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.
6 * Author: Alex Williamson <alex.williamson@redhat.com>
8 * Derived from original vfio:
9 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
10 * Author: Tom Lyon, pugs@cisco.com
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"
43 #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
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.
297 * For compound pages, any driver that sets the reserved bit in head
298 * page needs to set the reserved bit in all subpages to be safe.
300 static bool is_invalid_reserved_pfn(unsigned long pfn)
303 return PageReserved(pfn_to_page(pfn));
308 static int put_pfn(unsigned long pfn, int prot)
310 if (!is_invalid_reserved_pfn(pfn)) {
311 struct page *page = pfn_to_page(pfn);
312 if (prot & IOMMU_WRITE)
320 static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,
321 int prot, unsigned long *pfn)
323 struct page *page[1];
324 struct vm_area_struct *vma;
325 struct vm_area_struct *vmas[1];
326 unsigned int flags = 0;
329 if (prot & IOMMU_WRITE)
332 down_read(&mm->mmap_sem);
333 if (mm == current->mm) {
334 ret = get_user_pages(vaddr, 1, flags | FOLL_LONGTERM, page,
337 ret = get_user_pages_remote(NULL, mm, vaddr, 1, flags, page,
340 * The lifetime of a vaddr_get_pfn() page pin is
341 * userspace-controlled. In the fs-dax case this could
342 * lead to indefinite stalls in filesystem operations.
343 * Disallow attempts to pin fs-dax pages via this
346 if (ret > 0 && vma_is_fsdax(vmas[0])) {
351 up_read(&mm->mmap_sem);
354 *pfn = page_to_pfn(page[0]);
358 down_read(&mm->mmap_sem);
360 vaddr = untagged_addr(vaddr);
362 vma = find_vma_intersection(mm, vaddr, vaddr + 1);
364 if (vma && vma->vm_flags & VM_PFNMAP) {
365 *pfn = ((vaddr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
366 if (is_invalid_reserved_pfn(*pfn))
370 up_read(&mm->mmap_sem);
375 * Attempt to pin pages. We really don't want to track all the pfns and
376 * the iommu can only map chunks of consecutive pfns anyway, so get the
377 * first page and all consecutive pages with the same locking.
379 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
380 long npage, unsigned long *pfn_base,
383 unsigned long pfn = 0;
384 long ret, pinned = 0, lock_acct = 0;
386 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
388 /* This code path is only user initiated */
392 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base);
397 rsvd = is_invalid_reserved_pfn(*pfn_base);
400 * Reserved pages aren't counted against the user, externally pinned
401 * pages are already counted against the user.
403 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
404 if (!dma->lock_cap && current->mm->locked_vm + 1 > limit) {
405 put_pfn(*pfn_base, dma->prot);
406 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__,
407 limit << PAGE_SHIFT);
413 if (unlikely(disable_hugepages))
416 /* Lock all the consecutive pages from pfn_base */
417 for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage;
418 pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) {
419 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn);
423 if (pfn != *pfn_base + pinned ||
424 rsvd != is_invalid_reserved_pfn(pfn)) {
425 put_pfn(pfn, dma->prot);
429 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
430 if (!dma->lock_cap &&
431 current->mm->locked_vm + lock_acct + 1 > limit) {
432 put_pfn(pfn, dma->prot);
433 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
434 __func__, limit << PAGE_SHIFT);
443 ret = vfio_lock_acct(dma, lock_acct, false);
448 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
449 put_pfn(pfn, dma->prot);
458 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
459 unsigned long pfn, long npage,
462 long unlocked = 0, locked = 0;
465 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
466 if (put_pfn(pfn++, dma->prot)) {
468 if (vfio_find_vpfn(dma, iova))
474 vfio_lock_acct(dma, locked - unlocked, true);
479 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
480 unsigned long *pfn_base, bool do_accounting)
482 struct mm_struct *mm;
485 mm = get_task_mm(dma->task);
489 ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base);
490 if (!ret && do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
491 ret = vfio_lock_acct(dma, 1, true);
493 put_pfn(*pfn_base, dma->prot);
495 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
496 "(%ld) exceeded\n", __func__,
497 dma->task->comm, task_pid_nr(dma->task),
498 task_rlimit(dma->task, RLIMIT_MEMLOCK));
506 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
510 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
515 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
518 vfio_lock_acct(dma, -unlocked, true);
523 static int vfio_iommu_type1_pin_pages(void *iommu_data,
524 unsigned long *user_pfn,
526 unsigned long *phys_pfn)
528 struct vfio_iommu *iommu = iommu_data;
530 unsigned long remote_vaddr;
531 struct vfio_dma *dma;
534 if (!iommu || !user_pfn || !phys_pfn)
537 /* Supported for v2 version only */
541 mutex_lock(&iommu->lock);
543 /* Fail if notifier list is empty */
544 if (!iommu->notifier.head) {
550 * If iommu capable domain exist in the container then all pages are
551 * already pinned and accounted. Accouting should be done if there is no
552 * iommu capable domain in the container.
554 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
556 for (i = 0; i < npage; i++) {
558 struct vfio_pfn *vpfn;
560 iova = user_pfn[i] << PAGE_SHIFT;
561 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
567 if ((dma->prot & prot) != prot) {
572 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
574 phys_pfn[i] = vpfn->pfn;
578 remote_vaddr = dma->vaddr + iova - dma->iova;
579 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
584 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
586 vfio_unpin_page_external(dma, iova, do_accounting);
596 for (j = 0; j < i; j++) {
599 iova = user_pfn[j] << PAGE_SHIFT;
600 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
601 vfio_unpin_page_external(dma, iova, do_accounting);
605 mutex_unlock(&iommu->lock);
609 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
610 unsigned long *user_pfn,
613 struct vfio_iommu *iommu = iommu_data;
617 if (!iommu || !user_pfn)
620 /* Supported for v2 version only */
624 mutex_lock(&iommu->lock);
626 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
627 for (i = 0; i < npage; i++) {
628 struct vfio_dma *dma;
631 iova = user_pfn[i] << PAGE_SHIFT;
632 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
635 vfio_unpin_page_external(dma, iova, do_accounting);
639 mutex_unlock(&iommu->lock);
640 return i > npage ? npage : (i > 0 ? i : -EINVAL);
643 static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
644 struct list_head *regions,
645 struct iommu_iotlb_gather *iotlb_gather)
648 struct vfio_regions *entry, *next;
650 iommu_tlb_sync(domain->domain, iotlb_gather);
652 list_for_each_entry_safe(entry, next, regions, list) {
653 unlocked += vfio_unpin_pages_remote(dma,
655 entry->phys >> PAGE_SHIFT,
656 entry->len >> PAGE_SHIFT,
658 list_del(&entry->list);
668 * Generally, VFIO needs to unpin remote pages after each IOTLB flush.
669 * Therefore, when using IOTLB flush sync interface, VFIO need to keep track
670 * of these regions (currently using a list).
672 * This value specifies maximum number of regions for each IOTLB flush sync.
674 #define VFIO_IOMMU_TLB_SYNC_MAX 512
676 static size_t unmap_unpin_fast(struct vfio_domain *domain,
677 struct vfio_dma *dma, dma_addr_t *iova,
678 size_t len, phys_addr_t phys, long *unlocked,
679 struct list_head *unmapped_list,
681 struct iommu_iotlb_gather *iotlb_gather)
684 struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
687 unmapped = iommu_unmap_fast(domain->domain, *iova, len,
695 entry->len = unmapped;
696 list_add_tail(&entry->list, unmapped_list);
704 * Sync if the number of fast-unmap regions hits the limit
705 * or in case of errors.
707 if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
708 *unlocked += vfio_sync_unpin(dma, domain, unmapped_list,
716 static size_t unmap_unpin_slow(struct vfio_domain *domain,
717 struct vfio_dma *dma, dma_addr_t *iova,
718 size_t len, phys_addr_t phys,
721 size_t unmapped = iommu_unmap(domain->domain, *iova, len);
724 *unlocked += vfio_unpin_pages_remote(dma, *iova,
726 unmapped >> PAGE_SHIFT,
734 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
737 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
738 struct vfio_domain *domain, *d;
739 LIST_HEAD(unmapped_region_list);
740 struct iommu_iotlb_gather iotlb_gather;
741 int unmapped_region_cnt = 0;
747 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
751 * We use the IOMMU to track the physical addresses, otherwise we'd
752 * need a much more complicated tracking system. Unfortunately that
753 * means we need to use one of the iommu domains to figure out the
754 * pfns to unpin. The rest need to be unmapped in advance so we have
755 * no iommu translations remaining when the pages are unpinned.
757 domain = d = list_first_entry(&iommu->domain_list,
758 struct vfio_domain, next);
760 list_for_each_entry_continue(d, &iommu->domain_list, next) {
761 iommu_unmap(d->domain, dma->iova, dma->size);
765 iommu_iotlb_gather_init(&iotlb_gather);
767 size_t unmapped, len;
768 phys_addr_t phys, next;
770 phys = iommu_iova_to_phys(domain->domain, iova);
771 if (WARN_ON(!phys)) {
777 * To optimize for fewer iommu_unmap() calls, each of which
778 * may require hardware cache flushing, try to find the
779 * largest contiguous physical memory chunk to unmap.
781 for (len = PAGE_SIZE;
782 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
783 next = iommu_iova_to_phys(domain->domain, iova + len);
784 if (next != phys + len)
789 * First, try to use fast unmap/unpin. In case of failure,
790 * switch to slow unmap/unpin path.
792 unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys,
793 &unlocked, &unmapped_region_list,
794 &unmapped_region_cnt,
797 unmapped = unmap_unpin_slow(domain, dma, &iova, len,
799 if (WARN_ON(!unmapped))
804 dma->iommu_mapped = false;
806 if (unmapped_region_cnt) {
807 unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list,
812 vfio_lock_acct(dma, -unlocked, true);
818 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
820 vfio_unmap_unpin(iommu, dma, true);
821 vfio_unlink_dma(iommu, dma);
822 put_task_struct(dma->task);
827 static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu)
829 struct vfio_domain *domain;
830 unsigned long bitmap = ULONG_MAX;
832 mutex_lock(&iommu->lock);
833 list_for_each_entry(domain, &iommu->domain_list, next)
834 bitmap &= domain->domain->pgsize_bitmap;
835 mutex_unlock(&iommu->lock);
838 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
839 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
840 * That way the user will be able to map/unmap buffers whose size/
841 * start address is aligned with PAGE_SIZE. Pinning code uses that
842 * granularity while iommu driver can use the sub-PAGE_SIZE size
845 if (bitmap & ~PAGE_MASK) {
853 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
854 struct vfio_iommu_type1_dma_unmap *unmap)
857 struct vfio_dma *dma, *dma_last = NULL;
859 int ret = 0, retries = 0;
861 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
863 if (unmap->iova & mask)
865 if (!unmap->size || unmap->size & mask)
867 if (unmap->iova + unmap->size - 1 < unmap->iova ||
868 unmap->size > SIZE_MAX)
871 WARN_ON(mask & PAGE_MASK);
873 mutex_lock(&iommu->lock);
876 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
877 * avoid tracking individual mappings. This means that the granularity
878 * of the original mapping was lost and the user was allowed to attempt
879 * to unmap any range. Depending on the contiguousness of physical
880 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
881 * or may not have worked. We only guaranteed unmap granularity
882 * matching the original mapping; even though it was untracked here,
883 * the original mappings are reflected in IOMMU mappings. This
884 * resulted in a couple unusual behaviors. First, if a range is not
885 * able to be unmapped, ex. a set of 4k pages that was mapped as a
886 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
887 * a zero sized unmap. Also, if an unmap request overlaps the first
888 * address of a hugepage, the IOMMU will unmap the entire hugepage.
889 * This also returns success and the returned unmap size reflects the
890 * actual size unmapped.
892 * We attempt to maintain compatibility with this "v1" interface, but
893 * we take control out of the hands of the IOMMU. Therefore, an unmap
894 * request offset from the beginning of the original mapping will
895 * return success with zero sized unmap. And an unmap request covering
896 * the first iova of mapping will unmap the entire range.
898 * The v2 version of this interface intends to be more deterministic.
899 * Unmap requests must fully cover previous mappings. Multiple
900 * mappings may still be unmaped by specifying large ranges, but there
901 * must not be any previous mappings bisected by the range. An error
902 * will be returned if these conditions are not met. The v2 interface
903 * will only return success and a size of zero if there were no
904 * mappings within the range.
907 dma = vfio_find_dma(iommu, unmap->iova, 1);
908 if (dma && dma->iova != unmap->iova) {
912 dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0);
913 if (dma && dma->iova + dma->size != unmap->iova + unmap->size) {
919 while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
920 if (!iommu->v2 && unmap->iova > dma->iova)
923 * Task with same address space who mapped this iova range is
924 * allowed to unmap the iova range.
926 if (dma->task->mm != current->mm)
929 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
930 struct vfio_iommu_type1_dma_unmap nb_unmap;
932 if (dma_last == dma) {
933 BUG_ON(++retries > 10);
939 nb_unmap.iova = dma->iova;
940 nb_unmap.size = dma->size;
943 * Notify anyone (mdev vendor drivers) to invalidate and
944 * unmap iovas within the range we're about to unmap.
945 * Vendor drivers MUST unpin pages in response to an
948 mutex_unlock(&iommu->lock);
949 blocking_notifier_call_chain(&iommu->notifier,
950 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
954 unmapped += dma->size;
955 vfio_remove_dma(iommu, dma);
959 mutex_unlock(&iommu->lock);
961 /* Report how much was unmapped */
962 unmap->size = unmapped;
967 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
968 unsigned long pfn, long npage, int prot)
970 struct vfio_domain *d;
973 list_for_each_entry(d, &iommu->domain_list, next) {
974 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
975 npage << PAGE_SHIFT, prot | d->prot);
985 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next)
986 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
991 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
994 dma_addr_t iova = dma->iova;
995 unsigned long vaddr = dma->vaddr;
996 size_t size = map_size;
998 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1002 /* Pin a contiguous chunk of memory */
1003 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
1004 size >> PAGE_SHIFT, &pfn, limit);
1012 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
1015 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
1020 size -= npage << PAGE_SHIFT;
1021 dma->size += npage << PAGE_SHIFT;
1024 dma->iommu_mapped = true;
1027 vfio_remove_dma(iommu, dma);
1033 * Check dma map request is within a valid iova range
1035 static bool vfio_iommu_iova_dma_valid(struct vfio_iommu *iommu,
1036 dma_addr_t start, dma_addr_t end)
1038 struct list_head *iova = &iommu->iova_list;
1039 struct vfio_iova *node;
1041 list_for_each_entry(node, iova, list) {
1042 if (start >= node->start && end <= node->end)
1047 * Check for list_empty() as well since a container with
1048 * a single mdev device will have an empty list.
1050 return list_empty(iova);
1053 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1054 struct vfio_iommu_type1_dma_map *map)
1056 dma_addr_t iova = map->iova;
1057 unsigned long vaddr = map->vaddr;
1058 size_t size = map->size;
1059 int ret = 0, prot = 0;
1061 struct vfio_dma *dma;
1063 /* Verify that none of our __u64 fields overflow */
1064 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1067 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
1069 WARN_ON(mask & PAGE_MASK);
1071 /* READ/WRITE from device perspective */
1072 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1073 prot |= IOMMU_WRITE;
1074 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1077 if (!prot || !size || (size | iova | vaddr) & mask)
1080 /* Don't allow IOVA or virtual address wrap */
1081 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr)
1084 mutex_lock(&iommu->lock);
1086 if (vfio_find_dma(iommu, iova, size)) {
1091 if (!iommu->dma_avail) {
1096 if (!vfio_iommu_iova_dma_valid(iommu, iova, iova + size - 1)) {
1101 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1113 * We need to be able to both add to a task's locked memory and test
1114 * against the locked memory limit and we need to be able to do both
1115 * outside of this call path as pinning can be asynchronous via the
1116 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1117 * task_struct and VM locked pages requires an mm_struct, however
1118 * holding an indefinite mm reference is not recommended, therefore we
1119 * only hold a reference to a task. We could hold a reference to
1120 * current, however QEMU uses this call path through vCPU threads,
1121 * which can be killed resulting in a NULL mm and failure in the unmap
1122 * path when called via a different thread. Avoid this problem by
1123 * using the group_leader as threads within the same group require
1124 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1127 * Previously we also used the task for testing CAP_IPC_LOCK at the
1128 * time of pinning and accounting, however has_capability() makes use
1129 * of real_cred, a copy-on-write field, so we can't guarantee that it
1130 * matches group_leader, or in fact that it might not change by the
1131 * time it's evaluated. If a process were to call MAP_DMA with
1132 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1133 * possibly see different results for an iommu_mapped vfio_dma vs
1134 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1135 * time of calling MAP_DMA.
1137 get_task_struct(current->group_leader);
1138 dma->task = current->group_leader;
1139 dma->lock_cap = capable(CAP_IPC_LOCK);
1141 dma->pfn_list = RB_ROOT;
1143 /* Insert zero-sized and grow as we map chunks of it */
1144 vfio_link_dma(iommu, dma);
1146 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1147 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1150 ret = vfio_pin_map_dma(iommu, dma, size);
1153 mutex_unlock(&iommu->lock);
1157 static int vfio_bus_type(struct device *dev, void *data)
1159 struct bus_type **bus = data;
1161 if (*bus && *bus != dev->bus)
1169 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1170 struct vfio_domain *domain)
1172 struct vfio_domain *d;
1174 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1177 /* Arbitrarily pick the first domain in the list for lookups */
1178 d = list_first_entry(&iommu->domain_list, struct vfio_domain, next);
1179 n = rb_first(&iommu->dma_list);
1181 for (; n; n = rb_next(n)) {
1182 struct vfio_dma *dma;
1185 dma = rb_entry(n, struct vfio_dma, node);
1188 while (iova < dma->iova + dma->size) {
1192 if (dma->iommu_mapped) {
1196 phys = iommu_iova_to_phys(d->domain, iova);
1198 if (WARN_ON(!phys)) {
1206 while (i < dma->iova + dma->size &&
1207 p == iommu_iova_to_phys(d->domain, i)) {
1214 unsigned long vaddr = dma->vaddr +
1216 size_t n = dma->iova + dma->size - iova;
1219 npage = vfio_pin_pages_remote(dma, vaddr,
1228 phys = pfn << PAGE_SHIFT;
1229 size = npage << PAGE_SHIFT;
1232 ret = iommu_map(domain->domain, iova, phys,
1233 size, dma->prot | domain->prot);
1239 dma->iommu_mapped = true;
1245 * We change our unmap behavior slightly depending on whether the IOMMU
1246 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1247 * for practically any contiguous power-of-two mapping we give it. This means
1248 * we don't need to look for contiguous chunks ourselves to make unmapping
1249 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1250 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1251 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1252 * hugetlbfs is in use.
1254 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1257 int ret, order = get_order(PAGE_SIZE * 2);
1259 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1263 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1264 IOMMU_READ | IOMMU_WRITE | domain->prot);
1266 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1268 if (unmapped == PAGE_SIZE)
1269 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1271 domain->fgsp = true;
1274 __free_pages(pages, order);
1277 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1278 struct iommu_group *iommu_group)
1280 struct vfio_group *g;
1282 list_for_each_entry(g, &domain->group_list, next) {
1283 if (g->iommu_group == iommu_group)
1290 static bool vfio_iommu_has_sw_msi(struct list_head *group_resv_regions,
1293 struct iommu_resv_region *region;
1296 list_for_each_entry(region, group_resv_regions, list) {
1298 * The presence of any 'real' MSI regions should take
1299 * precedence over the software-managed one if the
1300 * IOMMU driver happens to advertise both types.
1302 if (region->type == IOMMU_RESV_MSI) {
1307 if (region->type == IOMMU_RESV_SW_MSI) {
1308 *base = region->start;
1316 static struct device *vfio_mdev_get_iommu_device(struct device *dev)
1318 struct device *(*fn)(struct device *dev);
1319 struct device *iommu_device;
1321 fn = symbol_get(mdev_get_iommu_device);
1323 iommu_device = fn(dev);
1324 symbol_put(mdev_get_iommu_device);
1326 return iommu_device;
1332 static int vfio_mdev_attach_domain(struct device *dev, void *data)
1334 struct iommu_domain *domain = data;
1335 struct device *iommu_device;
1337 iommu_device = vfio_mdev_get_iommu_device(dev);
1339 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1340 return iommu_aux_attach_device(domain, iommu_device);
1342 return iommu_attach_device(domain, iommu_device);
1348 static int vfio_mdev_detach_domain(struct device *dev, void *data)
1350 struct iommu_domain *domain = data;
1351 struct device *iommu_device;
1353 iommu_device = vfio_mdev_get_iommu_device(dev);
1355 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1356 iommu_aux_detach_device(domain, iommu_device);
1358 iommu_detach_device(domain, iommu_device);
1364 static int vfio_iommu_attach_group(struct vfio_domain *domain,
1365 struct vfio_group *group)
1367 if (group->mdev_group)
1368 return iommu_group_for_each_dev(group->iommu_group,
1370 vfio_mdev_attach_domain);
1372 return iommu_attach_group(domain->domain, group->iommu_group);
1375 static void vfio_iommu_detach_group(struct vfio_domain *domain,
1376 struct vfio_group *group)
1378 if (group->mdev_group)
1379 iommu_group_for_each_dev(group->iommu_group, domain->domain,
1380 vfio_mdev_detach_domain);
1382 iommu_detach_group(domain->domain, group->iommu_group);
1385 static bool vfio_bus_is_mdev(struct bus_type *bus)
1387 struct bus_type *mdev_bus;
1390 mdev_bus = symbol_get(mdev_bus_type);
1392 ret = (bus == mdev_bus);
1393 symbol_put(mdev_bus_type);
1399 static int vfio_mdev_iommu_device(struct device *dev, void *data)
1401 struct device **old = data, *new;
1403 new = vfio_mdev_get_iommu_device(dev);
1404 if (!new || (*old && *old != new))
1413 * This is a helper function to insert an address range to iova list.
1414 * The list is initially created with a single entry corresponding to
1415 * the IOMMU domain geometry to which the device group is attached.
1416 * The list aperture gets modified when a new domain is added to the
1417 * container if the new aperture doesn't conflict with the current one
1418 * or with any existing dma mappings. The list is also modified to
1419 * exclude any reserved regions associated with the device group.
1421 static int vfio_iommu_iova_insert(struct list_head *head,
1422 dma_addr_t start, dma_addr_t end)
1424 struct vfio_iova *region;
1426 region = kmalloc(sizeof(*region), GFP_KERNEL);
1430 INIT_LIST_HEAD(®ion->list);
1431 region->start = start;
1434 list_add_tail(®ion->list, head);
1439 * Check the new iommu aperture conflicts with existing aper or with any
1440 * existing dma mappings.
1442 static bool vfio_iommu_aper_conflict(struct vfio_iommu *iommu,
1443 dma_addr_t start, dma_addr_t end)
1445 struct vfio_iova *first, *last;
1446 struct list_head *iova = &iommu->iova_list;
1448 if (list_empty(iova))
1451 /* Disjoint sets, return conflict */
1452 first = list_first_entry(iova, struct vfio_iova, list);
1453 last = list_last_entry(iova, struct vfio_iova, list);
1454 if (start > last->end || end < first->start)
1457 /* Check for any existing dma mappings below the new start */
1458 if (start > first->start) {
1459 if (vfio_find_dma(iommu, first->start, start - first->start))
1463 /* Check for any existing dma mappings beyond the new end */
1464 if (end < last->end) {
1465 if (vfio_find_dma(iommu, end + 1, last->end - end))
1473 * Resize iommu iova aperture window. This is called only if the new
1474 * aperture has no conflict with existing aperture and dma mappings.
1476 static int vfio_iommu_aper_resize(struct list_head *iova,
1477 dma_addr_t start, dma_addr_t end)
1479 struct vfio_iova *node, *next;
1481 if (list_empty(iova))
1482 return vfio_iommu_iova_insert(iova, start, end);
1484 /* Adjust iova list start */
1485 list_for_each_entry_safe(node, next, iova, list) {
1486 if (start < node->start)
1488 if (start >= node->start && start < node->end) {
1489 node->start = start;
1492 /* Delete nodes before new start */
1493 list_del(&node->list);
1497 /* Adjust iova list end */
1498 list_for_each_entry_safe(node, next, iova, list) {
1499 if (end > node->end)
1501 if (end > node->start && end <= node->end) {
1505 /* Delete nodes after new end */
1506 list_del(&node->list);
1514 * Check reserved region conflicts with existing dma mappings
1516 static bool vfio_iommu_resv_conflict(struct vfio_iommu *iommu,
1517 struct list_head *resv_regions)
1519 struct iommu_resv_region *region;
1521 /* Check for conflict with existing dma mappings */
1522 list_for_each_entry(region, resv_regions, list) {
1523 if (region->type == IOMMU_RESV_DIRECT_RELAXABLE)
1526 if (vfio_find_dma(iommu, region->start, region->length))
1534 * Check iova region overlap with reserved regions and
1535 * exclude them from the iommu iova range
1537 static int vfio_iommu_resv_exclude(struct list_head *iova,
1538 struct list_head *resv_regions)
1540 struct iommu_resv_region *resv;
1541 struct vfio_iova *n, *next;
1543 list_for_each_entry(resv, resv_regions, list) {
1544 phys_addr_t start, end;
1546 if (resv->type == IOMMU_RESV_DIRECT_RELAXABLE)
1549 start = resv->start;
1550 end = resv->start + resv->length - 1;
1552 list_for_each_entry_safe(n, next, iova, list) {
1556 if (start > n->end || end < n->start)
1559 * Insert a new node if current node overlaps with the
1560 * reserve region to exlude that from valid iova range.
1561 * Note that, new node is inserted before the current
1562 * node and finally the current node is deleted keeping
1563 * the list updated and sorted.
1565 if (start > n->start)
1566 ret = vfio_iommu_iova_insert(&n->list, n->start,
1568 if (!ret && end < n->end)
1569 ret = vfio_iommu_iova_insert(&n->list, end + 1,
1579 if (list_empty(iova))
1585 static void vfio_iommu_resv_free(struct list_head *resv_regions)
1587 struct iommu_resv_region *n, *next;
1589 list_for_each_entry_safe(n, next, resv_regions, list) {
1595 static void vfio_iommu_iova_free(struct list_head *iova)
1597 struct vfio_iova *n, *next;
1599 list_for_each_entry_safe(n, next, iova, list) {
1605 static int vfio_iommu_iova_get_copy(struct vfio_iommu *iommu,
1606 struct list_head *iova_copy)
1608 struct list_head *iova = &iommu->iova_list;
1609 struct vfio_iova *n;
1612 list_for_each_entry(n, iova, list) {
1613 ret = vfio_iommu_iova_insert(iova_copy, n->start, n->end);
1621 vfio_iommu_iova_free(iova_copy);
1625 static void vfio_iommu_iova_insert_copy(struct vfio_iommu *iommu,
1626 struct list_head *iova_copy)
1628 struct list_head *iova = &iommu->iova_list;
1630 vfio_iommu_iova_free(iova);
1632 list_splice_tail(iova_copy, iova);
1634 static int vfio_iommu_type1_attach_group(void *iommu_data,
1635 struct iommu_group *iommu_group)
1637 struct vfio_iommu *iommu = iommu_data;
1638 struct vfio_group *group;
1639 struct vfio_domain *domain, *d;
1640 struct bus_type *bus = NULL;
1642 bool resv_msi, msi_remap;
1643 phys_addr_t resv_msi_base = 0;
1644 struct iommu_domain_geometry geo;
1645 LIST_HEAD(iova_copy);
1646 LIST_HEAD(group_resv_regions);
1648 mutex_lock(&iommu->lock);
1650 list_for_each_entry(d, &iommu->domain_list, next) {
1651 if (find_iommu_group(d, iommu_group)) {
1652 mutex_unlock(&iommu->lock);
1657 if (iommu->external_domain) {
1658 if (find_iommu_group(iommu->external_domain, iommu_group)) {
1659 mutex_unlock(&iommu->lock);
1664 group = kzalloc(sizeof(*group), GFP_KERNEL);
1665 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
1666 if (!group || !domain) {
1671 group->iommu_group = iommu_group;
1673 /* Determine bus_type in order to allocate a domain */
1674 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
1678 if (vfio_bus_is_mdev(bus)) {
1679 struct device *iommu_device = NULL;
1681 group->mdev_group = true;
1683 /* Determine the isolation type */
1684 ret = iommu_group_for_each_dev(iommu_group, &iommu_device,
1685 vfio_mdev_iommu_device);
1686 if (ret || !iommu_device) {
1687 if (!iommu->external_domain) {
1688 INIT_LIST_HEAD(&domain->group_list);
1689 iommu->external_domain = domain;
1694 list_add(&group->next,
1695 &iommu->external_domain->group_list);
1696 mutex_unlock(&iommu->lock);
1701 bus = iommu_device->bus;
1704 domain->domain = iommu_domain_alloc(bus);
1705 if (!domain->domain) {
1710 if (iommu->nesting) {
1713 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
1719 ret = vfio_iommu_attach_group(domain, group);
1723 /* Get aperture info */
1724 iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY, &geo);
1726 if (vfio_iommu_aper_conflict(iommu, geo.aperture_start,
1727 geo.aperture_end)) {
1732 ret = iommu_get_group_resv_regions(iommu_group, &group_resv_regions);
1736 if (vfio_iommu_resv_conflict(iommu, &group_resv_regions)) {
1742 * We don't want to work on the original iova list as the list
1743 * gets modified and in case of failure we have to retain the
1744 * original list. Get a copy here.
1746 ret = vfio_iommu_iova_get_copy(iommu, &iova_copy);
1750 ret = vfio_iommu_aper_resize(&iova_copy, geo.aperture_start,
1755 ret = vfio_iommu_resv_exclude(&iova_copy, &group_resv_regions);
1759 resv_msi = vfio_iommu_has_sw_msi(&group_resv_regions, &resv_msi_base);
1761 INIT_LIST_HEAD(&domain->group_list);
1762 list_add(&group->next, &domain->group_list);
1764 msi_remap = irq_domain_check_msi_remap() ||
1765 iommu_capable(bus, IOMMU_CAP_INTR_REMAP);
1767 if (!allow_unsafe_interrupts && !msi_remap) {
1768 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
1774 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
1775 domain->prot |= IOMMU_CACHE;
1778 * Try to match an existing compatible domain. We don't want to
1779 * preclude an IOMMU driver supporting multiple bus_types and being
1780 * able to include different bus_types in the same IOMMU domain, so
1781 * we test whether the domains use the same iommu_ops rather than
1782 * testing if they're on the same bus_type.
1784 list_for_each_entry(d, &iommu->domain_list, next) {
1785 if (d->domain->ops == domain->domain->ops &&
1786 d->prot == domain->prot) {
1787 vfio_iommu_detach_group(domain, group);
1788 if (!vfio_iommu_attach_group(d, group)) {
1789 list_add(&group->next, &d->group_list);
1790 iommu_domain_free(domain->domain);
1795 ret = vfio_iommu_attach_group(domain, group);
1801 vfio_test_domain_fgsp(domain);
1803 /* replay mappings on new domains */
1804 ret = vfio_iommu_replay(iommu, domain);
1809 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
1814 list_add(&domain->next, &iommu->domain_list);
1816 /* Delete the old one and insert new iova list */
1817 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
1818 mutex_unlock(&iommu->lock);
1819 vfio_iommu_resv_free(&group_resv_regions);
1824 vfio_iommu_detach_group(domain, group);
1826 iommu_domain_free(domain->domain);
1827 vfio_iommu_iova_free(&iova_copy);
1828 vfio_iommu_resv_free(&group_resv_regions);
1832 mutex_unlock(&iommu->lock);
1836 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
1838 struct rb_node *node;
1840 while ((node = rb_first(&iommu->dma_list)))
1841 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
1844 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
1846 struct rb_node *n, *p;
1848 n = rb_first(&iommu->dma_list);
1849 for (; n; n = rb_next(n)) {
1850 struct vfio_dma *dma;
1851 long locked = 0, unlocked = 0;
1853 dma = rb_entry(n, struct vfio_dma, node);
1854 unlocked += vfio_unmap_unpin(iommu, dma, false);
1855 p = rb_first(&dma->pfn_list);
1856 for (; p; p = rb_next(p)) {
1857 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
1860 if (!is_invalid_reserved_pfn(vpfn->pfn))
1863 vfio_lock_acct(dma, locked - unlocked, true);
1867 static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu)
1871 n = rb_first(&iommu->dma_list);
1872 for (; n; n = rb_next(n)) {
1873 struct vfio_dma *dma;
1875 dma = rb_entry(n, struct vfio_dma, node);
1877 if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list)))
1880 /* mdev vendor driver must unregister notifier */
1881 WARN_ON(iommu->notifier.head);
1885 * Called when a domain is removed in detach. It is possible that
1886 * the removed domain decided the iova aperture window. Modify the
1887 * iova aperture with the smallest window among existing domains.
1889 static void vfio_iommu_aper_expand(struct vfio_iommu *iommu,
1890 struct list_head *iova_copy)
1892 struct vfio_domain *domain;
1893 struct iommu_domain_geometry geo;
1894 struct vfio_iova *node;
1895 dma_addr_t start = 0;
1896 dma_addr_t end = (dma_addr_t)~0;
1898 if (list_empty(iova_copy))
1901 list_for_each_entry(domain, &iommu->domain_list, next) {
1902 iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY,
1904 if (geo.aperture_start > start)
1905 start = geo.aperture_start;
1906 if (geo.aperture_end < end)
1907 end = geo.aperture_end;
1910 /* Modify aperture limits. The new aper is either same or bigger */
1911 node = list_first_entry(iova_copy, struct vfio_iova, list);
1912 node->start = start;
1913 node = list_last_entry(iova_copy, struct vfio_iova, list);
1918 * Called when a group is detached. The reserved regions for that
1919 * group can be part of valid iova now. But since reserved regions
1920 * may be duplicated among groups, populate the iova valid regions
1923 static int vfio_iommu_resv_refresh(struct vfio_iommu *iommu,
1924 struct list_head *iova_copy)
1926 struct vfio_domain *d;
1927 struct vfio_group *g;
1928 struct vfio_iova *node;
1929 dma_addr_t start, end;
1930 LIST_HEAD(resv_regions);
1933 if (list_empty(iova_copy))
1936 list_for_each_entry(d, &iommu->domain_list, next) {
1937 list_for_each_entry(g, &d->group_list, next) {
1938 ret = iommu_get_group_resv_regions(g->iommu_group,
1945 node = list_first_entry(iova_copy, struct vfio_iova, list);
1946 start = node->start;
1947 node = list_last_entry(iova_copy, struct vfio_iova, list);
1950 /* purge the iova list and create new one */
1951 vfio_iommu_iova_free(iova_copy);
1953 ret = vfio_iommu_aper_resize(iova_copy, start, end);
1957 /* Exclude current reserved regions from iova ranges */
1958 ret = vfio_iommu_resv_exclude(iova_copy, &resv_regions);
1960 vfio_iommu_resv_free(&resv_regions);
1964 static void vfio_iommu_type1_detach_group(void *iommu_data,
1965 struct iommu_group *iommu_group)
1967 struct vfio_iommu *iommu = iommu_data;
1968 struct vfio_domain *domain;
1969 struct vfio_group *group;
1970 LIST_HEAD(iova_copy);
1972 mutex_lock(&iommu->lock);
1974 if (iommu->external_domain) {
1975 group = find_iommu_group(iommu->external_domain, iommu_group);
1977 list_del(&group->next);
1980 if (list_empty(&iommu->external_domain->group_list)) {
1981 vfio_sanity_check_pfn_list(iommu);
1983 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1984 vfio_iommu_unmap_unpin_all(iommu);
1986 kfree(iommu->external_domain);
1987 iommu->external_domain = NULL;
1989 goto detach_group_done;
1994 * Get a copy of iova list. This will be used to update
1995 * and to replace the current one later. Please note that
1996 * we will leave the original list as it is if update fails.
1998 vfio_iommu_iova_get_copy(iommu, &iova_copy);
2000 list_for_each_entry(domain, &iommu->domain_list, next) {
2001 group = find_iommu_group(domain, iommu_group);
2005 vfio_iommu_detach_group(domain, group);
2006 list_del(&group->next);
2009 * Group ownership provides privilege, if the group list is
2010 * empty, the domain goes away. If it's the last domain with
2011 * iommu and external domain doesn't exist, then all the
2012 * mappings go away too. If it's the last domain with iommu and
2013 * external domain exist, update accounting
2015 if (list_empty(&domain->group_list)) {
2016 if (list_is_singular(&iommu->domain_list)) {
2017 if (!iommu->external_domain)
2018 vfio_iommu_unmap_unpin_all(iommu);
2020 vfio_iommu_unmap_unpin_reaccount(iommu);
2022 iommu_domain_free(domain->domain);
2023 list_del(&domain->next);
2025 vfio_iommu_aper_expand(iommu, &iova_copy);
2030 if (!vfio_iommu_resv_refresh(iommu, &iova_copy))
2031 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2033 vfio_iommu_iova_free(&iova_copy);
2036 mutex_unlock(&iommu->lock);
2039 static void *vfio_iommu_type1_open(unsigned long arg)
2041 struct vfio_iommu *iommu;
2043 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
2045 return ERR_PTR(-ENOMEM);
2048 case VFIO_TYPE1_IOMMU:
2050 case VFIO_TYPE1_NESTING_IOMMU:
2051 iommu->nesting = true;
2053 case VFIO_TYPE1v2_IOMMU:
2058 return ERR_PTR(-EINVAL);
2061 INIT_LIST_HEAD(&iommu->domain_list);
2062 INIT_LIST_HEAD(&iommu->iova_list);
2063 iommu->dma_list = RB_ROOT;
2064 iommu->dma_avail = dma_entry_limit;
2065 mutex_init(&iommu->lock);
2066 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
2071 static void vfio_release_domain(struct vfio_domain *domain, bool external)
2073 struct vfio_group *group, *group_tmp;
2075 list_for_each_entry_safe(group, group_tmp,
2076 &domain->group_list, next) {
2078 vfio_iommu_detach_group(domain, group);
2079 list_del(&group->next);
2084 iommu_domain_free(domain->domain);
2087 static void vfio_iommu_type1_release(void *iommu_data)
2089 struct vfio_iommu *iommu = iommu_data;
2090 struct vfio_domain *domain, *domain_tmp;
2092 if (iommu->external_domain) {
2093 vfio_release_domain(iommu->external_domain, true);
2094 vfio_sanity_check_pfn_list(iommu);
2095 kfree(iommu->external_domain);
2098 vfio_iommu_unmap_unpin_all(iommu);
2100 list_for_each_entry_safe(domain, domain_tmp,
2101 &iommu->domain_list, next) {
2102 vfio_release_domain(domain, false);
2103 list_del(&domain->next);
2107 vfio_iommu_iova_free(&iommu->iova_list);
2112 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
2114 struct vfio_domain *domain;
2117 mutex_lock(&iommu->lock);
2118 list_for_each_entry(domain, &iommu->domain_list, next) {
2119 if (!(domain->prot & IOMMU_CACHE)) {
2124 mutex_unlock(&iommu->lock);
2129 static int vfio_iommu_iova_add_cap(struct vfio_info_cap *caps,
2130 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas,
2133 struct vfio_info_cap_header *header;
2134 struct vfio_iommu_type1_info_cap_iova_range *iova_cap;
2136 header = vfio_info_cap_add(caps, size,
2137 VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE, 1);
2139 return PTR_ERR(header);
2141 iova_cap = container_of(header,
2142 struct vfio_iommu_type1_info_cap_iova_range,
2144 iova_cap->nr_iovas = cap_iovas->nr_iovas;
2145 memcpy(iova_cap->iova_ranges, cap_iovas->iova_ranges,
2146 cap_iovas->nr_iovas * sizeof(*cap_iovas->iova_ranges));
2150 static int vfio_iommu_iova_build_caps(struct vfio_iommu *iommu,
2151 struct vfio_info_cap *caps)
2153 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas;
2154 struct vfio_iova *iova;
2156 int iovas = 0, i = 0, ret;
2158 mutex_lock(&iommu->lock);
2160 list_for_each_entry(iova, &iommu->iova_list, list)
2165 * Return 0 as a container with a single mdev device
2166 * will have an empty list
2172 size = sizeof(*cap_iovas) + (iovas * sizeof(*cap_iovas->iova_ranges));
2174 cap_iovas = kzalloc(size, GFP_KERNEL);
2180 cap_iovas->nr_iovas = iovas;
2182 list_for_each_entry(iova, &iommu->iova_list, list) {
2183 cap_iovas->iova_ranges[i].start = iova->start;
2184 cap_iovas->iova_ranges[i].end = iova->end;
2188 ret = vfio_iommu_iova_add_cap(caps, cap_iovas, size);
2192 mutex_unlock(&iommu->lock);
2196 static long vfio_iommu_type1_ioctl(void *iommu_data,
2197 unsigned int cmd, unsigned long arg)
2199 struct vfio_iommu *iommu = iommu_data;
2200 unsigned long minsz;
2202 if (cmd == VFIO_CHECK_EXTENSION) {
2204 case VFIO_TYPE1_IOMMU:
2205 case VFIO_TYPE1v2_IOMMU:
2206 case VFIO_TYPE1_NESTING_IOMMU:
2208 case VFIO_DMA_CC_IOMMU:
2211 return vfio_domains_have_iommu_cache(iommu);
2215 } else if (cmd == VFIO_IOMMU_GET_INFO) {
2216 struct vfio_iommu_type1_info info;
2217 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
2218 unsigned long capsz;
2221 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
2223 /* For backward compatibility, cannot require this */
2224 capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
2226 if (copy_from_user(&info, (void __user *)arg, minsz))
2229 if (info.argsz < minsz)
2232 if (info.argsz >= capsz) {
2234 info.cap_offset = 0; /* output, no-recopy necessary */
2237 info.flags = VFIO_IOMMU_INFO_PGSIZES;
2239 info.iova_pgsizes = vfio_pgsize_bitmap(iommu);
2241 ret = vfio_iommu_iova_build_caps(iommu, &caps);
2246 info.flags |= VFIO_IOMMU_INFO_CAPS;
2248 if (info.argsz < sizeof(info) + caps.size) {
2249 info.argsz = sizeof(info) + caps.size;
2251 vfio_info_cap_shift(&caps, sizeof(info));
2252 if (copy_to_user((void __user *)arg +
2253 sizeof(info), caps.buf,
2258 info.cap_offset = sizeof(info);
2264 return copy_to_user((void __user *)arg, &info, minsz) ?
2267 } else if (cmd == VFIO_IOMMU_MAP_DMA) {
2268 struct vfio_iommu_type1_dma_map map;
2269 uint32_t mask = VFIO_DMA_MAP_FLAG_READ |
2270 VFIO_DMA_MAP_FLAG_WRITE;
2272 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
2274 if (copy_from_user(&map, (void __user *)arg, minsz))
2277 if (map.argsz < minsz || map.flags & ~mask)
2280 return vfio_dma_do_map(iommu, &map);
2282 } else if (cmd == VFIO_IOMMU_UNMAP_DMA) {
2283 struct vfio_iommu_type1_dma_unmap unmap;
2286 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
2288 if (copy_from_user(&unmap, (void __user *)arg, minsz))
2291 if (unmap.argsz < minsz || unmap.flags)
2294 ret = vfio_dma_do_unmap(iommu, &unmap);
2298 return copy_to_user((void __user *)arg, &unmap, minsz) ?
2305 static int vfio_iommu_type1_register_notifier(void *iommu_data,
2306 unsigned long *events,
2307 struct notifier_block *nb)
2309 struct vfio_iommu *iommu = iommu_data;
2311 /* clear known events */
2312 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
2314 /* refuse to register if still events remaining */
2318 return blocking_notifier_chain_register(&iommu->notifier, nb);
2321 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
2322 struct notifier_block *nb)
2324 struct vfio_iommu *iommu = iommu_data;
2326 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
2329 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
2330 .name = "vfio-iommu-type1",
2331 .owner = THIS_MODULE,
2332 .open = vfio_iommu_type1_open,
2333 .release = vfio_iommu_type1_release,
2334 .ioctl = vfio_iommu_type1_ioctl,
2335 .attach_group = vfio_iommu_type1_attach_group,
2336 .detach_group = vfio_iommu_type1_detach_group,
2337 .pin_pages = vfio_iommu_type1_pin_pages,
2338 .unpin_pages = vfio_iommu_type1_unpin_pages,
2339 .register_notifier = vfio_iommu_type1_register_notifier,
2340 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
2343 static int __init vfio_iommu_type1_init(void)
2345 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
2348 static void __exit vfio_iommu_type1_cleanup(void)
2350 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
2353 module_init(vfio_iommu_type1_init);
2354 module_exit(vfio_iommu_type1_cleanup);
2356 MODULE_VERSION(DRIVER_VERSION);
2357 MODULE_LICENSE("GPL v2");
2358 MODULE_AUTHOR(DRIVER_AUTHOR);
2359 MODULE_DESCRIPTION(DRIVER_DESC);