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.
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);
313 unpin_user_pages_dirty_lock(&page, 1, prot & IOMMU_WRITE);
319 static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,
320 int prot, unsigned long *pfn)
322 struct page *page[1];
323 struct vm_area_struct *vma;
324 unsigned int flags = 0;
327 if (prot & IOMMU_WRITE)
330 down_read(&mm->mmap_sem);
331 ret = pin_user_pages_remote(NULL, mm, vaddr, 1, flags | FOLL_LONGTERM,
334 *pfn = page_to_pfn(page[0]);
339 vaddr = untagged_addr(vaddr);
341 vma = find_vma_intersection(mm, vaddr, vaddr + 1);
343 if (vma && vma->vm_flags & VM_PFNMAP) {
344 *pfn = ((vaddr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
345 if (is_invalid_reserved_pfn(*pfn))
349 up_read(&mm->mmap_sem);
354 * Attempt to pin pages. We really don't want to track all the pfns and
355 * the iommu can only map chunks of consecutive pfns anyway, so get the
356 * first page and all consecutive pages with the same locking.
358 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
359 long npage, unsigned long *pfn_base,
362 unsigned long pfn = 0;
363 long ret, pinned = 0, lock_acct = 0;
365 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
367 /* This code path is only user initiated */
371 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base);
376 rsvd = is_invalid_reserved_pfn(*pfn_base);
379 * Reserved pages aren't counted against the user, externally pinned
380 * pages are already counted against the user.
382 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
383 if (!dma->lock_cap && current->mm->locked_vm + 1 > limit) {
384 put_pfn(*pfn_base, dma->prot);
385 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__,
386 limit << PAGE_SHIFT);
392 if (unlikely(disable_hugepages))
395 /* Lock all the consecutive pages from pfn_base */
396 for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage;
397 pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) {
398 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn);
402 if (pfn != *pfn_base + pinned ||
403 rsvd != is_invalid_reserved_pfn(pfn)) {
404 put_pfn(pfn, dma->prot);
408 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
409 if (!dma->lock_cap &&
410 current->mm->locked_vm + lock_acct + 1 > limit) {
411 put_pfn(pfn, dma->prot);
412 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
413 __func__, limit << PAGE_SHIFT);
422 ret = vfio_lock_acct(dma, lock_acct, false);
427 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
428 put_pfn(pfn, dma->prot);
437 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
438 unsigned long pfn, long npage,
441 long unlocked = 0, locked = 0;
444 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
445 if (put_pfn(pfn++, dma->prot)) {
447 if (vfio_find_vpfn(dma, iova))
453 vfio_lock_acct(dma, locked - unlocked, true);
458 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
459 unsigned long *pfn_base, bool do_accounting)
461 struct mm_struct *mm;
464 mm = get_task_mm(dma->task);
468 ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base);
469 if (!ret && do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
470 ret = vfio_lock_acct(dma, 1, true);
472 put_pfn(*pfn_base, dma->prot);
474 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
475 "(%ld) exceeded\n", __func__,
476 dma->task->comm, task_pid_nr(dma->task),
477 task_rlimit(dma->task, RLIMIT_MEMLOCK));
485 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
489 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
494 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
497 vfio_lock_acct(dma, -unlocked, true);
502 static int vfio_iommu_type1_pin_pages(void *iommu_data,
503 unsigned long *user_pfn,
505 unsigned long *phys_pfn)
507 struct vfio_iommu *iommu = iommu_data;
509 unsigned long remote_vaddr;
510 struct vfio_dma *dma;
513 if (!iommu || !user_pfn || !phys_pfn)
516 /* Supported for v2 version only */
520 mutex_lock(&iommu->lock);
522 /* Fail if notifier list is empty */
523 if (!iommu->notifier.head) {
529 * If iommu capable domain exist in the container then all pages are
530 * already pinned and accounted. Accouting should be done if there is no
531 * iommu capable domain in the container.
533 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
535 for (i = 0; i < npage; i++) {
537 struct vfio_pfn *vpfn;
539 iova = user_pfn[i] << PAGE_SHIFT;
540 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
546 if ((dma->prot & prot) != prot) {
551 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
553 phys_pfn[i] = vpfn->pfn;
557 remote_vaddr = dma->vaddr + iova - dma->iova;
558 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
563 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
565 vfio_unpin_page_external(dma, iova, do_accounting);
575 for (j = 0; j < i; j++) {
578 iova = user_pfn[j] << PAGE_SHIFT;
579 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
580 vfio_unpin_page_external(dma, iova, do_accounting);
584 mutex_unlock(&iommu->lock);
588 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
589 unsigned long *user_pfn,
592 struct vfio_iommu *iommu = iommu_data;
596 if (!iommu || !user_pfn)
599 /* Supported for v2 version only */
603 mutex_lock(&iommu->lock);
605 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
606 for (i = 0; i < npage; i++) {
607 struct vfio_dma *dma;
610 iova = user_pfn[i] << PAGE_SHIFT;
611 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
614 vfio_unpin_page_external(dma, iova, do_accounting);
618 mutex_unlock(&iommu->lock);
619 return i > npage ? npage : (i > 0 ? i : -EINVAL);
622 static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
623 struct list_head *regions,
624 struct iommu_iotlb_gather *iotlb_gather)
627 struct vfio_regions *entry, *next;
629 iommu_tlb_sync(domain->domain, iotlb_gather);
631 list_for_each_entry_safe(entry, next, regions, list) {
632 unlocked += vfio_unpin_pages_remote(dma,
634 entry->phys >> PAGE_SHIFT,
635 entry->len >> PAGE_SHIFT,
637 list_del(&entry->list);
647 * Generally, VFIO needs to unpin remote pages after each IOTLB flush.
648 * Therefore, when using IOTLB flush sync interface, VFIO need to keep track
649 * of these regions (currently using a list).
651 * This value specifies maximum number of regions for each IOTLB flush sync.
653 #define VFIO_IOMMU_TLB_SYNC_MAX 512
655 static size_t unmap_unpin_fast(struct vfio_domain *domain,
656 struct vfio_dma *dma, dma_addr_t *iova,
657 size_t len, phys_addr_t phys, long *unlocked,
658 struct list_head *unmapped_list,
660 struct iommu_iotlb_gather *iotlb_gather)
663 struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
666 unmapped = iommu_unmap_fast(domain->domain, *iova, len,
674 entry->len = unmapped;
675 list_add_tail(&entry->list, unmapped_list);
683 * Sync if the number of fast-unmap regions hits the limit
684 * or in case of errors.
686 if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
687 *unlocked += vfio_sync_unpin(dma, domain, unmapped_list,
695 static size_t unmap_unpin_slow(struct vfio_domain *domain,
696 struct vfio_dma *dma, dma_addr_t *iova,
697 size_t len, phys_addr_t phys,
700 size_t unmapped = iommu_unmap(domain->domain, *iova, len);
703 *unlocked += vfio_unpin_pages_remote(dma, *iova,
705 unmapped >> PAGE_SHIFT,
713 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
716 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
717 struct vfio_domain *domain, *d;
718 LIST_HEAD(unmapped_region_list);
719 struct iommu_iotlb_gather iotlb_gather;
720 int unmapped_region_cnt = 0;
726 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
730 * We use the IOMMU to track the physical addresses, otherwise we'd
731 * need a much more complicated tracking system. Unfortunately that
732 * means we need to use one of the iommu domains to figure out the
733 * pfns to unpin. The rest need to be unmapped in advance so we have
734 * no iommu translations remaining when the pages are unpinned.
736 domain = d = list_first_entry(&iommu->domain_list,
737 struct vfio_domain, next);
739 list_for_each_entry_continue(d, &iommu->domain_list, next) {
740 iommu_unmap(d->domain, dma->iova, dma->size);
744 iommu_iotlb_gather_init(&iotlb_gather);
746 size_t unmapped, len;
747 phys_addr_t phys, next;
749 phys = iommu_iova_to_phys(domain->domain, iova);
750 if (WARN_ON(!phys)) {
756 * To optimize for fewer iommu_unmap() calls, each of which
757 * may require hardware cache flushing, try to find the
758 * largest contiguous physical memory chunk to unmap.
760 for (len = PAGE_SIZE;
761 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
762 next = iommu_iova_to_phys(domain->domain, iova + len);
763 if (next != phys + len)
768 * First, try to use fast unmap/unpin. In case of failure,
769 * switch to slow unmap/unpin path.
771 unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys,
772 &unlocked, &unmapped_region_list,
773 &unmapped_region_cnt,
776 unmapped = unmap_unpin_slow(domain, dma, &iova, len,
778 if (WARN_ON(!unmapped))
783 dma->iommu_mapped = false;
785 if (unmapped_region_cnt) {
786 unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list,
791 vfio_lock_acct(dma, -unlocked, true);
797 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
799 vfio_unmap_unpin(iommu, dma, true);
800 vfio_unlink_dma(iommu, dma);
801 put_task_struct(dma->task);
806 static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu)
808 struct vfio_domain *domain;
809 unsigned long bitmap = ULONG_MAX;
811 mutex_lock(&iommu->lock);
812 list_for_each_entry(domain, &iommu->domain_list, next)
813 bitmap &= domain->domain->pgsize_bitmap;
814 mutex_unlock(&iommu->lock);
817 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
818 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
819 * That way the user will be able to map/unmap buffers whose size/
820 * start address is aligned with PAGE_SIZE. Pinning code uses that
821 * granularity while iommu driver can use the sub-PAGE_SIZE size
824 if (bitmap & ~PAGE_MASK) {
832 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
833 struct vfio_iommu_type1_dma_unmap *unmap)
836 struct vfio_dma *dma, *dma_last = NULL;
838 int ret = 0, retries = 0;
840 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
842 if (unmap->iova & mask)
844 if (!unmap->size || unmap->size & mask)
846 if (unmap->iova + unmap->size - 1 < unmap->iova ||
847 unmap->size > SIZE_MAX)
850 WARN_ON(mask & PAGE_MASK);
852 mutex_lock(&iommu->lock);
855 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
856 * avoid tracking individual mappings. This means that the granularity
857 * of the original mapping was lost and the user was allowed to attempt
858 * to unmap any range. Depending on the contiguousness of physical
859 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
860 * or may not have worked. We only guaranteed unmap granularity
861 * matching the original mapping; even though it was untracked here,
862 * the original mappings are reflected in IOMMU mappings. This
863 * resulted in a couple unusual behaviors. First, if a range is not
864 * able to be unmapped, ex. a set of 4k pages that was mapped as a
865 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
866 * a zero sized unmap. Also, if an unmap request overlaps the first
867 * address of a hugepage, the IOMMU will unmap the entire hugepage.
868 * This also returns success and the returned unmap size reflects the
869 * actual size unmapped.
871 * We attempt to maintain compatibility with this "v1" interface, but
872 * we take control out of the hands of the IOMMU. Therefore, an unmap
873 * request offset from the beginning of the original mapping will
874 * return success with zero sized unmap. And an unmap request covering
875 * the first iova of mapping will unmap the entire range.
877 * The v2 version of this interface intends to be more deterministic.
878 * Unmap requests must fully cover previous mappings. Multiple
879 * mappings may still be unmaped by specifying large ranges, but there
880 * must not be any previous mappings bisected by the range. An error
881 * will be returned if these conditions are not met. The v2 interface
882 * will only return success and a size of zero if there were no
883 * mappings within the range.
886 dma = vfio_find_dma(iommu, unmap->iova, 1);
887 if (dma && dma->iova != unmap->iova) {
891 dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0);
892 if (dma && dma->iova + dma->size != unmap->iova + unmap->size) {
898 while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
899 if (!iommu->v2 && unmap->iova > dma->iova)
902 * Task with same address space who mapped this iova range is
903 * allowed to unmap the iova range.
905 if (dma->task->mm != current->mm)
908 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
909 struct vfio_iommu_type1_dma_unmap nb_unmap;
911 if (dma_last == dma) {
912 BUG_ON(++retries > 10);
918 nb_unmap.iova = dma->iova;
919 nb_unmap.size = dma->size;
922 * Notify anyone (mdev vendor drivers) to invalidate and
923 * unmap iovas within the range we're about to unmap.
924 * Vendor drivers MUST unpin pages in response to an
927 mutex_unlock(&iommu->lock);
928 blocking_notifier_call_chain(&iommu->notifier,
929 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
933 unmapped += dma->size;
934 vfio_remove_dma(iommu, dma);
938 mutex_unlock(&iommu->lock);
940 /* Report how much was unmapped */
941 unmap->size = unmapped;
946 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
947 unsigned long pfn, long npage, int prot)
949 struct vfio_domain *d;
952 list_for_each_entry(d, &iommu->domain_list, next) {
953 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
954 npage << PAGE_SHIFT, prot | d->prot);
964 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next)
965 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
970 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
973 dma_addr_t iova = dma->iova;
974 unsigned long vaddr = dma->vaddr;
975 size_t size = map_size;
977 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
981 /* Pin a contiguous chunk of memory */
982 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
983 size >> PAGE_SHIFT, &pfn, limit);
991 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
994 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
999 size -= npage << PAGE_SHIFT;
1000 dma->size += npage << PAGE_SHIFT;
1003 dma->iommu_mapped = true;
1006 vfio_remove_dma(iommu, dma);
1012 * Check dma map request is within a valid iova range
1014 static bool vfio_iommu_iova_dma_valid(struct vfio_iommu *iommu,
1015 dma_addr_t start, dma_addr_t end)
1017 struct list_head *iova = &iommu->iova_list;
1018 struct vfio_iova *node;
1020 list_for_each_entry(node, iova, list) {
1021 if (start >= node->start && end <= node->end)
1026 * Check for list_empty() as well since a container with
1027 * a single mdev device will have an empty list.
1029 return list_empty(iova);
1032 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1033 struct vfio_iommu_type1_dma_map *map)
1035 dma_addr_t iova = map->iova;
1036 unsigned long vaddr = map->vaddr;
1037 size_t size = map->size;
1038 int ret = 0, prot = 0;
1040 struct vfio_dma *dma;
1042 /* Verify that none of our __u64 fields overflow */
1043 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1046 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
1048 WARN_ON(mask & PAGE_MASK);
1050 /* READ/WRITE from device perspective */
1051 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1052 prot |= IOMMU_WRITE;
1053 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1056 if (!prot || !size || (size | iova | vaddr) & mask)
1059 /* Don't allow IOVA or virtual address wrap */
1060 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr)
1063 mutex_lock(&iommu->lock);
1065 if (vfio_find_dma(iommu, iova, size)) {
1070 if (!iommu->dma_avail) {
1075 if (!vfio_iommu_iova_dma_valid(iommu, iova, iova + size - 1)) {
1080 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1092 * We need to be able to both add to a task's locked memory and test
1093 * against the locked memory limit and we need to be able to do both
1094 * outside of this call path as pinning can be asynchronous via the
1095 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1096 * task_struct and VM locked pages requires an mm_struct, however
1097 * holding an indefinite mm reference is not recommended, therefore we
1098 * only hold a reference to a task. We could hold a reference to
1099 * current, however QEMU uses this call path through vCPU threads,
1100 * which can be killed resulting in a NULL mm and failure in the unmap
1101 * path when called via a different thread. Avoid this problem by
1102 * using the group_leader as threads within the same group require
1103 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1106 * Previously we also used the task for testing CAP_IPC_LOCK at the
1107 * time of pinning and accounting, however has_capability() makes use
1108 * of real_cred, a copy-on-write field, so we can't guarantee that it
1109 * matches group_leader, or in fact that it might not change by the
1110 * time it's evaluated. If a process were to call MAP_DMA with
1111 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1112 * possibly see different results for an iommu_mapped vfio_dma vs
1113 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1114 * time of calling MAP_DMA.
1116 get_task_struct(current->group_leader);
1117 dma->task = current->group_leader;
1118 dma->lock_cap = capable(CAP_IPC_LOCK);
1120 dma->pfn_list = RB_ROOT;
1122 /* Insert zero-sized and grow as we map chunks of it */
1123 vfio_link_dma(iommu, dma);
1125 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1126 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1129 ret = vfio_pin_map_dma(iommu, dma, size);
1132 mutex_unlock(&iommu->lock);
1136 static int vfio_bus_type(struct device *dev, void *data)
1138 struct bus_type **bus = data;
1140 if (*bus && *bus != dev->bus)
1148 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1149 struct vfio_domain *domain)
1151 struct vfio_domain *d;
1153 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1156 /* Arbitrarily pick the first domain in the list for lookups */
1157 d = list_first_entry(&iommu->domain_list, struct vfio_domain, next);
1158 n = rb_first(&iommu->dma_list);
1160 for (; n; n = rb_next(n)) {
1161 struct vfio_dma *dma;
1164 dma = rb_entry(n, struct vfio_dma, node);
1167 while (iova < dma->iova + dma->size) {
1171 if (dma->iommu_mapped) {
1175 phys = iommu_iova_to_phys(d->domain, iova);
1177 if (WARN_ON(!phys)) {
1185 while (i < dma->iova + dma->size &&
1186 p == iommu_iova_to_phys(d->domain, i)) {
1193 unsigned long vaddr = dma->vaddr +
1195 size_t n = dma->iova + dma->size - iova;
1198 npage = vfio_pin_pages_remote(dma, vaddr,
1207 phys = pfn << PAGE_SHIFT;
1208 size = npage << PAGE_SHIFT;
1211 ret = iommu_map(domain->domain, iova, phys,
1212 size, dma->prot | domain->prot);
1218 dma->iommu_mapped = true;
1224 * We change our unmap behavior slightly depending on whether the IOMMU
1225 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1226 * for practically any contiguous power-of-two mapping we give it. This means
1227 * we don't need to look for contiguous chunks ourselves to make unmapping
1228 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1229 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1230 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1231 * hugetlbfs is in use.
1233 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1236 int ret, order = get_order(PAGE_SIZE * 2);
1238 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1242 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1243 IOMMU_READ | IOMMU_WRITE | domain->prot);
1245 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1247 if (unmapped == PAGE_SIZE)
1248 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1250 domain->fgsp = true;
1253 __free_pages(pages, order);
1256 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1257 struct iommu_group *iommu_group)
1259 struct vfio_group *g;
1261 list_for_each_entry(g, &domain->group_list, next) {
1262 if (g->iommu_group == iommu_group)
1269 static bool vfio_iommu_has_sw_msi(struct list_head *group_resv_regions,
1272 struct iommu_resv_region *region;
1275 list_for_each_entry(region, group_resv_regions, list) {
1277 * The presence of any 'real' MSI regions should take
1278 * precedence over the software-managed one if the
1279 * IOMMU driver happens to advertise both types.
1281 if (region->type == IOMMU_RESV_MSI) {
1286 if (region->type == IOMMU_RESV_SW_MSI) {
1287 *base = region->start;
1295 static struct device *vfio_mdev_get_iommu_device(struct device *dev)
1297 struct device *(*fn)(struct device *dev);
1298 struct device *iommu_device;
1300 fn = symbol_get(mdev_get_iommu_device);
1302 iommu_device = fn(dev);
1303 symbol_put(mdev_get_iommu_device);
1305 return iommu_device;
1311 static int vfio_mdev_attach_domain(struct device *dev, void *data)
1313 struct iommu_domain *domain = data;
1314 struct device *iommu_device;
1316 iommu_device = vfio_mdev_get_iommu_device(dev);
1318 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1319 return iommu_aux_attach_device(domain, iommu_device);
1321 return iommu_attach_device(domain, iommu_device);
1327 static int vfio_mdev_detach_domain(struct device *dev, void *data)
1329 struct iommu_domain *domain = data;
1330 struct device *iommu_device;
1332 iommu_device = vfio_mdev_get_iommu_device(dev);
1334 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1335 iommu_aux_detach_device(domain, iommu_device);
1337 iommu_detach_device(domain, iommu_device);
1343 static int vfio_iommu_attach_group(struct vfio_domain *domain,
1344 struct vfio_group *group)
1346 if (group->mdev_group)
1347 return iommu_group_for_each_dev(group->iommu_group,
1349 vfio_mdev_attach_domain);
1351 return iommu_attach_group(domain->domain, group->iommu_group);
1354 static void vfio_iommu_detach_group(struct vfio_domain *domain,
1355 struct vfio_group *group)
1357 if (group->mdev_group)
1358 iommu_group_for_each_dev(group->iommu_group, domain->domain,
1359 vfio_mdev_detach_domain);
1361 iommu_detach_group(domain->domain, group->iommu_group);
1364 static bool vfio_bus_is_mdev(struct bus_type *bus)
1366 struct bus_type *mdev_bus;
1369 mdev_bus = symbol_get(mdev_bus_type);
1371 ret = (bus == mdev_bus);
1372 symbol_put(mdev_bus_type);
1378 static int vfio_mdev_iommu_device(struct device *dev, void *data)
1380 struct device **old = data, *new;
1382 new = vfio_mdev_get_iommu_device(dev);
1383 if (!new || (*old && *old != new))
1392 * This is a helper function to insert an address range to iova list.
1393 * The list is initially created with a single entry corresponding to
1394 * the IOMMU domain geometry to which the device group is attached.
1395 * The list aperture gets modified when a new domain is added to the
1396 * container if the new aperture doesn't conflict with the current one
1397 * or with any existing dma mappings. The list is also modified to
1398 * exclude any reserved regions associated with the device group.
1400 static int vfio_iommu_iova_insert(struct list_head *head,
1401 dma_addr_t start, dma_addr_t end)
1403 struct vfio_iova *region;
1405 region = kmalloc(sizeof(*region), GFP_KERNEL);
1409 INIT_LIST_HEAD(®ion->list);
1410 region->start = start;
1413 list_add_tail(®ion->list, head);
1418 * Check the new iommu aperture conflicts with existing aper or with any
1419 * existing dma mappings.
1421 static bool vfio_iommu_aper_conflict(struct vfio_iommu *iommu,
1422 dma_addr_t start, dma_addr_t end)
1424 struct vfio_iova *first, *last;
1425 struct list_head *iova = &iommu->iova_list;
1427 if (list_empty(iova))
1430 /* Disjoint sets, return conflict */
1431 first = list_first_entry(iova, struct vfio_iova, list);
1432 last = list_last_entry(iova, struct vfio_iova, list);
1433 if (start > last->end || end < first->start)
1436 /* Check for any existing dma mappings below the new start */
1437 if (start > first->start) {
1438 if (vfio_find_dma(iommu, first->start, start - first->start))
1442 /* Check for any existing dma mappings beyond the new end */
1443 if (end < last->end) {
1444 if (vfio_find_dma(iommu, end + 1, last->end - end))
1452 * Resize iommu iova aperture window. This is called only if the new
1453 * aperture has no conflict with existing aperture and dma mappings.
1455 static int vfio_iommu_aper_resize(struct list_head *iova,
1456 dma_addr_t start, dma_addr_t end)
1458 struct vfio_iova *node, *next;
1460 if (list_empty(iova))
1461 return vfio_iommu_iova_insert(iova, start, end);
1463 /* Adjust iova list start */
1464 list_for_each_entry_safe(node, next, iova, list) {
1465 if (start < node->start)
1467 if (start >= node->start && start < node->end) {
1468 node->start = start;
1471 /* Delete nodes before new start */
1472 list_del(&node->list);
1476 /* Adjust iova list end */
1477 list_for_each_entry_safe(node, next, iova, list) {
1478 if (end > node->end)
1480 if (end > node->start && end <= node->end) {
1484 /* Delete nodes after new end */
1485 list_del(&node->list);
1493 * Check reserved region conflicts with existing dma mappings
1495 static bool vfio_iommu_resv_conflict(struct vfio_iommu *iommu,
1496 struct list_head *resv_regions)
1498 struct iommu_resv_region *region;
1500 /* Check for conflict with existing dma mappings */
1501 list_for_each_entry(region, resv_regions, list) {
1502 if (region->type == IOMMU_RESV_DIRECT_RELAXABLE)
1505 if (vfio_find_dma(iommu, region->start, region->length))
1513 * Check iova region overlap with reserved regions and
1514 * exclude them from the iommu iova range
1516 static int vfio_iommu_resv_exclude(struct list_head *iova,
1517 struct list_head *resv_regions)
1519 struct iommu_resv_region *resv;
1520 struct vfio_iova *n, *next;
1522 list_for_each_entry(resv, resv_regions, list) {
1523 phys_addr_t start, end;
1525 if (resv->type == IOMMU_RESV_DIRECT_RELAXABLE)
1528 start = resv->start;
1529 end = resv->start + resv->length - 1;
1531 list_for_each_entry_safe(n, next, iova, list) {
1535 if (start > n->end || end < n->start)
1538 * Insert a new node if current node overlaps with the
1539 * reserve region to exlude that from valid iova range.
1540 * Note that, new node is inserted before the current
1541 * node and finally the current node is deleted keeping
1542 * the list updated and sorted.
1544 if (start > n->start)
1545 ret = vfio_iommu_iova_insert(&n->list, n->start,
1547 if (!ret && end < n->end)
1548 ret = vfio_iommu_iova_insert(&n->list, end + 1,
1558 if (list_empty(iova))
1564 static void vfio_iommu_resv_free(struct list_head *resv_regions)
1566 struct iommu_resv_region *n, *next;
1568 list_for_each_entry_safe(n, next, resv_regions, list) {
1574 static void vfio_iommu_iova_free(struct list_head *iova)
1576 struct vfio_iova *n, *next;
1578 list_for_each_entry_safe(n, next, iova, list) {
1584 static int vfio_iommu_iova_get_copy(struct vfio_iommu *iommu,
1585 struct list_head *iova_copy)
1587 struct list_head *iova = &iommu->iova_list;
1588 struct vfio_iova *n;
1591 list_for_each_entry(n, iova, list) {
1592 ret = vfio_iommu_iova_insert(iova_copy, n->start, n->end);
1600 vfio_iommu_iova_free(iova_copy);
1604 static void vfio_iommu_iova_insert_copy(struct vfio_iommu *iommu,
1605 struct list_head *iova_copy)
1607 struct list_head *iova = &iommu->iova_list;
1609 vfio_iommu_iova_free(iova);
1611 list_splice_tail(iova_copy, iova);
1613 static int vfio_iommu_type1_attach_group(void *iommu_data,
1614 struct iommu_group *iommu_group)
1616 struct vfio_iommu *iommu = iommu_data;
1617 struct vfio_group *group;
1618 struct vfio_domain *domain, *d;
1619 struct bus_type *bus = NULL;
1621 bool resv_msi, msi_remap;
1622 phys_addr_t resv_msi_base = 0;
1623 struct iommu_domain_geometry geo;
1624 LIST_HEAD(iova_copy);
1625 LIST_HEAD(group_resv_regions);
1627 mutex_lock(&iommu->lock);
1629 list_for_each_entry(d, &iommu->domain_list, next) {
1630 if (find_iommu_group(d, iommu_group)) {
1631 mutex_unlock(&iommu->lock);
1636 if (iommu->external_domain) {
1637 if (find_iommu_group(iommu->external_domain, iommu_group)) {
1638 mutex_unlock(&iommu->lock);
1643 group = kzalloc(sizeof(*group), GFP_KERNEL);
1644 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
1645 if (!group || !domain) {
1650 group->iommu_group = iommu_group;
1652 /* Determine bus_type in order to allocate a domain */
1653 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
1657 if (vfio_bus_is_mdev(bus)) {
1658 struct device *iommu_device = NULL;
1660 group->mdev_group = true;
1662 /* Determine the isolation type */
1663 ret = iommu_group_for_each_dev(iommu_group, &iommu_device,
1664 vfio_mdev_iommu_device);
1665 if (ret || !iommu_device) {
1666 if (!iommu->external_domain) {
1667 INIT_LIST_HEAD(&domain->group_list);
1668 iommu->external_domain = domain;
1673 list_add(&group->next,
1674 &iommu->external_domain->group_list);
1675 mutex_unlock(&iommu->lock);
1680 bus = iommu_device->bus;
1683 domain->domain = iommu_domain_alloc(bus);
1684 if (!domain->domain) {
1689 if (iommu->nesting) {
1692 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
1698 ret = vfio_iommu_attach_group(domain, group);
1702 /* Get aperture info */
1703 iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY, &geo);
1705 if (vfio_iommu_aper_conflict(iommu, geo.aperture_start,
1706 geo.aperture_end)) {
1711 ret = iommu_get_group_resv_regions(iommu_group, &group_resv_regions);
1715 if (vfio_iommu_resv_conflict(iommu, &group_resv_regions)) {
1721 * We don't want to work on the original iova list as the list
1722 * gets modified and in case of failure we have to retain the
1723 * original list. Get a copy here.
1725 ret = vfio_iommu_iova_get_copy(iommu, &iova_copy);
1729 ret = vfio_iommu_aper_resize(&iova_copy, geo.aperture_start,
1734 ret = vfio_iommu_resv_exclude(&iova_copy, &group_resv_regions);
1738 resv_msi = vfio_iommu_has_sw_msi(&group_resv_regions, &resv_msi_base);
1740 INIT_LIST_HEAD(&domain->group_list);
1741 list_add(&group->next, &domain->group_list);
1743 msi_remap = irq_domain_check_msi_remap() ||
1744 iommu_capable(bus, IOMMU_CAP_INTR_REMAP);
1746 if (!allow_unsafe_interrupts && !msi_remap) {
1747 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
1753 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
1754 domain->prot |= IOMMU_CACHE;
1757 * Try to match an existing compatible domain. We don't want to
1758 * preclude an IOMMU driver supporting multiple bus_types and being
1759 * able to include different bus_types in the same IOMMU domain, so
1760 * we test whether the domains use the same iommu_ops rather than
1761 * testing if they're on the same bus_type.
1763 list_for_each_entry(d, &iommu->domain_list, next) {
1764 if (d->domain->ops == domain->domain->ops &&
1765 d->prot == domain->prot) {
1766 vfio_iommu_detach_group(domain, group);
1767 if (!vfio_iommu_attach_group(d, group)) {
1768 list_add(&group->next, &d->group_list);
1769 iommu_domain_free(domain->domain);
1774 ret = vfio_iommu_attach_group(domain, group);
1780 vfio_test_domain_fgsp(domain);
1782 /* replay mappings on new domains */
1783 ret = vfio_iommu_replay(iommu, domain);
1788 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
1793 list_add(&domain->next, &iommu->domain_list);
1795 /* Delete the old one and insert new iova list */
1796 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
1797 mutex_unlock(&iommu->lock);
1798 vfio_iommu_resv_free(&group_resv_regions);
1803 vfio_iommu_detach_group(domain, group);
1805 iommu_domain_free(domain->domain);
1806 vfio_iommu_iova_free(&iova_copy);
1807 vfio_iommu_resv_free(&group_resv_regions);
1811 mutex_unlock(&iommu->lock);
1815 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
1817 struct rb_node *node;
1819 while ((node = rb_first(&iommu->dma_list)))
1820 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
1823 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
1825 struct rb_node *n, *p;
1827 n = rb_first(&iommu->dma_list);
1828 for (; n; n = rb_next(n)) {
1829 struct vfio_dma *dma;
1830 long locked = 0, unlocked = 0;
1832 dma = rb_entry(n, struct vfio_dma, node);
1833 unlocked += vfio_unmap_unpin(iommu, dma, false);
1834 p = rb_first(&dma->pfn_list);
1835 for (; p; p = rb_next(p)) {
1836 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
1839 if (!is_invalid_reserved_pfn(vpfn->pfn))
1842 vfio_lock_acct(dma, locked - unlocked, true);
1846 static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu)
1850 n = rb_first(&iommu->dma_list);
1851 for (; n; n = rb_next(n)) {
1852 struct vfio_dma *dma;
1854 dma = rb_entry(n, struct vfio_dma, node);
1856 if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list)))
1859 /* mdev vendor driver must unregister notifier */
1860 WARN_ON(iommu->notifier.head);
1864 * Called when a domain is removed in detach. It is possible that
1865 * the removed domain decided the iova aperture window. Modify the
1866 * iova aperture with the smallest window among existing domains.
1868 static void vfio_iommu_aper_expand(struct vfio_iommu *iommu,
1869 struct list_head *iova_copy)
1871 struct vfio_domain *domain;
1872 struct iommu_domain_geometry geo;
1873 struct vfio_iova *node;
1874 dma_addr_t start = 0;
1875 dma_addr_t end = (dma_addr_t)~0;
1877 if (list_empty(iova_copy))
1880 list_for_each_entry(domain, &iommu->domain_list, next) {
1881 iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY,
1883 if (geo.aperture_start > start)
1884 start = geo.aperture_start;
1885 if (geo.aperture_end < end)
1886 end = geo.aperture_end;
1889 /* Modify aperture limits. The new aper is either same or bigger */
1890 node = list_first_entry(iova_copy, struct vfio_iova, list);
1891 node->start = start;
1892 node = list_last_entry(iova_copy, struct vfio_iova, list);
1897 * Called when a group is detached. The reserved regions for that
1898 * group can be part of valid iova now. But since reserved regions
1899 * may be duplicated among groups, populate the iova valid regions
1902 static int vfio_iommu_resv_refresh(struct vfio_iommu *iommu,
1903 struct list_head *iova_copy)
1905 struct vfio_domain *d;
1906 struct vfio_group *g;
1907 struct vfio_iova *node;
1908 dma_addr_t start, end;
1909 LIST_HEAD(resv_regions);
1912 if (list_empty(iova_copy))
1915 list_for_each_entry(d, &iommu->domain_list, next) {
1916 list_for_each_entry(g, &d->group_list, next) {
1917 ret = iommu_get_group_resv_regions(g->iommu_group,
1924 node = list_first_entry(iova_copy, struct vfio_iova, list);
1925 start = node->start;
1926 node = list_last_entry(iova_copy, struct vfio_iova, list);
1929 /* purge the iova list and create new one */
1930 vfio_iommu_iova_free(iova_copy);
1932 ret = vfio_iommu_aper_resize(iova_copy, start, end);
1936 /* Exclude current reserved regions from iova ranges */
1937 ret = vfio_iommu_resv_exclude(iova_copy, &resv_regions);
1939 vfio_iommu_resv_free(&resv_regions);
1943 static void vfio_iommu_type1_detach_group(void *iommu_data,
1944 struct iommu_group *iommu_group)
1946 struct vfio_iommu *iommu = iommu_data;
1947 struct vfio_domain *domain;
1948 struct vfio_group *group;
1949 LIST_HEAD(iova_copy);
1951 mutex_lock(&iommu->lock);
1953 if (iommu->external_domain) {
1954 group = find_iommu_group(iommu->external_domain, iommu_group);
1956 list_del(&group->next);
1959 if (list_empty(&iommu->external_domain->group_list)) {
1960 vfio_sanity_check_pfn_list(iommu);
1962 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1963 vfio_iommu_unmap_unpin_all(iommu);
1965 kfree(iommu->external_domain);
1966 iommu->external_domain = NULL;
1968 goto detach_group_done;
1973 * Get a copy of iova list. This will be used to update
1974 * and to replace the current one later. Please note that
1975 * we will leave the original list as it is if update fails.
1977 vfio_iommu_iova_get_copy(iommu, &iova_copy);
1979 list_for_each_entry(domain, &iommu->domain_list, next) {
1980 group = find_iommu_group(domain, iommu_group);
1984 vfio_iommu_detach_group(domain, group);
1985 list_del(&group->next);
1988 * Group ownership provides privilege, if the group list is
1989 * empty, the domain goes away. If it's the last domain with
1990 * iommu and external domain doesn't exist, then all the
1991 * mappings go away too. If it's the last domain with iommu and
1992 * external domain exist, update accounting
1994 if (list_empty(&domain->group_list)) {
1995 if (list_is_singular(&iommu->domain_list)) {
1996 if (!iommu->external_domain)
1997 vfio_iommu_unmap_unpin_all(iommu);
1999 vfio_iommu_unmap_unpin_reaccount(iommu);
2001 iommu_domain_free(domain->domain);
2002 list_del(&domain->next);
2004 vfio_iommu_aper_expand(iommu, &iova_copy);
2009 if (!vfio_iommu_resv_refresh(iommu, &iova_copy))
2010 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2012 vfio_iommu_iova_free(&iova_copy);
2015 mutex_unlock(&iommu->lock);
2018 static void *vfio_iommu_type1_open(unsigned long arg)
2020 struct vfio_iommu *iommu;
2022 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
2024 return ERR_PTR(-ENOMEM);
2027 case VFIO_TYPE1_IOMMU:
2029 case VFIO_TYPE1_NESTING_IOMMU:
2030 iommu->nesting = true;
2032 case VFIO_TYPE1v2_IOMMU:
2037 return ERR_PTR(-EINVAL);
2040 INIT_LIST_HEAD(&iommu->domain_list);
2041 INIT_LIST_HEAD(&iommu->iova_list);
2042 iommu->dma_list = RB_ROOT;
2043 iommu->dma_avail = dma_entry_limit;
2044 mutex_init(&iommu->lock);
2045 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
2050 static void vfio_release_domain(struct vfio_domain *domain, bool external)
2052 struct vfio_group *group, *group_tmp;
2054 list_for_each_entry_safe(group, group_tmp,
2055 &domain->group_list, next) {
2057 vfio_iommu_detach_group(domain, group);
2058 list_del(&group->next);
2063 iommu_domain_free(domain->domain);
2066 static void vfio_iommu_type1_release(void *iommu_data)
2068 struct vfio_iommu *iommu = iommu_data;
2069 struct vfio_domain *domain, *domain_tmp;
2071 if (iommu->external_domain) {
2072 vfio_release_domain(iommu->external_domain, true);
2073 vfio_sanity_check_pfn_list(iommu);
2074 kfree(iommu->external_domain);
2077 vfio_iommu_unmap_unpin_all(iommu);
2079 list_for_each_entry_safe(domain, domain_tmp,
2080 &iommu->domain_list, next) {
2081 vfio_release_domain(domain, false);
2082 list_del(&domain->next);
2086 vfio_iommu_iova_free(&iommu->iova_list);
2091 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
2093 struct vfio_domain *domain;
2096 mutex_lock(&iommu->lock);
2097 list_for_each_entry(domain, &iommu->domain_list, next) {
2098 if (!(domain->prot & IOMMU_CACHE)) {
2103 mutex_unlock(&iommu->lock);
2108 static int vfio_iommu_iova_add_cap(struct vfio_info_cap *caps,
2109 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas,
2112 struct vfio_info_cap_header *header;
2113 struct vfio_iommu_type1_info_cap_iova_range *iova_cap;
2115 header = vfio_info_cap_add(caps, size,
2116 VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE, 1);
2118 return PTR_ERR(header);
2120 iova_cap = container_of(header,
2121 struct vfio_iommu_type1_info_cap_iova_range,
2123 iova_cap->nr_iovas = cap_iovas->nr_iovas;
2124 memcpy(iova_cap->iova_ranges, cap_iovas->iova_ranges,
2125 cap_iovas->nr_iovas * sizeof(*cap_iovas->iova_ranges));
2129 static int vfio_iommu_iova_build_caps(struct vfio_iommu *iommu,
2130 struct vfio_info_cap *caps)
2132 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas;
2133 struct vfio_iova *iova;
2135 int iovas = 0, i = 0, ret;
2137 mutex_lock(&iommu->lock);
2139 list_for_each_entry(iova, &iommu->iova_list, list)
2144 * Return 0 as a container with a single mdev device
2145 * will have an empty list
2151 size = sizeof(*cap_iovas) + (iovas * sizeof(*cap_iovas->iova_ranges));
2153 cap_iovas = kzalloc(size, GFP_KERNEL);
2159 cap_iovas->nr_iovas = iovas;
2161 list_for_each_entry(iova, &iommu->iova_list, list) {
2162 cap_iovas->iova_ranges[i].start = iova->start;
2163 cap_iovas->iova_ranges[i].end = iova->end;
2167 ret = vfio_iommu_iova_add_cap(caps, cap_iovas, size);
2171 mutex_unlock(&iommu->lock);
2175 static long vfio_iommu_type1_ioctl(void *iommu_data,
2176 unsigned int cmd, unsigned long arg)
2178 struct vfio_iommu *iommu = iommu_data;
2179 unsigned long minsz;
2181 if (cmd == VFIO_CHECK_EXTENSION) {
2183 case VFIO_TYPE1_IOMMU:
2184 case VFIO_TYPE1v2_IOMMU:
2185 case VFIO_TYPE1_NESTING_IOMMU:
2187 case VFIO_DMA_CC_IOMMU:
2190 return vfio_domains_have_iommu_cache(iommu);
2194 } else if (cmd == VFIO_IOMMU_GET_INFO) {
2195 struct vfio_iommu_type1_info info;
2196 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
2197 unsigned long capsz;
2200 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
2202 /* For backward compatibility, cannot require this */
2203 capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
2205 if (copy_from_user(&info, (void __user *)arg, minsz))
2208 if (info.argsz < minsz)
2211 if (info.argsz >= capsz) {
2213 info.cap_offset = 0; /* output, no-recopy necessary */
2216 info.flags = VFIO_IOMMU_INFO_PGSIZES;
2218 info.iova_pgsizes = vfio_pgsize_bitmap(iommu);
2220 ret = vfio_iommu_iova_build_caps(iommu, &caps);
2225 info.flags |= VFIO_IOMMU_INFO_CAPS;
2227 if (info.argsz < sizeof(info) + caps.size) {
2228 info.argsz = sizeof(info) + caps.size;
2230 vfio_info_cap_shift(&caps, sizeof(info));
2231 if (copy_to_user((void __user *)arg +
2232 sizeof(info), caps.buf,
2237 info.cap_offset = sizeof(info);
2243 return copy_to_user((void __user *)arg, &info, minsz) ?
2246 } else if (cmd == VFIO_IOMMU_MAP_DMA) {
2247 struct vfio_iommu_type1_dma_map map;
2248 uint32_t mask = VFIO_DMA_MAP_FLAG_READ |
2249 VFIO_DMA_MAP_FLAG_WRITE;
2251 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
2253 if (copy_from_user(&map, (void __user *)arg, minsz))
2256 if (map.argsz < minsz || map.flags & ~mask)
2259 return vfio_dma_do_map(iommu, &map);
2261 } else if (cmd == VFIO_IOMMU_UNMAP_DMA) {
2262 struct vfio_iommu_type1_dma_unmap unmap;
2265 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
2267 if (copy_from_user(&unmap, (void __user *)arg, minsz))
2270 if (unmap.argsz < minsz || unmap.flags)
2273 ret = vfio_dma_do_unmap(iommu, &unmap);
2277 return copy_to_user((void __user *)arg, &unmap, minsz) ?
2284 static int vfio_iommu_type1_register_notifier(void *iommu_data,
2285 unsigned long *events,
2286 struct notifier_block *nb)
2288 struct vfio_iommu *iommu = iommu_data;
2290 /* clear known events */
2291 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
2293 /* refuse to register if still events remaining */
2297 return blocking_notifier_chain_register(&iommu->notifier, nb);
2300 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
2301 struct notifier_block *nb)
2303 struct vfio_iommu *iommu = iommu_data;
2305 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
2308 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
2309 .name = "vfio-iommu-type1",
2310 .owner = THIS_MODULE,
2311 .open = vfio_iommu_type1_open,
2312 .release = vfio_iommu_type1_release,
2313 .ioctl = vfio_iommu_type1_ioctl,
2314 .attach_group = vfio_iommu_type1_attach_group,
2315 .detach_group = vfio_iommu_type1_detach_group,
2316 .pin_pages = vfio_iommu_type1_pin_pages,
2317 .unpin_pages = vfio_iommu_type1_unpin_pages,
2318 .register_notifier = vfio_iommu_type1_register_notifier,
2319 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
2322 static int __init vfio_iommu_type1_init(void)
2324 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
2327 static void __exit vfio_iommu_type1_cleanup(void)
2329 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
2332 module_init(vfio_iommu_type1_init);
2333 module_exit(vfio_iommu_type1_cleanup);
2335 MODULE_VERSION(DRIVER_VERSION);
2336 MODULE_LICENSE("GPL v2");
2337 MODULE_AUTHOR(DRIVER_AUTHOR);
2338 MODULE_DESCRIPTION(DRIVER_DESC);