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];
348 if (mm == current->mm) {
349 ret = get_user_pages_longterm(vaddr, 1, !!(prot & IOMMU_WRITE),
352 unsigned int flags = 0;
354 if (prot & IOMMU_WRITE)
357 down_read(&mm->mmap_sem);
358 ret = get_user_pages_remote(NULL, mm, vaddr, 1, flags, page,
361 * The lifetime of a vaddr_get_pfn() page pin is
362 * userspace-controlled. In the fs-dax case this could
363 * lead to indefinite stalls in filesystem operations.
364 * Disallow attempts to pin fs-dax pages via this
367 if (ret > 0 && vma_is_fsdax(vmas[0])) {
371 up_read(&mm->mmap_sem);
375 *pfn = page_to_pfn(page[0]);
379 down_read(&mm->mmap_sem);
381 vma = find_vma_intersection(mm, vaddr, vaddr + 1);
383 if (vma && vma->vm_flags & VM_PFNMAP) {
384 *pfn = ((vaddr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
385 if (is_invalid_reserved_pfn(*pfn))
389 up_read(&mm->mmap_sem);
394 * Attempt to pin pages. We really don't want to track all the pfns and
395 * the iommu can only map chunks of consecutive pfns anyway, so get the
396 * first page and all consecutive pages with the same locking.
398 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
399 long npage, unsigned long *pfn_base,
402 unsigned long pfn = 0;
403 long ret, pinned = 0, lock_acct = 0;
405 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
407 /* This code path is only user initiated */
411 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base);
416 rsvd = is_invalid_reserved_pfn(*pfn_base);
419 * Reserved pages aren't counted against the user, externally pinned
420 * pages are already counted against the user.
422 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
423 if (!dma->lock_cap && current->mm->locked_vm + 1 > limit) {
424 put_pfn(*pfn_base, dma->prot);
425 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__,
426 limit << PAGE_SHIFT);
432 if (unlikely(disable_hugepages))
435 /* Lock all the consecutive pages from pfn_base */
436 for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage;
437 pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) {
438 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn);
442 if (pfn != *pfn_base + pinned ||
443 rsvd != is_invalid_reserved_pfn(pfn)) {
444 put_pfn(pfn, dma->prot);
448 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
449 if (!dma->lock_cap &&
450 current->mm->locked_vm + lock_acct + 1 > limit) {
451 put_pfn(pfn, dma->prot);
452 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
453 __func__, limit << PAGE_SHIFT);
462 ret = vfio_lock_acct(dma, lock_acct, false);
467 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
468 put_pfn(pfn, dma->prot);
477 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
478 unsigned long pfn, long npage,
481 long unlocked = 0, locked = 0;
484 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
485 if (put_pfn(pfn++, dma->prot)) {
487 if (vfio_find_vpfn(dma, iova))
493 vfio_lock_acct(dma, locked - unlocked, true);
498 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
499 unsigned long *pfn_base, bool do_accounting)
501 struct mm_struct *mm;
504 mm = get_task_mm(dma->task);
508 ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base);
509 if (!ret && do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
510 ret = vfio_lock_acct(dma, 1, true);
512 put_pfn(*pfn_base, dma->prot);
514 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
515 "(%ld) exceeded\n", __func__,
516 dma->task->comm, task_pid_nr(dma->task),
517 task_rlimit(dma->task, RLIMIT_MEMLOCK));
525 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
529 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
534 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
537 vfio_lock_acct(dma, -unlocked, true);
542 static int vfio_iommu_type1_pin_pages(void *iommu_data,
543 unsigned long *user_pfn,
545 unsigned long *phys_pfn)
547 struct vfio_iommu *iommu = iommu_data;
549 unsigned long remote_vaddr;
550 struct vfio_dma *dma;
553 if (!iommu || !user_pfn || !phys_pfn)
556 /* Supported for v2 version only */
560 mutex_lock(&iommu->lock);
562 /* Fail if notifier list is empty */
563 if ((!iommu->external_domain) || (!iommu->notifier.head)) {
569 * If iommu capable domain exist in the container then all pages are
570 * already pinned and accounted. Accouting should be done if there is no
571 * iommu capable domain in the container.
573 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
575 for (i = 0; i < npage; i++) {
577 struct vfio_pfn *vpfn;
579 iova = user_pfn[i] << PAGE_SHIFT;
580 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
586 if ((dma->prot & prot) != prot) {
591 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
593 phys_pfn[i] = vpfn->pfn;
597 remote_vaddr = dma->vaddr + iova - dma->iova;
598 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
603 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
605 vfio_unpin_page_external(dma, iova, do_accounting);
615 for (j = 0; j < i; j++) {
618 iova = user_pfn[j] << PAGE_SHIFT;
619 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
620 vfio_unpin_page_external(dma, iova, do_accounting);
624 mutex_unlock(&iommu->lock);
628 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
629 unsigned long *user_pfn,
632 struct vfio_iommu *iommu = iommu_data;
636 if (!iommu || !user_pfn)
639 /* Supported for v2 version only */
643 mutex_lock(&iommu->lock);
645 if (!iommu->external_domain) {
646 mutex_unlock(&iommu->lock);
650 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
651 for (i = 0; i < npage; i++) {
652 struct vfio_dma *dma;
655 iova = user_pfn[i] << PAGE_SHIFT;
656 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
659 vfio_unpin_page_external(dma, iova, do_accounting);
663 mutex_unlock(&iommu->lock);
664 return i > npage ? npage : (i > 0 ? i : -EINVAL);
667 static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
668 struct list_head *regions)
671 struct vfio_regions *entry, *next;
673 iommu_tlb_sync(domain->domain);
675 list_for_each_entry_safe(entry, next, regions, list) {
676 unlocked += vfio_unpin_pages_remote(dma,
678 entry->phys >> PAGE_SHIFT,
679 entry->len >> PAGE_SHIFT,
681 list_del(&entry->list);
691 * Generally, VFIO needs to unpin remote pages after each IOTLB flush.
692 * Therefore, when using IOTLB flush sync interface, VFIO need to keep track
693 * of these regions (currently using a list).
695 * This value specifies maximum number of regions for each IOTLB flush sync.
697 #define VFIO_IOMMU_TLB_SYNC_MAX 512
699 static size_t unmap_unpin_fast(struct vfio_domain *domain,
700 struct vfio_dma *dma, dma_addr_t *iova,
701 size_t len, phys_addr_t phys, long *unlocked,
702 struct list_head *unmapped_list,
706 struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
709 unmapped = iommu_unmap_fast(domain->domain, *iova, len);
714 iommu_tlb_range_add(domain->domain, *iova, unmapped);
717 entry->len = unmapped;
718 list_add_tail(&entry->list, unmapped_list);
726 * Sync if the number of fast-unmap regions hits the limit
727 * or in case of errors.
729 if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
730 *unlocked += vfio_sync_unpin(dma, domain,
738 static size_t unmap_unpin_slow(struct vfio_domain *domain,
739 struct vfio_dma *dma, dma_addr_t *iova,
740 size_t len, phys_addr_t phys,
743 size_t unmapped = iommu_unmap(domain->domain, *iova, len);
746 *unlocked += vfio_unpin_pages_remote(dma, *iova,
748 unmapped >> PAGE_SHIFT,
756 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
759 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
760 struct vfio_domain *domain, *d;
761 LIST_HEAD(unmapped_region_list);
762 int unmapped_region_cnt = 0;
768 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
772 * We use the IOMMU to track the physical addresses, otherwise we'd
773 * need a much more complicated tracking system. Unfortunately that
774 * means we need to use one of the iommu domains to figure out the
775 * pfns to unpin. The rest need to be unmapped in advance so we have
776 * no iommu translations remaining when the pages are unpinned.
778 domain = d = list_first_entry(&iommu->domain_list,
779 struct vfio_domain, next);
781 list_for_each_entry_continue(d, &iommu->domain_list, next) {
782 iommu_unmap(d->domain, dma->iova, dma->size);
787 size_t unmapped, len;
788 phys_addr_t phys, next;
790 phys = iommu_iova_to_phys(domain->domain, iova);
791 if (WARN_ON(!phys)) {
797 * To optimize for fewer iommu_unmap() calls, each of which
798 * may require hardware cache flushing, try to find the
799 * largest contiguous physical memory chunk to unmap.
801 for (len = PAGE_SIZE;
802 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
803 next = iommu_iova_to_phys(domain->domain, iova + len);
804 if (next != phys + len)
809 * First, try to use fast unmap/unpin. In case of failure,
810 * switch to slow unmap/unpin path.
812 unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys,
813 &unlocked, &unmapped_region_list,
814 &unmapped_region_cnt);
816 unmapped = unmap_unpin_slow(domain, dma, &iova, len,
818 if (WARN_ON(!unmapped))
823 dma->iommu_mapped = false;
825 if (unmapped_region_cnt)
826 unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list);
829 vfio_lock_acct(dma, -unlocked, true);
835 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
837 vfio_unmap_unpin(iommu, dma, true);
838 vfio_unlink_dma(iommu, dma);
839 put_task_struct(dma->task);
843 static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu)
845 struct vfio_domain *domain;
846 unsigned long bitmap = ULONG_MAX;
848 mutex_lock(&iommu->lock);
849 list_for_each_entry(domain, &iommu->domain_list, next)
850 bitmap &= domain->domain->pgsize_bitmap;
851 mutex_unlock(&iommu->lock);
854 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
855 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
856 * That way the user will be able to map/unmap buffers whose size/
857 * start address is aligned with PAGE_SIZE. Pinning code uses that
858 * granularity while iommu driver can use the sub-PAGE_SIZE size
861 if (bitmap & ~PAGE_MASK) {
869 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
870 struct vfio_iommu_type1_dma_unmap *unmap)
873 struct vfio_dma *dma, *dma_last = NULL;
875 int ret = 0, retries = 0;
877 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
879 if (unmap->iova & mask)
881 if (!unmap->size || unmap->size & mask)
883 if (unmap->iova + unmap->size < unmap->iova ||
884 unmap->size > SIZE_MAX)
887 WARN_ON(mask & PAGE_MASK);
889 mutex_lock(&iommu->lock);
892 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
893 * avoid tracking individual mappings. This means that the granularity
894 * of the original mapping was lost and the user was allowed to attempt
895 * to unmap any range. Depending on the contiguousness of physical
896 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
897 * or may not have worked. We only guaranteed unmap granularity
898 * matching the original mapping; even though it was untracked here,
899 * the original mappings are reflected in IOMMU mappings. This
900 * resulted in a couple unusual behaviors. First, if a range is not
901 * able to be unmapped, ex. a set of 4k pages that was mapped as a
902 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
903 * a zero sized unmap. Also, if an unmap request overlaps the first
904 * address of a hugepage, the IOMMU will unmap the entire hugepage.
905 * This also returns success and the returned unmap size reflects the
906 * actual size unmapped.
908 * We attempt to maintain compatibility with this "v1" interface, but
909 * we take control out of the hands of the IOMMU. Therefore, an unmap
910 * request offset from the beginning of the original mapping will
911 * return success with zero sized unmap. And an unmap request covering
912 * the first iova of mapping will unmap the entire range.
914 * The v2 version of this interface intends to be more deterministic.
915 * Unmap requests must fully cover previous mappings. Multiple
916 * mappings may still be unmaped by specifying large ranges, but there
917 * must not be any previous mappings bisected by the range. An error
918 * will be returned if these conditions are not met. The v2 interface
919 * will only return success and a size of zero if there were no
920 * mappings within the range.
923 dma = vfio_find_dma(iommu, unmap->iova, 1);
924 if (dma && dma->iova != unmap->iova) {
928 dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0);
929 if (dma && dma->iova + dma->size != unmap->iova + unmap->size) {
935 while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
936 if (!iommu->v2 && unmap->iova > dma->iova)
939 * Task with same address space who mapped this iova range is
940 * allowed to unmap the iova range.
942 if (dma->task->mm != current->mm)
945 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
946 struct vfio_iommu_type1_dma_unmap nb_unmap;
948 if (dma_last == dma) {
949 BUG_ON(++retries > 10);
955 nb_unmap.iova = dma->iova;
956 nb_unmap.size = dma->size;
959 * Notify anyone (mdev vendor drivers) to invalidate and
960 * unmap iovas within the range we're about to unmap.
961 * Vendor drivers MUST unpin pages in response to an
964 mutex_unlock(&iommu->lock);
965 blocking_notifier_call_chain(&iommu->notifier,
966 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
970 unmapped += dma->size;
971 vfio_remove_dma(iommu, dma);
975 mutex_unlock(&iommu->lock);
977 /* Report how much was unmapped */
978 unmap->size = unmapped;
984 * Turns out AMD IOMMU has a page table bug where it won't map large pages
985 * to a region that previously mapped smaller pages. This should be fixed
986 * soon, so this is just a temporary workaround to break mappings down into
987 * PAGE_SIZE. Better to map smaller pages than nothing.
989 static int map_try_harder(struct vfio_domain *domain, dma_addr_t iova,
990 unsigned long pfn, long npage, int prot)
995 for (i = 0; i < npage; i++, pfn++, iova += PAGE_SIZE) {
996 ret = iommu_map(domain->domain, iova,
997 (phys_addr_t)pfn << PAGE_SHIFT,
998 PAGE_SIZE, prot | domain->prot);
1003 for (; i < npage && i > 0; i--, iova -= PAGE_SIZE)
1004 iommu_unmap(domain->domain, iova, PAGE_SIZE);
1009 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
1010 unsigned long pfn, long npage, int prot)
1012 struct vfio_domain *d;
1015 list_for_each_entry(d, &iommu->domain_list, next) {
1016 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
1017 npage << PAGE_SHIFT, prot | d->prot);
1019 if (ret != -EBUSY ||
1020 map_try_harder(d, iova, pfn, npage, prot))
1030 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next)
1031 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
1036 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
1039 dma_addr_t iova = dma->iova;
1040 unsigned long vaddr = dma->vaddr;
1041 size_t size = map_size;
1043 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1047 /* Pin a contiguous chunk of memory */
1048 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
1049 size >> PAGE_SHIFT, &pfn, limit);
1057 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
1060 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
1065 size -= npage << PAGE_SHIFT;
1066 dma->size += npage << PAGE_SHIFT;
1069 dma->iommu_mapped = true;
1072 vfio_remove_dma(iommu, dma);
1077 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1078 struct vfio_iommu_type1_dma_map *map)
1080 dma_addr_t iova = map->iova;
1081 unsigned long vaddr = map->vaddr;
1082 size_t size = map->size;
1083 int ret = 0, prot = 0;
1085 struct vfio_dma *dma;
1087 /* Verify that none of our __u64 fields overflow */
1088 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1091 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
1093 WARN_ON(mask & PAGE_MASK);
1095 /* READ/WRITE from device perspective */
1096 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1097 prot |= IOMMU_WRITE;
1098 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1101 if (!prot || !size || (size | iova | vaddr) & mask)
1104 /* Don't allow IOVA or virtual address wrap */
1105 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr)
1108 mutex_lock(&iommu->lock);
1110 if (vfio_find_dma(iommu, iova, size)) {
1115 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1126 * We need to be able to both add to a task's locked memory and test
1127 * against the locked memory limit and we need to be able to do both
1128 * outside of this call path as pinning can be asynchronous via the
1129 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1130 * task_struct and VM locked pages requires an mm_struct, however
1131 * holding an indefinite mm reference is not recommended, therefore we
1132 * only hold a reference to a task. We could hold a reference to
1133 * current, however QEMU uses this call path through vCPU threads,
1134 * which can be killed resulting in a NULL mm and failure in the unmap
1135 * path when called via a different thread. Avoid this problem by
1136 * using the group_leader as threads within the same group require
1137 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1140 * Previously we also used the task for testing CAP_IPC_LOCK at the
1141 * time of pinning and accounting, however has_capability() makes use
1142 * of real_cred, a copy-on-write field, so we can't guarantee that it
1143 * matches group_leader, or in fact that it might not change by the
1144 * time it's evaluated. If a process were to call MAP_DMA with
1145 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1146 * possibly see different results for an iommu_mapped vfio_dma vs
1147 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1148 * time of calling MAP_DMA.
1150 get_task_struct(current->group_leader);
1151 dma->task = current->group_leader;
1152 dma->lock_cap = capable(CAP_IPC_LOCK);
1154 dma->pfn_list = RB_ROOT;
1156 /* Insert zero-sized and grow as we map chunks of it */
1157 vfio_link_dma(iommu, dma);
1159 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1160 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1163 ret = vfio_pin_map_dma(iommu, dma, size);
1166 mutex_unlock(&iommu->lock);
1170 static int vfio_bus_type(struct device *dev, void *data)
1172 struct bus_type **bus = data;
1174 if (*bus && *bus != dev->bus)
1182 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1183 struct vfio_domain *domain)
1185 struct vfio_domain *d;
1187 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1190 /* Arbitrarily pick the first domain in the list for lookups */
1191 d = list_first_entry(&iommu->domain_list, struct vfio_domain, next);
1192 n = rb_first(&iommu->dma_list);
1194 for (; n; n = rb_next(n)) {
1195 struct vfio_dma *dma;
1198 dma = rb_entry(n, struct vfio_dma, node);
1201 while (iova < dma->iova + dma->size) {
1205 if (dma->iommu_mapped) {
1209 phys = iommu_iova_to_phys(d->domain, iova);
1211 if (WARN_ON(!phys)) {
1219 while (i < dma->iova + dma->size &&
1220 p == iommu_iova_to_phys(d->domain, i)) {
1227 unsigned long vaddr = dma->vaddr +
1229 size_t n = dma->iova + dma->size - iova;
1232 npage = vfio_pin_pages_remote(dma, vaddr,
1241 phys = pfn << PAGE_SHIFT;
1242 size = npage << PAGE_SHIFT;
1245 ret = iommu_map(domain->domain, iova, phys,
1246 size, dma->prot | domain->prot);
1252 dma->iommu_mapped = true;
1258 * We change our unmap behavior slightly depending on whether the IOMMU
1259 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1260 * for practically any contiguous power-of-two mapping we give it. This means
1261 * we don't need to look for contiguous chunks ourselves to make unmapping
1262 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1263 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1264 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1265 * hugetlbfs is in use.
1267 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1270 int ret, order = get_order(PAGE_SIZE * 2);
1272 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1276 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1277 IOMMU_READ | IOMMU_WRITE | domain->prot);
1279 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1281 if (unmapped == PAGE_SIZE)
1282 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1284 domain->fgsp = true;
1287 __free_pages(pages, order);
1290 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1291 struct iommu_group *iommu_group)
1293 struct vfio_group *g;
1295 list_for_each_entry(g, &domain->group_list, next) {
1296 if (g->iommu_group == iommu_group)
1303 static bool vfio_iommu_has_sw_msi(struct iommu_group *group, phys_addr_t *base)
1305 struct list_head group_resv_regions;
1306 struct iommu_resv_region *region, *next;
1309 INIT_LIST_HEAD(&group_resv_regions);
1310 iommu_get_group_resv_regions(group, &group_resv_regions);
1311 list_for_each_entry(region, &group_resv_regions, list) {
1313 * The presence of any 'real' MSI regions should take
1314 * precedence over the software-managed one if the
1315 * IOMMU driver happens to advertise both types.
1317 if (region->type == IOMMU_RESV_MSI) {
1322 if (region->type == IOMMU_RESV_SW_MSI) {
1323 *base = region->start;
1327 list_for_each_entry_safe(region, next, &group_resv_regions, list)
1332 static int vfio_iommu_type1_attach_group(void *iommu_data,
1333 struct iommu_group *iommu_group)
1335 struct vfio_iommu *iommu = iommu_data;
1336 struct vfio_group *group;
1337 struct vfio_domain *domain, *d;
1338 struct bus_type *bus = NULL, *mdev_bus;
1340 bool resv_msi, msi_remap;
1341 phys_addr_t resv_msi_base;
1343 mutex_lock(&iommu->lock);
1345 list_for_each_entry(d, &iommu->domain_list, next) {
1346 if (find_iommu_group(d, iommu_group)) {
1347 mutex_unlock(&iommu->lock);
1352 if (iommu->external_domain) {
1353 if (find_iommu_group(iommu->external_domain, iommu_group)) {
1354 mutex_unlock(&iommu->lock);
1359 group = kzalloc(sizeof(*group), GFP_KERNEL);
1360 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
1361 if (!group || !domain) {
1366 group->iommu_group = iommu_group;
1368 /* Determine bus_type in order to allocate a domain */
1369 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
1373 mdev_bus = symbol_get(mdev_bus_type);
1376 if ((bus == mdev_bus) && !iommu_present(bus)) {
1377 symbol_put(mdev_bus_type);
1378 if (!iommu->external_domain) {
1379 INIT_LIST_HEAD(&domain->group_list);
1380 iommu->external_domain = domain;
1384 list_add(&group->next,
1385 &iommu->external_domain->group_list);
1386 mutex_unlock(&iommu->lock);
1389 symbol_put(mdev_bus_type);
1392 domain->domain = iommu_domain_alloc(bus);
1393 if (!domain->domain) {
1398 if (iommu->nesting) {
1401 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
1407 ret = iommu_attach_group(domain->domain, iommu_group);
1411 resv_msi = vfio_iommu_has_sw_msi(iommu_group, &resv_msi_base);
1413 INIT_LIST_HEAD(&domain->group_list);
1414 list_add(&group->next, &domain->group_list);
1416 msi_remap = irq_domain_check_msi_remap() ||
1417 iommu_capable(bus, IOMMU_CAP_INTR_REMAP);
1419 if (!allow_unsafe_interrupts && !msi_remap) {
1420 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
1426 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
1427 domain->prot |= IOMMU_CACHE;
1430 * Try to match an existing compatible domain. We don't want to
1431 * preclude an IOMMU driver supporting multiple bus_types and being
1432 * able to include different bus_types in the same IOMMU domain, so
1433 * we test whether the domains use the same iommu_ops rather than
1434 * testing if they're on the same bus_type.
1436 list_for_each_entry(d, &iommu->domain_list, next) {
1437 if (d->domain->ops == domain->domain->ops &&
1438 d->prot == domain->prot) {
1439 iommu_detach_group(domain->domain, iommu_group);
1440 if (!iommu_attach_group(d->domain, iommu_group)) {
1441 list_add(&group->next, &d->group_list);
1442 iommu_domain_free(domain->domain);
1444 mutex_unlock(&iommu->lock);
1448 ret = iommu_attach_group(domain->domain, iommu_group);
1454 vfio_test_domain_fgsp(domain);
1456 /* replay mappings on new domains */
1457 ret = vfio_iommu_replay(iommu, domain);
1462 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
1467 list_add(&domain->next, &iommu->domain_list);
1469 mutex_unlock(&iommu->lock);
1474 iommu_detach_group(domain->domain, iommu_group);
1476 iommu_domain_free(domain->domain);
1480 mutex_unlock(&iommu->lock);
1484 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
1486 struct rb_node *node;
1488 while ((node = rb_first(&iommu->dma_list)))
1489 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
1492 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
1494 struct rb_node *n, *p;
1496 n = rb_first(&iommu->dma_list);
1497 for (; n; n = rb_next(n)) {
1498 struct vfio_dma *dma;
1499 long locked = 0, unlocked = 0;
1501 dma = rb_entry(n, struct vfio_dma, node);
1502 unlocked += vfio_unmap_unpin(iommu, dma, false);
1503 p = rb_first(&dma->pfn_list);
1504 for (; p; p = rb_next(p)) {
1505 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
1508 if (!is_invalid_reserved_pfn(vpfn->pfn))
1511 vfio_lock_acct(dma, locked - unlocked, true);
1515 static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu)
1519 n = rb_first(&iommu->dma_list);
1520 for (; n; n = rb_next(n)) {
1521 struct vfio_dma *dma;
1523 dma = rb_entry(n, struct vfio_dma, node);
1525 if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list)))
1528 /* mdev vendor driver must unregister notifier */
1529 WARN_ON(iommu->notifier.head);
1532 static void vfio_iommu_type1_detach_group(void *iommu_data,
1533 struct iommu_group *iommu_group)
1535 struct vfio_iommu *iommu = iommu_data;
1536 struct vfio_domain *domain;
1537 struct vfio_group *group;
1539 mutex_lock(&iommu->lock);
1541 if (iommu->external_domain) {
1542 group = find_iommu_group(iommu->external_domain, iommu_group);
1544 list_del(&group->next);
1547 if (list_empty(&iommu->external_domain->group_list)) {
1548 vfio_sanity_check_pfn_list(iommu);
1550 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1551 vfio_iommu_unmap_unpin_all(iommu);
1553 kfree(iommu->external_domain);
1554 iommu->external_domain = NULL;
1556 goto detach_group_done;
1560 list_for_each_entry(domain, &iommu->domain_list, next) {
1561 group = find_iommu_group(domain, iommu_group);
1565 iommu_detach_group(domain->domain, iommu_group);
1566 list_del(&group->next);
1569 * Group ownership provides privilege, if the group list is
1570 * empty, the domain goes away. If it's the last domain with
1571 * iommu and external domain doesn't exist, then all the
1572 * mappings go away too. If it's the last domain with iommu and
1573 * external domain exist, update accounting
1575 if (list_empty(&domain->group_list)) {
1576 if (list_is_singular(&iommu->domain_list)) {
1577 if (!iommu->external_domain)
1578 vfio_iommu_unmap_unpin_all(iommu);
1580 vfio_iommu_unmap_unpin_reaccount(iommu);
1582 iommu_domain_free(domain->domain);
1583 list_del(&domain->next);
1590 mutex_unlock(&iommu->lock);
1593 static void *vfio_iommu_type1_open(unsigned long arg)
1595 struct vfio_iommu *iommu;
1597 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
1599 return ERR_PTR(-ENOMEM);
1602 case VFIO_TYPE1_IOMMU:
1604 case VFIO_TYPE1_NESTING_IOMMU:
1605 iommu->nesting = true;
1606 case VFIO_TYPE1v2_IOMMU:
1611 return ERR_PTR(-EINVAL);
1614 INIT_LIST_HEAD(&iommu->domain_list);
1615 iommu->dma_list = RB_ROOT;
1616 mutex_init(&iommu->lock);
1617 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
1622 static void vfio_release_domain(struct vfio_domain *domain, bool external)
1624 struct vfio_group *group, *group_tmp;
1626 list_for_each_entry_safe(group, group_tmp,
1627 &domain->group_list, next) {
1629 iommu_detach_group(domain->domain, group->iommu_group);
1630 list_del(&group->next);
1635 iommu_domain_free(domain->domain);
1638 static void vfio_iommu_type1_release(void *iommu_data)
1640 struct vfio_iommu *iommu = iommu_data;
1641 struct vfio_domain *domain, *domain_tmp;
1643 if (iommu->external_domain) {
1644 vfio_release_domain(iommu->external_domain, true);
1645 vfio_sanity_check_pfn_list(iommu);
1646 kfree(iommu->external_domain);
1649 vfio_iommu_unmap_unpin_all(iommu);
1651 list_for_each_entry_safe(domain, domain_tmp,
1652 &iommu->domain_list, next) {
1653 vfio_release_domain(domain, false);
1654 list_del(&domain->next);
1660 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
1662 struct vfio_domain *domain;
1665 mutex_lock(&iommu->lock);
1666 list_for_each_entry(domain, &iommu->domain_list, next) {
1667 if (!(domain->prot & IOMMU_CACHE)) {
1672 mutex_unlock(&iommu->lock);
1677 static long vfio_iommu_type1_ioctl(void *iommu_data,
1678 unsigned int cmd, unsigned long arg)
1680 struct vfio_iommu *iommu = iommu_data;
1681 unsigned long minsz;
1683 if (cmd == VFIO_CHECK_EXTENSION) {
1685 case VFIO_TYPE1_IOMMU:
1686 case VFIO_TYPE1v2_IOMMU:
1687 case VFIO_TYPE1_NESTING_IOMMU:
1689 case VFIO_DMA_CC_IOMMU:
1692 return vfio_domains_have_iommu_cache(iommu);
1696 } else if (cmd == VFIO_IOMMU_GET_INFO) {
1697 struct vfio_iommu_type1_info info;
1699 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
1701 if (copy_from_user(&info, (void __user *)arg, minsz))
1704 if (info.argsz < minsz)
1707 info.flags = VFIO_IOMMU_INFO_PGSIZES;
1709 info.iova_pgsizes = vfio_pgsize_bitmap(iommu);
1711 return copy_to_user((void __user *)arg, &info, minsz) ?
1714 } else if (cmd == VFIO_IOMMU_MAP_DMA) {
1715 struct vfio_iommu_type1_dma_map map;
1716 uint32_t mask = VFIO_DMA_MAP_FLAG_READ |
1717 VFIO_DMA_MAP_FLAG_WRITE;
1719 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
1721 if (copy_from_user(&map, (void __user *)arg, minsz))
1724 if (map.argsz < minsz || map.flags & ~mask)
1727 return vfio_dma_do_map(iommu, &map);
1729 } else if (cmd == VFIO_IOMMU_UNMAP_DMA) {
1730 struct vfio_iommu_type1_dma_unmap unmap;
1733 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
1735 if (copy_from_user(&unmap, (void __user *)arg, minsz))
1738 if (unmap.argsz < minsz || unmap.flags)
1741 ret = vfio_dma_do_unmap(iommu, &unmap);
1745 return copy_to_user((void __user *)arg, &unmap, minsz) ?
1752 static int vfio_iommu_type1_register_notifier(void *iommu_data,
1753 unsigned long *events,
1754 struct notifier_block *nb)
1756 struct vfio_iommu *iommu = iommu_data;
1758 /* clear known events */
1759 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
1761 /* refuse to register if still events remaining */
1765 return blocking_notifier_chain_register(&iommu->notifier, nb);
1768 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
1769 struct notifier_block *nb)
1771 struct vfio_iommu *iommu = iommu_data;
1773 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
1776 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
1777 .name = "vfio-iommu-type1",
1778 .owner = THIS_MODULE,
1779 .open = vfio_iommu_type1_open,
1780 .release = vfio_iommu_type1_release,
1781 .ioctl = vfio_iommu_type1_ioctl,
1782 .attach_group = vfio_iommu_type1_attach_group,
1783 .detach_group = vfio_iommu_type1_detach_group,
1784 .pin_pages = vfio_iommu_type1_pin_pages,
1785 .unpin_pages = vfio_iommu_type1_unpin_pages,
1786 .register_notifier = vfio_iommu_type1_register_notifier,
1787 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
1790 static int __init vfio_iommu_type1_init(void)
1792 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
1795 static void __exit vfio_iommu_type1_cleanup(void)
1797 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
1800 module_init(vfio_iommu_type1_init);
1801 module_exit(vfio_iommu_type1_cleanup);
1803 MODULE_VERSION(DRIVER_VERSION);
1804 MODULE_LICENSE("GPL v2");
1805 MODULE_AUTHOR(DRIVER_AUTHOR);
1806 MODULE_DESCRIPTION(DRIVER_DESC);