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 bool lock_cap; /* capable(CAP_IPC_LOCK) */
87 struct task_struct *task;
88 struct rb_root pfn_list; /* Ex-user pinned pfn list */
92 struct iommu_group *iommu_group;
93 struct list_head next;
97 * Guest RAM pinning working set or DMA target
101 dma_addr_t iova; /* Device address */
102 unsigned long pfn; /* Host pfn */
106 struct vfio_regions {
107 struct list_head list;
113 #define IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu) \
114 (!list_empty(&iommu->domain_list))
116 static int put_pfn(unsigned long pfn, int prot);
119 * This code handles mapping and unmapping of user data buffers
120 * into DMA'ble space using the IOMMU
123 static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
124 dma_addr_t start, size_t size)
126 struct rb_node *node = iommu->dma_list.rb_node;
129 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
131 if (start + size <= dma->iova)
132 node = node->rb_left;
133 else if (start >= dma->iova + dma->size)
134 node = node->rb_right;
142 static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new)
144 struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL;
145 struct vfio_dma *dma;
149 dma = rb_entry(parent, struct vfio_dma, node);
151 if (new->iova + new->size <= dma->iova)
152 link = &(*link)->rb_left;
154 link = &(*link)->rb_right;
157 rb_link_node(&new->node, parent, link);
158 rb_insert_color(&new->node, &iommu->dma_list);
161 static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
163 rb_erase(&old->node, &iommu->dma_list);
167 * Helper Functions for host iova-pfn list
169 static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
171 struct vfio_pfn *vpfn;
172 struct rb_node *node = dma->pfn_list.rb_node;
175 vpfn = rb_entry(node, struct vfio_pfn, node);
177 if (iova < vpfn->iova)
178 node = node->rb_left;
179 else if (iova > vpfn->iova)
180 node = node->rb_right;
187 static void vfio_link_pfn(struct vfio_dma *dma,
188 struct vfio_pfn *new)
190 struct rb_node **link, *parent = NULL;
191 struct vfio_pfn *vpfn;
193 link = &dma->pfn_list.rb_node;
196 vpfn = rb_entry(parent, struct vfio_pfn, node);
198 if (new->iova < vpfn->iova)
199 link = &(*link)->rb_left;
201 link = &(*link)->rb_right;
204 rb_link_node(&new->node, parent, link);
205 rb_insert_color(&new->node, &dma->pfn_list);
208 static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old)
210 rb_erase(&old->node, &dma->pfn_list);
213 static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova,
216 struct vfio_pfn *vpfn;
218 vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL);
224 atomic_set(&vpfn->ref_count, 1);
225 vfio_link_pfn(dma, vpfn);
229 static void vfio_remove_from_pfn_list(struct vfio_dma *dma,
230 struct vfio_pfn *vpfn)
232 vfio_unlink_pfn(dma, vpfn);
236 static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma,
239 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
242 atomic_inc(&vpfn->ref_count);
246 static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn)
250 if (atomic_dec_and_test(&vpfn->ref_count)) {
251 ret = put_pfn(vpfn->pfn, dma->prot);
252 vfio_remove_from_pfn_list(dma, vpfn);
257 static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
259 struct mm_struct *mm;
265 mm = async ? get_task_mm(dma->task) : dma->task->mm;
267 return -ESRCH; /* process exited */
269 ret = down_write_killable(&mm->mmap_sem);
272 if (!dma->lock_cap) {
275 limit = task_rlimit(dma->task,
276 RLIMIT_MEMLOCK) >> PAGE_SHIFT;
278 if (mm->locked_vm + npage > limit)
284 mm->locked_vm += npage;
286 up_write(&mm->mmap_sem);
296 * Some mappings aren't backed by a struct page, for example an mmap'd
297 * MMIO range for our own or another device. These use a different
298 * pfn conversion and shouldn't be tracked as locked pages.
300 static bool is_invalid_reserved_pfn(unsigned long pfn)
302 if (pfn_valid(pfn)) {
304 struct page *tail = pfn_to_page(pfn);
305 struct page *head = compound_head(tail);
306 reserved = !!(PageReserved(head));
309 * "head" is not a dangling pointer
310 * (compound_head takes care of that)
311 * but the hugepage may have been split
312 * from under us (and we may not hold a
313 * reference count on the head page so it can
314 * be reused before we run PageReferenced), so
315 * we've to check PageTail before returning
322 return PageReserved(tail);
328 static int put_pfn(unsigned long pfn, int prot)
330 if (!is_invalid_reserved_pfn(pfn)) {
331 struct page *page = pfn_to_page(pfn);
332 if (prot & IOMMU_WRITE)
340 static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,
341 int prot, unsigned long *pfn)
343 struct page *page[1];
344 struct vm_area_struct *vma;
345 struct vm_area_struct *vmas[1];
346 unsigned int flags = 0;
349 if (prot & IOMMU_WRITE)
352 down_read(&mm->mmap_sem);
353 if (mm == current->mm) {
354 ret = get_user_pages_longterm(vaddr, 1, flags, page, vmas);
356 ret = get_user_pages_remote(NULL, mm, vaddr, 1, flags, page,
359 * The lifetime of a vaddr_get_pfn() page pin is
360 * userspace-controlled. In the fs-dax case this could
361 * lead to indefinite stalls in filesystem operations.
362 * Disallow attempts to pin fs-dax pages via this
365 if (ret > 0 && vma_is_fsdax(vmas[0])) {
370 up_read(&mm->mmap_sem);
373 *pfn = page_to_pfn(page[0]);
377 down_read(&mm->mmap_sem);
379 vma = find_vma_intersection(mm, vaddr, vaddr + 1);
381 if (vma && vma->vm_flags & VM_PFNMAP) {
382 *pfn = ((vaddr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
383 if (is_invalid_reserved_pfn(*pfn))
387 up_read(&mm->mmap_sem);
392 * Attempt to pin pages. We really don't want to track all the pfns and
393 * the iommu can only map chunks of consecutive pfns anyway, so get the
394 * first page and all consecutive pages with the same locking.
396 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
397 long npage, unsigned long *pfn_base,
400 unsigned long pfn = 0;
401 long ret, pinned = 0, lock_acct = 0;
403 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
405 /* This code path is only user initiated */
409 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base);
414 rsvd = is_invalid_reserved_pfn(*pfn_base);
417 * Reserved pages aren't counted against the user, externally pinned
418 * pages are already counted against the user.
420 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
421 if (!dma->lock_cap && current->mm->locked_vm + 1 > limit) {
422 put_pfn(*pfn_base, dma->prot);
423 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__,
424 limit << PAGE_SHIFT);
430 if (unlikely(disable_hugepages))
433 /* Lock all the consecutive pages from pfn_base */
434 for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage;
435 pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) {
436 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn);
440 if (pfn != *pfn_base + pinned ||
441 rsvd != is_invalid_reserved_pfn(pfn)) {
442 put_pfn(pfn, dma->prot);
446 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
447 if (!dma->lock_cap &&
448 current->mm->locked_vm + lock_acct + 1 > limit) {
449 put_pfn(pfn, dma->prot);
450 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
451 __func__, limit << PAGE_SHIFT);
460 ret = vfio_lock_acct(dma, lock_acct, false);
465 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
466 put_pfn(pfn, dma->prot);
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, locked - unlocked, true);
496 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
497 unsigned long *pfn_base, bool do_accounting)
499 struct mm_struct *mm;
502 mm = get_task_mm(dma->task);
506 ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base);
507 if (!ret && do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
508 ret = vfio_lock_acct(dma, 1, true);
510 put_pfn(*pfn_base, dma->prot);
512 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
513 "(%ld) exceeded\n", __func__,
514 dma->task->comm, task_pid_nr(dma->task),
515 task_rlimit(dma->task, RLIMIT_MEMLOCK));
523 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
527 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
532 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
535 vfio_lock_acct(dma, -unlocked, true);
540 static int vfio_iommu_type1_pin_pages(void *iommu_data,
541 unsigned long *user_pfn,
543 unsigned long *phys_pfn)
545 struct vfio_iommu *iommu = iommu_data;
547 unsigned long remote_vaddr;
548 struct vfio_dma *dma;
551 if (!iommu || !user_pfn || !phys_pfn)
554 /* Supported for v2 version only */
558 mutex_lock(&iommu->lock);
560 /* Fail if notifier list is empty */
561 if ((!iommu->external_domain) || (!iommu->notifier.head)) {
567 * If iommu capable domain exist in the container then all pages are
568 * already pinned and accounted. Accouting should be done if there is no
569 * iommu capable domain in the container.
571 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
573 for (i = 0; i < npage; i++) {
575 struct vfio_pfn *vpfn;
577 iova = user_pfn[i] << PAGE_SHIFT;
578 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
584 if ((dma->prot & prot) != prot) {
589 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
591 phys_pfn[i] = vpfn->pfn;
595 remote_vaddr = dma->vaddr + iova - dma->iova;
596 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
601 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
603 vfio_unpin_page_external(dma, iova, do_accounting);
613 for (j = 0; j < i; j++) {
616 iova = user_pfn[j] << PAGE_SHIFT;
617 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
618 vfio_unpin_page_external(dma, iova, do_accounting);
622 mutex_unlock(&iommu->lock);
626 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
627 unsigned long *user_pfn,
630 struct vfio_iommu *iommu = iommu_data;
634 if (!iommu || !user_pfn)
637 /* Supported for v2 version only */
641 mutex_lock(&iommu->lock);
643 if (!iommu->external_domain) {
644 mutex_unlock(&iommu->lock);
648 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
649 for (i = 0; i < npage; i++) {
650 struct vfio_dma *dma;
653 iova = user_pfn[i] << PAGE_SHIFT;
654 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
657 vfio_unpin_page_external(dma, iova, do_accounting);
661 mutex_unlock(&iommu->lock);
662 return i > npage ? npage : (i > 0 ? i : -EINVAL);
665 static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
666 struct list_head *regions)
669 struct vfio_regions *entry, *next;
671 iommu_tlb_sync(domain->domain);
673 list_for_each_entry_safe(entry, next, regions, list) {
674 unlocked += vfio_unpin_pages_remote(dma,
676 entry->phys >> PAGE_SHIFT,
677 entry->len >> PAGE_SHIFT,
679 list_del(&entry->list);
689 * Generally, VFIO needs to unpin remote pages after each IOTLB flush.
690 * Therefore, when using IOTLB flush sync interface, VFIO need to keep track
691 * of these regions (currently using a list).
693 * This value specifies maximum number of regions for each IOTLB flush sync.
695 #define VFIO_IOMMU_TLB_SYNC_MAX 512
697 static size_t unmap_unpin_fast(struct vfio_domain *domain,
698 struct vfio_dma *dma, dma_addr_t *iova,
699 size_t len, phys_addr_t phys, long *unlocked,
700 struct list_head *unmapped_list,
704 struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
707 unmapped = iommu_unmap_fast(domain->domain, *iova, len);
712 iommu_tlb_range_add(domain->domain, *iova, unmapped);
715 entry->len = unmapped;
716 list_add_tail(&entry->list, unmapped_list);
724 * Sync if the number of fast-unmap regions hits the limit
725 * or in case of errors.
727 if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
728 *unlocked += vfio_sync_unpin(dma, domain,
736 static size_t unmap_unpin_slow(struct vfio_domain *domain,
737 struct vfio_dma *dma, dma_addr_t *iova,
738 size_t len, phys_addr_t phys,
741 size_t unmapped = iommu_unmap(domain->domain, *iova, len);
744 *unlocked += vfio_unpin_pages_remote(dma, *iova,
746 unmapped >> PAGE_SHIFT,
754 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
757 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
758 struct vfio_domain *domain, *d;
759 LIST_HEAD(unmapped_region_list);
760 int unmapped_region_cnt = 0;
766 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
770 * We use the IOMMU to track the physical addresses, otherwise we'd
771 * need a much more complicated tracking system. Unfortunately that
772 * means we need to use one of the iommu domains to figure out the
773 * pfns to unpin. The rest need to be unmapped in advance so we have
774 * no iommu translations remaining when the pages are unpinned.
776 domain = d = list_first_entry(&iommu->domain_list,
777 struct vfio_domain, next);
779 list_for_each_entry_continue(d, &iommu->domain_list, next) {
780 iommu_unmap(d->domain, dma->iova, dma->size);
785 size_t unmapped, len;
786 phys_addr_t phys, next;
788 phys = iommu_iova_to_phys(domain->domain, iova);
789 if (WARN_ON(!phys)) {
795 * To optimize for fewer iommu_unmap() calls, each of which
796 * may require hardware cache flushing, try to find the
797 * largest contiguous physical memory chunk to unmap.
799 for (len = PAGE_SIZE;
800 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
801 next = iommu_iova_to_phys(domain->domain, iova + len);
802 if (next != phys + len)
807 * First, try to use fast unmap/unpin. In case of failure,
808 * switch to slow unmap/unpin path.
810 unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys,
811 &unlocked, &unmapped_region_list,
812 &unmapped_region_cnt);
814 unmapped = unmap_unpin_slow(domain, dma, &iova, len,
816 if (WARN_ON(!unmapped))
821 dma->iommu_mapped = false;
823 if (unmapped_region_cnt)
824 unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list);
827 vfio_lock_acct(dma, -unlocked, true);
833 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
835 vfio_unmap_unpin(iommu, dma, true);
836 vfio_unlink_dma(iommu, dma);
837 put_task_struct(dma->task);
841 static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu)
843 struct vfio_domain *domain;
844 unsigned long bitmap = ULONG_MAX;
846 mutex_lock(&iommu->lock);
847 list_for_each_entry(domain, &iommu->domain_list, next)
848 bitmap &= domain->domain->pgsize_bitmap;
849 mutex_unlock(&iommu->lock);
852 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
853 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
854 * That way the user will be able to map/unmap buffers whose size/
855 * start address is aligned with PAGE_SIZE. Pinning code uses that
856 * granularity while iommu driver can use the sub-PAGE_SIZE size
859 if (bitmap & ~PAGE_MASK) {
867 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
868 struct vfio_iommu_type1_dma_unmap *unmap)
871 struct vfio_dma *dma, *dma_last = NULL;
873 int ret = 0, retries = 0;
875 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
877 if (unmap->iova & mask)
879 if (!unmap->size || unmap->size & mask)
881 if (unmap->iova + unmap->size - 1 < unmap->iova ||
882 unmap->size > SIZE_MAX)
885 WARN_ON(mask & PAGE_MASK);
887 mutex_lock(&iommu->lock);
890 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
891 * avoid tracking individual mappings. This means that the granularity
892 * of the original mapping was lost and the user was allowed to attempt
893 * to unmap any range. Depending on the contiguousness of physical
894 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
895 * or may not have worked. We only guaranteed unmap granularity
896 * matching the original mapping; even though it was untracked here,
897 * the original mappings are reflected in IOMMU mappings. This
898 * resulted in a couple unusual behaviors. First, if a range is not
899 * able to be unmapped, ex. a set of 4k pages that was mapped as a
900 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
901 * a zero sized unmap. Also, if an unmap request overlaps the first
902 * address of a hugepage, the IOMMU will unmap the entire hugepage.
903 * This also returns success and the returned unmap size reflects the
904 * actual size unmapped.
906 * We attempt to maintain compatibility with this "v1" interface, but
907 * we take control out of the hands of the IOMMU. Therefore, an unmap
908 * request offset from the beginning of the original mapping will
909 * return success with zero sized unmap. And an unmap request covering
910 * the first iova of mapping will unmap the entire range.
912 * The v2 version of this interface intends to be more deterministic.
913 * Unmap requests must fully cover previous mappings. Multiple
914 * mappings may still be unmaped by specifying large ranges, but there
915 * must not be any previous mappings bisected by the range. An error
916 * will be returned if these conditions are not met. The v2 interface
917 * will only return success and a size of zero if there were no
918 * mappings within the range.
921 dma = vfio_find_dma(iommu, unmap->iova, 1);
922 if (dma && dma->iova != unmap->iova) {
926 dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0);
927 if (dma && dma->iova + dma->size != unmap->iova + unmap->size) {
933 while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
934 if (!iommu->v2 && unmap->iova > dma->iova)
937 * Task with same address space who mapped this iova range is
938 * allowed to unmap the iova range.
940 if (dma->task->mm != current->mm)
943 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
944 struct vfio_iommu_type1_dma_unmap nb_unmap;
946 if (dma_last == dma) {
947 BUG_ON(++retries > 10);
953 nb_unmap.iova = dma->iova;
954 nb_unmap.size = dma->size;
957 * Notify anyone (mdev vendor drivers) to invalidate and
958 * unmap iovas within the range we're about to unmap.
959 * Vendor drivers MUST unpin pages in response to an
962 mutex_unlock(&iommu->lock);
963 blocking_notifier_call_chain(&iommu->notifier,
964 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
968 unmapped += dma->size;
969 vfio_remove_dma(iommu, dma);
973 mutex_unlock(&iommu->lock);
975 /* Report how much was unmapped */
976 unmap->size = unmapped;
981 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
982 unsigned long pfn, long npage, int prot)
984 struct vfio_domain *d;
987 list_for_each_entry(d, &iommu->domain_list, next) {
988 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
989 npage << PAGE_SHIFT, prot | d->prot);
999 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next)
1000 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
1005 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
1008 dma_addr_t iova = dma->iova;
1009 unsigned long vaddr = dma->vaddr;
1010 size_t size = map_size;
1012 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1016 /* Pin a contiguous chunk of memory */
1017 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
1018 size >> PAGE_SHIFT, &pfn, limit);
1026 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
1029 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
1034 size -= npage << PAGE_SHIFT;
1035 dma->size += npage << PAGE_SHIFT;
1038 dma->iommu_mapped = true;
1041 vfio_remove_dma(iommu, dma);
1046 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1047 struct vfio_iommu_type1_dma_map *map)
1049 dma_addr_t iova = map->iova;
1050 unsigned long vaddr = map->vaddr;
1051 size_t size = map->size;
1052 int ret = 0, prot = 0;
1054 struct vfio_dma *dma;
1056 /* Verify that none of our __u64 fields overflow */
1057 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1060 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
1062 WARN_ON(mask & PAGE_MASK);
1064 /* READ/WRITE from device perspective */
1065 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1066 prot |= IOMMU_WRITE;
1067 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1070 if (!prot || !size || (size | iova | vaddr) & mask)
1073 /* Don't allow IOVA or virtual address wrap */
1074 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr)
1077 mutex_lock(&iommu->lock);
1079 if (vfio_find_dma(iommu, iova, size)) {
1084 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1095 * We need to be able to both add to a task's locked memory and test
1096 * against the locked memory limit and we need to be able to do both
1097 * outside of this call path as pinning can be asynchronous via the
1098 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1099 * task_struct and VM locked pages requires an mm_struct, however
1100 * holding an indefinite mm reference is not recommended, therefore we
1101 * only hold a reference to a task. We could hold a reference to
1102 * current, however QEMU uses this call path through vCPU threads,
1103 * which can be killed resulting in a NULL mm and failure in the unmap
1104 * path when called via a different thread. Avoid this problem by
1105 * using the group_leader as threads within the same group require
1106 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1109 * Previously we also used the task for testing CAP_IPC_LOCK at the
1110 * time of pinning and accounting, however has_capability() makes use
1111 * of real_cred, a copy-on-write field, so we can't guarantee that it
1112 * matches group_leader, or in fact that it might not change by the
1113 * time it's evaluated. If a process were to call MAP_DMA with
1114 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1115 * possibly see different results for an iommu_mapped vfio_dma vs
1116 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1117 * time of calling MAP_DMA.
1119 get_task_struct(current->group_leader);
1120 dma->task = current->group_leader;
1121 dma->lock_cap = capable(CAP_IPC_LOCK);
1123 dma->pfn_list = RB_ROOT;
1125 /* Insert zero-sized and grow as we map chunks of it */
1126 vfio_link_dma(iommu, dma);
1128 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1129 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1132 ret = vfio_pin_map_dma(iommu, dma, size);
1135 mutex_unlock(&iommu->lock);
1139 static int vfio_bus_type(struct device *dev, void *data)
1141 struct bus_type **bus = data;
1143 if (*bus && *bus != dev->bus)
1151 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1152 struct vfio_domain *domain)
1154 struct vfio_domain *d;
1156 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1159 /* Arbitrarily pick the first domain in the list for lookups */
1160 d = list_first_entry(&iommu->domain_list, struct vfio_domain, next);
1161 n = rb_first(&iommu->dma_list);
1163 for (; n; n = rb_next(n)) {
1164 struct vfio_dma *dma;
1167 dma = rb_entry(n, struct vfio_dma, node);
1170 while (iova < dma->iova + dma->size) {
1174 if (dma->iommu_mapped) {
1178 phys = iommu_iova_to_phys(d->domain, iova);
1180 if (WARN_ON(!phys)) {
1188 while (i < dma->iova + dma->size &&
1189 p == iommu_iova_to_phys(d->domain, i)) {
1196 unsigned long vaddr = dma->vaddr +
1198 size_t n = dma->iova + dma->size - iova;
1201 npage = vfio_pin_pages_remote(dma, vaddr,
1210 phys = pfn << PAGE_SHIFT;
1211 size = npage << PAGE_SHIFT;
1214 ret = iommu_map(domain->domain, iova, phys,
1215 size, dma->prot | domain->prot);
1221 dma->iommu_mapped = true;
1227 * We change our unmap behavior slightly depending on whether the IOMMU
1228 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1229 * for practically any contiguous power-of-two mapping we give it. This means
1230 * we don't need to look for contiguous chunks ourselves to make unmapping
1231 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1232 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1233 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1234 * hugetlbfs is in use.
1236 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1239 int ret, order = get_order(PAGE_SIZE * 2);
1241 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1245 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1246 IOMMU_READ | IOMMU_WRITE | domain->prot);
1248 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1250 if (unmapped == PAGE_SIZE)
1251 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1253 domain->fgsp = true;
1256 __free_pages(pages, order);
1259 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1260 struct iommu_group *iommu_group)
1262 struct vfio_group *g;
1264 list_for_each_entry(g, &domain->group_list, next) {
1265 if (g->iommu_group == iommu_group)
1272 static bool vfio_iommu_has_sw_msi(struct iommu_group *group, phys_addr_t *base)
1274 struct list_head group_resv_regions;
1275 struct iommu_resv_region *region, *next;
1278 INIT_LIST_HEAD(&group_resv_regions);
1279 iommu_get_group_resv_regions(group, &group_resv_regions);
1280 list_for_each_entry(region, &group_resv_regions, list) {
1282 * The presence of any 'real' MSI regions should take
1283 * precedence over the software-managed one if the
1284 * IOMMU driver happens to advertise both types.
1286 if (region->type == IOMMU_RESV_MSI) {
1291 if (region->type == IOMMU_RESV_SW_MSI) {
1292 *base = region->start;
1296 list_for_each_entry_safe(region, next, &group_resv_regions, list)
1301 static int vfio_iommu_type1_attach_group(void *iommu_data,
1302 struct iommu_group *iommu_group)
1304 struct vfio_iommu *iommu = iommu_data;
1305 struct vfio_group *group;
1306 struct vfio_domain *domain, *d;
1307 struct bus_type *bus = NULL, *mdev_bus;
1309 bool resv_msi, msi_remap;
1310 phys_addr_t resv_msi_base;
1312 mutex_lock(&iommu->lock);
1314 list_for_each_entry(d, &iommu->domain_list, next) {
1315 if (find_iommu_group(d, iommu_group)) {
1316 mutex_unlock(&iommu->lock);
1321 if (iommu->external_domain) {
1322 if (find_iommu_group(iommu->external_domain, iommu_group)) {
1323 mutex_unlock(&iommu->lock);
1328 group = kzalloc(sizeof(*group), GFP_KERNEL);
1329 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
1330 if (!group || !domain) {
1335 group->iommu_group = iommu_group;
1337 /* Determine bus_type in order to allocate a domain */
1338 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
1342 mdev_bus = symbol_get(mdev_bus_type);
1345 if ((bus == mdev_bus) && !iommu_present(bus)) {
1346 symbol_put(mdev_bus_type);
1347 if (!iommu->external_domain) {
1348 INIT_LIST_HEAD(&domain->group_list);
1349 iommu->external_domain = domain;
1353 list_add(&group->next,
1354 &iommu->external_domain->group_list);
1355 mutex_unlock(&iommu->lock);
1358 symbol_put(mdev_bus_type);
1361 domain->domain = iommu_domain_alloc(bus);
1362 if (!domain->domain) {
1367 if (iommu->nesting) {
1370 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
1376 ret = iommu_attach_group(domain->domain, iommu_group);
1380 resv_msi = vfio_iommu_has_sw_msi(iommu_group, &resv_msi_base);
1382 INIT_LIST_HEAD(&domain->group_list);
1383 list_add(&group->next, &domain->group_list);
1385 msi_remap = irq_domain_check_msi_remap() ||
1386 iommu_capable(bus, IOMMU_CAP_INTR_REMAP);
1388 if (!allow_unsafe_interrupts && !msi_remap) {
1389 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
1395 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
1396 domain->prot |= IOMMU_CACHE;
1399 * Try to match an existing compatible domain. We don't want to
1400 * preclude an IOMMU driver supporting multiple bus_types and being
1401 * able to include different bus_types in the same IOMMU domain, so
1402 * we test whether the domains use the same iommu_ops rather than
1403 * testing if they're on the same bus_type.
1405 list_for_each_entry(d, &iommu->domain_list, next) {
1406 if (d->domain->ops == domain->domain->ops &&
1407 d->prot == domain->prot) {
1408 iommu_detach_group(domain->domain, iommu_group);
1409 if (!iommu_attach_group(d->domain, iommu_group)) {
1410 list_add(&group->next, &d->group_list);
1411 iommu_domain_free(domain->domain);
1413 mutex_unlock(&iommu->lock);
1417 ret = iommu_attach_group(domain->domain, iommu_group);
1423 vfio_test_domain_fgsp(domain);
1425 /* replay mappings on new domains */
1426 ret = vfio_iommu_replay(iommu, domain);
1431 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
1436 list_add(&domain->next, &iommu->domain_list);
1438 mutex_unlock(&iommu->lock);
1443 iommu_detach_group(domain->domain, iommu_group);
1445 iommu_domain_free(domain->domain);
1449 mutex_unlock(&iommu->lock);
1453 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
1455 struct rb_node *node;
1457 while ((node = rb_first(&iommu->dma_list)))
1458 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
1461 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
1463 struct rb_node *n, *p;
1465 n = rb_first(&iommu->dma_list);
1466 for (; n; n = rb_next(n)) {
1467 struct vfio_dma *dma;
1468 long locked = 0, unlocked = 0;
1470 dma = rb_entry(n, struct vfio_dma, node);
1471 unlocked += vfio_unmap_unpin(iommu, dma, false);
1472 p = rb_first(&dma->pfn_list);
1473 for (; p; p = rb_next(p)) {
1474 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
1477 if (!is_invalid_reserved_pfn(vpfn->pfn))
1480 vfio_lock_acct(dma, locked - unlocked, true);
1484 static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu)
1488 n = rb_first(&iommu->dma_list);
1489 for (; n; n = rb_next(n)) {
1490 struct vfio_dma *dma;
1492 dma = rb_entry(n, struct vfio_dma, node);
1494 if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list)))
1497 /* mdev vendor driver must unregister notifier */
1498 WARN_ON(iommu->notifier.head);
1501 static void vfio_iommu_type1_detach_group(void *iommu_data,
1502 struct iommu_group *iommu_group)
1504 struct vfio_iommu *iommu = iommu_data;
1505 struct vfio_domain *domain;
1506 struct vfio_group *group;
1508 mutex_lock(&iommu->lock);
1510 if (iommu->external_domain) {
1511 group = find_iommu_group(iommu->external_domain, iommu_group);
1513 list_del(&group->next);
1516 if (list_empty(&iommu->external_domain->group_list)) {
1517 vfio_sanity_check_pfn_list(iommu);
1519 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1520 vfio_iommu_unmap_unpin_all(iommu);
1522 kfree(iommu->external_domain);
1523 iommu->external_domain = NULL;
1525 goto detach_group_done;
1529 list_for_each_entry(domain, &iommu->domain_list, next) {
1530 group = find_iommu_group(domain, iommu_group);
1534 iommu_detach_group(domain->domain, iommu_group);
1535 list_del(&group->next);
1538 * Group ownership provides privilege, if the group list is
1539 * empty, the domain goes away. If it's the last domain with
1540 * iommu and external domain doesn't exist, then all the
1541 * mappings go away too. If it's the last domain with iommu and
1542 * external domain exist, update accounting
1544 if (list_empty(&domain->group_list)) {
1545 if (list_is_singular(&iommu->domain_list)) {
1546 if (!iommu->external_domain)
1547 vfio_iommu_unmap_unpin_all(iommu);
1549 vfio_iommu_unmap_unpin_reaccount(iommu);
1551 iommu_domain_free(domain->domain);
1552 list_del(&domain->next);
1559 mutex_unlock(&iommu->lock);
1562 static void *vfio_iommu_type1_open(unsigned long arg)
1564 struct vfio_iommu *iommu;
1566 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
1568 return ERR_PTR(-ENOMEM);
1571 case VFIO_TYPE1_IOMMU:
1573 case VFIO_TYPE1_NESTING_IOMMU:
1574 iommu->nesting = true;
1576 case VFIO_TYPE1v2_IOMMU:
1581 return ERR_PTR(-EINVAL);
1584 INIT_LIST_HEAD(&iommu->domain_list);
1585 iommu->dma_list = RB_ROOT;
1586 mutex_init(&iommu->lock);
1587 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
1592 static void vfio_release_domain(struct vfio_domain *domain, bool external)
1594 struct vfio_group *group, *group_tmp;
1596 list_for_each_entry_safe(group, group_tmp,
1597 &domain->group_list, next) {
1599 iommu_detach_group(domain->domain, group->iommu_group);
1600 list_del(&group->next);
1605 iommu_domain_free(domain->domain);
1608 static void vfio_iommu_type1_release(void *iommu_data)
1610 struct vfio_iommu *iommu = iommu_data;
1611 struct vfio_domain *domain, *domain_tmp;
1613 if (iommu->external_domain) {
1614 vfio_release_domain(iommu->external_domain, true);
1615 vfio_sanity_check_pfn_list(iommu);
1616 kfree(iommu->external_domain);
1619 vfio_iommu_unmap_unpin_all(iommu);
1621 list_for_each_entry_safe(domain, domain_tmp,
1622 &iommu->domain_list, next) {
1623 vfio_release_domain(domain, false);
1624 list_del(&domain->next);
1630 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
1632 struct vfio_domain *domain;
1635 mutex_lock(&iommu->lock);
1636 list_for_each_entry(domain, &iommu->domain_list, next) {
1637 if (!(domain->prot & IOMMU_CACHE)) {
1642 mutex_unlock(&iommu->lock);
1647 static long vfio_iommu_type1_ioctl(void *iommu_data,
1648 unsigned int cmd, unsigned long arg)
1650 struct vfio_iommu *iommu = iommu_data;
1651 unsigned long minsz;
1653 if (cmd == VFIO_CHECK_EXTENSION) {
1655 case VFIO_TYPE1_IOMMU:
1656 case VFIO_TYPE1v2_IOMMU:
1657 case VFIO_TYPE1_NESTING_IOMMU:
1659 case VFIO_DMA_CC_IOMMU:
1662 return vfio_domains_have_iommu_cache(iommu);
1666 } else if (cmd == VFIO_IOMMU_GET_INFO) {
1667 struct vfio_iommu_type1_info info;
1669 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
1671 if (copy_from_user(&info, (void __user *)arg, minsz))
1674 if (info.argsz < minsz)
1677 info.flags = VFIO_IOMMU_INFO_PGSIZES;
1679 info.iova_pgsizes = vfio_pgsize_bitmap(iommu);
1681 return copy_to_user((void __user *)arg, &info, minsz) ?
1684 } else if (cmd == VFIO_IOMMU_MAP_DMA) {
1685 struct vfio_iommu_type1_dma_map map;
1686 uint32_t mask = VFIO_DMA_MAP_FLAG_READ |
1687 VFIO_DMA_MAP_FLAG_WRITE;
1689 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
1691 if (copy_from_user(&map, (void __user *)arg, minsz))
1694 if (map.argsz < minsz || map.flags & ~mask)
1697 return vfio_dma_do_map(iommu, &map);
1699 } else if (cmd == VFIO_IOMMU_UNMAP_DMA) {
1700 struct vfio_iommu_type1_dma_unmap unmap;
1703 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
1705 if (copy_from_user(&unmap, (void __user *)arg, minsz))
1708 if (unmap.argsz < minsz || unmap.flags)
1711 ret = vfio_dma_do_unmap(iommu, &unmap);
1715 return copy_to_user((void __user *)arg, &unmap, minsz) ?
1722 static int vfio_iommu_type1_register_notifier(void *iommu_data,
1723 unsigned long *events,
1724 struct notifier_block *nb)
1726 struct vfio_iommu *iommu = iommu_data;
1728 /* clear known events */
1729 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
1731 /* refuse to register if still events remaining */
1735 return blocking_notifier_chain_register(&iommu->notifier, nb);
1738 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
1739 struct notifier_block *nb)
1741 struct vfio_iommu *iommu = iommu_data;
1743 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
1746 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
1747 .name = "vfio-iommu-type1",
1748 .owner = THIS_MODULE,
1749 .open = vfio_iommu_type1_open,
1750 .release = vfio_iommu_type1_release,
1751 .ioctl = vfio_iommu_type1_ioctl,
1752 .attach_group = vfio_iommu_type1_attach_group,
1753 .detach_group = vfio_iommu_type1_detach_group,
1754 .pin_pages = vfio_iommu_type1_pin_pages,
1755 .unpin_pages = vfio_iommu_type1_unpin_pages,
1756 .register_notifier = vfio_iommu_type1_register_notifier,
1757 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
1760 static int __init vfio_iommu_type1_init(void)
1762 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
1765 static void __exit vfio_iommu_type1_cleanup(void)
1767 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
1770 module_init(vfio_iommu_type1_init);
1771 module_exit(vfio_iommu_type1_cleanup);
1773 MODULE_VERSION(DRIVER_VERSION);
1774 MODULE_LICENSE("GPL v2");
1775 MODULE_AUTHOR(DRIVER_AUTHOR);
1776 MODULE_DESCRIPTION(DRIVER_DESC);