2 * VFIO: IOMMU DMA mapping support for Type1 IOMMU
4 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
5 * Author: Alex Williamson <alex.williamson@redhat.com>
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
11 * Derived from original vfio:
12 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
13 * Author: Tom Lyon, pugs@cisco.com
15 * We arbitrarily define a Type1 IOMMU as one matching the below code.
16 * It could be called the x86 IOMMU as it's designed for AMD-Vi & Intel
17 * VT-d, but that makes it harder to re-use as theoretically anyone
18 * implementing a similar IOMMU could make use of this. We expect the
19 * IOMMU to support the IOMMU API and have few to no restrictions around
20 * the IOVA range that can be mapped. The Type1 IOMMU is currently
21 * optimized for relatively static mappings of a userspace process with
22 * userpsace pages pinned into memory. We also assume devices and IOMMU
23 * domains are PCI based as the IOMMU API is still centered around a
24 * device/bus interface rather than a group interface.
27 #include <linux/compat.h>
28 #include <linux/device.h>
30 #include <linux/iommu.h>
31 #include <linux/module.h>
33 #include <linux/rbtree.h>
34 #include <linux/sched/signal.h>
35 #include <linux/sched/mm.h>
36 #include <linux/slab.h>
37 #include <linux/uaccess.h>
38 #include <linux/vfio.h>
39 #include <linux/workqueue.h>
40 #include <linux/mdev.h>
41 #include <linux/notifier.h>
42 #include <linux/dma-iommu.h>
43 #include <linux/irqdomain.h>
45 #define DRIVER_VERSION "0.2"
46 #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
47 #define DRIVER_DESC "Type1 IOMMU driver for VFIO"
49 static bool allow_unsafe_interrupts;
50 module_param_named(allow_unsafe_interrupts,
51 allow_unsafe_interrupts, bool, S_IRUGO | S_IWUSR);
52 MODULE_PARM_DESC(allow_unsafe_interrupts,
53 "Enable VFIO IOMMU support for on platforms without interrupt remapping support.");
55 static bool disable_hugepages;
56 module_param_named(disable_hugepages,
57 disable_hugepages, bool, S_IRUGO | S_IWUSR);
58 MODULE_PARM_DESC(disable_hugepages,
59 "Disable VFIO IOMMU support for IOMMU hugepages.");
61 static unsigned int dma_entry_limit __read_mostly = U16_MAX;
62 module_param_named(dma_entry_limit, dma_entry_limit, uint, 0644);
63 MODULE_PARM_DESC(dma_entry_limit,
64 "Maximum number of user DMA mappings per container (65535).");
67 struct list_head domain_list;
68 struct vfio_domain *external_domain; /* domain for external user */
70 struct rb_root dma_list;
71 struct blocking_notifier_head notifier;
72 unsigned int dma_avail;
78 struct iommu_domain *domain;
79 struct list_head next;
80 struct list_head group_list;
81 int prot; /* IOMMU_CACHE */
82 bool fgsp; /* Fine-grained super pages */
87 dma_addr_t iova; /* Device address */
88 unsigned long vaddr; /* Process virtual addr */
89 size_t size; /* Map size (bytes) */
90 int prot; /* IOMMU_READ/WRITE */
92 bool lock_cap; /* capable(CAP_IPC_LOCK) */
93 struct task_struct *task;
94 struct rb_root pfn_list; /* Ex-user pinned pfn list */
98 struct iommu_group *iommu_group;
99 struct list_head next;
103 * Guest RAM pinning working set or DMA target
107 dma_addr_t iova; /* Device address */
108 unsigned long pfn; /* Host pfn */
112 struct vfio_regions {
113 struct list_head list;
119 #define IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu) \
120 (!list_empty(&iommu->domain_list))
122 static int put_pfn(unsigned long pfn, int prot);
125 * This code handles mapping and unmapping of user data buffers
126 * into DMA'ble space using the IOMMU
129 static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
130 dma_addr_t start, size_t size)
132 struct rb_node *node = iommu->dma_list.rb_node;
135 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
137 if (start + size <= dma->iova)
138 node = node->rb_left;
139 else if (start >= dma->iova + dma->size)
140 node = node->rb_right;
148 static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new)
150 struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL;
151 struct vfio_dma *dma;
155 dma = rb_entry(parent, struct vfio_dma, node);
157 if (new->iova + new->size <= dma->iova)
158 link = &(*link)->rb_left;
160 link = &(*link)->rb_right;
163 rb_link_node(&new->node, parent, link);
164 rb_insert_color(&new->node, &iommu->dma_list);
167 static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
169 rb_erase(&old->node, &iommu->dma_list);
173 * Helper Functions for host iova-pfn list
175 static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
177 struct vfio_pfn *vpfn;
178 struct rb_node *node = dma->pfn_list.rb_node;
181 vpfn = rb_entry(node, struct vfio_pfn, node);
183 if (iova < vpfn->iova)
184 node = node->rb_left;
185 else if (iova > vpfn->iova)
186 node = node->rb_right;
193 static void vfio_link_pfn(struct vfio_dma *dma,
194 struct vfio_pfn *new)
196 struct rb_node **link, *parent = NULL;
197 struct vfio_pfn *vpfn;
199 link = &dma->pfn_list.rb_node;
202 vpfn = rb_entry(parent, struct vfio_pfn, node);
204 if (new->iova < vpfn->iova)
205 link = &(*link)->rb_left;
207 link = &(*link)->rb_right;
210 rb_link_node(&new->node, parent, link);
211 rb_insert_color(&new->node, &dma->pfn_list);
214 static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old)
216 rb_erase(&old->node, &dma->pfn_list);
219 static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova,
222 struct vfio_pfn *vpfn;
224 vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL);
230 atomic_set(&vpfn->ref_count, 1);
231 vfio_link_pfn(dma, vpfn);
235 static void vfio_remove_from_pfn_list(struct vfio_dma *dma,
236 struct vfio_pfn *vpfn)
238 vfio_unlink_pfn(dma, vpfn);
242 static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma,
245 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
248 atomic_inc(&vpfn->ref_count);
252 static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn)
256 if (atomic_dec_and_test(&vpfn->ref_count)) {
257 ret = put_pfn(vpfn->pfn, dma->prot);
258 vfio_remove_from_pfn_list(dma, vpfn);
263 static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
265 struct mm_struct *mm;
271 mm = async ? get_task_mm(dma->task) : dma->task->mm;
273 return -ESRCH; /* process exited */
275 ret = down_write_killable(&mm->mmap_sem);
278 if (!dma->lock_cap) {
281 limit = task_rlimit(dma->task,
282 RLIMIT_MEMLOCK) >> PAGE_SHIFT;
284 if (mm->locked_vm + npage > limit)
290 mm->locked_vm += npage;
292 up_write(&mm->mmap_sem);
302 * Some mappings aren't backed by a struct page, for example an mmap'd
303 * MMIO range for our own or another device. These use a different
304 * pfn conversion and shouldn't be tracked as locked pages.
306 static bool is_invalid_reserved_pfn(unsigned long pfn)
308 if (pfn_valid(pfn)) {
310 struct page *tail = pfn_to_page(pfn);
311 struct page *head = compound_head(tail);
312 reserved = !!(PageReserved(head));
315 * "head" is not a dangling pointer
316 * (compound_head takes care of that)
317 * but the hugepage may have been split
318 * from under us (and we may not hold a
319 * reference count on the head page so it can
320 * be reused before we run PageReferenced), so
321 * we've to check PageTail before returning
328 return PageReserved(tail);
334 static int put_pfn(unsigned long pfn, int prot)
336 if (!is_invalid_reserved_pfn(pfn)) {
337 struct page *page = pfn_to_page(pfn);
338 if (prot & IOMMU_WRITE)
346 static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,
347 int prot, unsigned long *pfn)
349 struct page *page[1];
350 struct vm_area_struct *vma;
351 struct vm_area_struct *vmas[1];
352 unsigned int flags = 0;
355 if (prot & IOMMU_WRITE)
358 down_read(&mm->mmap_sem);
359 if (mm == current->mm) {
360 ret = get_user_pages_longterm(vaddr, 1, flags, page, vmas);
362 ret = get_user_pages_remote(NULL, mm, vaddr, 1, flags, page,
365 * The lifetime of a vaddr_get_pfn() page pin is
366 * userspace-controlled. In the fs-dax case this could
367 * lead to indefinite stalls in filesystem operations.
368 * Disallow attempts to pin fs-dax pages via this
371 if (ret > 0 && vma_is_fsdax(vmas[0])) {
376 up_read(&mm->mmap_sem);
379 *pfn = page_to_pfn(page[0]);
383 down_read(&mm->mmap_sem);
385 vma = find_vma_intersection(mm, vaddr, vaddr + 1);
387 if (vma && vma->vm_flags & VM_PFNMAP) {
388 *pfn = ((vaddr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
389 if (is_invalid_reserved_pfn(*pfn))
393 up_read(&mm->mmap_sem);
398 * Attempt to pin pages. We really don't want to track all the pfns and
399 * the iommu can only map chunks of consecutive pfns anyway, so get the
400 * first page and all consecutive pages with the same locking.
402 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
403 long npage, unsigned long *pfn_base,
406 unsigned long pfn = 0;
407 long ret, pinned = 0, lock_acct = 0;
409 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
411 /* This code path is only user initiated */
415 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base);
420 rsvd = is_invalid_reserved_pfn(*pfn_base);
423 * Reserved pages aren't counted against the user, externally pinned
424 * pages are already counted against the user.
426 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
427 if (!dma->lock_cap && current->mm->locked_vm + 1 > limit) {
428 put_pfn(*pfn_base, dma->prot);
429 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__,
430 limit << PAGE_SHIFT);
436 if (unlikely(disable_hugepages))
439 /* Lock all the consecutive pages from pfn_base */
440 for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage;
441 pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) {
442 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn);
446 if (pfn != *pfn_base + pinned ||
447 rsvd != is_invalid_reserved_pfn(pfn)) {
448 put_pfn(pfn, dma->prot);
452 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
453 if (!dma->lock_cap &&
454 current->mm->locked_vm + lock_acct + 1 > limit) {
455 put_pfn(pfn, dma->prot);
456 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
457 __func__, limit << PAGE_SHIFT);
466 ret = vfio_lock_acct(dma, lock_acct, false);
471 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
472 put_pfn(pfn, dma->prot);
481 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
482 unsigned long pfn, long npage,
485 long unlocked = 0, locked = 0;
488 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
489 if (put_pfn(pfn++, dma->prot)) {
491 if (vfio_find_vpfn(dma, iova))
497 vfio_lock_acct(dma, locked - unlocked, true);
502 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
503 unsigned long *pfn_base, bool do_accounting)
505 struct mm_struct *mm;
508 mm = get_task_mm(dma->task);
512 ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base);
513 if (!ret && do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
514 ret = vfio_lock_acct(dma, 1, true);
516 put_pfn(*pfn_base, dma->prot);
518 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
519 "(%ld) exceeded\n", __func__,
520 dma->task->comm, task_pid_nr(dma->task),
521 task_rlimit(dma->task, RLIMIT_MEMLOCK));
529 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
533 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
538 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
541 vfio_lock_acct(dma, -unlocked, true);
546 static int vfio_iommu_type1_pin_pages(void *iommu_data,
547 unsigned long *user_pfn,
549 unsigned long *phys_pfn)
551 struct vfio_iommu *iommu = iommu_data;
553 unsigned long remote_vaddr;
554 struct vfio_dma *dma;
557 if (!iommu || !user_pfn || !phys_pfn)
560 /* Supported for v2 version only */
564 mutex_lock(&iommu->lock);
566 /* Fail if notifier list is empty */
567 if ((!iommu->external_domain) || (!iommu->notifier.head)) {
573 * If iommu capable domain exist in the container then all pages are
574 * already pinned and accounted. Accouting should be done if there is no
575 * iommu capable domain in the container.
577 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
579 for (i = 0; i < npage; i++) {
581 struct vfio_pfn *vpfn;
583 iova = user_pfn[i] << PAGE_SHIFT;
584 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
590 if ((dma->prot & prot) != prot) {
595 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
597 phys_pfn[i] = vpfn->pfn;
601 remote_vaddr = dma->vaddr + iova - dma->iova;
602 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
607 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
609 vfio_unpin_page_external(dma, iova, do_accounting);
619 for (j = 0; j < i; j++) {
622 iova = user_pfn[j] << PAGE_SHIFT;
623 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
624 vfio_unpin_page_external(dma, iova, do_accounting);
628 mutex_unlock(&iommu->lock);
632 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
633 unsigned long *user_pfn,
636 struct vfio_iommu *iommu = iommu_data;
640 if (!iommu || !user_pfn)
643 /* Supported for v2 version only */
647 mutex_lock(&iommu->lock);
649 if (!iommu->external_domain) {
650 mutex_unlock(&iommu->lock);
654 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
655 for (i = 0; i < npage; i++) {
656 struct vfio_dma *dma;
659 iova = user_pfn[i] << PAGE_SHIFT;
660 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
663 vfio_unpin_page_external(dma, iova, do_accounting);
667 mutex_unlock(&iommu->lock);
668 return i > npage ? npage : (i > 0 ? i : -EINVAL);
671 static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
672 struct list_head *regions)
675 struct vfio_regions *entry, *next;
677 iommu_tlb_sync(domain->domain);
679 list_for_each_entry_safe(entry, next, regions, list) {
680 unlocked += vfio_unpin_pages_remote(dma,
682 entry->phys >> PAGE_SHIFT,
683 entry->len >> PAGE_SHIFT,
685 list_del(&entry->list);
695 * Generally, VFIO needs to unpin remote pages after each IOTLB flush.
696 * Therefore, when using IOTLB flush sync interface, VFIO need to keep track
697 * of these regions (currently using a list).
699 * This value specifies maximum number of regions for each IOTLB flush sync.
701 #define VFIO_IOMMU_TLB_SYNC_MAX 512
703 static size_t unmap_unpin_fast(struct vfio_domain *domain,
704 struct vfio_dma *dma, dma_addr_t *iova,
705 size_t len, phys_addr_t phys, long *unlocked,
706 struct list_head *unmapped_list,
710 struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
713 unmapped = iommu_unmap_fast(domain->domain, *iova, len);
718 iommu_tlb_range_add(domain->domain, *iova, unmapped);
721 entry->len = unmapped;
722 list_add_tail(&entry->list, unmapped_list);
730 * Sync if the number of fast-unmap regions hits the limit
731 * or in case of errors.
733 if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
734 *unlocked += vfio_sync_unpin(dma, domain,
742 static size_t unmap_unpin_slow(struct vfio_domain *domain,
743 struct vfio_dma *dma, dma_addr_t *iova,
744 size_t len, phys_addr_t phys,
747 size_t unmapped = iommu_unmap(domain->domain, *iova, len);
750 *unlocked += vfio_unpin_pages_remote(dma, *iova,
752 unmapped >> PAGE_SHIFT,
760 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
763 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
764 struct vfio_domain *domain, *d;
765 LIST_HEAD(unmapped_region_list);
766 int unmapped_region_cnt = 0;
772 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
776 * We use the IOMMU to track the physical addresses, otherwise we'd
777 * need a much more complicated tracking system. Unfortunately that
778 * means we need to use one of the iommu domains to figure out the
779 * pfns to unpin. The rest need to be unmapped in advance so we have
780 * no iommu translations remaining when the pages are unpinned.
782 domain = d = list_first_entry(&iommu->domain_list,
783 struct vfio_domain, next);
785 list_for_each_entry_continue(d, &iommu->domain_list, next) {
786 iommu_unmap(d->domain, dma->iova, dma->size);
791 size_t unmapped, len;
792 phys_addr_t phys, next;
794 phys = iommu_iova_to_phys(domain->domain, iova);
795 if (WARN_ON(!phys)) {
801 * To optimize for fewer iommu_unmap() calls, each of which
802 * may require hardware cache flushing, try to find the
803 * largest contiguous physical memory chunk to unmap.
805 for (len = PAGE_SIZE;
806 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
807 next = iommu_iova_to_phys(domain->domain, iova + len);
808 if (next != phys + len)
813 * First, try to use fast unmap/unpin. In case of failure,
814 * switch to slow unmap/unpin path.
816 unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys,
817 &unlocked, &unmapped_region_list,
818 &unmapped_region_cnt);
820 unmapped = unmap_unpin_slow(domain, dma, &iova, len,
822 if (WARN_ON(!unmapped))
827 dma->iommu_mapped = false;
829 if (unmapped_region_cnt)
830 unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list);
833 vfio_lock_acct(dma, -unlocked, true);
839 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
841 vfio_unmap_unpin(iommu, dma, true);
842 vfio_unlink_dma(iommu, dma);
843 put_task_struct(dma->task);
848 static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu)
850 struct vfio_domain *domain;
851 unsigned long bitmap = ULONG_MAX;
853 mutex_lock(&iommu->lock);
854 list_for_each_entry(domain, &iommu->domain_list, next)
855 bitmap &= domain->domain->pgsize_bitmap;
856 mutex_unlock(&iommu->lock);
859 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
860 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
861 * That way the user will be able to map/unmap buffers whose size/
862 * start address is aligned with PAGE_SIZE. Pinning code uses that
863 * granularity while iommu driver can use the sub-PAGE_SIZE size
866 if (bitmap & ~PAGE_MASK) {
874 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
875 struct vfio_iommu_type1_dma_unmap *unmap)
878 struct vfio_dma *dma, *dma_last = NULL;
880 int ret = 0, retries = 0;
882 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
884 if (unmap->iova & mask)
886 if (!unmap->size || unmap->size & mask)
888 if (unmap->iova + unmap->size - 1 < unmap->iova ||
889 unmap->size > SIZE_MAX)
892 WARN_ON(mask & PAGE_MASK);
894 mutex_lock(&iommu->lock);
897 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
898 * avoid tracking individual mappings. This means that the granularity
899 * of the original mapping was lost and the user was allowed to attempt
900 * to unmap any range. Depending on the contiguousness of physical
901 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
902 * or may not have worked. We only guaranteed unmap granularity
903 * matching the original mapping; even though it was untracked here,
904 * the original mappings are reflected in IOMMU mappings. This
905 * resulted in a couple unusual behaviors. First, if a range is not
906 * able to be unmapped, ex. a set of 4k pages that was mapped as a
907 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
908 * a zero sized unmap. Also, if an unmap request overlaps the first
909 * address of a hugepage, the IOMMU will unmap the entire hugepage.
910 * This also returns success and the returned unmap size reflects the
911 * actual size unmapped.
913 * We attempt to maintain compatibility with this "v1" interface, but
914 * we take control out of the hands of the IOMMU. Therefore, an unmap
915 * request offset from the beginning of the original mapping will
916 * return success with zero sized unmap. And an unmap request covering
917 * the first iova of mapping will unmap the entire range.
919 * The v2 version of this interface intends to be more deterministic.
920 * Unmap requests must fully cover previous mappings. Multiple
921 * mappings may still be unmaped by specifying large ranges, but there
922 * must not be any previous mappings bisected by the range. An error
923 * will be returned if these conditions are not met. The v2 interface
924 * will only return success and a size of zero if there were no
925 * mappings within the range.
928 dma = vfio_find_dma(iommu, unmap->iova, 1);
929 if (dma && dma->iova != unmap->iova) {
933 dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0);
934 if (dma && dma->iova + dma->size != unmap->iova + unmap->size) {
940 while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
941 if (!iommu->v2 && unmap->iova > dma->iova)
944 * Task with same address space who mapped this iova range is
945 * allowed to unmap the iova range.
947 if (dma->task->mm != current->mm)
950 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
951 struct vfio_iommu_type1_dma_unmap nb_unmap;
953 if (dma_last == dma) {
954 BUG_ON(++retries > 10);
960 nb_unmap.iova = dma->iova;
961 nb_unmap.size = dma->size;
964 * Notify anyone (mdev vendor drivers) to invalidate and
965 * unmap iovas within the range we're about to unmap.
966 * Vendor drivers MUST unpin pages in response to an
969 mutex_unlock(&iommu->lock);
970 blocking_notifier_call_chain(&iommu->notifier,
971 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
975 unmapped += dma->size;
976 vfio_remove_dma(iommu, dma);
980 mutex_unlock(&iommu->lock);
982 /* Report how much was unmapped */
983 unmap->size = unmapped;
988 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
989 unsigned long pfn, long npage, int prot)
991 struct vfio_domain *d;
994 list_for_each_entry(d, &iommu->domain_list, next) {
995 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
996 npage << PAGE_SHIFT, prot | d->prot);
1006 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next)
1007 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
1012 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
1015 dma_addr_t iova = dma->iova;
1016 unsigned long vaddr = dma->vaddr;
1017 size_t size = map_size;
1019 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1023 /* Pin a contiguous chunk of memory */
1024 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
1025 size >> PAGE_SHIFT, &pfn, limit);
1033 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
1036 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
1041 size -= npage << PAGE_SHIFT;
1042 dma->size += npage << PAGE_SHIFT;
1045 dma->iommu_mapped = true;
1048 vfio_remove_dma(iommu, dma);
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 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1108 * We need to be able to both add to a task's locked memory and test
1109 * against the locked memory limit and we need to be able to do both
1110 * outside of this call path as pinning can be asynchronous via the
1111 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1112 * task_struct and VM locked pages requires an mm_struct, however
1113 * holding an indefinite mm reference is not recommended, therefore we
1114 * only hold a reference to a task. We could hold a reference to
1115 * current, however QEMU uses this call path through vCPU threads,
1116 * which can be killed resulting in a NULL mm and failure in the unmap
1117 * path when called via a different thread. Avoid this problem by
1118 * using the group_leader as threads within the same group require
1119 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1122 * Previously we also used the task for testing CAP_IPC_LOCK at the
1123 * time of pinning and accounting, however has_capability() makes use
1124 * of real_cred, a copy-on-write field, so we can't guarantee that it
1125 * matches group_leader, or in fact that it might not change by the
1126 * time it's evaluated. If a process were to call MAP_DMA with
1127 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1128 * possibly see different results for an iommu_mapped vfio_dma vs
1129 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1130 * time of calling MAP_DMA.
1132 get_task_struct(current->group_leader);
1133 dma->task = current->group_leader;
1134 dma->lock_cap = capable(CAP_IPC_LOCK);
1136 dma->pfn_list = RB_ROOT;
1138 /* Insert zero-sized and grow as we map chunks of it */
1139 vfio_link_dma(iommu, dma);
1141 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1142 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1145 ret = vfio_pin_map_dma(iommu, dma, size);
1148 mutex_unlock(&iommu->lock);
1152 static int vfio_bus_type(struct device *dev, void *data)
1154 struct bus_type **bus = data;
1156 if (*bus && *bus != dev->bus)
1164 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1165 struct vfio_domain *domain)
1167 struct vfio_domain *d;
1169 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1172 /* Arbitrarily pick the first domain in the list for lookups */
1173 d = list_first_entry(&iommu->domain_list, struct vfio_domain, next);
1174 n = rb_first(&iommu->dma_list);
1176 for (; n; n = rb_next(n)) {
1177 struct vfio_dma *dma;
1180 dma = rb_entry(n, struct vfio_dma, node);
1183 while (iova < dma->iova + dma->size) {
1187 if (dma->iommu_mapped) {
1191 phys = iommu_iova_to_phys(d->domain, iova);
1193 if (WARN_ON(!phys)) {
1201 while (i < dma->iova + dma->size &&
1202 p == iommu_iova_to_phys(d->domain, i)) {
1209 unsigned long vaddr = dma->vaddr +
1211 size_t n = dma->iova + dma->size - iova;
1214 npage = vfio_pin_pages_remote(dma, vaddr,
1223 phys = pfn << PAGE_SHIFT;
1224 size = npage << PAGE_SHIFT;
1227 ret = iommu_map(domain->domain, iova, phys,
1228 size, dma->prot | domain->prot);
1234 dma->iommu_mapped = true;
1240 * We change our unmap behavior slightly depending on whether the IOMMU
1241 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1242 * for practically any contiguous power-of-two mapping we give it. This means
1243 * we don't need to look for contiguous chunks ourselves to make unmapping
1244 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1245 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1246 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1247 * hugetlbfs is in use.
1249 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1252 int ret, order = get_order(PAGE_SIZE * 2);
1254 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1258 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1259 IOMMU_READ | IOMMU_WRITE | domain->prot);
1261 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1263 if (unmapped == PAGE_SIZE)
1264 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1266 domain->fgsp = true;
1269 __free_pages(pages, order);
1272 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1273 struct iommu_group *iommu_group)
1275 struct vfio_group *g;
1277 list_for_each_entry(g, &domain->group_list, next) {
1278 if (g->iommu_group == iommu_group)
1285 static bool vfio_iommu_has_sw_msi(struct iommu_group *group, phys_addr_t *base)
1287 struct list_head group_resv_regions;
1288 struct iommu_resv_region *region, *next;
1291 INIT_LIST_HEAD(&group_resv_regions);
1292 iommu_get_group_resv_regions(group, &group_resv_regions);
1293 list_for_each_entry(region, &group_resv_regions, list) {
1295 * The presence of any 'real' MSI regions should take
1296 * precedence over the software-managed one if the
1297 * IOMMU driver happens to advertise both types.
1299 if (region->type == IOMMU_RESV_MSI) {
1304 if (region->type == IOMMU_RESV_SW_MSI) {
1305 *base = region->start;
1309 list_for_each_entry_safe(region, next, &group_resv_regions, list)
1314 static int vfio_iommu_type1_attach_group(void *iommu_data,
1315 struct iommu_group *iommu_group)
1317 struct vfio_iommu *iommu = iommu_data;
1318 struct vfio_group *group;
1319 struct vfio_domain *domain, *d;
1320 struct bus_type *bus = NULL, *mdev_bus;
1322 bool resv_msi, msi_remap;
1323 phys_addr_t resv_msi_base;
1325 mutex_lock(&iommu->lock);
1327 list_for_each_entry(d, &iommu->domain_list, next) {
1328 if (find_iommu_group(d, iommu_group)) {
1329 mutex_unlock(&iommu->lock);
1334 if (iommu->external_domain) {
1335 if (find_iommu_group(iommu->external_domain, iommu_group)) {
1336 mutex_unlock(&iommu->lock);
1341 group = kzalloc(sizeof(*group), GFP_KERNEL);
1342 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
1343 if (!group || !domain) {
1348 group->iommu_group = iommu_group;
1350 /* Determine bus_type in order to allocate a domain */
1351 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
1355 mdev_bus = symbol_get(mdev_bus_type);
1358 if ((bus == mdev_bus) && !iommu_present(bus)) {
1359 symbol_put(mdev_bus_type);
1360 if (!iommu->external_domain) {
1361 INIT_LIST_HEAD(&domain->group_list);
1362 iommu->external_domain = domain;
1366 list_add(&group->next,
1367 &iommu->external_domain->group_list);
1368 mutex_unlock(&iommu->lock);
1371 symbol_put(mdev_bus_type);
1374 domain->domain = iommu_domain_alloc(bus);
1375 if (!domain->domain) {
1380 if (iommu->nesting) {
1383 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
1389 ret = iommu_attach_group(domain->domain, iommu_group);
1393 resv_msi = vfio_iommu_has_sw_msi(iommu_group, &resv_msi_base);
1395 INIT_LIST_HEAD(&domain->group_list);
1396 list_add(&group->next, &domain->group_list);
1398 msi_remap = irq_domain_check_msi_remap() ||
1399 iommu_capable(bus, IOMMU_CAP_INTR_REMAP);
1401 if (!allow_unsafe_interrupts && !msi_remap) {
1402 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
1408 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
1409 domain->prot |= IOMMU_CACHE;
1412 * Try to match an existing compatible domain. We don't want to
1413 * preclude an IOMMU driver supporting multiple bus_types and being
1414 * able to include different bus_types in the same IOMMU domain, so
1415 * we test whether the domains use the same iommu_ops rather than
1416 * testing if they're on the same bus_type.
1418 list_for_each_entry(d, &iommu->domain_list, next) {
1419 if (d->domain->ops == domain->domain->ops &&
1420 d->prot == domain->prot) {
1421 iommu_detach_group(domain->domain, iommu_group);
1422 if (!iommu_attach_group(d->domain, iommu_group)) {
1423 list_add(&group->next, &d->group_list);
1424 iommu_domain_free(domain->domain);
1426 mutex_unlock(&iommu->lock);
1430 ret = iommu_attach_group(domain->domain, iommu_group);
1436 vfio_test_domain_fgsp(domain);
1438 /* replay mappings on new domains */
1439 ret = vfio_iommu_replay(iommu, domain);
1444 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
1449 list_add(&domain->next, &iommu->domain_list);
1451 mutex_unlock(&iommu->lock);
1456 iommu_detach_group(domain->domain, iommu_group);
1458 iommu_domain_free(domain->domain);
1462 mutex_unlock(&iommu->lock);
1466 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
1468 struct rb_node *node;
1470 while ((node = rb_first(&iommu->dma_list)))
1471 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
1474 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
1476 struct rb_node *n, *p;
1478 n = rb_first(&iommu->dma_list);
1479 for (; n; n = rb_next(n)) {
1480 struct vfio_dma *dma;
1481 long locked = 0, unlocked = 0;
1483 dma = rb_entry(n, struct vfio_dma, node);
1484 unlocked += vfio_unmap_unpin(iommu, dma, false);
1485 p = rb_first(&dma->pfn_list);
1486 for (; p; p = rb_next(p)) {
1487 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
1490 if (!is_invalid_reserved_pfn(vpfn->pfn))
1493 vfio_lock_acct(dma, locked - unlocked, true);
1497 static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu)
1501 n = rb_first(&iommu->dma_list);
1502 for (; n; n = rb_next(n)) {
1503 struct vfio_dma *dma;
1505 dma = rb_entry(n, struct vfio_dma, node);
1507 if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list)))
1510 /* mdev vendor driver must unregister notifier */
1511 WARN_ON(iommu->notifier.head);
1514 static void vfio_iommu_type1_detach_group(void *iommu_data,
1515 struct iommu_group *iommu_group)
1517 struct vfio_iommu *iommu = iommu_data;
1518 struct vfio_domain *domain;
1519 struct vfio_group *group;
1521 mutex_lock(&iommu->lock);
1523 if (iommu->external_domain) {
1524 group = find_iommu_group(iommu->external_domain, iommu_group);
1526 list_del(&group->next);
1529 if (list_empty(&iommu->external_domain->group_list)) {
1530 vfio_sanity_check_pfn_list(iommu);
1532 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1533 vfio_iommu_unmap_unpin_all(iommu);
1535 kfree(iommu->external_domain);
1536 iommu->external_domain = NULL;
1538 goto detach_group_done;
1542 list_for_each_entry(domain, &iommu->domain_list, next) {
1543 group = find_iommu_group(domain, iommu_group);
1547 iommu_detach_group(domain->domain, iommu_group);
1548 list_del(&group->next);
1551 * Group ownership provides privilege, if the group list is
1552 * empty, the domain goes away. If it's the last domain with
1553 * iommu and external domain doesn't exist, then all the
1554 * mappings go away too. If it's the last domain with iommu and
1555 * external domain exist, update accounting
1557 if (list_empty(&domain->group_list)) {
1558 if (list_is_singular(&iommu->domain_list)) {
1559 if (!iommu->external_domain)
1560 vfio_iommu_unmap_unpin_all(iommu);
1562 vfio_iommu_unmap_unpin_reaccount(iommu);
1564 iommu_domain_free(domain->domain);
1565 list_del(&domain->next);
1572 mutex_unlock(&iommu->lock);
1575 static void *vfio_iommu_type1_open(unsigned long arg)
1577 struct vfio_iommu *iommu;
1579 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
1581 return ERR_PTR(-ENOMEM);
1584 case VFIO_TYPE1_IOMMU:
1586 case VFIO_TYPE1_NESTING_IOMMU:
1587 iommu->nesting = true;
1589 case VFIO_TYPE1v2_IOMMU:
1594 return ERR_PTR(-EINVAL);
1597 INIT_LIST_HEAD(&iommu->domain_list);
1598 iommu->dma_list = RB_ROOT;
1599 iommu->dma_avail = dma_entry_limit;
1600 mutex_init(&iommu->lock);
1601 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
1606 static void vfio_release_domain(struct vfio_domain *domain, bool external)
1608 struct vfio_group *group, *group_tmp;
1610 list_for_each_entry_safe(group, group_tmp,
1611 &domain->group_list, next) {
1613 iommu_detach_group(domain->domain, group->iommu_group);
1614 list_del(&group->next);
1619 iommu_domain_free(domain->domain);
1622 static void vfio_iommu_type1_release(void *iommu_data)
1624 struct vfio_iommu *iommu = iommu_data;
1625 struct vfio_domain *domain, *domain_tmp;
1627 if (iommu->external_domain) {
1628 vfio_release_domain(iommu->external_domain, true);
1629 vfio_sanity_check_pfn_list(iommu);
1630 kfree(iommu->external_domain);
1633 vfio_iommu_unmap_unpin_all(iommu);
1635 list_for_each_entry_safe(domain, domain_tmp,
1636 &iommu->domain_list, next) {
1637 vfio_release_domain(domain, false);
1638 list_del(&domain->next);
1644 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
1646 struct vfio_domain *domain;
1649 mutex_lock(&iommu->lock);
1650 list_for_each_entry(domain, &iommu->domain_list, next) {
1651 if (!(domain->prot & IOMMU_CACHE)) {
1656 mutex_unlock(&iommu->lock);
1661 static long vfio_iommu_type1_ioctl(void *iommu_data,
1662 unsigned int cmd, unsigned long arg)
1664 struct vfio_iommu *iommu = iommu_data;
1665 unsigned long minsz;
1667 if (cmd == VFIO_CHECK_EXTENSION) {
1669 case VFIO_TYPE1_IOMMU:
1670 case VFIO_TYPE1v2_IOMMU:
1671 case VFIO_TYPE1_NESTING_IOMMU:
1673 case VFIO_DMA_CC_IOMMU:
1676 return vfio_domains_have_iommu_cache(iommu);
1680 } else if (cmd == VFIO_IOMMU_GET_INFO) {
1681 struct vfio_iommu_type1_info info;
1683 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
1685 if (copy_from_user(&info, (void __user *)arg, minsz))
1688 if (info.argsz < minsz)
1691 info.flags = VFIO_IOMMU_INFO_PGSIZES;
1693 info.iova_pgsizes = vfio_pgsize_bitmap(iommu);
1695 return copy_to_user((void __user *)arg, &info, minsz) ?
1698 } else if (cmd == VFIO_IOMMU_MAP_DMA) {
1699 struct vfio_iommu_type1_dma_map map;
1700 uint32_t mask = VFIO_DMA_MAP_FLAG_READ |
1701 VFIO_DMA_MAP_FLAG_WRITE;
1703 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
1705 if (copy_from_user(&map, (void __user *)arg, minsz))
1708 if (map.argsz < minsz || map.flags & ~mask)
1711 return vfio_dma_do_map(iommu, &map);
1713 } else if (cmd == VFIO_IOMMU_UNMAP_DMA) {
1714 struct vfio_iommu_type1_dma_unmap unmap;
1717 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
1719 if (copy_from_user(&unmap, (void __user *)arg, minsz))
1722 if (unmap.argsz < minsz || unmap.flags)
1725 ret = vfio_dma_do_unmap(iommu, &unmap);
1729 return copy_to_user((void __user *)arg, &unmap, minsz) ?
1736 static int vfio_iommu_type1_register_notifier(void *iommu_data,
1737 unsigned long *events,
1738 struct notifier_block *nb)
1740 struct vfio_iommu *iommu = iommu_data;
1742 /* clear known events */
1743 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
1745 /* refuse to register if still events remaining */
1749 return blocking_notifier_chain_register(&iommu->notifier, nb);
1752 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
1753 struct notifier_block *nb)
1755 struct vfio_iommu *iommu = iommu_data;
1757 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
1760 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
1761 .name = "vfio-iommu-type1",
1762 .owner = THIS_MODULE,
1763 .open = vfio_iommu_type1_open,
1764 .release = vfio_iommu_type1_release,
1765 .ioctl = vfio_iommu_type1_ioctl,
1766 .attach_group = vfio_iommu_type1_attach_group,
1767 .detach_group = vfio_iommu_type1_detach_group,
1768 .pin_pages = vfio_iommu_type1_pin_pages,
1769 .unpin_pages = vfio_iommu_type1_unpin_pages,
1770 .register_notifier = vfio_iommu_type1_register_notifier,
1771 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
1774 static int __init vfio_iommu_type1_init(void)
1776 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
1779 static void __exit vfio_iommu_type1_cleanup(void)
1781 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
1784 module_init(vfio_iommu_type1_init);
1785 module_exit(vfio_iommu_type1_cleanup);
1787 MODULE_VERSION(DRIVER_VERSION);
1788 MODULE_LICENSE("GPL v2");
1789 MODULE_AUTHOR(DRIVER_AUTHOR);
1790 MODULE_DESCRIPTION(DRIVER_DESC);