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;
276 is_current = (task->mm == current->mm);
278 mm = is_current ? task->mm : get_task_mm(task);
280 return; /* process exited */
282 if (down_write_trylock(&mm->mmap_sem)) {
283 mm->locked_vm += npage;
284 up_write(&mm->mmap_sem);
291 mm = get_task_mm(task);
297 * Couldn't get mmap_sem lock, so must setup to update
298 * mm->locked_vm later. If locked_vm were atomic, we
299 * wouldn't need this silliness
301 vwork = kmalloc(sizeof(struct vwork), GFP_KERNEL);
302 if (WARN_ON(!vwork)) {
306 INIT_WORK(&vwork->work, vfio_lock_acct_bg);
308 vwork->npage = npage;
309 schedule_work(&vwork->work);
313 * Some mappings aren't backed by a struct page, for example an mmap'd
314 * MMIO range for our own or another device. These use a different
315 * pfn conversion and shouldn't be tracked as locked pages.
317 static bool is_invalid_reserved_pfn(unsigned long pfn)
319 if (pfn_valid(pfn)) {
321 struct page *tail = pfn_to_page(pfn);
322 struct page *head = compound_head(tail);
323 reserved = !!(PageReserved(head));
326 * "head" is not a dangling pointer
327 * (compound_head takes care of that)
328 * but the hugepage may have been split
329 * from under us (and we may not hold a
330 * reference count on the head page so it can
331 * be reused before we run PageReferenced), so
332 * we've to check PageTail before returning
339 return PageReserved(tail);
345 static int put_pfn(unsigned long pfn, int prot)
347 if (!is_invalid_reserved_pfn(pfn)) {
348 struct page *page = pfn_to_page(pfn);
349 if (prot & IOMMU_WRITE)
357 static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,
358 int prot, unsigned long *pfn)
360 struct page *page[1];
361 struct vm_area_struct *vma;
364 if (mm == current->mm) {
365 ret = get_user_pages_fast(vaddr, 1, !!(prot & IOMMU_WRITE),
368 unsigned int flags = 0;
370 if (prot & IOMMU_WRITE)
373 down_read(&mm->mmap_sem);
374 ret = get_user_pages_remote(NULL, mm, vaddr, 1, flags, page,
376 up_read(&mm->mmap_sem);
380 *pfn = page_to_pfn(page[0]);
384 down_read(&mm->mmap_sem);
386 vma = find_vma_intersection(mm, vaddr, vaddr + 1);
388 if (vma && vma->vm_flags & VM_PFNMAP) {
389 *pfn = ((vaddr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
390 if (is_invalid_reserved_pfn(*pfn))
394 up_read(&mm->mmap_sem);
399 * Attempt to pin pages. We really don't want to track all the pfns and
400 * the iommu can only map chunks of consecutive pfns anyway, so get the
401 * first page and all consecutive pages with the same locking.
403 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
404 long npage, unsigned long *pfn_base)
406 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
407 bool lock_cap = capable(CAP_IPC_LOCK);
408 long ret, pinned = 0, lock_acct = 0;
410 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
412 /* This code path is only user initiated */
416 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base);
421 rsvd = is_invalid_reserved_pfn(*pfn_base);
424 * Reserved pages aren't counted against the user, externally pinned
425 * pages are already counted against the user.
427 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
428 if (!lock_cap && current->mm->locked_vm + 1 > limit) {
429 put_pfn(*pfn_base, dma->prot);
430 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__,
431 limit << PAGE_SHIFT);
437 if (unlikely(disable_hugepages))
440 /* Lock all the consecutive pages from pfn_base */
441 for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage;
442 pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) {
443 unsigned long pfn = 0;
445 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn);
449 if (pfn != *pfn_base + pinned ||
450 rsvd != is_invalid_reserved_pfn(pfn)) {
451 put_pfn(pfn, dma->prot);
455 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
457 current->mm->locked_vm + lock_acct + 1 > limit) {
458 put_pfn(pfn, dma->prot);
459 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
460 __func__, limit << PAGE_SHIFT);
468 vfio_lock_acct(current, lock_acct);
473 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
474 unsigned long pfn, long npage,
477 long unlocked = 0, locked = 0;
480 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
481 if (put_pfn(pfn++, dma->prot)) {
483 if (vfio_find_vpfn(dma, iova))
489 vfio_lock_acct(dma->task, locked - unlocked);
494 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
495 unsigned long *pfn_base, bool do_accounting)
498 bool lock_cap = ns_capable(task_active_pid_ns(dma->task)->user_ns,
500 struct mm_struct *mm;
504 mm = get_task_mm(dma->task);
508 ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base);
512 rsvd = is_invalid_reserved_pfn(*pfn_base);
513 limit = task_rlimit(dma->task, RLIMIT_MEMLOCK) >> PAGE_SHIFT;
515 if (!rsvd && !lock_cap && mm->locked_vm + 1 > limit) {
516 put_pfn(*pfn_base, dma->prot);
517 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK (%ld) exceeded\n",
518 __func__, dma->task->comm, task_pid_nr(dma->task),
519 limit << PAGE_SHIFT);
524 if (!rsvd && do_accounting)
525 vfio_lock_acct(dma->task, 1);
533 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
537 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
542 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
545 vfio_lock_acct(dma->task, -unlocked);
550 static int vfio_iommu_type1_pin_pages(void *iommu_data,
551 unsigned long *user_pfn,
553 unsigned long *phys_pfn)
555 struct vfio_iommu *iommu = iommu_data;
557 unsigned long remote_vaddr;
558 struct vfio_dma *dma;
561 if (!iommu || !user_pfn || !phys_pfn)
564 /* Supported for v2 version only */
568 mutex_lock(&iommu->lock);
570 /* Fail if notifier list is empty */
571 if ((!iommu->external_domain) || (!iommu->notifier.head)) {
577 * If iommu capable domain exist in the container then all pages are
578 * already pinned and accounted. Accouting should be done if there is no
579 * iommu capable domain in the container.
581 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
583 for (i = 0; i < npage; i++) {
585 struct vfio_pfn *vpfn;
587 iova = user_pfn[i] << PAGE_SHIFT;
588 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
594 if ((dma->prot & prot) != prot) {
599 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
601 phys_pfn[i] = vpfn->pfn;
605 remote_vaddr = dma->vaddr + iova - dma->iova;
606 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
613 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
615 vfio_unpin_page_external(dma, iova, do_accounting);
625 for (j = 0; j < i; j++) {
628 iova = user_pfn[j] << PAGE_SHIFT;
629 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
630 vfio_unpin_page_external(dma, iova, do_accounting);
634 mutex_unlock(&iommu->lock);
638 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
639 unsigned long *user_pfn,
642 struct vfio_iommu *iommu = iommu_data;
646 if (!iommu || !user_pfn)
649 /* Supported for v2 version only */
653 mutex_lock(&iommu->lock);
655 if (!iommu->external_domain) {
656 mutex_unlock(&iommu->lock);
660 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
661 for (i = 0; i < npage; i++) {
662 struct vfio_dma *dma;
665 iova = user_pfn[i] << PAGE_SHIFT;
666 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
669 vfio_unpin_page_external(dma, iova, do_accounting);
673 mutex_unlock(&iommu->lock);
674 return i > npage ? npage : (i > 0 ? i : -EINVAL);
677 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
680 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
681 struct vfio_domain *domain, *d;
687 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
691 * We use the IOMMU to track the physical addresses, otherwise we'd
692 * need a much more complicated tracking system. Unfortunately that
693 * means we need to use one of the iommu domains to figure out the
694 * pfns to unpin. The rest need to be unmapped in advance so we have
695 * no iommu translations remaining when the pages are unpinned.
697 domain = d = list_first_entry(&iommu->domain_list,
698 struct vfio_domain, next);
700 list_for_each_entry_continue(d, &iommu->domain_list, next) {
701 iommu_unmap(d->domain, dma->iova, dma->size);
706 size_t unmapped, len;
707 phys_addr_t phys, next;
709 phys = iommu_iova_to_phys(domain->domain, iova);
710 if (WARN_ON(!phys)) {
716 * To optimize for fewer iommu_unmap() calls, each of which
717 * may require hardware cache flushing, try to find the
718 * largest contiguous physical memory chunk to unmap.
720 for (len = PAGE_SIZE;
721 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
722 next = iommu_iova_to_phys(domain->domain, iova + len);
723 if (next != phys + len)
727 unmapped = iommu_unmap(domain->domain, iova, len);
728 if (WARN_ON(!unmapped))
731 unlocked += vfio_unpin_pages_remote(dma, iova,
733 unmapped >> PAGE_SHIFT,
740 dma->iommu_mapped = false;
742 vfio_lock_acct(dma->task, -unlocked);
748 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
750 vfio_unmap_unpin(iommu, dma, true);
751 vfio_unlink_dma(iommu, dma);
752 put_task_struct(dma->task);
756 static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu)
758 struct vfio_domain *domain;
759 unsigned long bitmap = ULONG_MAX;
761 mutex_lock(&iommu->lock);
762 list_for_each_entry(domain, &iommu->domain_list, next)
763 bitmap &= domain->domain->pgsize_bitmap;
764 mutex_unlock(&iommu->lock);
767 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
768 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
769 * That way the user will be able to map/unmap buffers whose size/
770 * start address is aligned with PAGE_SIZE. Pinning code uses that
771 * granularity while iommu driver can use the sub-PAGE_SIZE size
774 if (bitmap & ~PAGE_MASK) {
782 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
783 struct vfio_iommu_type1_dma_unmap *unmap)
786 struct vfio_dma *dma, *dma_last = NULL;
788 int ret = 0, retries = 0;
790 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
792 if (unmap->iova & mask)
794 if (!unmap->size || unmap->size & mask)
797 WARN_ON(mask & PAGE_MASK);
799 mutex_lock(&iommu->lock);
802 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
803 * avoid tracking individual mappings. This means that the granularity
804 * of the original mapping was lost and the user was allowed to attempt
805 * to unmap any range. Depending on the contiguousness of physical
806 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
807 * or may not have worked. We only guaranteed unmap granularity
808 * matching the original mapping; even though it was untracked here,
809 * the original mappings are reflected in IOMMU mappings. This
810 * resulted in a couple unusual behaviors. First, if a range is not
811 * able to be unmapped, ex. a set of 4k pages that was mapped as a
812 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
813 * a zero sized unmap. Also, if an unmap request overlaps the first
814 * address of a hugepage, the IOMMU will unmap the entire hugepage.
815 * This also returns success and the returned unmap size reflects the
816 * actual size unmapped.
818 * We attempt to maintain compatibility with this "v1" interface, but
819 * we take control out of the hands of the IOMMU. Therefore, an unmap
820 * request offset from the beginning of the original mapping will
821 * return success with zero sized unmap. And an unmap request covering
822 * the first iova of mapping will unmap the entire range.
824 * The v2 version of this interface intends to be more deterministic.
825 * Unmap requests must fully cover previous mappings. Multiple
826 * mappings may still be unmaped by specifying large ranges, but there
827 * must not be any previous mappings bisected by the range. An error
828 * will be returned if these conditions are not met. The v2 interface
829 * will only return success and a size of zero if there were no
830 * mappings within the range.
833 dma = vfio_find_dma(iommu, unmap->iova, 1);
834 if (dma && dma->iova != unmap->iova) {
838 dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0);
839 if (dma && dma->iova + dma->size != unmap->iova + unmap->size) {
845 while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
846 if (!iommu->v2 && unmap->iova > dma->iova)
849 * Task with same address space who mapped this iova range is
850 * allowed to unmap the iova range.
852 if (dma->task->mm != current->mm)
855 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
856 struct vfio_iommu_type1_dma_unmap nb_unmap;
858 if (dma_last == dma) {
859 BUG_ON(++retries > 10);
865 nb_unmap.iova = dma->iova;
866 nb_unmap.size = dma->size;
869 * Notify anyone (mdev vendor drivers) to invalidate and
870 * unmap iovas within the range we're about to unmap.
871 * Vendor drivers MUST unpin pages in response to an
874 mutex_unlock(&iommu->lock);
875 blocking_notifier_call_chain(&iommu->notifier,
876 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
880 unmapped += dma->size;
881 vfio_remove_dma(iommu, dma);
885 mutex_unlock(&iommu->lock);
887 /* Report how much was unmapped */
888 unmap->size = unmapped;
894 * Turns out AMD IOMMU has a page table bug where it won't map large pages
895 * to a region that previously mapped smaller pages. This should be fixed
896 * soon, so this is just a temporary workaround to break mappings down into
897 * PAGE_SIZE. Better to map smaller pages than nothing.
899 static int map_try_harder(struct vfio_domain *domain, dma_addr_t iova,
900 unsigned long pfn, long npage, int prot)
905 for (i = 0; i < npage; i++, pfn++, iova += PAGE_SIZE) {
906 ret = iommu_map(domain->domain, iova,
907 (phys_addr_t)pfn << PAGE_SHIFT,
908 PAGE_SIZE, prot | domain->prot);
913 for (; i < npage && i > 0; i--, iova -= PAGE_SIZE)
914 iommu_unmap(domain->domain, iova, PAGE_SIZE);
919 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
920 unsigned long pfn, long npage, int prot)
922 struct vfio_domain *d;
925 list_for_each_entry(d, &iommu->domain_list, next) {
926 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
927 npage << PAGE_SHIFT, prot | d->prot);
930 map_try_harder(d, iova, pfn, npage, prot))
940 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next)
941 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
946 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
949 dma_addr_t iova = dma->iova;
950 unsigned long vaddr = dma->vaddr;
951 size_t size = map_size;
957 /* Pin a contiguous chunk of memory */
958 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
959 size >> PAGE_SHIFT, &pfn);
967 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
970 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
975 size -= npage << PAGE_SHIFT;
976 dma->size += npage << PAGE_SHIFT;
979 dma->iommu_mapped = true;
982 vfio_remove_dma(iommu, dma);
987 static int vfio_dma_do_map(struct vfio_iommu *iommu,
988 struct vfio_iommu_type1_dma_map *map)
990 dma_addr_t iova = map->iova;
991 unsigned long vaddr = map->vaddr;
992 size_t size = map->size;
993 int ret = 0, prot = 0;
995 struct vfio_dma *dma;
997 /* Verify that none of our __u64 fields overflow */
998 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1001 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
1003 WARN_ON(mask & PAGE_MASK);
1005 /* READ/WRITE from device perspective */
1006 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1007 prot |= IOMMU_WRITE;
1008 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1011 if (!prot || !size || (size | iova | vaddr) & mask)
1014 /* Don't allow IOVA or virtual address wrap */
1015 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr)
1018 mutex_lock(&iommu->lock);
1020 if (vfio_find_dma(iommu, iova, size)) {
1025 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1034 get_task_struct(current);
1035 dma->task = current;
1036 dma->pfn_list = RB_ROOT;
1038 /* Insert zero-sized and grow as we map chunks of it */
1039 vfio_link_dma(iommu, dma);
1041 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1042 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1045 ret = vfio_pin_map_dma(iommu, dma, size);
1048 mutex_unlock(&iommu->lock);
1052 static int vfio_bus_type(struct device *dev, void *data)
1054 struct bus_type **bus = data;
1056 if (*bus && *bus != dev->bus)
1064 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1065 struct vfio_domain *domain)
1067 struct vfio_domain *d;
1071 /* Arbitrarily pick the first domain in the list for lookups */
1072 d = list_first_entry(&iommu->domain_list, struct vfio_domain, next);
1073 n = rb_first(&iommu->dma_list);
1075 for (; n; n = rb_next(n)) {
1076 struct vfio_dma *dma;
1079 dma = rb_entry(n, struct vfio_dma, node);
1082 while (iova < dma->iova + dma->size) {
1086 if (dma->iommu_mapped) {
1090 phys = iommu_iova_to_phys(d->domain, iova);
1092 if (WARN_ON(!phys)) {
1100 while (i < dma->iova + dma->size &&
1101 p == iommu_iova_to_phys(d->domain, i)) {
1108 unsigned long vaddr = dma->vaddr +
1110 size_t n = dma->iova + dma->size - iova;
1113 npage = vfio_pin_pages_remote(dma, vaddr,
1122 phys = pfn << PAGE_SHIFT;
1123 size = npage << PAGE_SHIFT;
1126 ret = iommu_map(domain->domain, iova, phys,
1127 size, dma->prot | domain->prot);
1133 dma->iommu_mapped = true;
1139 * We change our unmap behavior slightly depending on whether the IOMMU
1140 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1141 * for practically any contiguous power-of-two mapping we give it. This means
1142 * we don't need to look for contiguous chunks ourselves to make unmapping
1143 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1144 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1145 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1146 * hugetlbfs is in use.
1148 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1151 int ret, order = get_order(PAGE_SIZE * 2);
1153 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1157 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1158 IOMMU_READ | IOMMU_WRITE | domain->prot);
1160 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1162 if (unmapped == PAGE_SIZE)
1163 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1165 domain->fgsp = true;
1168 __free_pages(pages, order);
1171 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1172 struct iommu_group *iommu_group)
1174 struct vfio_group *g;
1176 list_for_each_entry(g, &domain->group_list, next) {
1177 if (g->iommu_group == iommu_group)
1184 static int vfio_iommu_type1_attach_group(void *iommu_data,
1185 struct iommu_group *iommu_group)
1187 struct vfio_iommu *iommu = iommu_data;
1188 struct vfio_group *group;
1189 struct vfio_domain *domain, *d;
1190 struct bus_type *bus = NULL, *mdev_bus;
1193 mutex_lock(&iommu->lock);
1195 list_for_each_entry(d, &iommu->domain_list, next) {
1196 if (find_iommu_group(d, iommu_group)) {
1197 mutex_unlock(&iommu->lock);
1202 if (iommu->external_domain) {
1203 if (find_iommu_group(iommu->external_domain, iommu_group)) {
1204 mutex_unlock(&iommu->lock);
1209 group = kzalloc(sizeof(*group), GFP_KERNEL);
1210 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
1211 if (!group || !domain) {
1216 group->iommu_group = iommu_group;
1218 /* Determine bus_type in order to allocate a domain */
1219 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
1223 mdev_bus = symbol_get(mdev_bus_type);
1226 if ((bus == mdev_bus) && !iommu_present(bus)) {
1227 symbol_put(mdev_bus_type);
1228 if (!iommu->external_domain) {
1229 INIT_LIST_HEAD(&domain->group_list);
1230 iommu->external_domain = domain;
1234 list_add(&group->next,
1235 &iommu->external_domain->group_list);
1236 mutex_unlock(&iommu->lock);
1239 symbol_put(mdev_bus_type);
1242 domain->domain = iommu_domain_alloc(bus);
1243 if (!domain->domain) {
1248 if (iommu->nesting) {
1251 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
1257 ret = iommu_attach_group(domain->domain, iommu_group);
1261 INIT_LIST_HEAD(&domain->group_list);
1262 list_add(&group->next, &domain->group_list);
1264 if (!allow_unsafe_interrupts &&
1265 !iommu_capable(bus, IOMMU_CAP_INTR_REMAP)) {
1266 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
1272 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
1273 domain->prot |= IOMMU_CACHE;
1276 * Try to match an existing compatible domain. We don't want to
1277 * preclude an IOMMU driver supporting multiple bus_types and being
1278 * able to include different bus_types in the same IOMMU domain, so
1279 * we test whether the domains use the same iommu_ops rather than
1280 * testing if they're on the same bus_type.
1282 list_for_each_entry(d, &iommu->domain_list, next) {
1283 if (d->domain->ops == domain->domain->ops &&
1284 d->prot == domain->prot) {
1285 iommu_detach_group(domain->domain, iommu_group);
1286 if (!iommu_attach_group(d->domain, iommu_group)) {
1287 list_add(&group->next, &d->group_list);
1288 iommu_domain_free(domain->domain);
1290 mutex_unlock(&iommu->lock);
1294 ret = iommu_attach_group(domain->domain, iommu_group);
1300 vfio_test_domain_fgsp(domain);
1302 /* replay mappings on new domains */
1303 ret = vfio_iommu_replay(iommu, domain);
1307 list_add(&domain->next, &iommu->domain_list);
1309 mutex_unlock(&iommu->lock);
1314 iommu_detach_group(domain->domain, iommu_group);
1316 iommu_domain_free(domain->domain);
1320 mutex_unlock(&iommu->lock);
1324 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
1326 struct rb_node *node;
1328 while ((node = rb_first(&iommu->dma_list)))
1329 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
1332 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
1334 struct rb_node *n, *p;
1336 n = rb_first(&iommu->dma_list);
1337 for (; n; n = rb_next(n)) {
1338 struct vfio_dma *dma;
1339 long locked = 0, unlocked = 0;
1341 dma = rb_entry(n, struct vfio_dma, node);
1342 unlocked += vfio_unmap_unpin(iommu, dma, false);
1343 p = rb_first(&dma->pfn_list);
1344 for (; p; p = rb_next(p)) {
1345 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
1348 if (!is_invalid_reserved_pfn(vpfn->pfn))
1351 vfio_lock_acct(dma->task, locked - unlocked);
1355 static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu)
1359 n = rb_first(&iommu->dma_list);
1360 for (; n; n = rb_next(n)) {
1361 struct vfio_dma *dma;
1363 dma = rb_entry(n, struct vfio_dma, node);
1365 if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list)))
1368 /* mdev vendor driver must unregister notifier */
1369 WARN_ON(iommu->notifier.head);
1372 static void vfio_iommu_type1_detach_group(void *iommu_data,
1373 struct iommu_group *iommu_group)
1375 struct vfio_iommu *iommu = iommu_data;
1376 struct vfio_domain *domain;
1377 struct vfio_group *group;
1379 mutex_lock(&iommu->lock);
1381 if (iommu->external_domain) {
1382 group = find_iommu_group(iommu->external_domain, iommu_group);
1384 list_del(&group->next);
1387 if (list_empty(&iommu->external_domain->group_list)) {
1388 vfio_sanity_check_pfn_list(iommu);
1390 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1391 vfio_iommu_unmap_unpin_all(iommu);
1393 kfree(iommu->external_domain);
1394 iommu->external_domain = NULL;
1396 goto detach_group_done;
1400 list_for_each_entry(domain, &iommu->domain_list, next) {
1401 group = find_iommu_group(domain, iommu_group);
1405 iommu_detach_group(domain->domain, iommu_group);
1406 list_del(&group->next);
1409 * Group ownership provides privilege, if the group list is
1410 * empty, the domain goes away. If it's the last domain with
1411 * iommu and external domain doesn't exist, then all the
1412 * mappings go away too. If it's the last domain with iommu and
1413 * external domain exist, update accounting
1415 if (list_empty(&domain->group_list)) {
1416 if (list_is_singular(&iommu->domain_list)) {
1417 if (!iommu->external_domain)
1418 vfio_iommu_unmap_unpin_all(iommu);
1420 vfio_iommu_unmap_unpin_reaccount(iommu);
1422 iommu_domain_free(domain->domain);
1423 list_del(&domain->next);
1430 mutex_unlock(&iommu->lock);
1433 static void *vfio_iommu_type1_open(unsigned long arg)
1435 struct vfio_iommu *iommu;
1437 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
1439 return ERR_PTR(-ENOMEM);
1442 case VFIO_TYPE1_IOMMU:
1444 case VFIO_TYPE1_NESTING_IOMMU:
1445 iommu->nesting = true;
1446 case VFIO_TYPE1v2_IOMMU:
1451 return ERR_PTR(-EINVAL);
1454 INIT_LIST_HEAD(&iommu->domain_list);
1455 iommu->dma_list = RB_ROOT;
1456 mutex_init(&iommu->lock);
1457 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
1462 static void vfio_release_domain(struct vfio_domain *domain, bool external)
1464 struct vfio_group *group, *group_tmp;
1466 list_for_each_entry_safe(group, group_tmp,
1467 &domain->group_list, next) {
1469 iommu_detach_group(domain->domain, group->iommu_group);
1470 list_del(&group->next);
1475 iommu_domain_free(domain->domain);
1478 static void vfio_iommu_type1_release(void *iommu_data)
1480 struct vfio_iommu *iommu = iommu_data;
1481 struct vfio_domain *domain, *domain_tmp;
1483 if (iommu->external_domain) {
1484 vfio_release_domain(iommu->external_domain, true);
1485 vfio_sanity_check_pfn_list(iommu);
1486 kfree(iommu->external_domain);
1489 vfio_iommu_unmap_unpin_all(iommu);
1491 list_for_each_entry_safe(domain, domain_tmp,
1492 &iommu->domain_list, next) {
1493 vfio_release_domain(domain, false);
1494 list_del(&domain->next);
1500 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
1502 struct vfio_domain *domain;
1505 mutex_lock(&iommu->lock);
1506 list_for_each_entry(domain, &iommu->domain_list, next) {
1507 if (!(domain->prot & IOMMU_CACHE)) {
1512 mutex_unlock(&iommu->lock);
1517 static long vfio_iommu_type1_ioctl(void *iommu_data,
1518 unsigned int cmd, unsigned long arg)
1520 struct vfio_iommu *iommu = iommu_data;
1521 unsigned long minsz;
1523 if (cmd == VFIO_CHECK_EXTENSION) {
1525 case VFIO_TYPE1_IOMMU:
1526 case VFIO_TYPE1v2_IOMMU:
1527 case VFIO_TYPE1_NESTING_IOMMU:
1529 case VFIO_DMA_CC_IOMMU:
1532 return vfio_domains_have_iommu_cache(iommu);
1536 } else if (cmd == VFIO_IOMMU_GET_INFO) {
1537 struct vfio_iommu_type1_info info;
1539 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
1541 if (copy_from_user(&info, (void __user *)arg, minsz))
1544 if (info.argsz < minsz)
1547 info.flags = VFIO_IOMMU_INFO_PGSIZES;
1549 info.iova_pgsizes = vfio_pgsize_bitmap(iommu);
1551 return copy_to_user((void __user *)arg, &info, minsz) ?
1554 } else if (cmd == VFIO_IOMMU_MAP_DMA) {
1555 struct vfio_iommu_type1_dma_map map;
1556 uint32_t mask = VFIO_DMA_MAP_FLAG_READ |
1557 VFIO_DMA_MAP_FLAG_WRITE;
1559 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
1561 if (copy_from_user(&map, (void __user *)arg, minsz))
1564 if (map.argsz < minsz || map.flags & ~mask)
1567 return vfio_dma_do_map(iommu, &map);
1569 } else if (cmd == VFIO_IOMMU_UNMAP_DMA) {
1570 struct vfio_iommu_type1_dma_unmap unmap;
1573 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
1575 if (copy_from_user(&unmap, (void __user *)arg, minsz))
1578 if (unmap.argsz < minsz || unmap.flags)
1581 ret = vfio_dma_do_unmap(iommu, &unmap);
1585 return copy_to_user((void __user *)arg, &unmap, minsz) ?
1592 static int vfio_iommu_type1_register_notifier(void *iommu_data,
1593 unsigned long *events,
1594 struct notifier_block *nb)
1596 struct vfio_iommu *iommu = iommu_data;
1598 /* clear known events */
1599 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
1601 /* refuse to register if still events remaining */
1605 return blocking_notifier_chain_register(&iommu->notifier, nb);
1608 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
1609 struct notifier_block *nb)
1611 struct vfio_iommu *iommu = iommu_data;
1613 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
1616 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
1617 .name = "vfio-iommu-type1",
1618 .owner = THIS_MODULE,
1619 .open = vfio_iommu_type1_open,
1620 .release = vfio_iommu_type1_release,
1621 .ioctl = vfio_iommu_type1_ioctl,
1622 .attach_group = vfio_iommu_type1_attach_group,
1623 .detach_group = vfio_iommu_type1_detach_group,
1624 .pin_pages = vfio_iommu_type1_pin_pages,
1625 .unpin_pages = vfio_iommu_type1_unpin_pages,
1626 .register_notifier = vfio_iommu_type1_register_notifier,
1627 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
1630 static int __init vfio_iommu_type1_init(void)
1632 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
1635 static void __exit vfio_iommu_type1_cleanup(void)
1637 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
1640 module_init(vfio_iommu_type1_init);
1641 module_exit(vfio_iommu_type1_cleanup);
1643 MODULE_VERSION(DRIVER_VERSION);
1644 MODULE_LICENSE("GPL v2");
1645 MODULE_AUTHOR(DRIVER_AUTHOR);
1646 MODULE_DESCRIPTION(DRIVER_DESC);