 dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0);
840 if (dma && dma->iova + dma->size != unmap->iova + unmap->size) {
846 while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
847 if (!iommu->v2 && unmap->iova > dma->iova)
850 * Task with same address space who mapped this iova range is
851 * allowed to unmap the iova range.
853 if (dma->task->mm != current->mm)
856 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
857 struct vfio_iommu_type1_dma_unmap nb_unmap;
859 if (dma_last == dma) {
860 BUG_ON(++retries > 10);
866 nb_unmap.iova = dma->iova;
867 nb_unmap.size = dma->size;
870 * Notify anyone (mdev vendor drivers) to invalidate and
871 * unmap iovas within the range we're about to unmap.
872 * Vendor drivers MUST unpin pages in response to an
875 mutex_unlock(&iommu->lock);
876 blocking_notifier_call_chain(&iommu->notifier,
877 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
881 unmapped += dma->size;
882 vfio_remove_dma(iommu, dma);
886 mutex_unlock(&iommu->lock);
888 /* Report how much was unmapped */
889 unmap->size = unmapped;
895 * Turns out AMD IOMMU has a page table bug where it won't map large pages
896 * to a region that previously mapped smaller pages. This should be fixed
897 * soon, so this is just a temporary workaround to break mappings down into
898 * PAGE_SIZE. Better to map smaller pages than nothing.
900 static int map_try_harder(struct vfio_domain *domain, dma_addr_t iova,
901 unsigned long pfn, long npage, int prot)
906 for (i = 0; i < npage; i++, pfn++, iova += PAGE_SIZE) {
907 ret = iommu_map(domain->domain, iova,
908 (phys_addr_t)pfn << PAGE_SHIFT,
909 PAGE_SIZE, prot | domain->prot);
914 for (; i < npage && i > 0; i--, iova -= PAGE_SIZE)
915 iommu_unmap(domain->domain, iova, PAGE_SIZE);
920 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
921 unsigned long pfn, long npage, int prot)
923 struct vfio_domain *d;
926 list_for_each_entry(d, &iommu->domain_list, next) {
927 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
928 npage << PAGE_SHIFT, prot | d->prot);
931 map_try_harder(d, iova, pfn, npage, prot))
941 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next)
942 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
947 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
950 dma_addr_t iova = dma->iova;
951 unsigned long vaddr = dma->vaddr;
952 size_t size = map_size;
958 /* Pin a contiguous chunk of memory */
959 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
960 size >> PAGE_SHIFT, &pfn);
968 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
971 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
976 size -= npage << PAGE_SHIFT;
977 dma->size += npage << PAGE_SHIFT;
980 dma->iommu_mapped = true;
983 vfio_remove_dma(iommu, dma);
988 static int vfio_dma_do_map(struct vfio_iommu *iommu,
989 struct vfio_iommu_type1_dma_map *map)
991 dma_addr_t iova = map->iova;
992 unsigned long vaddr = map->vaddr;
993 size_t size = map->size;
994 int ret = 0, prot = 0;
996 struct vfio_dma *dma;
998 /* Verify that none of our __u64 fields overflow */
999 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1002 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
1004 WARN_ON(mask & PAGE_MASK);
1006 /* READ/WRITE from device perspective */
1007 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1008 prot |= IOMMU_WRITE;
1009 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1012 if (!prot || !size || (size | iova | vaddr) & mask)
1015 /* Don't allow IOVA or virtual address wrap */
1016 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr)
1019 mutex_lock(&iommu->lock);
1021 if (vfio_find_dma(iommu, iova, size)) {
1026 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1035 get_task_struct(current);
1036 dma->task = current;
1037 dma->pfn_list = RB_ROOT;
1039 /* Insert zero-sized and grow as we map chunks of it */
1040 vfio_link_dma(iommu, dma);
1042 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1043 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1046 ret = vfio_pin_map_dma(iommu, dma, size);
1049 mutex_unlock(&iommu->lock);
1053 static int vfio_bus_type(struct device *dev, void *data)
1055 struct bus_type **bus = data;
1057 if (*bus && *bus != dev->bus)
1065 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1066 struct vfio_domain *domain)
1068 struct vfio_domain *d;
1072 /* Arbitrarily pick the first domain in the list for lookups */
1073 d = list_first_entry(&iommu->domain_list, struct vfio_domain, next);
1074 n = rb_first(&iommu->dma_list);
1076 for (; n; n = rb_next(n)) {
1077 struct vfio_dma *dma;
1080 dma = rb_entry(n, struct vfio_dma, node);
1083 while (iova < dma->iova + dma->size) {
1087 if (dma->iommu_mapped) {
1091 phys = iommu_iova_to_phys(d->domain, iova);
1093 if (WARN_ON(!phys)) {
1101 while (i < dma->iova + dma->size &&
1102 p == iommu_iova_to_phys(d->domain, i)) {
1109 unsigned long vaddr = dma->vaddr +
1111 size_t n = dma->iova + dma->size - iova;
1114 npage = vfio_pin_pages_remote(dma, vaddr,
1123 phys = pfn << PAGE_SHIFT;
1124 size = npage << PAGE_SHIFT;
1127 ret = iommu_map(domain->domain, iova, phys,
1128 size, dma->prot | domain->prot);
1134 dma->iommu_mapped = true;
1140 * We change our unmap behavior slightly depending on whether the IOMMU
1141 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1142 * for practically any contiguous power-of-two mapping we give it. This means
1143 * we don't need to look for contiguous chunks ourselves to make unmapping
1144 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1145 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1146 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1147 * hugetlbfs is in use.
1149 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1152 int ret, order = get_order(PAGE_SIZE * 2);
1154 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1158 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1159 IOMMU_READ | IOMMU_WRITE | domain->prot);
1161 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1163 if (unmapped == PAGE_SIZE)
1164 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1166 domain->fgsp = true;
1169 __free_pages(pages, order);
1172 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1173 struct iommu_group *iommu_group)
1175 struct vfio_group *g;
1177 list_for_each_entry(g, &domain->group_list, next) {
1178 if (g->iommu_group == iommu_group)
1185 static bool vfio_iommu_has_resv_msi(struct iommu_group *group,
1188 struct list_head group_resv_regions;
1189 struct iommu_resv_region *region, *next;
1192 INIT_LIST_HEAD(&group_resv_regions);
1193 iommu_get_group_resv_regions(group, &group_resv_regions);
1194 list_for_each_entry(region, &group_resv_regions, list) {
1195 if (region->type & IOMMU_RESV_MSI) {
1196 *base = region->start;
1202 list_for_each_entry_safe(region, next, &group_resv_regions, list)
1207 static int vfio_iommu_type1_attach_group(void *iommu_data,
1208 struct iommu_group *iommu_group)
1210 struct vfio_iommu *iommu = iommu_data;
1211 struct vfio_group *group;
1212 struct vfio_domain *domain, *d;
1213 struct bus_type *bus = NULL, *mdev_bus;
1215 bool resv_msi, msi_remap;
1216 phys_addr_t resv_msi_base;
1218 mutex_lock(&iommu->lock);
1220 list_for_each_entry(d, &iommu->domain_list, next) {
1221 if (find_iommu_group(d, iommu_group)) {
1222 mutex_unlock(&iommu->lock);
1227 if (iommu->external_domain) {
1228 if (find_iommu_group(iommu->external_domain, iommu_group)) {
1229 mutex_unlock(&iommu->lock);
1234 group = kzalloc(sizeof(*group), GFP_KERNEL);
1235 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
1236 if (!group || !domain) {
1241 group->iommu_group = iommu_group;
1243 /* Determine bus_type in order to allocate a domain */
1244 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
1248 mdev_bus = symbol_get(mdev_bus_type);
1251 if ((bus == mdev_bus) && !iommu_present(bus)) {
1252 symbol_put(mdev_bus_type);
1253 if (!iommu->external_domain) {
1254 INIT_LIST_HEAD(&domain->group_list);
1255 iommu->external_domain = domain;
1259 list_add(&group->next,
1260 &iommu->external_domain->group_list);
1261 mutex_unlock(&iommu->lock);
1264 symbol_put(mdev_bus_type);
1267 domain->domain = iommu_domain_alloc(bus);
1268 if (!domain->domain) {
1273 if (iommu->nesting) {
1276 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
1282 ret = iommu_attach_group(domain->domain, iommu_group);
1286 resv_msi = vfio_iommu_has_resv_msi(iommu_group, &resv_msi_base);
1288 INIT_LIST_HEAD(&domain->group_list);
1289 list_add(&group->next, &domain->group_list);
1291 msi_remap = resv_msi ? irq_domain_check_msi_remap() :
1292 iommu_capable(bus, IOMMU_CAP_INTR_REMAP);
1294 if (!allow_unsafe_interrupts && !msi_remap) {
1295 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
1301 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
1302 domain->prot |= IOMMU_CACHE;
1305 * Try to match an existing compatible domain. We don't want to
1306 * preclude an IOMMU driver supporting multiple bus_types and being
1307 * able to include different bus_types in the same IOMMU domain, so
1308 * we test whether the domains use the same iommu_ops rather than
1309 * testing if they're on the same bus_type.
1311 list_for_each_entry(d, &iommu->domain_list, next) {
1312 if (d->domain->ops == domain->domain->ops &&
1313 d->prot == domain->prot) {
1314 iommu_detach_group(domain->domain, iommu_group);
1315 if (!iommu_attach_group(d->domain, iommu_group)) {
1316 list_add(&group->next, &d->group_list);
1317 iommu_domain_free(domain->domain);
1319 mutex_unlock(&iommu->lock);
1323 ret = iommu_attach_group(domain->domain, iommu_group);
1329 vfio_test_domain_fgsp(domain);
1331 /* replay mappings on new domains */
1332 ret = vfio_iommu_replay(iommu, domain);
1337 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
1342 list_add(&domain->next, &iommu->domain_list);
1344 mutex_unlock(&iommu->lock);
1349 iommu_detach_group(domain->domain, iommu_group);
1351 iommu_domain_free(domain->domain);
1355 mutex_unlock(&iommu->lock);
1359 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
1361 struct rb_node *node;
1363 while ((node = rb_first(&iommu->dma_list)))
1364 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
1367 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
1369 struct rb_node *n, *p;
1371 n = rb_first(&iommu->dma_list);
1372 for (; n; n = rb_next(n)) {
1373 struct vfio_dma *dma;
1374 long locked = 0, unlocked = 0;
1376 dma = rb_entry(n, struct vfio_dma, node);
1377 unlocked += vfio_unmap_unpin(iommu, dma, false);
1378 p = rb_first(&dma->pfn_list);
1379 for (; p; p = rb_next(p)) {
1380 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
1383 if (!is_invalid_reserved_pfn(vpfn->pfn))
1386 vfio_lock_acct(dma->task, locked - unlocked);
1390 static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu)
1394 n = rb_first(&iommu->dma_list);
1395 for (; n; n = rb_next(n)) {
1396 struct vfio_dma *dma;
1398 dma = rb_entry(n, struct vfio_dma, node);
1400 if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list)))
1403 /* mdev vendor driver must unregister notifier */
1404 WARN_ON(iommu->notifier.head);
1407 static void vfio_iommu_type1_detach_group(void *iommu_data,
1408 struct iommu_group *iommu_group)
1410 struct vfio_iommu *iommu = iommu_data;
1411 struct vfio_domain *domain;
1412 struct vfio_group *group;
1414 mutex_lock(&iommu->lock);
1416 if (iommu->external_domain) {
1417 group = find_iommu_group(iommu->external_domain, iommu_group);
1419 list_del(&group->next);
1422 if (list_empty(&iommu->external_domain->group_list)) {
1423 vfio_sanity_check_pfn_list(iommu);
1425 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1426 vfio_iommu_unmap_unpin_all(iommu);
1428 kfree(iommu->external_domain);
1429 iommu->external_domain = NULL;
1431 goto detach_group_done;
1435 list_for_each_entry(domain, &iommu->domain_list, next) {
1436 group = find_iommu_group(domain, iommu_group);
1440 iommu_detach_group(domain->domain, iommu_group);
1441 list_del(&group->next);
1444 * Group ownership provides privilege, if the group list is
1445 * empty, the domain goes away. If it's the last domain with
1446 * iommu and external domain doesn't exist, then all the
1447 * mappings go away too. If it's the last domain with iommu and
1448 * external domain exist, update accounting
1450 if (list_empty(&domain->group_list)) {
1451 if (list_is_singular(&iommu->domain_list)) {
1452 if (!iommu->external_domain)
1453 vfio_iommu_unmap_unpin_all(iommu);
1455 vfio_iommu_unmap_unpin_reaccount(iommu);
1457 iommu_domain_free(domain->domain);
1458 list_del(&domain->next);
1465 mutex_unlock(&iommu->lock);
1468 static void *vfio_iommu_type1_open(unsigned long arg)
1470 struct vfio_iommu *iommu;
1472 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
1474 return ERR_PTR(-ENOMEM);
1477 case VFIO_TYPE1_IOMMU:
1479 case VFIO_TYPE1_NESTING_IOMMU:
1480 iommu->nesting = true;
1481 case VFIO_TYPE1v2_IOMMU:
1486 return ERR_PTR(-EINVAL);
1489 INIT_LIST_HEAD(&iommu->domain_list);
1490 iommu->dma_list = RB_ROOT;
1491 mutex_init(&iommu->lock);
1492 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
1497 static void vfio_release_domain(struct vfio_domain *domain, bool external)
1499 struct vfio_group *group, *group_tmp;
1501 list_for_each_entry_safe(group, group_tmp,
1502 &domain->group_list, next) {
1504 iommu_detach_group(domain->domain, group->iommu_group);
1505 list_del(&group->next);
1510 iommu_domain_free(domain->domain);
1513 static void vfio_iommu_type1_release(void *iommu_data)
1515 struct vfio_iommu *iommu = iommu_data;
1516 struct vfio_domain *domain, *domain_tmp;
1518 if (iommu->external_domain) {
1519 vfio_release_domain(iommu->external_domain, true);
1520 vfio_sanity_check_pfn_list(iommu);
1521 kfree(iommu->external_domain);
1524 vfio_iommu_unmap_unpin_all(iommu);
1526 list_for_each_entry_safe(domain, domain_tmp,
1527 &iommu->domain_list, next) {
1528 vfio_release_domain(domain, false);
1529 list_del(&domain->next);
1535 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
1537 struct vfio_domain *domain;
1540 mutex_lock(&iommu->lock);
1541 list_for_each_entry(domain, &iommu->domain_list, next) {
1542 if (!(domain->prot & IOMMU_CACHE)) {
1547 mutex_unlock(&iommu->lock);
1552 static long vfio_iommu_type1_ioctl(void *iommu_data,
1553 unsigned int cmd, unsigned long arg)
1555 struct vfio_iommu *iommu = iommu_data;
1556 unsigned long minsz;
1558 if (cmd == VFIO_CHECK_EXTENSION) {
1560 case VFIO_TYPE1_IOMMU:
1561 case VFIO_TYPE1v2_IOMMU:
1562 case VFIO_TYPE1_NESTING_IOMMU:
1564 case VFIO_DMA_CC_IOMMU:
1567 return vfio_domains_have_iommu_cache(iommu);
1571 } else if (cmd == VFIO_IOMMU_GET_INFO) {
1572 struct vfio_iommu_type1_info info;
1574 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
1576 if (copy_from_user(&info, (void __user *)arg, minsz))
1579 if (info.argsz < minsz)
1582 info.flags = VFIO_IOMMU_INFO_PGSIZES;
1584 info.iova_pgsizes = vfio_pgsize_bitmap(iommu);
1586 return copy_to_user((void __user *)arg, &info, minsz) ?
1589 } else if (cmd == VFIO_IOMMU_MAP_DMA) {
1590 struct vfio_iommu_type1_dma_map map;
1591 uint32_t mask = VFIO_DMA_MAP_FLAG_READ |
1592 VFIO_DMA_MAP_FLAG_WRITE;
1594 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
1596 if (copy_from_user(&map, (void __user *)arg, minsz))
1599 if (map.argsz < minsz || map.flags & ~mask)
1602 return vfio_dma_do_map(iommu, &map);
1604 } else if (cmd == VFIO_IOMMU_UNMAP_DMA) {
1605 struct vfio_iommu_type1_dma_unmap unmap;
1608 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
1610 if (copy_from_user(&unmap, (void __user *)arg, minsz))
1613 if (unmap.argsz < minsz || unmap.flags)
1616 ret = vfio_dma_do_unmap(iommu, &unmap);
1620 return copy_to_user((void __user *)arg, &unmap, minsz) ?
1627 static int vfio_iommu_type1_register_notifier(void *iommu_data,
1628 unsigned long *events,
1629 struct notifier_block *nb)
1631 struct vfio_iommu *iommu = iommu_data;
1633 /* clear known events */
1634 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
1636 /* refuse to register if still events remaining */
1640 return blocking_notifier_chain_register(&iommu->notifier, nb);
1643 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
1644 struct notifier_block *nb)
1646 struct vfio_iommu *iommu = iommu_data;
1648 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
1651 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
1652 .name = "vfio-iommu-type1",
1653 .owner = THIS_MODULE,
1654 .open = vfio_iommu_type1_open,
1655 .release = vfio_iommu_type1_release,
1656 .ioctl = vfio_iommu_type1_ioctl,
1657 .attach_group = vfio_iommu_type1_attach_group,
1658 .detach_group = vfio_iommu_type1_detach_group,
1659 .pin_pages = vfio_iommu_type1_pin_pages,
1660 .unpin_pages = vfio_iommu_type1_unpin_pages,
1661 .register_notifier = vfio_iommu_type1_register_notifier,
1662 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
1665 static int __init vfio_iommu_type1_init(void)
1667 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
1670 static void __exit vfio_iommu_type1_cleanup(void)
1672 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
1675 module_init(vfio_iommu_type1_init);
1676 module_exit(vfio_iommu_type1_cleanup);
1678 MODULE_VERSION(DRIVER_VERSION);
1679 MODULE_LICENSE("GPL v2");
1680 MODULE_AUTHOR(DRIVER_AUTHOR);
1681 MODULE_DESCRIPTION(DRIVER_DESC);