1 // SPDX-License-Identifier: GPL-2.0-only
3 * VFIO: IOMMU DMA mapping support for Type1 IOMMU
5 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
8 * Derived from original vfio:
9 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
12 * We arbitrarily define a Type1 IOMMU as one matching the below code.
13 * It could be called the x86 IOMMU as it's designed for AMD-Vi & Intel
14 * VT-d, but that makes it harder to re-use as theoretically anyone
15 * implementing a similar IOMMU could make use of this. We expect the
16 * IOMMU to support the IOMMU API and have few to no restrictions around
17 * the IOVA range that can be mapped. The Type1 IOMMU is currently
18 * optimized for relatively static mappings of a userspace process with
19 * userpsace pages pinned into memory. We also assume devices and IOMMU
20 * domains are PCI based as the IOMMU API is still centered around a
21 * device/bus interface rather than a group interface.
24 #include <linux/compat.h>
25 #include <linux/device.h>
27 #include <linux/iommu.h>
28 #include <linux/module.h>
30 #include <linux/rbtree.h>
31 #include <linux/sched/signal.h>
32 #include <linux/sched/mm.h>
33 #include <linux/slab.h>
34 #include <linux/uaccess.h>
35 #include <linux/vfio.h>
36 #include <linux/workqueue.h>
37 #include <linux/mdev.h>
38 #include <linux/notifier.h>
39 #include <linux/dma-iommu.h>
40 #include <linux/irqdomain.h>
42 #define DRIVER_VERSION "0.2"
44 #define DRIVER_DESC "Type1 IOMMU driver for VFIO"
46 static bool allow_unsafe_interrupts;
47 module_param_named(allow_unsafe_interrupts,
48 allow_unsafe_interrupts, bool, S_IRUGO | S_IWUSR);
49 MODULE_PARM_DESC(allow_unsafe_interrupts,
50 "Enable VFIO IOMMU support for on platforms without interrupt remapping support.");
52 static bool disable_hugepages;
53 module_param_named(disable_hugepages,
54 disable_hugepages, bool, S_IRUGO | S_IWUSR);
55 MODULE_PARM_DESC(disable_hugepages,
56 "Disable VFIO IOMMU support for IOMMU hugepages.");
58 static unsigned int dma_entry_limit __read_mostly = U16_MAX;
59 module_param_named(dma_entry_limit, dma_entry_limit, uint, 0644);
60 MODULE_PARM_DESC(dma_entry_limit,
61 "Maximum number of user DMA mappings per container (65535).");
64 struct list_head domain_list;
65 struct list_head iova_list;
66 struct vfio_domain *external_domain; /* domain for external user */
68 struct rb_root dma_list;
69 struct blocking_notifier_head notifier;
70 unsigned int dma_avail;
76 struct iommu_domain *domain;
77 struct list_head next;
78 struct list_head group_list;
79 int prot; /* IOMMU_CACHE */
80 bool fgsp; /* Fine-grained super pages */
85 dma_addr_t iova; /* Device address */
86 unsigned long vaddr; /* Process virtual addr */
87 size_t size; /* Map size (bytes) */
88 int prot; /* IOMMU_READ/WRITE */
90 bool lock_cap; /* capable(CAP_IPC_LOCK) */
91 struct task_struct *task;
92 struct rb_root pfn_list; /* Ex-user pinned pfn list */
96 struct iommu_group *iommu_group;
97 struct list_head next;
98 bool mdev_group; /* An mdev group */
102 struct list_head list;
108 * Guest RAM pinning working set or DMA target
112 dma_addr_t iova; /* Device address */
113 unsigned long pfn; /* Host pfn */
117 struct vfio_regions {
118 struct list_head list;
124 #define IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu) \
125 (!list_empty(&iommu->domain_list))
127 static int put_pfn(unsigned long pfn, int prot);
130 * This code handles mapping and unmapping of user data buffers
131 * into DMA'ble space using the IOMMU
134 static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
135 dma_addr_t start, size_t size)
137 struct rb_node *node = iommu->dma_list.rb_node;
140 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
142 if (start + size <= dma->iova)
143 node = node->rb_left;
144 else if (start >= dma->iova + dma->size)
145 node = node->rb_right;
153 static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new)
155 struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL;
156 struct vfio_dma *dma;
160 dma = rb_entry(parent, struct vfio_dma, node);
162 if (new->iova + new->size <= dma->iova)
163 link = &(*link)->rb_left;
165 link = &(*link)->rb_right;
168 rb_link_node(&new->node, parent, link);
169 rb_insert_color(&new->node, &iommu->dma_list);
172 static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
174 rb_erase(&old->node, &iommu->dma_list);
178 * Helper Functions for host iova-pfn list
180 static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
182 struct vfio_pfn *vpfn;
183 struct rb_node *node = dma->pfn_list.rb_node;
186 vpfn = rb_entry(node, struct vfio_pfn, node);
188 if (iova < vpfn->iova)
189 node = node->rb_left;
190 else if (iova > vpfn->iova)
191 node = node->rb_right;
198 static void vfio_link_pfn(struct vfio_dma *dma,
199 struct vfio_pfn *new)
201 struct rb_node **link, *parent = NULL;
202 struct vfio_pfn *vpfn;
204 link = &dma->pfn_list.rb_node;
207 vpfn = rb_entry(parent, struct vfio_pfn, node);
209 if (new->iova < vpfn->iova)
210 link = &(*link)->rb_left;
212 link = &(*link)->rb_right;
215 rb_link_node(&new->node, parent, link);
216 rb_insert_color(&new->node, &dma->pfn_list);
219 static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old)
221 rb_erase(&old->node, &dma->pfn_list);
224 static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova,
227 struct vfio_pfn *vpfn;
229 vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL);
235 atomic_set(&vpfn->ref_count, 1);
236 vfio_link_pfn(dma, vpfn);
240 static void vfio_remove_from_pfn_list(struct vfio_dma *dma,
241 struct vfio_pfn *vpfn)
243 vfio_unlink_pfn(dma, vpfn);
247 static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma,
250 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
253 atomic_inc(&vpfn->ref_count);
257 static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn)
261 if (atomic_dec_and_test(&vpfn->ref_count)) {
262 ret = put_pfn(vpfn->pfn, dma->prot);
263 vfio_remove_from_pfn_list(dma, vpfn);
268 static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
270 struct mm_struct *mm;
276 mm = async ? get_task_mm(dma->task) : dma->task->mm;
278 return -ESRCH; /* process exited */
280 ret = down_write_killable(&mm->mmap_sem);
282 ret = __account_locked_vm(mm, abs(npage), npage > 0, dma->task,
284 up_write(&mm->mmap_sem);
294 * Some mappings aren't backed by a struct page, for example an mmap'd
295 * MMIO range for our own or another device. These use a different
296 * pfn conversion and shouldn't be tracked as locked pages.
298 static bool is_invalid_reserved_pfn(unsigned long pfn)
300 if (pfn_valid(pfn)) {
302 struct page *tail = pfn_to_page(pfn);
303 struct page *head = compound_head(tail);
304 reserved = !!(PageReserved(head));
307 * "head" is not a dangling pointer
308 * (compound_head takes care of that)
309 * but the hugepage may have been split
310 * from under us (and we may not hold a
311 * reference count on the head page so it can
312 * be reused before we run PageReferenced), so
313 * we've to check PageTail before returning
320 return PageReserved(tail);
326 static int put_pfn(unsigned long pfn, int prot)
328 if (!is_invalid_reserved_pfn(pfn)) {
329 struct page *page = pfn_to_page(pfn);
330 if (prot & IOMMU_WRITE)
338 static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,
339 int prot, unsigned long *pfn)
341 struct page *page[1];
342 struct vm_area_struct *vma;
343 struct vm_area_struct *vmas[1];
344 unsigned int flags = 0;
347 if (prot & IOMMU_WRITE)
350 down_read(&mm->mmap_sem);
351 if (mm == current->mm) {
352 ret = get_user_pages(vaddr, 1, flags | FOLL_LONGTERM, page,
355 ret = get_user_pages_remote(NULL, mm, vaddr, 1, flags, page,
358 * The lifetime of a vaddr_get_pfn() page pin is
359 * userspace-controlled. In the fs-dax case this could
360 * lead to indefinite stalls in filesystem operations.
361 * Disallow attempts to pin fs-dax pages via this
364 if (ret > 0 && vma_is_fsdax(vmas[0])) {
369 up_read(&mm->mmap_sem);
372 *pfn = page_to_pfn(page[0]);
376 down_read(&mm->mmap_sem);
378 vaddr = untagged_addr(vaddr);
380 vma = find_vma_intersection(mm, vaddr, vaddr + 1);
382 if (vma && vma->vm_flags & VM_PFNMAP) {
383 *pfn = ((vaddr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
384 if (is_invalid_reserved_pfn(*pfn))
388 up_read(&mm->mmap_sem);
393 * Attempt to pin pages. We really don't want to track all the pfns and
394 * the iommu can only map chunks of consecutive pfns anyway, so get the
395 * first page and all consecutive pages with the same locking.
397 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
398 long npage, unsigned long *pfn_base,
401 unsigned long pfn = 0;
402 long ret, pinned = 0, lock_acct = 0;
404 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
406 /* This code path is only user initiated */
410 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base);
415 rsvd = is_invalid_reserved_pfn(*pfn_base);
418 * Reserved pages aren't counted against the user, externally pinned
419 * pages are already counted against the user.
421 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
422 if (!dma->lock_cap && current->mm->locked_vm + 1 > limit) {
423 put_pfn(*pfn_base, dma->prot);
424 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__,
425 limit << PAGE_SHIFT);
431 if (unlikely(disable_hugepages))
434 /* Lock all the consecutive pages from pfn_base */
435 for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage;
436 pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) {
437 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn);
441 if (pfn != *pfn_base + pinned ||
442 rsvd != is_invalid_reserved_pfn(pfn)) {
443 put_pfn(pfn, dma->prot);
447 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
448 if (!dma->lock_cap &&
449 current->mm->locked_vm + lock_acct + 1 > limit) {
450 put_pfn(pfn, dma->prot);
451 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
452 __func__, limit << PAGE_SHIFT);
461 ret = vfio_lock_acct(dma, lock_acct, false);
466 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
467 put_pfn(pfn, dma->prot);
476 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
477 unsigned long pfn, long npage,
480 long unlocked = 0, locked = 0;
483 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
484 if (put_pfn(pfn++, dma->prot)) {
486 if (vfio_find_vpfn(dma, iova))
492 vfio_lock_acct(dma, locked - unlocked, true);
497 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
498 unsigned long *pfn_base, bool do_accounting)
500 struct mm_struct *mm;
503 mm = get_task_mm(dma->task);
507 ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base);
508 if (!ret && do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
509 ret = vfio_lock_acct(dma, 1, true);
511 put_pfn(*pfn_base, dma->prot);
513 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
514 "(%ld) exceeded\n", __func__,
515 dma->task->comm, task_pid_nr(dma->task),
516 task_rlimit(dma->task, RLIMIT_MEMLOCK));
524 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
528 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
533 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
536 vfio_lock_acct(dma, -unlocked, true);
541 static int vfio_iommu_type1_pin_pages(void *iommu_data,
542 unsigned long *user_pfn,
544 unsigned long *phys_pfn)
546 struct vfio_iommu *iommu = iommu_data;
548 unsigned long remote_vaddr;
549 struct vfio_dma *dma;
552 if (!iommu || !user_pfn || !phys_pfn)
555 /* Supported for v2 version only */
559 mutex_lock(&iommu->lock);
561 /* Fail if notifier list is empty */
562 if (!iommu->notifier.head) {
568 * If iommu capable domain exist in the container then all pages are
569 * already pinned and accounted. Accouting should be done if there is no
570 * iommu capable domain in the container.
572 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
574 for (i = 0; i < npage; i++) {
576 struct vfio_pfn *vpfn;
578 iova = user_pfn[i] << PAGE_SHIFT;
579 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
585 if ((dma->prot & prot) != prot) {
590 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
592 phys_pfn[i] = vpfn->pfn;
596 remote_vaddr = dma->vaddr + iova - dma->iova;
597 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
602 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
604 vfio_unpin_page_external(dma, iova, do_accounting);
614 for (j = 0; j < i; j++) {
617 iova = user_pfn[j] << PAGE_SHIFT;
618 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
619 vfio_unpin_page_external(dma, iova, do_accounting);
623 mutex_unlock(&iommu->lock);
627 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
628 unsigned long *user_pfn,
631 struct vfio_iommu *iommu = iommu_data;
635 if (!iommu || !user_pfn)
638 /* Supported for v2 version only */
642 mutex_lock(&iommu->lock);
644 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
645 for (i = 0; i < npage; i++) {
646 struct vfio_dma *dma;
649 iova = user_pfn[i] << PAGE_SHIFT;
650 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
653 vfio_unpin_page_external(dma, iova, do_accounting);
657 mutex_unlock(&iommu->lock);
658 return i > npage ? npage : (i > 0 ? i : -EINVAL);
661 static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
662 struct list_head *regions,
663 struct iommu_iotlb_gather *iotlb_gather)
666 struct vfio_regions *entry, *next;
668 iommu_tlb_sync(domain->domain, iotlb_gather);
670 list_for_each_entry_safe(entry, next, regions, list) {
671 unlocked += vfio_unpin_pages_remote(dma,
673 entry->phys >> PAGE_SHIFT,
674 entry->len >> PAGE_SHIFT,
676 list_del(&entry->list);
686 * Generally, VFIO needs to unpin remote pages after each IOTLB flush.
687 * Therefore, when using IOTLB flush sync interface, VFIO need to keep track
688 * of these regions (currently using a list).
690 * This value specifies maximum number of regions for each IOTLB flush sync.
692 #define VFIO_IOMMU_TLB_SYNC_MAX 512
694 static size_t unmap_unpin_fast(struct vfio_domain *domain,
695 struct vfio_dma *dma, dma_addr_t *iova,
696 size_t len, phys_addr_t phys, long *unlocked,
697 struct list_head *unmapped_list,
699 struct iommu_iotlb_gather *iotlb_gather)
702 struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
705 unmapped = iommu_unmap_fast(domain->domain, *iova, len,
713 entry->len = unmapped;
714 list_add_tail(&entry->list, unmapped_list);
722 * Sync if the number of fast-unmap regions hits the limit
723 * or in case of errors.
725 if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
726 *unlocked += vfio_sync_unpin(dma, domain, unmapped_list,
734 static size_t unmap_unpin_slow(struct vfio_domain *domain,
735 struct vfio_dma *dma, dma_addr_t *iova,
736 size_t len, phys_addr_t phys,
739 size_t unmapped = iommu_unmap(domain->domain, *iova, len);
742 *unlocked += vfio_unpin_pages_remote(dma, *iova,
744 unmapped >> PAGE_SHIFT,
752 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
755 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
756 struct vfio_domain *domain, *d;
757 LIST_HEAD(unmapped_region_list);
758 struct iommu_iotlb_gather iotlb_gather;
759 int unmapped_region_cnt = 0;
765 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
769 * We use the IOMMU to track the physical addresses, otherwise we'd
770 * need a much more complicated tracking system. Unfortunately that
771 * means we need to use one of the iommu domains to figure out the
772 * pfns to unpin. The rest need to be unmapped in advance so we have
773 * no iommu translations remaining when the pages are unpinned.
775 domain = d = list_first_entry(&iommu->domain_list,
776 struct vfio_domain, next);
778 list_for_each_entry_continue(d, &iommu->domain_list, next) {
779 iommu_unmap(d->domain, dma->iova, dma->size);
783 iommu_iotlb_gather_init(&iotlb_gather);
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,
815 unmapped = unmap_unpin_slow(domain, dma, &iova, len,
817 if (WARN_ON(!unmapped))
822 dma->iommu_mapped = false;
824 if (unmapped_region_cnt) {
825 unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list,
830 vfio_lock_acct(dma, -unlocked, true);
836 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
838 vfio_unmap_unpin(iommu, dma, true);
839 vfio_unlink_dma(iommu, dma);
840 put_task_struct(dma->task);
845 static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu)
847 struct vfio_domain *domain;
848 unsigned long bitmap = ULONG_MAX;
850 mutex_lock(&iommu->lock);
851 list_for_each_entry(domain, &iommu->domain_list, next)
852 bitmap &= domain->domain->pgsize_bitmap;
853 mutex_unlock(&iommu->lock);
856 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
857 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
858 * That way the user will be able to map/unmap buffers whose size/
859 * start address is aligned with PAGE_SIZE. Pinning code uses that
860 * granularity while iommu driver can use the sub-PAGE_SIZE size
863 if (bitmap & ~PAGE_MASK) {
871 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
872 struct vfio_iommu_type1_dma_unmap *unmap)
875 struct vfio_dma *dma, *dma_last = NULL;
877 int ret = 0, retries = 0;
879 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
881 if (unmap->iova & mask)
883 if (!unmap->size || unmap->size & mask)
885 if (unmap->iova + unmap->size - 1 < unmap->iova ||
886 unmap->size > SIZE_MAX)
889 WARN_ON(mask & PAGE_MASK);
891 mutex_lock(&iommu->lock);
894 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
895 * avoid tracking individual mappings. This means that the granularity
896 * of the original mapping was lost and the user was allowed to attempt
897 * to unmap any range. Depending on the contiguousness of physical
898 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
899 * or may not have worked. We only guaranteed unmap granularity
900 * matching the original mapping; even though it was untracked here,
901 * the original mappings are reflected in IOMMU mappings. This
902 * resulted in a couple unusual behaviors. First, if a range is not
903 * able to be unmapped, ex. a set of 4k pages that was mapped as a
904 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
905 * a zero sized unmap. Also, if an unmap request overlaps the first
906 * address of a hugepage, the IOMMU will unmap the entire hugepage.
907 * This also returns success and the returned unmap size reflects the
908 * actual size unmapped.
910 * We attempt to maintain compatibility with this "v1" interface, but
911 * we take control out of the hands of the IOMMU. Therefore, an unmap
912 * request offset from the beginning of the original mapping will
913 * return success with zero sized unmap. And an unmap request covering
914 * the first iova of mapping will unmap the entire range.
916 * The v2 version of this interface intends to be more deterministic.
917 * Unmap requests must fully cover previous mappings. Multiple
918 * mappings may still be unmaped by specifying large ranges, but there
919 * must not be any previous mappings bisected by the range. An error
920 * will be returned if these conditions are not met. The v2 interface
921 * will only return success and a size of zero if there were no
922 * mappings within the range.
925 dma = vfio_find_dma(iommu, unmap->iova, 1);
926 if (dma && dma->iova != unmap->iova) {
930 dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0);
931 if (dma && dma->iova + dma->size != unmap->iova + unmap->size) {
937 while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
938 if (!iommu->v2 && unmap->iova > dma->iova)
941 * Task with same address space who mapped this iova range is
942 * allowed to unmap the iova range.
944 if (dma->task->mm != current->mm)
947 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
948 struct vfio_iommu_type1_dma_unmap nb_unmap;
950 if (dma_last == dma) {
951 BUG_ON(++retries > 10);
957 nb_unmap.iova = dma->iova;
958 nb_unmap.size = dma->size;
961 * Notify anyone (mdev vendor drivers) to invalidate and
962 * unmap iovas within the range we're about to unmap.
963 * Vendor drivers MUST unpin pages in response to an
966 mutex_unlock(&iommu->lock);
967 blocking_notifier_call_chain(&iommu->notifier,
968 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
972 unmapped += dma->size;
973 vfio_remove_dma(iommu, dma);
977 mutex_unlock(&iommu->lock);
979 /* Report how much was unmapped */
980 unmap->size = unmapped;
985 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
986 unsigned long pfn, long npage, int prot)
988 struct vfio_domain *d;
991 list_for_each_entry(d, &iommu->domain_list, next) {
992 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
993 npage << PAGE_SHIFT, prot | d->prot);
1003 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next)
1004 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
1009 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
1012 dma_addr_t iova = dma->iova;
1013 unsigned long vaddr = dma->vaddr;
1014 size_t size = map_size;
1016 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1020 /* Pin a contiguous chunk of memory */
1021 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
1022 size >> PAGE_SHIFT, &pfn, limit);
1030 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
1033 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
1038 size -= npage << PAGE_SHIFT;
1039 dma->size += npage << PAGE_SHIFT;
1042 dma->iommu_mapped = true;
1045 vfio_remove_dma(iommu, dma);
1051 * Check dma map request is within a valid iova range
1053 static bool vfio_iommu_iova_dma_valid(struct vfio_iommu *iommu,
1054 dma_addr_t start, dma_addr_t end)
1056 struct list_head *iova = &iommu->iova_list;
1057 struct vfio_iova *node;
1059 list_for_each_entry(node, iova, list) {
1060 if (start >= node->start && end <= node->end)
1065 * Check for list_empty() as well since a container with
1066 * a single mdev device will have an empty list.
1068 return list_empty(iova);
1071 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1072 struct vfio_iommu_type1_dma_map *map)
1074 dma_addr_t iova = map->iova;
1075 unsigned long vaddr = map->vaddr;
1076 size_t size = map->size;
1077 int ret = 0, prot = 0;
1079 struct vfio_dma *dma;
1081 /* Verify that none of our __u64 fields overflow */
1082 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1085 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
1087 WARN_ON(mask & PAGE_MASK);
1089 /* READ/WRITE from device perspective */
1090 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1091 prot |= IOMMU_WRITE;
1092 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1095 if (!prot || !size || (size | iova | vaddr) & mask)
1098 /* Don't allow IOVA or virtual address wrap */
1099 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr)
1102 mutex_lock(&iommu->lock);
1104 if (vfio_find_dma(iommu, iova, size)) {
1109 if (!iommu->dma_avail) {
1114 if (!vfio_iommu_iova_dma_valid(iommu, iova, iova + size - 1)) {
1119 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1131 * We need to be able to both add to a task's locked memory and test
1132 * against the locked memory limit and we need to be able to do both
1133 * outside of this call path as pinning can be asynchronous via the
1134 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1135 * task_struct and VM locked pages requires an mm_struct, however
1136 * holding an indefinite mm reference is not recommended, therefore we
1137 * only hold a reference to a task. We could hold a reference to
1138 * current, however QEMU uses this call path through vCPU threads,
1139 * which can be killed resulting in a NULL mm and failure in the unmap
1140 * path when called via a different thread. Avoid this problem by
1141 * using the group_leader as threads within the same group require
1142 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1145 * Previously we also used the task for testing CAP_IPC_LOCK at the
1146 * time of pinning and accounting, however has_capability() makes use
1147 * of real_cred, a copy-on-write field, so we can't guarantee that it
1148 * matches group_leader, or in fact that it might not change by the
1149 * time it's evaluated. If a process were to call MAP_DMA with
1150 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1151 * possibly see different results for an iommu_mapped vfio_dma vs
1152 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1153 * time of calling MAP_DMA.
1155 get_task_struct(current->group_leader);
1156 dma->task = current->group_leader;
1157 dma->lock_cap = capable(CAP_IPC_LOCK);
1159 dma->pfn_list = RB_ROOT;
1161 /* Insert zero-sized and grow as we map chunks of it */
1162 vfio_link_dma(iommu, dma);
1164 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1165 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1168 ret = vfio_pin_map_dma(iommu, dma, size);
1171 mutex_unlock(&iommu->lock);
1175 static int vfio_bus_type(struct device *dev, void *data)
1177 struct bus_type **bus = data;
1179 if (*bus && *bus != dev->bus)
1187 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1188 struct vfio_domain *domain)
1190 struct vfio_domain *d;
1192 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1195 /* Arbitrarily pick the first domain in the list for lookups */
1196 d = list_first_entry(&iommu->domain_list, struct vfio_domain, next);
1197 n = rb_first(&iommu->dma_list);
1199 for (; n; n = rb_next(n)) {
1200 struct vfio_dma *dma;
1203 dma = rb_entry(n, struct vfio_dma, node);
1206 while (iova < dma->iova + dma->size) {
1210 if (dma->iommu_mapped) {
1214 phys = iommu_iova_to_phys(d->domain, iova);
1216 if (WARN_ON(!phys)) {
1224 while (i < dma->iova + dma->size &&
1225 p == iommu_iova_to_phys(d->domain, i)) {
1232 unsigned long vaddr = dma->vaddr +
1234 size_t n = dma->iova + dma->size - iova;
1237 npage = vfio_pin_pages_remote(dma, vaddr,
1246 phys = pfn << PAGE_SHIFT;
1247 size = npage << PAGE_SHIFT;
1250 ret = iommu_map(domain->domain, iova, phys,
1251 size, dma->prot | domain->prot);
1257 dma->iommu_mapped = true;
1263 * We change our unmap behavior slightly depending on whether the IOMMU
1264 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1265 * for practically any contiguous power-of-two mapping we give it. This means
1266 * we don't need to look for contiguous chunks ourselves to make unmapping
1267 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1268 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1269 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1270 * hugetlbfs is in use.
1272 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1275 int ret, order = get_order(PAGE_SIZE * 2);
1277 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1281 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1282 IOMMU_READ | IOMMU_WRITE | domain->prot);
1284 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1286 if (unmapped == PAGE_SIZE)
1287 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1289 domain->fgsp = true;
1292 __free_pages(pages, order);
1295 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1296 struct iommu_group *iommu_group)
1298 struct vfio_group *g;
1300 list_for_each_entry(g, &domain->group_list, next) {
1301 if (g->iommu_group == iommu_group)
1308 static bool vfio_iommu_has_sw_msi(struct list_head *group_resv_regions,
1311 struct iommu_resv_region *region;
1314 list_for_each_entry(region, group_resv_regions, list) {
1316 * The presence of any 'real' MSI regions should take
1317 * precedence over the software-managed one if the
1318 * IOMMU driver happens to advertise both types.
1320 if (region->type == IOMMU_RESV_MSI) {
1325 if (region->type == IOMMU_RESV_SW_MSI) {
1326 *base = region->start;
1334 static struct device *vfio_mdev_get_iommu_device(struct device *dev)
1336 struct device *(*fn)(struct device *dev);
1337 struct device *iommu_device;
1339 fn = symbol_get(mdev_get_iommu_device);
1341 iommu_device = fn(dev);
1342 symbol_put(mdev_get_iommu_device);
1344 return iommu_device;
1350 static int vfio_mdev_attach_domain(struct device *dev, void *data)
1352 struct iommu_domain *domain = data;
1353 struct device *iommu_device;
1355 iommu_device = vfio_mdev_get_iommu_device(dev);
1357 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1358 return iommu_aux_attach_device(domain, iommu_device);
1360 return iommu_attach_device(domain, iommu_device);
1366 static int vfio_mdev_detach_domain(struct device *dev, void *data)
1368 struct iommu_domain *domain = data;
1369 struct device *iommu_device;
1371 iommu_device = vfio_mdev_get_iommu_device(dev);
1373 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1374 iommu_aux_detach_device(domain, iommu_device);
1376 iommu_detach_device(domain, iommu_device);
1382 static int vfio_iommu_attach_group(struct vfio_domain *domain,
1383 struct vfio_group *group)
1385 if (group->mdev_group)
1386 return iommu_group_for_each_dev(group->iommu_group,
1388 vfio_mdev_attach_domain);
1390 return iommu_attach_group(domain->domain, group->iommu_group);
1393 static void vfio_iommu_detach_group(struct vfio_domain *domain,
1394 struct vfio_group *group)
1396 if (group->mdev_group)
1397 iommu_group_for_each_dev(group->iommu_group, domain->domain,
1398 vfio_mdev_detach_domain);
1400 iommu_detach_group(domain->domain, group->iommu_group);
1403 static bool vfio_bus_is_mdev(struct bus_type *bus)
1405 struct bus_type *mdev_bus;
1408 mdev_bus = symbol_get(mdev_bus_type);
1410 ret = (bus == mdev_bus);
1411 symbol_put(mdev_bus_type);
1417 static int vfio_mdev_iommu_device(struct device *dev, void *data)
1419 struct device **old = data, *new;
1421 new = vfio_mdev_get_iommu_device(dev);
1422 if (!new || (*old && *old != new))
1431 * This is a helper function to insert an address range to iova list.
1432 * The list is initially created with a single entry corresponding to
1433 * the IOMMU domain geometry to which the device group is attached.
1434 * The list aperture gets modified when a new domain is added to the
1435 * container if the new aperture doesn't conflict with the current one
1436 * or with any existing dma mappings. The list is also modified to
1437 * exclude any reserved regions associated with the device group.
1439 static int vfio_iommu_iova_insert(struct list_head *head,
1440 dma_addr_t start, dma_addr_t end)
1442 struct vfio_iova *region;
1444 region = kmalloc(sizeof(*region), GFP_KERNEL);
1448 INIT_LIST_HEAD(®ion->list);
1449 region->start = start;
1452 list_add_tail(®ion->list, head);
1457 * Check the new iommu aperture conflicts with existing aper or with any
1458 * existing dma mappings.
1460 static bool vfio_iommu_aper_conflict(struct vfio_iommu *iommu,
1461 dma_addr_t start, dma_addr_t end)
1463 struct vfio_iova *first, *last;
1464 struct list_head *iova = &iommu->iova_list;
1466 if (list_empty(iova))
1469 /* Disjoint sets, return conflict */
1470 first = list_first_entry(iova, struct vfio_iova, list);
1471 last = list_last_entry(iova, struct vfio_iova, list);
1472 if (start > last->end || end < first->start)
1475 /* Check for any existing dma mappings below the new start */
1476 if (start > first->start) {
1477 if (vfio_find_dma(iommu, first->start, start - first->start))
1481 /* Check for any existing dma mappings beyond the new end */
1482 if (end < last->end) {
1483 if (vfio_find_dma(iommu, end + 1, last->end - end))
1491 * Resize iommu iova aperture window. This is called only if the new
1492 * aperture has no conflict with existing aperture and dma mappings.
1494 static int vfio_iommu_aper_resize(struct list_head *iova,
1495 dma_addr_t start, dma_addr_t end)
1497 struct vfio_iova *node, *next;
1499 if (list_empty(iova))
1500 return vfio_iommu_iova_insert(iova, start, end);
1502 /* Adjust iova list start */
1503 list_for_each_entry_safe(node, next, iova, list) {
1504 if (start < node->start)
1506 if (start >= node->start && start < node->end) {
1507 node->start = start;
1510 /* Delete nodes before new start */
1511 list_del(&node->list);
1515 /* Adjust iova list end */
1516 list_for_each_entry_safe(node, next, iova, list) {
1517 if (end > node->end)
1519 if (end > node->start && end <= node->end) {
1523 /* Delete nodes after new end */
1524 list_del(&node->list);
1532 * Check reserved region conflicts with existing dma mappings
1534 static bool vfio_iommu_resv_conflict(struct vfio_iommu *iommu,
1535 struct list_head *resv_regions)
1537 struct iommu_resv_region *region;
1539 /* Check for conflict with existing dma mappings */
1540 list_for_each_entry(region, resv_regions, list) {
1541 if (region->type == IOMMU_RESV_DIRECT_RELAXABLE)
1544 if (vfio_find_dma(iommu, region->start, region->length))
1552 * Check iova region overlap with reserved regions and
1553 * exclude them from the iommu iova range
1555 static int vfio_iommu_resv_exclude(struct list_head *iova,
1556 struct list_head *resv_regions)
1558 struct iommu_resv_region *resv;
1559 struct vfio_iova *n, *next;
1561 list_for_each_entry(resv, resv_regions, list) {
1562 phys_addr_t start, end;
1564 if (resv->type == IOMMU_RESV_DIRECT_RELAXABLE)
1567 start = resv->start;
1568 end = resv->start + resv->length - 1;
1570 list_for_each_entry_safe(n, next, iova, list) {
1574 if (start > n->end || end < n->start)
1577 * Insert a new node if current node overlaps with the
1578 * reserve region to exlude that from valid iova range.
1579 * Note that, new node is inserted before the current
1580 * node and finally the current node is deleted keeping
1581 * the list updated and sorted.
1583 if (start > n->start)
1584 ret = vfio_iommu_iova_insert(&n->list, n->start,
1586 if (!ret && end < n->end)
1587 ret = vfio_iommu_iova_insert(&n->list, end + 1,
1597 if (list_empty(iova))
1603 static void vfio_iommu_resv_free(struct list_head *resv_regions)
1605 struct iommu_resv_region *n, *next;
1607 list_for_each_entry_safe(n, next, resv_regions, list) {
1613 static void vfio_iommu_iova_free(struct list_head *iova)
1615 struct vfio_iova *n, *next;
1617 list_for_each_entry_safe(n, next, iova, list) {
1623 static int vfio_iommu_iova_get_copy(struct vfio_iommu *iommu,
1624 struct list_head *iova_copy)
1626 struct list_head *iova = &iommu->iova_list;
1627 struct vfio_iova *n;
1630 list_for_each_entry(n, iova, list) {
1631 ret = vfio_iommu_iova_insert(iova_copy, n->start, n->end);
1639 vfio_iommu_iova_free(iova_copy);
1643 static void vfio_iommu_iova_insert_copy(struct vfio_iommu *iommu,
1644 struct list_head *iova_copy)
1646 struct list_head *iova = &iommu->iova_list;
1648 vfio_iommu_iova_free(iova);
1650 list_splice_tail(iova_copy, iova);
1652 static int vfio_iommu_type1_attach_group(void *iommu_data,
1653 struct iommu_group *iommu_group)
1655 struct vfio_iommu *iommu = iommu_data;
1656 struct vfio_group *group;
1657 struct vfio_domain *domain, *d;
1658 struct bus_type *bus = NULL;
1660 bool resv_msi, msi_remap;
1661 phys_addr_t resv_msi_base;
1662 struct iommu_domain_geometry geo;
1663 LIST_HEAD(iova_copy);
1664 LIST_HEAD(group_resv_regions);
1666 mutex_lock(&iommu->lock);
1668 list_for_each_entry(d, &iommu->domain_list, next) {
1669 if (find_iommu_group(d, iommu_group)) {
1670 mutex_unlock(&iommu->lock);
1675 if (iommu->external_domain) {
1676 if (find_iommu_group(iommu->external_domain, iommu_group)) {
1677 mutex_unlock(&iommu->lock);
1682 group = kzalloc(sizeof(*group), GFP_KERNEL);
1683 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
1684 if (!group || !domain) {
1689 group->iommu_group = iommu_group;
1691 /* Determine bus_type in order to allocate a domain */
1692 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
1696 if (vfio_bus_is_mdev(bus)) {
1697 struct device *iommu_device = NULL;
1699 group->mdev_group = true;
1701 /* Determine the isolation type */
1702 ret = iommu_group_for_each_dev(iommu_group, &iommu_device,
1703 vfio_mdev_iommu_device);
1704 if (ret || !iommu_device) {
1705 if (!iommu->external_domain) {
1706 INIT_LIST_HEAD(&domain->group_list);
1707 iommu->external_domain = domain;
1712 list_add(&group->next,
1713 &iommu->external_domain->group_list);
1714 mutex_unlock(&iommu->lock);
1719 bus = iommu_device->bus;
1722 domain->domain = iommu_domain_alloc(bus);
1723 if (!domain->domain) {
1728 if (iommu->nesting) {
1731 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
1737 ret = vfio_iommu_attach_group(domain, group);
1741 /* Get aperture info */
1742 iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY, &geo);
1744 if (vfio_iommu_aper_conflict(iommu, geo.aperture_start,
1745 geo.aperture_end)) {
1750 ret = iommu_get_group_resv_regions(iommu_group, &group_resv_regions);
1754 if (vfio_iommu_resv_conflict(iommu, &group_resv_regions)) {
1760 * We don't want to work on the original iova list as the list
1761 * gets modified and in case of failure we have to retain the
1762 * original list. Get a copy here.
1764 ret = vfio_iommu_iova_get_copy(iommu, &iova_copy);
1768 ret = vfio_iommu_aper_resize(&iova_copy, geo.aperture_start,
1773 ret = vfio_iommu_resv_exclude(&iova_copy, &group_resv_regions);
1777 resv_msi = vfio_iommu_has_sw_msi(&group_resv_regions, &resv_msi_base);
1779 INIT_LIST_HEAD(&domain->group_list);
1780 list_add(&group->next, &domain->group_list);
1782 msi_remap = irq_domain_check_msi_remap() ||
1783 iommu_capable(bus, IOMMU_CAP_INTR_REMAP);
1785 if (!allow_unsafe_interrupts && !msi_remap) {
1786 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
1792 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
1793 domain->prot |= IOMMU_CACHE;
1796 * Try to match an existing compatible domain. We don't want to
1797 * preclude an IOMMU driver supporting multiple bus_types and being
1798 * able to include different bus_types in the same IOMMU domain, so
1799 * we test whether the domains use the same iommu_ops rather than
1800 * testing if they're on the same bus_type.
1802 list_for_each_entry(d, &iommu->domain_list, next) {
1803 if (d->domain->ops == domain->domain->ops &&
1804 d->prot == domain->prot) {
1805 vfio_iommu_detach_group(domain, group);
1806 if (!vfio_iommu_attach_group(d, group)) {
1807 list_add(&group->next, &d->group_list);
1808 iommu_domain_free(domain->domain);
1813 ret = vfio_iommu_attach_group(domain, group);
1819 vfio_test_domain_fgsp(domain);
1821 /* replay mappings on new domains */
1822 ret = vfio_iommu_replay(iommu, domain);
1827 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
1832 list_add(&domain->next, &iommu->domain_list);
1834 /* Delete the old one and insert new iova list */
1835 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
1836 mutex_unlock(&iommu->lock);
1837 vfio_iommu_resv_free(&group_resv_regions);
1842 vfio_iommu_detach_group(domain, group);
1844 iommu_domain_free(domain->domain);
1845 vfio_iommu_iova_free(&iova_copy);
1846 vfio_iommu_resv_free(&group_resv_regions);
1850 mutex_unlock(&iommu->lock);
1854 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
1856 struct rb_node *node;
1858 while ((node = rb_first(&iommu->dma_list)))
1859 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
1862 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
1864 struct rb_node *n, *p;
1866 n = rb_first(&iommu->dma_list);
1867 for (; n; n = rb_next(n)) {
1868 struct vfio_dma *dma;
1869 long locked = 0, unlocked = 0;
1871 dma = rb_entry(n, struct vfio_dma, node);
1872 unlocked += vfio_unmap_unpin(iommu, dma, false);
1873 p = rb_first(&dma->pfn_list);
1874 for (; p; p = rb_next(p)) {
1875 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
1878 if (!is_invalid_reserved_pfn(vpfn->pfn))
1881 vfio_lock_acct(dma, locked - unlocked, true);
1885 static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu)
1889 n = rb_first(&iommu->dma_list);
1890 for (; n; n = rb_next(n)) {
1891 struct vfio_dma *dma;
1893 dma = rb_entry(n, struct vfio_dma, node);
1895 if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list)))
1898 /* mdev vendor driver must unregister notifier */
1899 WARN_ON(iommu->notifier.head);
1903 * Called when a domain is removed in detach. It is possible that
1904 * the removed domain decided the iova aperture window. Modify the
1905 * iova aperture with the smallest window among existing domains.
1907 static void vfio_iommu_aper_expand(struct vfio_iommu *iommu,
1908 struct list_head *iova_copy)
1910 struct vfio_domain *domain;
1911 struct iommu_domain_geometry geo;
1912 struct vfio_iova *node;
1913 dma_addr_t start = 0;
1914 dma_addr_t end = (dma_addr_t)~0;
1916 if (list_empty(iova_copy))
1919 list_for_each_entry(domain, &iommu->domain_list, next) {
1920 iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY,
1922 if (geo.aperture_start > start)
1923 start = geo.aperture_start;
1924 if (geo.aperture_end < end)
1925 end = geo.aperture_end;
1928 /* Modify aperture limits. The new aper is either same or bigger */
1929 node = list_first_entry(iova_copy, struct vfio_iova, list);
1930 node->start = start;
1931 node = list_last_entry(iova_copy, struct vfio_iova, list);
1936 * Called when a group is detached. The reserved regions for that
1937 * group can be part of valid iova now. But since reserved regions
1938 * may be duplicated among groups, populate the iova valid regions
1941 static int vfio_iommu_resv_refresh(struct vfio_iommu *iommu,
1942 struct list_head *iova_copy)
1944 struct vfio_domain *d;
1945 struct vfio_group *g;
1946 struct vfio_iova *node;
1947 dma_addr_t start, end;
1948 LIST_HEAD(resv_regions);
1951 if (list_empty(iova_copy))
1954 list_for_each_entry(d, &iommu->domain_list, next) {
1955 list_for_each_entry(g, &d->group_list, next) {
1956 ret = iommu_get_group_resv_regions(g->iommu_group,
1963 node = list_first_entry(iova_copy, struct vfio_iova, list);
1964 start = node->start;
1965 node = list_last_entry(iova_copy, struct vfio_iova, list);
1968 /* purge the iova list and create new one */
1969 vfio_iommu_iova_free(iova_copy);
1971 ret = vfio_iommu_aper_resize(iova_copy, start, end);
1975 /* Exclude current reserved regions from iova ranges */
1976 ret = vfio_iommu_resv_exclude(iova_copy, &resv_regions);
1978 vfio_iommu_resv_free(&resv_regions);
1982 static void vfio_iommu_type1_detach_group(void *iommu_data,
1983 struct iommu_group *iommu_group)
1985 struct vfio_iommu *iommu = iommu_data;
1986 struct vfio_domain *domain;
1987 struct vfio_group *group;
1988 LIST_HEAD(iova_copy);
1990 mutex_lock(&iommu->lock);
1992 if (iommu->external_domain) {
1993 group = find_iommu_group(iommu->external_domain, iommu_group);
1995 list_del(&group->next);
1998 if (list_empty(&iommu->external_domain->group_list)) {
1999 vfio_sanity_check_pfn_list(iommu);
2001 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
2002 vfio_iommu_unmap_unpin_all(iommu);
2004 kfree(iommu->external_domain);
2005 iommu->external_domain = NULL;
2007 goto detach_group_done;
2012 * Get a copy of iova list. This will be used to update
2013 * and to replace the current one later. Please note that
2014 * we will leave the original list as it is if update fails.
2016 vfio_iommu_iova_get_copy(iommu, &iova_copy);
2018 list_for_each_entry(domain, &iommu->domain_list, next) {
2019 group = find_iommu_group(domain, iommu_group);
2023 vfio_iommu_detach_group(domain, group);
2024 list_del(&group->next);
2027 * Group ownership provides privilege, if the group list is
2028 * empty, the domain goes away. If it's the last domain with
2029 * iommu and external domain doesn't exist, then all the
2030 * mappings go away too. If it's the last domain with iommu and
2031 * external domain exist, update accounting
2033 if (list_empty(&domain->group_list)) {
2034 if (list_is_singular(&iommu->domain_list)) {
2035 if (!iommu->external_domain)
2036 vfio_iommu_unmap_unpin_all(iommu);
2038 vfio_iommu_unmap_unpin_reaccount(iommu);
2040 iommu_domain_free(domain->domain);
2041 list_del(&domain->next);
2043 vfio_iommu_aper_expand(iommu, &iova_copy);
2048 if (!vfio_iommu_resv_refresh(iommu, &iova_copy))
2049 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2051 vfio_iommu_iova_free(&iova_copy);
2054 mutex_unlock(&iommu->lock);
2057 static void *vfio_iommu_type1_open(unsigned long arg)
2059 struct vfio_iommu *iommu;
2061 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
2063 return ERR_PTR(-ENOMEM);
2066 case VFIO_TYPE1_IOMMU:
2068 case VFIO_TYPE1_NESTING_IOMMU:
2069 iommu->nesting = true;
2071 case VFIO_TYPE1v2_IOMMU:
2076 return ERR_PTR(-EINVAL);
2079 INIT_LIST_HEAD(&iommu->domain_list);
2080 INIT_LIST_HEAD(&iommu->iova_list);
2081 iommu->dma_list = RB_ROOT;
2082 iommu->dma_avail = dma_entry_limit;
2083 mutex_init(&iommu->lock);
2084 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
2089 static void vfio_release_domain(struct vfio_domain *domain, bool external)
2091 struct vfio_group *group, *group_tmp;
2093 list_for_each_entry_safe(group, group_tmp,
2094 &domain->group_list, next) {
2096 vfio_iommu_detach_group(domain, group);
2097 list_del(&group->next);
2102 iommu_domain_free(domain->domain);
2105 static void vfio_iommu_type1_release(void *iommu_data)
2107 struct vfio_iommu *iommu = iommu_data;
2108 struct vfio_domain *domain, *domain_tmp;
2110 if (iommu->external_domain) {
2111 vfio_release_domain(iommu->external_domain, true);
2112 vfio_sanity_check_pfn_list(iommu);
2113 kfree(iommu->external_domain);
2116 vfio_iommu_unmap_unpin_all(iommu);
2118 list_for_each_entry_safe(domain, domain_tmp,
2119 &iommu->domain_list, next) {
2120 vfio_release_domain(domain, false);
2121 list_del(&domain->next);
2125 vfio_iommu_iova_free(&iommu->iova_list);
2130 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
2132 struct vfio_domain *domain;
2135 mutex_lock(&iommu->lock);
2136 list_for_each_entry(domain, &iommu->domain_list, next) {
2137 if (!(domain->prot & IOMMU_CACHE)) {
2142 mutex_unlock(&iommu->lock);
2147 static int vfio_iommu_iova_add_cap(struct vfio_info_cap *caps,
2148 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas,
2151 struct vfio_info_cap_header *header;
2152 struct vfio_iommu_type1_info_cap_iova_range *iova_cap;
2154 header = vfio_info_cap_add(caps, size,
2155 VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE, 1);
2157 return PTR_ERR(header);
2159 iova_cap = container_of(header,
2160 struct vfio_iommu_type1_info_cap_iova_range,
2162 iova_cap->nr_iovas = cap_iovas->nr_iovas;
2163 memcpy(iova_cap->iova_ranges, cap_iovas->iova_ranges,
2164 cap_iovas->nr_iovas * sizeof(*cap_iovas->iova_ranges));
2168 static int vfio_iommu_iova_build_caps(struct vfio_iommu *iommu,
2169 struct vfio_info_cap *caps)
2171 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas;
2172 struct vfio_iova *iova;
2174 int iovas = 0, i = 0, ret;
2176 mutex_lock(&iommu->lock);
2178 list_for_each_entry(iova, &iommu->iova_list, list)
2183 * Return 0 as a container with a single mdev device
2184 * will have an empty list
2190 size = sizeof(*cap_iovas) + (iovas * sizeof(*cap_iovas->iova_ranges));
2192 cap_iovas = kzalloc(size, GFP_KERNEL);
2198 cap_iovas->nr_iovas = iovas;
2200 list_for_each_entry(iova, &iommu->iova_list, list) {
2201 cap_iovas->iova_ranges[i].start = iova->start;
2202 cap_iovas->iova_ranges[i].end = iova->end;
2206 ret = vfio_iommu_iova_add_cap(caps, cap_iovas, size);
2210 mutex_unlock(&iommu->lock);
2214 static long vfio_iommu_type1_ioctl(void *iommu_data,
2215 unsigned int cmd, unsigned long arg)
2217 struct vfio_iommu *iommu = iommu_data;
2218 unsigned long minsz;
2220 if (cmd == VFIO_CHECK_EXTENSION) {
2222 case VFIO_TYPE1_IOMMU:
2223 case VFIO_TYPE1v2_IOMMU:
2224 case VFIO_TYPE1_NESTING_IOMMU:
2226 case VFIO_DMA_CC_IOMMU:
2229 return vfio_domains_have_iommu_cache(iommu);
2233 } else if (cmd == VFIO_IOMMU_GET_INFO) {
2234 struct vfio_iommu_type1_info info;
2235 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
2236 unsigned long capsz;
2239 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
2241 /* For backward compatibility, cannot require this */
2242 capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
2244 if (copy_from_user(&info, (void __user *)arg, minsz))
2247 if (info.argsz < minsz)
2250 if (info.argsz >= capsz) {
2252 info.cap_offset = 0; /* output, no-recopy necessary */
2255 info.flags = VFIO_IOMMU_INFO_PGSIZES;
2257 info.iova_pgsizes = vfio_pgsize_bitmap(iommu);
2259 ret = vfio_iommu_iova_build_caps(iommu, &caps);
2264 info.flags |= VFIO_IOMMU_INFO_CAPS;
2266 if (info.argsz < sizeof(info) + caps.size) {
2267 info.argsz = sizeof(info) + caps.size;
2269 vfio_info_cap_shift(&caps, sizeof(info));
2270 if (copy_to_user((void __user *)arg +
2271 sizeof(info), caps.buf,
2276 info.cap_offset = sizeof(info);
2282 return copy_to_user((void __user *)arg, &info, minsz) ?
2285 } else if (cmd == VFIO_IOMMU_MAP_DMA) {
2286 struct vfio_iommu_type1_dma_map map;
2287 uint32_t mask = VFIO_DMA_MAP_FLAG_READ |
2288 VFIO_DMA_MAP_FLAG_WRITE;
2290 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
2292 if (copy_from_user(&map, (void __user *)arg, minsz))
2295 if (map.argsz < minsz || map.flags & ~mask)
2298 return vfio_dma_do_map(iommu, &map);
2300 } else if (cmd == VFIO_IOMMU_UNMAP_DMA) {
2301 struct vfio_iommu_type1_dma_unmap unmap;
2304 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
2306 if (copy_from_user(&unmap, (void __user *)arg, minsz))
2309 if (unmap.argsz < minsz || unmap.flags)
2312 ret = vfio_dma_do_unmap(iommu, &unmap);
2316 return copy_to_user((void __user *)arg, &unmap, minsz) ?
2323 static int vfio_iommu_type1_register_notifier(void *iommu_data,
2324 unsigned long *events,
2325 struct notifier_block *nb)
2327 struct vfio_iommu *iommu = iommu_data;
2329 /* clear known events */
2330 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
2332 /* refuse to register if still events remaining */
2336 return blocking_notifier_chain_register(&iommu->notifier, nb);
2339 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
2340 struct notifier_block *nb)
2342 struct vfio_iommu *iommu = iommu_data;
2344 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
2347 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
2348 .name = "vfio-iommu-type1",
2349 .owner = THIS_MODULE,
2350 .open = vfio_iommu_type1_open,
2351 .release = vfio_iommu_type1_release,
2352 .ioctl = vfio_iommu_type1_ioctl,
2353 .attach_group = vfio_iommu_type1_attach_group,
2354 .detach_group = vfio_iommu_type1_detach_group,
2355 .pin_pages = vfio_iommu_type1_pin_pages,
2356 .unpin_pages = vfio_iommu_type1_unpin_pages,
2357 .register_notifier = vfio_iommu_type1_register_notifier,
2358 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
2361 static int __init vfio_iommu_type1_init(void)
2363 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
2366 static void __exit vfio_iommu_type1_cleanup(void)
2368 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
2371 module_init(vfio_iommu_type1_init);
2372 module_exit(vfio_iommu_type1_cleanup);
2374 MODULE_VERSION(DRIVER_VERSION);
2375 MODULE_LICENSE("GPL v2");
2376 MODULE_AUTHOR(DRIVER_AUTHOR);
2377 MODULE_DESCRIPTION(DRIVER_DESC);