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 #define IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu) \
113 (!list_empty(&iommu->domain_list))
115 static int put_pfn(unsigned long pfn, int prot);
118 * This code handles mapping and unmapping of user data buffers
119 * into DMA'ble space using the IOMMU
122 static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
123 dma_addr_t start, size_t size)
125 struct rb_node *node = iommu->dma_list.rb_node;
128 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
130 if (start + size <= dma->iova)
131 node = node->rb_left;
132 else if (start >= dma->iova + dma->size)
133 node = node->rb_right;
141 static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new)
143 struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL;
144 struct vfio_dma *dma;
148 dma = rb_entry(parent, struct vfio_dma, node);
150 if (new->iova + new->size <= dma->iova)
151 link = &(*link)->rb_left;
153 link = &(*link)->rb_right;
156 rb_link_node(&new->node, parent, link);
157 rb_insert_color(&new->node, &iommu->dma_list);
160 static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
162 rb_erase(&old->node, &iommu->dma_list);
166 * Helper Functions for host iova-pfn list
168 static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
170 struct vfio_pfn *vpfn;
171 struct rb_node *node = dma->pfn_list.rb_node;
174 vpfn = rb_entry(node, struct vfio_pfn, node);
176 if (iova < vpfn->iova)
177 node = node->rb_left;
178 else if (iova > vpfn->iova)
179 node = node->rb_right;
186 static void vfio_link_pfn(struct vfio_dma *dma,
187 struct vfio_pfn *new)
189 struct rb_node **link, *parent = NULL;
190 struct vfio_pfn *vpfn;
192 link = &dma->pfn_list.rb_node;
195 vpfn = rb_entry(parent, struct vfio_pfn, node);
197 if (new->iova < vpfn->iova)
198 link = &(*link)->rb_left;
200 link = &(*link)->rb_right;
203 rb_link_node(&new->node, parent, link);
204 rb_insert_color(&new->node, &dma->pfn_list);
207 static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old)
209 rb_erase(&old->node, &dma->pfn_list);
212 static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova,
215 struct vfio_pfn *vpfn;
217 vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL);
223 atomic_set(&vpfn->ref_count, 1);
224 vfio_link_pfn(dma, vpfn);
228 static void vfio_remove_from_pfn_list(struct vfio_dma *dma,
229 struct vfio_pfn *vpfn)
231 vfio_unlink_pfn(dma, vpfn);
235 static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma,
238 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
241 atomic_inc(&vpfn->ref_count);
245 static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn)
249 if (atomic_dec_and_test(&vpfn->ref_count)) {
250 ret = put_pfn(vpfn->pfn, dma->prot);
251 vfio_remove_from_pfn_list(dma, vpfn);
256 static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
258 struct mm_struct *mm;
264 mm = async ? get_task_mm(dma->task) : dma->task->mm;
266 return -ESRCH; /* process exited */
268 ret = down_write_killable(&mm->mmap_sem);
271 if (!dma->lock_cap) {
274 limit = task_rlimit(dma->task,
275 RLIMIT_MEMLOCK) >> PAGE_SHIFT;
277 if (mm->locked_vm + npage > limit)
283 mm->locked_vm += npage;
285 up_write(&mm->mmap_sem);
295 * Some mappings aren't backed by a struct page, for example an mmap'd
296 * MMIO range for our own or another device. These use a different
297 * pfn conversion and shouldn't be tracked as locked pages.
299 static bool is_invalid_reserved_pfn(unsigned long pfn)
301 if (pfn_valid(pfn)) {
303 struct page *tail = pfn_to_page(pfn);
304 struct page *head = compound_head(tail);
305 reserved = !!(PageReserved(head));
308 * "head" is not a dangling pointer
309 * (compound_head takes care of that)
310 * but the hugepage may have been split
311 * from under us (and we may not hold a
312 * reference count on the head page so it can
313 * be reused before we run PageReferenced), so
314 * we've to check PageTail before returning
321 return PageReserved(tail);
327 static int put_pfn(unsigned long pfn, int prot)
329 if (!is_invalid_reserved_pfn(pfn)) {
330 struct page *page = pfn_to_page(pfn);
331 if (prot & IOMMU_WRITE)
339 static int follow_fault_pfn(struct vm_area_struct *vma, struct mm_struct *mm,
340 unsigned long vaddr, unsigned long *pfn,
345 ret = follow_pfn(vma, vaddr, pfn);
347 bool unlocked = false;
349 ret = fixup_user_fault(NULL, mm, vaddr,
351 (write_fault ? FAULT_FLAG_WRITE : 0),
359 ret = follow_pfn(vma, vaddr, pfn);
365 static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,
366 int prot, unsigned long *pfn)
368 struct page *page[1];
369 struct vm_area_struct *vma;
370 struct vm_area_struct *vmas[1];
371 unsigned int flags = 0;
374 if (prot & IOMMU_WRITE)
377 down_read(&mm->mmap_sem);
378 if (mm == current->mm) {
379 ret = get_user_pages_longterm(vaddr, 1, flags, page, vmas);
381 ret = get_user_pages_remote(NULL, mm, vaddr, 1, flags, page,
384 * The lifetime of a vaddr_get_pfn() page pin is
385 * userspace-controlled. In the fs-dax case this could
386 * lead to indefinite stalls in filesystem operations.
387 * Disallow attempts to pin fs-dax pages via this
390 if (ret > 0 && vma_is_fsdax(vmas[0])) {
395 up_read(&mm->mmap_sem);
398 *pfn = page_to_pfn(page[0]);
402 down_read(&mm->mmap_sem);
405 vma = find_vma_intersection(mm, vaddr, vaddr + 1);
407 if (vma && vma->vm_flags & VM_PFNMAP) {
408 ret = follow_fault_pfn(vma, mm, vaddr, pfn, prot & IOMMU_WRITE);
412 if (!ret && !is_invalid_reserved_pfn(*pfn))
416 up_read(&mm->mmap_sem);
421 * Attempt to pin pages. We really don't want to track all the pfns and
422 * the iommu can only map chunks of consecutive pfns anyway, so get the
423 * first page and all consecutive pages with the same locking.
425 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
426 long npage, unsigned long *pfn_base,
429 unsigned long pfn = 0;
430 long ret, pinned = 0, lock_acct = 0;
432 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
434 /* This code path is only user initiated */
438 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base);
443 rsvd = is_invalid_reserved_pfn(*pfn_base);
446 * Reserved pages aren't counted against the user, externally pinned
447 * pages are already counted against the user.
449 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
450 if (!dma->lock_cap && current->mm->locked_vm + 1 > limit) {
451 put_pfn(*pfn_base, dma->prot);
452 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__,
453 limit << PAGE_SHIFT);
459 if (unlikely(disable_hugepages))
462 /* Lock all the consecutive pages from pfn_base */
463 for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage;
464 pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) {
465 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn);
469 if (pfn != *pfn_base + pinned ||
470 rsvd != is_invalid_reserved_pfn(pfn)) {
471 put_pfn(pfn, dma->prot);
475 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
476 if (!dma->lock_cap &&
477 current->mm->locked_vm + lock_acct + 1 > limit) {
478 put_pfn(pfn, dma->prot);
479 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
480 __func__, limit << PAGE_SHIFT);
489 ret = vfio_lock_acct(dma, lock_acct, false);
494 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
495 put_pfn(pfn, dma->prot);
504 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
505 unsigned long pfn, long npage,
508 long unlocked = 0, locked = 0;
511 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
512 if (put_pfn(pfn++, dma->prot)) {
514 if (vfio_find_vpfn(dma, iova))
520 vfio_lock_acct(dma, locked - unlocked, true);
525 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
526 unsigned long *pfn_base, bool do_accounting)
528 struct mm_struct *mm;
531 mm = get_task_mm(dma->task);
535 ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base);
536 if (!ret && do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
537 ret = vfio_lock_acct(dma, 1, true);
539 put_pfn(*pfn_base, dma->prot);
541 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
542 "(%ld) exceeded\n", __func__,
543 dma->task->comm, task_pid_nr(dma->task),
544 task_rlimit(dma->task, RLIMIT_MEMLOCK));
552 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
556 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
561 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
564 vfio_lock_acct(dma, -unlocked, true);
569 static int vfio_iommu_type1_pin_pages(void *iommu_data,
570 unsigned long *user_pfn,
572 unsigned long *phys_pfn)
574 struct vfio_iommu *iommu = iommu_data;
576 unsigned long remote_vaddr;
577 struct vfio_dma *dma;
580 if (!iommu || !user_pfn || !phys_pfn)
583 /* Supported for v2 version only */
587 mutex_lock(&iommu->lock);
589 /* Fail if notifier list is empty */
590 if ((!iommu->external_domain) || (!iommu->notifier.head)) {
596 * If iommu capable domain exist in the container then all pages are
597 * already pinned and accounted. Accouting should be done if there is no
598 * iommu capable domain in the container.
600 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
602 for (i = 0; i < npage; i++) {
604 struct vfio_pfn *vpfn;
606 iova = user_pfn[i] << PAGE_SHIFT;
607 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
613 if ((dma->prot & prot) != prot) {
618 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
620 phys_pfn[i] = vpfn->pfn;
624 remote_vaddr = dma->vaddr + (iova - dma->iova);
625 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
630 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
632 if (put_pfn(phys_pfn[i], dma->prot) && do_accounting)
633 vfio_lock_acct(dma, -1, true);
643 for (j = 0; j < i; j++) {
646 iova = user_pfn[j] << PAGE_SHIFT;
647 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
648 vfio_unpin_page_external(dma, iova, do_accounting);
652 mutex_unlock(&iommu->lock);
656 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
657 unsigned long *user_pfn,
660 struct vfio_iommu *iommu = iommu_data;
664 if (!iommu || !user_pfn)
667 /* Supported for v2 version only */
671 mutex_lock(&iommu->lock);
673 if (!iommu->external_domain) {
674 mutex_unlock(&iommu->lock);
678 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
679 for (i = 0; i < npage; i++) {
680 struct vfio_dma *dma;
683 iova = user_pfn[i] << PAGE_SHIFT;
684 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
687 vfio_unpin_page_external(dma, iova, do_accounting);
691 mutex_unlock(&iommu->lock);
692 return i > npage ? npage : (i > 0 ? i : -EINVAL);
695 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
698 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
699 struct vfio_domain *domain, *d;
705 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
709 * We use the IOMMU to track the physical addresses, otherwise we'd
710 * need a much more complicated tracking system. Unfortunately that
711 * means we need to use one of the iommu domains to figure out the
712 * pfns to unpin. The rest need to be unmapped in advance so we have
713 * no iommu translations remaining when the pages are unpinned.
715 domain = d = list_first_entry(&iommu->domain_list,
716 struct vfio_domain, next);
718 list_for_each_entry_continue(d, &iommu->domain_list, next) {
719 iommu_unmap(d->domain, dma->iova, dma->size);
724 size_t unmapped, len;
725 phys_addr_t phys, next;
727 phys = iommu_iova_to_phys(domain->domain, iova);
728 if (WARN_ON(!phys)) {
734 * To optimize for fewer iommu_unmap() calls, each of which
735 * may require hardware cache flushing, try to find the
736 * largest contiguous physical memory chunk to unmap.
738 for (len = PAGE_SIZE;
739 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
740 next = iommu_iova_to_phys(domain->domain, iova + len);
741 if (next != phys + len)
745 unmapped = iommu_unmap(domain->domain, iova, len);
746 if (WARN_ON(!unmapped))
749 unlocked += vfio_unpin_pages_remote(dma, iova,
751 unmapped >> PAGE_SHIFT,
758 dma->iommu_mapped = false;
760 vfio_lock_acct(dma, -unlocked, true);
766 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
768 vfio_unmap_unpin(iommu, dma, true);
769 vfio_unlink_dma(iommu, dma);
770 put_task_struct(dma->task);
775 static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu)
777 struct vfio_domain *domain;
778 unsigned long bitmap = ULONG_MAX;
780 mutex_lock(&iommu->lock);
781 list_for_each_entry(domain, &iommu->domain_list, next)
782 bitmap &= domain->domain->pgsize_bitmap;
783 mutex_unlock(&iommu->lock);
786 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
787 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
788 * That way the user will be able to map/unmap buffers whose size/
789 * start address is aligned with PAGE_SIZE. Pinning code uses that
790 * granularity while iommu driver can use the sub-PAGE_SIZE size
793 if (bitmap & ~PAGE_MASK) {
801 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
802 struct vfio_iommu_type1_dma_unmap *unmap)
805 struct vfio_dma *dma, *dma_last = NULL;
807 int ret = 0, retries = 0;
809 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
811 if (unmap->iova & mask)
813 if (!unmap->size || unmap->size & mask)
816 WARN_ON(mask & PAGE_MASK);
818 mutex_lock(&iommu->lock);
821 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
822 * avoid tracking individual mappings. This means that the granularity
823 * of the original mapping was lost and the user was allowed to attempt
824 * to unmap any range. Depending on the contiguousness of physical
825 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
826 * or may not have worked. We only guaranteed unmap granularity
827 * matching the original mapping; even though it was untracked here,
828 * the original mappings are reflected in IOMMU mappings. This
829 * resulted in a couple unusual behaviors. First, if a range is not
830 * able to be unmapped, ex. a set of 4k pages that was mapped as a
831 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
832 * a zero sized unmap. Also, if an unmap request overlaps the first
833 * address of a hugepage, the IOMMU will unmap the entire hugepage.
834 * This also returns success and the returned unmap size reflects the
835 * actual size unmapped.
837 * We attempt to maintain compatibility with this "v1" interface, but
838 * we take control out of the hands of the IOMMU. Therefore, an unmap
839 * request offset from the beginning of the original mapping will
840 * return success with zero sized unmap. And an unmap request covering
841 * the first iova of mapping will unmap the entire range.
843 * The v2 version of this interface intends to be more deterministic.
844 * Unmap requests must fully cover previous mappings. Multiple
845 * mappings may still be unmaped by specifying large ranges, but there
846 * must not be any previous mappings bisected by the range. An error
847 * will be returned if these conditions are not met. The v2 interface
848 * will only return success and a size of zero if there were no
849 * mappings within the range.
852 dma = vfio_find_dma(iommu, unmap->iova, 1);
853 if (dma && dma->iova != unmap->iova) {
857 dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0);
858 if (dma && dma->iova + dma->size != unmap->iova + unmap->size) {
864 while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
865 if (!iommu->v2 && unmap->iova > dma->iova)
868 * Task with same address space who mapped this iova range is
869 * allowed to unmap the iova range.
871 if (dma->task->mm != current->mm)
874 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
875 struct vfio_iommu_type1_dma_unmap nb_unmap;
877 if (dma_last == dma) {
878 BUG_ON(++retries > 10);
884 nb_unmap.iova = dma->iova;
885 nb_unmap.size = dma->size;
888 * Notify anyone (mdev vendor drivers) to invalidate and
889 * unmap iovas within the range we're about to unmap.
890 * Vendor drivers MUST unpin pages in response to an
893 mutex_unlock(&iommu->lock);
894 blocking_notifier_call_chain(&iommu->notifier,
895 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
899 unmapped += dma->size;
900 vfio_remove_dma(iommu, dma);
904 mutex_unlock(&iommu->lock);
906 /* Report how much was unmapped */
907 unmap->size = unmapped;
913 * Turns out AMD IOMMU has a page table bug where it won't map large pages
914 * to a region that previously mapped smaller pages. This should be fixed
915 * soon, so this is just a temporary workaround to break mappings down into
916 * PAGE_SIZE. Better to map smaller pages than nothing.
918 static int map_try_harder(struct vfio_domain *domain, dma_addr_t iova,
919 unsigned long pfn, long npage, int prot)
924 for (i = 0; i < npage; i++, pfn++, iova += PAGE_SIZE) {
925 ret = iommu_map(domain->domain, iova,
926 (phys_addr_t)pfn << PAGE_SHIFT,
927 PAGE_SIZE, prot | domain->prot);
932 for (; i < npage && i > 0; i--, iova -= PAGE_SIZE)
933 iommu_unmap(domain->domain, iova, PAGE_SIZE);
938 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
939 unsigned long pfn, long npage, int prot)
941 struct vfio_domain *d;
944 list_for_each_entry(d, &iommu->domain_list, next) {
945 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
946 npage << PAGE_SHIFT, prot | d->prot);
949 map_try_harder(d, iova, pfn, npage, prot))
959 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next)
960 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
965 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
968 dma_addr_t iova = dma->iova;
969 unsigned long vaddr = dma->vaddr;
970 size_t size = map_size;
972 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
976 /* Pin a contiguous chunk of memory */
977 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
978 size >> PAGE_SHIFT, &pfn, limit);
986 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
989 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
994 size -= npage << PAGE_SHIFT;
995 dma->size += npage << PAGE_SHIFT;
998 dma->iommu_mapped = true;
1001 vfio_remove_dma(iommu, dma);
1006 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1007 struct vfio_iommu_type1_dma_map *map)
1009 dma_addr_t iova = map->iova;
1010 unsigned long vaddr = map->vaddr;
1011 size_t size = map->size;
1012 int ret = 0, prot = 0;
1014 struct vfio_dma *dma;
1016 /* Verify that none of our __u64 fields overflow */
1017 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1020 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
1022 WARN_ON(mask & PAGE_MASK);
1024 /* READ/WRITE from device perspective */
1025 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1026 prot |= IOMMU_WRITE;
1027 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1030 if (!prot || !size || (size | iova | vaddr) & mask)
1033 /* Don't allow IOVA or virtual address wrap */
1034 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr)
1037 mutex_lock(&iommu->lock);
1039 if (vfio_find_dma(iommu, iova, size)) {
1044 if (!iommu->dma_avail) {
1049 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1061 * We need to be able to both add to a task's locked memory and test
1062 * against the locked memory limit and we need to be able to do both
1063 * outside of this call path as pinning can be asynchronous via the
1064 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1065 * task_struct and VM locked pages requires an mm_struct, however
1066 * holding an indefinite mm reference is not recommended, therefore we
1067 * only hold a reference to a task. We could hold a reference to
1068 * current, however QEMU uses this call path through vCPU threads,
1069 * which can be killed resulting in a NULL mm and failure in the unmap
1070 * path when called via a different thread. Avoid this problem by
1071 * using the group_leader as threads within the same group require
1072 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1075 * Previously we also used the task for testing CAP_IPC_LOCK at the
1076 * time of pinning and accounting, however has_capability() makes use
1077 * of real_cred, a copy-on-write field, so we can't guarantee that it
1078 * matches group_leader, or in fact that it might not change by the
1079 * time it's evaluated. If a process were to call MAP_DMA with
1080 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1081 * possibly see different results for an iommu_mapped vfio_dma vs
1082 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1083 * time of calling MAP_DMA.
1085 get_task_struct(current->group_leader);
1086 dma->task = current->group_leader;
1087 dma->lock_cap = capable(CAP_IPC_LOCK);
1089 dma->pfn_list = RB_ROOT;
1091 /* Insert zero-sized and grow as we map chunks of it */
1092 vfio_link_dma(iommu, dma);
1094 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1095 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1098 ret = vfio_pin_map_dma(iommu, dma, size);
1101 mutex_unlock(&iommu->lock);
1105 static int vfio_bus_type(struct device *dev, void *data)
1107 struct bus_type **bus = data;
1109 if (*bus && *bus != dev->bus)
1117 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1118 struct vfio_domain *domain)
1120 struct vfio_domain *d = NULL;
1122 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1125 /* Arbitrarily pick the first domain in the list for lookups */
1126 if (!list_empty(&iommu->domain_list))
1127 d = list_first_entry(&iommu->domain_list,
1128 struct vfio_domain, next);
1130 n = rb_first(&iommu->dma_list);
1132 for (; n; n = rb_next(n)) {
1133 struct vfio_dma *dma;
1136 dma = rb_entry(n, struct vfio_dma, node);
1139 while (iova < dma->iova + dma->size) {
1143 if (dma->iommu_mapped) {
1147 if (WARN_ON(!d)) { /* mapped w/o a domain?! */
1152 phys = iommu_iova_to_phys(d->domain, iova);
1154 if (WARN_ON(!phys)) {
1162 while (i < dma->iova + dma->size &&
1163 p == iommu_iova_to_phys(d->domain, i)) {
1170 unsigned long vaddr = dma->vaddr +
1172 size_t n = dma->iova + dma->size - iova;
1175 npage = vfio_pin_pages_remote(dma, vaddr,
1184 phys = pfn << PAGE_SHIFT;
1185 size = npage << PAGE_SHIFT;
1188 ret = iommu_map(domain->domain, iova, phys,
1189 size, dma->prot | domain->prot);
1191 if (!dma->iommu_mapped)
1192 vfio_unpin_pages_remote(dma, iova,
1203 /* All dmas are now mapped, defer to second tree walk for unwind */
1204 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1205 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1207 dma->iommu_mapped = true;
1213 for (; n; n = rb_prev(n)) {
1214 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1217 if (dma->iommu_mapped) {
1218 iommu_unmap(domain->domain, dma->iova, dma->size);
1223 while (iova < dma->iova + dma->size) {
1224 phys_addr_t phys, p;
1228 phys = iommu_iova_to_phys(domain->domain, iova);
1237 while (i < dma->iova + dma->size &&
1238 p == iommu_iova_to_phys(domain->domain, i)) {
1244 iommu_unmap(domain->domain, iova, size);
1245 vfio_unpin_pages_remote(dma, iova, phys >> PAGE_SHIFT,
1246 size >> PAGE_SHIFT, true);
1254 * We change our unmap behavior slightly depending on whether the IOMMU
1255 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1256 * for practically any contiguous power-of-two mapping we give it. This means
1257 * we don't need to look for contiguous chunks ourselves to make unmapping
1258 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1259 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1260 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1261 * hugetlbfs is in use.
1263 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1266 int ret, order = get_order(PAGE_SIZE * 2);
1268 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1272 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1273 IOMMU_READ | IOMMU_WRITE | domain->prot);
1275 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1277 if (unmapped == PAGE_SIZE)
1278 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1280 domain->fgsp = true;
1283 __free_pages(pages, order);
1286 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1287 struct iommu_group *iommu_group)
1289 struct vfio_group *g;
1291 list_for_each_entry(g, &domain->group_list, next) {
1292 if (g->iommu_group == iommu_group)
1299 static bool vfio_iommu_has_sw_msi(struct iommu_group *group, phys_addr_t *base)
1301 struct list_head group_resv_regions;
1302 struct iommu_resv_region *region, *next;
1305 INIT_LIST_HEAD(&group_resv_regions);
1306 iommu_get_group_resv_regions(group, &group_resv_regions);
1307 list_for_each_entry(region, &group_resv_regions, list) {
1309 * The presence of any 'real' MSI regions should take
1310 * precedence over the software-managed one if the
1311 * IOMMU driver happens to advertise both types.
1313 if (region->type == IOMMU_RESV_MSI) {
1318 if (region->type == IOMMU_RESV_SW_MSI) {
1319 *base = region->start;
1323 list_for_each_entry_safe(region, next, &group_resv_regions, list)
1328 static int vfio_iommu_type1_attach_group(void *iommu_data,
1329 struct iommu_group *iommu_group)
1331 struct vfio_iommu *iommu = iommu_data;
1332 struct vfio_group *group;
1333 struct vfio_domain *domain, *d;
1334 struct bus_type *bus = NULL, *mdev_bus;
1336 bool resv_msi, msi_remap;
1337 phys_addr_t resv_msi_base;
1339 mutex_lock(&iommu->lock);
1341 list_for_each_entry(d, &iommu->domain_list, next) {
1342 if (find_iommu_group(d, iommu_group)) {
1343 mutex_unlock(&iommu->lock);
1348 if (iommu->external_domain) {
1349 if (find_iommu_group(iommu->external_domain, iommu_group)) {
1350 mutex_unlock(&iommu->lock);
1355 group = kzalloc(sizeof(*group), GFP_KERNEL);
1356 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
1357 if (!group || !domain) {
1362 group->iommu_group = iommu_group;
1364 /* Determine bus_type in order to allocate a domain */
1365 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
1369 mdev_bus = symbol_get(mdev_bus_type);
1372 if ((bus == mdev_bus) && !iommu_present(bus)) {
1373 symbol_put(mdev_bus_type);
1374 if (!iommu->external_domain) {
1375 INIT_LIST_HEAD(&domain->group_list);
1376 iommu->external_domain = domain;
1380 list_add(&group->next,
1381 &iommu->external_domain->group_list);
1382 mutex_unlock(&iommu->lock);
1385 symbol_put(mdev_bus_type);
1388 domain->domain = iommu_domain_alloc(bus);
1389 if (!domain->domain) {
1394 if (iommu->nesting) {
1397 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
1403 ret = iommu_attach_group(domain->domain, iommu_group);
1407 resv_msi = vfio_iommu_has_sw_msi(iommu_group, &resv_msi_base);
1409 INIT_LIST_HEAD(&domain->group_list);
1410 list_add(&group->next, &domain->group_list);
1412 msi_remap = irq_domain_check_msi_remap() ||
1413 iommu_capable(bus, IOMMU_CAP_INTR_REMAP);
1415 if (!allow_unsafe_interrupts && !msi_remap) {
1416 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
1422 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
1423 domain->prot |= IOMMU_CACHE;
1426 * Try to match an existing compatible domain. We don't want to
1427 * preclude an IOMMU driver supporting multiple bus_types and being
1428 * able to include different bus_types in the same IOMMU domain, so
1429 * we test whether the domains use the same iommu_ops rather than
1430 * testing if they're on the same bus_type.
1432 list_for_each_entry(d, &iommu->domain_list, next) {
1433 if (d->domain->ops == domain->domain->ops &&
1434 d->prot == domain->prot) {
1435 iommu_detach_group(domain->domain, iommu_group);
1436 if (!iommu_attach_group(d->domain, iommu_group)) {
1437 list_add(&group->next, &d->group_list);
1438 iommu_domain_free(domain->domain);
1440 mutex_unlock(&iommu->lock);
1444 ret = iommu_attach_group(domain->domain, iommu_group);
1450 vfio_test_domain_fgsp(domain);
1452 /* replay mappings on new domains */
1453 ret = vfio_iommu_replay(iommu, domain);
1458 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
1463 list_add(&domain->next, &iommu->domain_list);
1465 mutex_unlock(&iommu->lock);
1470 iommu_detach_group(domain->domain, iommu_group);
1472 iommu_domain_free(domain->domain);
1476 mutex_unlock(&iommu->lock);
1480 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
1482 struct rb_node *node;
1484 while ((node = rb_first(&iommu->dma_list)))
1485 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
1488 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
1490 struct rb_node *n, *p;
1492 n = rb_first(&iommu->dma_list);
1493 for (; n; n = rb_next(n)) {
1494 struct vfio_dma *dma;
1495 long locked = 0, unlocked = 0;
1497 dma = rb_entry(n, struct vfio_dma, node);
1498 unlocked += vfio_unmap_unpin(iommu, dma, false);
1499 p = rb_first(&dma->pfn_list);
1500 for (; p; p = rb_next(p)) {
1501 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
1504 if (!is_invalid_reserved_pfn(vpfn->pfn))
1507 vfio_lock_acct(dma, locked - unlocked, true);
1511 static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu)
1515 n = rb_first(&iommu->dma_list);
1516 for (; n; n = rb_next(n)) {
1517 struct vfio_dma *dma;
1519 dma = rb_entry(n, struct vfio_dma, node);
1521 if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list)))
1524 /* mdev vendor driver must unregister notifier */
1525 WARN_ON(iommu->notifier.head);
1528 static void vfio_iommu_type1_detach_group(void *iommu_data,
1529 struct iommu_group *iommu_group)
1531 struct vfio_iommu *iommu = iommu_data;
1532 struct vfio_domain *domain;
1533 struct vfio_group *group;
1535 mutex_lock(&iommu->lock);
1537 if (iommu->external_domain) {
1538 group = find_iommu_group(iommu->external_domain, iommu_group);
1540 list_del(&group->next);
1543 if (list_empty(&iommu->external_domain->group_list)) {
1544 vfio_sanity_check_pfn_list(iommu);
1546 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1547 vfio_iommu_unmap_unpin_all(iommu);
1549 kfree(iommu->external_domain);
1550 iommu->external_domain = NULL;
1552 goto detach_group_done;
1556 list_for_each_entry(domain, &iommu->domain_list, next) {
1557 group = find_iommu_group(domain, iommu_group);
1561 iommu_detach_group(domain->domain, iommu_group);
1562 list_del(&group->next);
1565 * Group ownership provides privilege, if the group list is
1566 * empty, the domain goes away. If it's the last domain with
1567 * iommu and external domain doesn't exist, then all the
1568 * mappings go away too. If it's the last domain with iommu and
1569 * external domain exist, update accounting
1571 if (list_empty(&domain->group_list)) {
1572 if (list_is_singular(&iommu->domain_list)) {
1573 if (!iommu->external_domain)
1574 vfio_iommu_unmap_unpin_all(iommu);
1576 vfio_iommu_unmap_unpin_reaccount(iommu);
1578 iommu_domain_free(domain->domain);
1579 list_del(&domain->next);
1586 mutex_unlock(&iommu->lock);
1589 static void *vfio_iommu_type1_open(unsigned long arg)
1591 struct vfio_iommu *iommu;
1593 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
1595 return ERR_PTR(-ENOMEM);
1598 case VFIO_TYPE1_IOMMU:
1600 case VFIO_TYPE1_NESTING_IOMMU:
1601 iommu->nesting = true;
1602 case VFIO_TYPE1v2_IOMMU:
1607 return ERR_PTR(-EINVAL);
1610 INIT_LIST_HEAD(&iommu->domain_list);
1611 iommu->dma_list = RB_ROOT;
1612 iommu->dma_avail = dma_entry_limit;
1613 mutex_init(&iommu->lock);
1614 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
1619 static void vfio_release_domain(struct vfio_domain *domain, bool external)
1621 struct vfio_group *group, *group_tmp;
1623 list_for_each_entry_safe(group, group_tmp,
1624 &domain->group_list, next) {
1626 iommu_detach_group(domain->domain, group->iommu_group);
1627 list_del(&group->next);
1632 iommu_domain_free(domain->domain);
1635 static void vfio_iommu_type1_release(void *iommu_data)
1637 struct vfio_iommu *iommu = iommu_data;
1638 struct vfio_domain *domain, *domain_tmp;
1640 if (iommu->external_domain) {
1641 vfio_release_domain(iommu->external_domain, true);
1642 vfio_sanity_check_pfn_list(iommu);
1643 kfree(iommu->external_domain);
1646 vfio_iommu_unmap_unpin_all(iommu);
1648 list_for_each_entry_safe(domain, domain_tmp,
1649 &iommu->domain_list, next) {
1650 vfio_release_domain(domain, false);
1651 list_del(&domain->next);
1657 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
1659 struct vfio_domain *domain;
1662 mutex_lock(&iommu->lock);
1663 list_for_each_entry(domain, &iommu->domain_list, next) {
1664 if (!(domain->prot & IOMMU_CACHE)) {
1669 mutex_unlock(&iommu->lock);
1674 static long vfio_iommu_type1_ioctl(void *iommu_data,
1675 unsigned int cmd, unsigned long arg)
1677 struct vfio_iommu *iommu = iommu_data;
1678 unsigned long minsz;
1680 if (cmd == VFIO_CHECK_EXTENSION) {
1682 case VFIO_TYPE1_IOMMU:
1683 case VFIO_TYPE1v2_IOMMU:
1684 case VFIO_TYPE1_NESTING_IOMMU:
1686 case VFIO_DMA_CC_IOMMU:
1689 return vfio_domains_have_iommu_cache(iommu);
1693 } else if (cmd == VFIO_IOMMU_GET_INFO) {
1694 struct vfio_iommu_type1_info info;
1696 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
1698 if (copy_from_user(&info, (void __user *)arg, minsz))
1701 if (info.argsz < minsz)
1704 info.flags = VFIO_IOMMU_INFO_PGSIZES;
1706 info.iova_pgsizes = vfio_pgsize_bitmap(iommu);
1708 return copy_to_user((void __user *)arg, &info, minsz) ?
1711 } else if (cmd == VFIO_IOMMU_MAP_DMA) {
1712 struct vfio_iommu_type1_dma_map map;
1713 uint32_t mask = VFIO_DMA_MAP_FLAG_READ |
1714 VFIO_DMA_MAP_FLAG_WRITE;
1716 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
1718 if (copy_from_user(&map, (void __user *)arg, minsz))
1721 if (map.argsz < minsz || map.flags & ~mask)
1724 return vfio_dma_do_map(iommu, &map);
1726 } else if (cmd == VFIO_IOMMU_UNMAP_DMA) {
1727 struct vfio_iommu_type1_dma_unmap unmap;
1730 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
1732 if (copy_from_user(&unmap, (void __user *)arg, minsz))
1735 if (unmap.argsz < minsz || unmap.flags)
1738 ret = vfio_dma_do_unmap(iommu, &unmap);
1742 return copy_to_user((void __user *)arg, &unmap, minsz) ?
1749 static int vfio_iommu_type1_register_notifier(void *iommu_data,
1750 unsigned long *events,
1751 struct notifier_block *nb)
1753 struct vfio_iommu *iommu = iommu_data;
1755 /* clear known events */
1756 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
1758 /* refuse to register if still events remaining */
1762 return blocking_notifier_chain_register(&iommu->notifier, nb);
1765 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
1766 struct notifier_block *nb)
1768 struct vfio_iommu *iommu = iommu_data;
1770 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
1773 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
1774 .name = "vfio-iommu-type1",
1775 .owner = THIS_MODULE,
1776 .open = vfio_iommu_type1_open,
1777 .release = vfio_iommu_type1_release,
1778 .ioctl = vfio_iommu_type1_ioctl,
1779 .attach_group = vfio_iommu_type1_attach_group,
1780 .detach_group = vfio_iommu_type1_detach_group,
1781 .pin_pages = vfio_iommu_type1_pin_pages,
1782 .unpin_pages = vfio_iommu_type1_unpin_pages,
1783 .register_notifier = vfio_iommu_type1_register_notifier,
1784 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
1787 static int __init vfio_iommu_type1_init(void)
1789 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
1792 static void __exit vfio_iommu_type1_cleanup(void)
1794 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
1797 module_init(vfio_iommu_type1_init);
1798 module_exit(vfio_iommu_type1_cleanup);
1800 MODULE_VERSION(DRIVER_VERSION);
1801 MODULE_LICENSE("GPL v2");
1802 MODULE_AUTHOR(DRIVER_AUTHOR);
1803 MODULE_DESCRIPTION(DRIVER_DESC);