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.h>
35 #include <linux/slab.h>
36 #include <linux/uaccess.h>
37 #include <linux/vfio.h>
38 #include <linux/workqueue.h>
39 #include <linux/mdev.h>
40 #include <linux/notifier.h>
42 #define DRIVER_VERSION "0.2"
43 #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
44 #define DRIVER_DESC "Type1 IOMMU driver for VFIO"
46 static bool allow_unsafe_interrupts;
47 module_param_named(allow_unsafe_interrupts,
48 allow_unsafe_interrupts, bool, S_IRUGO | S_IWUSR);
49 MODULE_PARM_DESC(allow_unsafe_interrupts,
50 "Enable VFIO IOMMU support for on platforms without interrupt remapping support.");
52 static bool disable_hugepages;
53 module_param_named(disable_hugepages,
54 disable_hugepages, bool, S_IRUGO | S_IWUSR);
55 MODULE_PARM_DESC(disable_hugepages,
56 "Disable VFIO IOMMU support for IOMMU hugepages.");
59 struct list_head domain_list;
60 struct vfio_domain *external_domain; /* domain for external user */
62 struct rb_root dma_list;
63 struct blocking_notifier_head notifier;
69 struct iommu_domain *domain;
70 struct list_head next;
71 struct list_head group_list;
72 int prot; /* IOMMU_CACHE */
73 bool fgsp; /* Fine-grained super pages */
78 dma_addr_t iova; /* Device address */
79 unsigned long vaddr; /* Process virtual addr */
80 size_t size; /* Map size (bytes) */
81 int prot; /* IOMMU_READ/WRITE */
83 struct task_struct *task;
84 struct rb_root pfn_list; /* Ex-user pinned pfn list */
88 struct iommu_group *iommu_group;
89 struct list_head next;
93 * Guest RAM pinning working set or DMA target
97 dma_addr_t iova; /* Device address */
98 unsigned long pfn; /* Host pfn */
102 #define IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu) \
103 (!list_empty(&iommu->domain_list))
105 static int put_pfn(unsigned long pfn, int prot);
108 * This code handles mapping and unmapping of user data buffers
109 * into DMA'ble space using the IOMMU
112 static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
113 dma_addr_t start, size_t size)
115 struct rb_node *node = iommu->dma_list.rb_node;
118 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
120 if (start + size <= dma->iova)
121 node = node->rb_left;
122 else if (start >= dma->iova + dma->size)
123 node = node->rb_right;
131 static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new)
133 struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL;
134 struct vfio_dma *dma;
138 dma = rb_entry(parent, struct vfio_dma, node);
140 if (new->iova + new->size <= dma->iova)
141 link = &(*link)->rb_left;
143 link = &(*link)->rb_right;
146 rb_link_node(&new->node, parent, link);
147 rb_insert_color(&new->node, &iommu->dma_list);
150 static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
152 rb_erase(&old->node, &iommu->dma_list);
156 * Helper Functions for host iova-pfn list
158 static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
160 struct vfio_pfn *vpfn;
161 struct rb_node *node = dma->pfn_list.rb_node;
164 vpfn = rb_entry(node, struct vfio_pfn, node);
166 if (iova < vpfn->iova)
167 node = node->rb_left;
168 else if (iova > vpfn->iova)
169 node = node->rb_right;
176 static void vfio_link_pfn(struct vfio_dma *dma,
177 struct vfio_pfn *new)
179 struct rb_node **link, *parent = NULL;
180 struct vfio_pfn *vpfn;
182 link = &dma->pfn_list.rb_node;
185 vpfn = rb_entry(parent, struct vfio_pfn, node);
187 if (new->iova < vpfn->iova)
188 link = &(*link)->rb_left;
190 link = &(*link)->rb_right;
193 rb_link_node(&new->node, parent, link);
194 rb_insert_color(&new->node, &dma->pfn_list);
197 static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old)
199 rb_erase(&old->node, &dma->pfn_list);
202 static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova,
205 struct vfio_pfn *vpfn;
207 vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL);
213 atomic_set(&vpfn->ref_count, 1);
214 vfio_link_pfn(dma, vpfn);
218 static void vfio_remove_from_pfn_list(struct vfio_dma *dma,
219 struct vfio_pfn *vpfn)
221 vfio_unlink_pfn(dma, vpfn);
225 static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma,
228 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
231 atomic_inc(&vpfn->ref_count);
235 static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn)
239 if (atomic_dec_and_test(&vpfn->ref_count)) {
240 ret = put_pfn(vpfn->pfn, dma->prot);
241 vfio_remove_from_pfn_list(dma, vpfn);
247 struct mm_struct *mm;
249 struct work_struct work;
252 /* delayed decrement/increment for locked_vm */
253 static void vfio_lock_acct_bg(struct work_struct *work)
255 struct vwork *vwork = container_of(work, struct vwork, work);
256 struct mm_struct *mm;
259 down_write(&mm->mmap_sem);
260 mm->locked_vm += vwork->npage;
261 up_write(&mm->mmap_sem);
266 static void vfio_lock_acct(struct task_struct *task, long npage)
269 struct mm_struct *mm;
275 is_current = (task->mm == current->mm);
277 mm = is_current ? task->mm : get_task_mm(task);
279 return; /* process exited */
281 if (down_write_trylock(&mm->mmap_sem)) {
282 mm->locked_vm += npage;
283 up_write(&mm->mmap_sem);
290 mm = get_task_mm(task);
296 * Couldn't get mmap_sem lock, so must setup to update
297 * mm->locked_vm later. If locked_vm were atomic, we
298 * wouldn't need this silliness
300 vwork = kmalloc(sizeof(struct vwork), GFP_KERNEL);
301 if (WARN_ON(!vwork)) {
305 INIT_WORK(&vwork->work, vfio_lock_acct_bg);
307 vwork->npage = npage;
308 schedule_work(&vwork->work);
312 * Some mappings aren't backed by a struct page, for example an mmap'd
313 * MMIO range for our own or another device. These use a different
314 * pfn conversion and shouldn't be tracked as locked pages.
316 static bool is_invalid_reserved_pfn(unsigned long pfn)
318 if (pfn_valid(pfn)) {
320 struct page *tail = pfn_to_page(pfn);
321 struct page *head = compound_head(tail);
322 reserved = !!(PageReserved(head));
325 * "head" is not a dangling pointer
326 * (compound_head takes care of that)
327 * but the hugepage may have been split
328 * from under us (and we may not hold a
329 * reference count on the head page so it can
330 * be reused before we run PageReferenced), so
331 * we've to check PageTail before returning
338 return PageReserved(tail);
344 static int put_pfn(unsigned long pfn, int prot)
346 if (!is_invalid_reserved_pfn(pfn)) {
347 struct page *page = pfn_to_page(pfn);
348 if (prot & IOMMU_WRITE)
356 static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,
357 int prot, unsigned long *pfn)
359 struct page *page[1];
360 struct vm_area_struct *vma;
363 if (mm == current->mm) {
364 ret = get_user_pages_fast(vaddr, 1, !!(prot & IOMMU_WRITE),
367 unsigned int flags = 0;
369 if (prot & IOMMU_WRITE)
372 down_read(&mm->mmap_sem);
373 ret = get_user_pages_remote(NULL, mm, vaddr, 1, flags, page,
375 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)
405 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
406 bool lock_cap = capable(CAP_IPC_LOCK);
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 (!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 unsigned long pfn = 0;
444 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn);
448 if (pfn != *pfn_base + pinned ||
449 rsvd != is_invalid_reserved_pfn(pfn)) {
450 put_pfn(pfn, dma->prot);
454 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
456 current->mm->locked_vm + lock_acct + 1 > limit) {
457 put_pfn(pfn, dma->prot);
458 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
459 __func__, limit << PAGE_SHIFT);
467 vfio_lock_acct(current, lock_acct);
472 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
473 unsigned long pfn, long npage,
476 long unlocked = 0, locked = 0;
479 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
480 if (put_pfn(pfn++, dma->prot)) {
482 if (vfio_find_vpfn(dma, iova))
488 vfio_lock_acct(dma->task, locked - unlocked);
493 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
494 unsigned long *pfn_base, bool do_accounting)
497 bool lock_cap = has_capability(dma->task, CAP_IPC_LOCK);
498 struct mm_struct *mm;
502 mm = get_task_mm(dma->task);
506 ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base);
510 rsvd = is_invalid_reserved_pfn(*pfn_base);
511 limit = task_rlimit(dma->task, RLIMIT_MEMLOCK) >> PAGE_SHIFT;
513 if (!rsvd && !lock_cap && mm->locked_vm + 1 > limit) {
514 put_pfn(*pfn_base, dma->prot);
515 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK (%ld) exceeded\n",
516 __func__, dma->task->comm, task_pid_nr(dma->task),
517 limit << PAGE_SHIFT);
522 if (!rsvd && do_accounting)
523 vfio_lock_acct(dma->task, 1);
531 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
535 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
540 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
543 vfio_lock_acct(dma->task, -unlocked);
548 static int vfio_iommu_type1_pin_pages(void *iommu_data,
549 unsigned long *user_pfn,
551 unsigned long *phys_pfn)
553 struct vfio_iommu *iommu = iommu_data;
555 unsigned long remote_vaddr;
556 struct vfio_dma *dma;
559 if (!iommu || !user_pfn || !phys_pfn)
562 /* Supported for v2 version only */
566 mutex_lock(&iommu->lock);
568 /* Fail if notifier list is empty */
569 if ((!iommu->external_domain) || (!iommu->notifier.head)) {
575 * If iommu capable domain exist in the container then all pages are
576 * already pinned and accounted. Accouting should be done if there is no
577 * iommu capable domain in the container.
579 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
581 for (i = 0; i < npage; i++) {
583 struct vfio_pfn *vpfn;
585 iova = user_pfn[i] << PAGE_SHIFT;
586 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
592 if ((dma->prot & prot) != prot) {
597 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
599 phys_pfn[i] = vpfn->pfn;
603 remote_vaddr = dma->vaddr + iova - dma->iova;
604 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
611 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
613 vfio_unpin_page_external(dma, iova, do_accounting);
623 for (j = 0; j < i; j++) {
626 iova = user_pfn[j] << PAGE_SHIFT;
627 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
628 vfio_unpin_page_external(dma, iova, do_accounting);
632 mutex_unlock(&iommu->lock);
636 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
637 unsigned long *user_pfn,
640 struct vfio_iommu *iommu = iommu_data;
644 if (!iommu || !user_pfn)
647 /* Supported for v2 version only */
651 mutex_lock(&iommu->lock);
653 if (!iommu->external_domain) {
654 mutex_unlock(&iommu->lock);
658 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
659 for (i = 0; i < npage; i++) {
660 struct vfio_dma *dma;
663 iova = user_pfn[i] << PAGE_SHIFT;
664 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
667 vfio_unpin_page_external(dma, iova, do_accounting);
671 mutex_unlock(&iommu->lock);
672 return i > npage ? npage : (i > 0 ? i : -EINVAL);
675 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
678 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
679 struct vfio_domain *domain, *d;
685 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
689 * We use the IOMMU to track the physical addresses, otherwise we'd
690 * need a much more complicated tracking system. Unfortunately that
691 * means we need to use one of the iommu domains to figure out the
692 * pfns to unpin. The rest need to be unmapped in advance so we have
693 * no iommu translations remaining when the pages are unpinned.
695 domain = d = list_first_entry(&iommu->domain_list,
696 struct vfio_domain, next);
698 list_for_each_entry_continue(d, &iommu->domain_list, next) {
699 iommu_unmap(d->domain, dma->iova, dma->size);
704 size_t unmapped, len;
705 phys_addr_t phys, next;
707 phys = iommu_iova_to_phys(domain->domain, iova);
708 if (WARN_ON(!phys)) {
714 * To optimize for fewer iommu_unmap() calls, each of which
715 * may require hardware cache flushing, try to find the
716 * largest contiguous physical memory chunk to unmap.
718 for (len = PAGE_SIZE;
719 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
720 next = iommu_iova_to_phys(domain->domain, iova + len);
721 if (next != phys + len)
725 unmapped = iommu_unmap(domain->domain, iova, len);
726 if (WARN_ON(!unmapped))
729 unlocked += vfio_unpin_pages_remote(dma, iova,
731 unmapped >> PAGE_SHIFT,
738 dma->iommu_mapped = false;
740 vfio_lock_acct(dma->task, -unlocked);
746 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
748 vfio_unmap_unpin(iommu, dma, true);
749 vfio_unlink_dma(iommu, dma);
750 put_task_struct(dma->task);
754 static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu)
756 struct vfio_domain *domain;
757 unsigned long bitmap = ULONG_MAX;
759 mutex_lock(&iommu->lock);
760 list_for_each_entry(domain, &iommu->domain_list, next)
761 bitmap &= domain->domain->pgsize_bitmap;
762 mutex_unlock(&iommu->lock);
765 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
766 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
767 * That way the user will be able to map/unmap buffers whose size/
768 * start address is aligned with PAGE_SIZE. Pinning code uses that
769 * granularity while iommu driver can use the sub-PAGE_SIZE size
772 if (bitmap & ~PAGE_MASK) {
780 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
781 struct vfio_iommu_type1_dma_unmap *unmap)
784 struct vfio_dma *dma, *dma_last = NULL;
786 int ret = 0, retries = 0;
788 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
790 if (unmap->iova & mask)
792 if (!unmap->size || unmap->size & mask)
795 WARN_ON(mask & PAGE_MASK);
797 mutex_lock(&iommu->lock);
800 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
801 * avoid tracking individual mappings. This means that the granularity
802 * of the original mapping was lost and the user was allowed to attempt
803 * to unmap any range. Depending on the contiguousness of physical
804 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
805 * or may not have worked. We only guaranteed unmap granularity
806 * matching the original mapping; even though it was untracked here,
807 * the original mappings are reflected in IOMMU mappings. This
808 * resulted in a couple unusual behaviors. First, if a range is not
809 * able to be unmapped, ex. a set of 4k pages that was mapped as a
810 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
811 * a zero sized unmap. Also, if an unmap request overlaps the first
812 * address of a hugepage, the IOMMU will unmap the entire hugepage.
813 * This also returns success and the returned unmap size reflects the
814 * actual size unmapped.
816 * We attempt to maintain compatibility with this "v1" interface, but
817 * we take control out of the hands of the IOMMU. Therefore, an unmap
818 * request offset from the beginning of the original mapping will
819 * return success with zero sized unmap. And an unmap request covering
820 * the first iova of mapping will unmap the entire range.
822 * The v2 version of this interface intends to be more deterministic.
823 * Unmap requests must fully cover previous mappings. Multiple
824 * mappings may still be unmaped by specifying large ranges, but there
825 * must not be any previous mappings bisected by the range. An error
826 * will be returned if these conditions are not met. The v2 interface
827 * will only return success and a size of zero if there were no
828 * mappings within the range.
831 dma = vfio_find_dma(iommu, unmap->iova, 1);
832 if (dma && dma->iova != unmap->iova) {
836 dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0);
837 if (dma && dma->iova + dma->size != unmap->iova + unmap->size) {
843 while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
844 if (!iommu->v2 && unmap->iova > dma->iova)
847 * Task with same address space who mapped this iova range is
848 * allowed to unmap the iova range.
850 if (dma->task->mm != current->mm)
853 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
854 struct vfio_iommu_type1_dma_unmap nb_unmap;
856 if (dma_last == dma) {
857 BUG_ON(++retries > 10);
863 nb_unmap.iova = dma->iova;
864 nb_unmap.size = dma->size;
867 * Notify anyone (mdev vendor drivers) to invalidate and
868 * unmap iovas within the range we're about to unmap.
869 * Vendor drivers MUST unpin pages in response to an
872 mutex_unlock(&iommu->lock);
873 blocking_notifier_call_chain(&iommu->notifier,
874 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
878 unmapped += dma->size;
879 vfio_remove_dma(iommu, dma);
883 mutex_unlock(&iommu->lock);
885 /* Report how much was unmapped */
886 unmap->size = unmapped;
892 * Turns out AMD IOMMU has a page table bug where it won't map large pages
893 * to a region that previously mapped smaller pages. This should be fixed
894 * soon, so this is just a temporary workaround to break mappings down into
895 * PAGE_SIZE. Better to map smaller pages than nothing.
897 static int map_try_harder(struct vfio_domain *domain, dma_addr_t iova,
898 unsigned long pfn, long npage, int prot)
903 for (i = 0; i < npage; i++, pfn++, iova += PAGE_SIZE) {
904 ret = iommu_map(domain->domain, iova,
905 (phys_addr_t)pfn << PAGE_SHIFT,
906 PAGE_SIZE, prot | domain->prot);
911 for (; i < npage && i > 0; i--, iova -= PAGE_SIZE)
912 iommu_unmap(domain->domain, iova, PAGE_SIZE);
917 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
918 unsigned long pfn, long npage, int prot)
920 struct vfio_domain *d;
923 list_for_each_entry(d, &iommu->domain_list, next) {
924 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
925 npage << PAGE_SHIFT, prot | d->prot);
928 map_try_harder(d, iova, pfn, npage, prot))
938 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next)
939 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
944 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
947 dma_addr_t iova = dma->iova;
948 unsigned long vaddr = dma->vaddr;
949 size_t size = map_size;
955 /* Pin a contiguous chunk of memory */
956 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
957 size >> PAGE_SHIFT, &pfn);
965 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
968 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
973 size -= npage << PAGE_SHIFT;
974 dma->size += npage << PAGE_SHIFT;
977 dma->iommu_mapped = true;
980 vfio_remove_dma(iommu, dma);
985 static int vfio_dma_do_map(struct vfio_iommu *iommu,
986 struct vfio_iommu_type1_dma_map *map)
988 dma_addr_t iova = map->iova;
989 unsigned long vaddr = map->vaddr;
990 size_t size = map->size;
991 int ret = 0, prot = 0;
993 struct vfio_dma *dma;
995 /* Verify that none of our __u64 fields overflow */
996 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
999 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
1001 WARN_ON(mask & PAGE_MASK);
1003 /* READ/WRITE from device perspective */
1004 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1005 prot |= IOMMU_WRITE;
1006 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1009 if (!prot || !size || (size | iova | vaddr) & mask)
1012 /* Don't allow IOVA or virtual address wrap */
1013 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr)
1016 mutex_lock(&iommu->lock);
1018 if (vfio_find_dma(iommu, iova, size)) {
1023 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1032 get_task_struct(current);
1033 dma->task = current;
1034 dma->pfn_list = RB_ROOT;
1036 /* Insert zero-sized and grow as we map chunks of it */
1037 vfio_link_dma(iommu, dma);
1039 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1040 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1043 ret = vfio_pin_map_dma(iommu, dma, size);
1046 mutex_unlock(&iommu->lock);
1050 static int vfio_bus_type(struct device *dev, void *data)
1052 struct bus_type **bus = data;
1054 if (*bus && *bus != dev->bus)
1062 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1063 struct vfio_domain *domain)
1065 struct vfio_domain *d;
1069 /* Arbitrarily pick the first domain in the list for lookups */
1070 d = list_first_entry(&iommu->domain_list, struct vfio_domain, next);
1071 n = rb_first(&iommu->dma_list);
1073 for (; n; n = rb_next(n)) {
1074 struct vfio_dma *dma;
1077 dma = rb_entry(n, struct vfio_dma, node);
1080 while (iova < dma->iova + dma->size) {
1084 if (dma->iommu_mapped) {
1088 phys = iommu_iova_to_phys(d->domain, iova);
1090 if (WARN_ON(!phys)) {
1098 while (i < dma->iova + dma->size &&
1099 p == iommu_iova_to_phys(d->domain, i)) {
1106 unsigned long vaddr = dma->vaddr +
1108 size_t n = dma->iova + dma->size - iova;
1111 npage = vfio_pin_pages_remote(dma, vaddr,
1120 phys = pfn << PAGE_SHIFT;
1121 size = npage << PAGE_SHIFT;
1124 ret = iommu_map(domain->domain, iova, phys,
1125 size, dma->prot | domain->prot);
1131 dma->iommu_mapped = true;
1137 * We change our unmap behavior slightly depending on whether the IOMMU
1138 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1139 * for practically any contiguous power-of-two mapping we give it. This means
1140 * we don't need to look for contiguous chunks ourselves to make unmapping
1141 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1142 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1143 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1144 * hugetlbfs is in use.
1146 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1149 int ret, order = get_order(PAGE_SIZE * 2);
1151 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1155 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1156 IOMMU_READ | IOMMU_WRITE | domain->prot);
1158 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1160 if (unmapped == PAGE_SIZE)
1161 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1163 domain->fgsp = true;
1166 __free_pages(pages, order);
1169 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1170 struct iommu_group *iommu_group)
1172 struct vfio_group *g;
1174 list_for_each_entry(g, &domain->group_list, next) {
1175 if (g->iommu_group == iommu_group)
1182 static int vfio_iommu_type1_attach_group(void *iommu_data,
1183 struct iommu_group *iommu_group)
1185 struct vfio_iommu *iommu = iommu_data;
1186 struct vfio_group *group;
1187 struct vfio_domain *domain, *d;
1188 struct bus_type *bus = NULL, *mdev_bus;
1191 mutex_lock(&iommu->lock);
1193 list_for_each_entry(d, &iommu->domain_list, next) {
1194 if (find_iommu_group(d, iommu_group)) {
1195 mutex_unlock(&iommu->lock);
1200 if (iommu->external_domain) {
1201 if (find_iommu_group(iommu->external_domain, iommu_group)) {
1202 mutex_unlock(&iommu->lock);
1207 group = kzalloc(sizeof(*group), GFP_KERNEL);
1208 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
1209 if (!group || !domain) {
1214 group->iommu_group = iommu_group;
1216 /* Determine bus_type in order to allocate a domain */
1217 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
1221 mdev_bus = symbol_get(mdev_bus_type);
1224 if ((bus == mdev_bus) && !iommu_present(bus)) {
1225 symbol_put(mdev_bus_type);
1226 if (!iommu->external_domain) {
1227 INIT_LIST_HEAD(&domain->group_list);
1228 iommu->external_domain = domain;
1232 list_add(&group->next,
1233 &iommu->external_domain->group_list);
1234 mutex_unlock(&iommu->lock);
1237 symbol_put(mdev_bus_type);
1240 domain->domain = iommu_domain_alloc(bus);
1241 if (!domain->domain) {
1246 if (iommu->nesting) {
1249 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
1255 ret = iommu_attach_group(domain->domain, iommu_group);
1259 INIT_LIST_HEAD(&domain->group_list);
1260 list_add(&group->next, &domain->group_list);
1262 if (!allow_unsafe_interrupts &&
1263 !iommu_capable(bus, IOMMU_CAP_INTR_REMAP)) {
1264 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
1270 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
1271 domain->prot |= IOMMU_CACHE;
1274 * Try to match an existing compatible domain. We don't want to
1275 * preclude an IOMMU driver supporting multiple bus_types and being
1276 * able to include different bus_types in the same IOMMU domain, so
1277 * we test whether the domains use the same iommu_ops rather than
1278 * testing if they're on the same bus_type.
1280 list_for_each_entry(d, &iommu->domain_list, next) {
1281 if (d->domain->ops == domain->domain->ops &&
1282 d->prot == domain->prot) {
1283 iommu_detach_group(domain->domain, iommu_group);
1284 if (!iommu_attach_group(d->domain, iommu_group)) {
1285 list_add(&group->next, &d->group_list);
1286 iommu_domain_free(domain->domain);
1288 mutex_unlock(&iommu->lock);
1292 ret = iommu_attach_group(domain->domain, iommu_group);
1298 vfio_test_domain_fgsp(domain);
1300 /* replay mappings on new domains */
1301 ret = vfio_iommu_replay(iommu, domain);
1305 list_add(&domain->next, &iommu->domain_list);
1307 mutex_unlock(&iommu->lock);
1312 iommu_detach_group(domain->domain, iommu_group);
1314 iommu_domain_free(domain->domain);
1318 mutex_unlock(&iommu->lock);
1322 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
1324 struct rb_node *node;
1326 while ((node = rb_first(&iommu->dma_list)))
1327 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
1330 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
1332 struct rb_node *n, *p;
1334 n = rb_first(&iommu->dma_list);
1335 for (; n; n = rb_next(n)) {
1336 struct vfio_dma *dma;
1337 long locked = 0, unlocked = 0;
1339 dma = rb_entry(n, struct vfio_dma, node);
1340 unlocked += vfio_unmap_unpin(iommu, dma, false);
1341 p = rb_first(&dma->pfn_list);
1342 for (; p; p = rb_next(p)) {
1343 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
1346 if (!is_invalid_reserved_pfn(vpfn->pfn))
1349 vfio_lock_acct(dma->task, locked - unlocked);
1353 static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu)
1357 n = rb_first(&iommu->dma_list);
1358 for (; n; n = rb_next(n)) {
1359 struct vfio_dma *dma;
1361 dma = rb_entry(n, struct vfio_dma, node);
1363 if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list)))
1366 /* mdev vendor driver must unregister notifier */
1367 WARN_ON(iommu->notifier.head);
1370 static void vfio_iommu_type1_detach_group(void *iommu_data,
1371 struct iommu_group *iommu_group)
1373 struct vfio_iommu *iommu = iommu_data;
1374 struct vfio_domain *domain;
1375 struct vfio_group *group;
1377 mutex_lock(&iommu->lock);
1379 if (iommu->external_domain) {
1380 group = find_iommu_group(iommu->external_domain, iommu_group);
1382 list_del(&group->next);
1385 if (list_empty(&iommu->external_domain->group_list)) {
1386 vfio_sanity_check_pfn_list(iommu);
1388 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1389 vfio_iommu_unmap_unpin_all(iommu);
1391 kfree(iommu->external_domain);
1392 iommu->external_domain = NULL;
1394 goto detach_group_done;
1398 list_for_each_entry(domain, &iommu->domain_list, next) {
1399 group = find_iommu_group(domain, iommu_group);
1403 iommu_detach_group(domain->domain, iommu_group);
1404 list_del(&group->next);
1407 * Group ownership provides privilege, if the group list is
1408 * empty, the domain goes away. If it's the last domain with
1409 * iommu and external domain doesn't exist, then all the
1410 * mappings go away too. If it's the last domain with iommu and
1411 * external domain exist, update accounting
1413 if (list_empty(&domain->group_list)) {
1414 if (list_is_singular(&iommu->domain_list)) {
1415 if (!iommu->external_domain)
1416 vfio_iommu_unmap_unpin_all(iommu);
1418 vfio_iommu_unmap_unpin_reaccount(iommu);
1420 iommu_domain_free(domain->domain);
1421 list_del(&domain->next);
1428 mutex_unlock(&iommu->lock);
1431 static void *vfio_iommu_type1_open(unsigned long arg)
1433 struct vfio_iommu *iommu;
1435 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
1437 return ERR_PTR(-ENOMEM);
1440 case VFIO_TYPE1_IOMMU:
1442 case VFIO_TYPE1_NESTING_IOMMU:
1443 iommu->nesting = true;
1444 case VFIO_TYPE1v2_IOMMU:
1449 return ERR_PTR(-EINVAL);
1452 INIT_LIST_HEAD(&iommu->domain_list);
1453 iommu->dma_list = RB_ROOT;
1454 mutex_init(&iommu->lock);
1455 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
1460 static void vfio_release_domain(struct vfio_domain *domain, bool external)
1462 struct vfio_group *group, *group_tmp;
1464 list_for_each_entry_safe(group, group_tmp,
1465 &domain->group_list, next) {
1467 iommu_detach_group(domain->domain, group->iommu_group);
1468 list_del(&group->next);
1473 iommu_domain_free(domain->domain);
1476 static void vfio_iommu_type1_release(void *iommu_data)
1478 struct vfio_iommu *iommu = iommu_data;
1479 struct vfio_domain *domain, *domain_tmp;
1481 if (iommu->external_domain) {
1482 vfio_release_domain(iommu->external_domain, true);
1483 vfio_sanity_check_pfn_list(iommu);
1484 kfree(iommu->external_domain);
1487 vfio_iommu_unmap_unpin_all(iommu);
1489 list_for_each_entry_safe(domain, domain_tmp,
1490 &iommu->domain_list, next) {
1491 vfio_release_domain(domain, false);
1492 list_del(&domain->next);
1498 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
1500 struct vfio_domain *domain;
1503 mutex_lock(&iommu->lock);
1504 list_for_each_entry(domain, &iommu->domain_list, next) {
1505 if (!(domain->prot & IOMMU_CACHE)) {
1510 mutex_unlock(&iommu->lock);
1515 static long vfio_iommu_type1_ioctl(void *iommu_data,
1516 unsigned int cmd, unsigned long arg)
1518 struct vfio_iommu *iommu = iommu_data;
1519 unsigned long minsz;
1521 if (cmd == VFIO_CHECK_EXTENSION) {
1523 case VFIO_TYPE1_IOMMU:
1524 case VFIO_TYPE1v2_IOMMU:
1525 case VFIO_TYPE1_NESTING_IOMMU:
1527 case VFIO_DMA_CC_IOMMU:
1530 return vfio_domains_have_iommu_cache(iommu);
1534 } else if (cmd == VFIO_IOMMU_GET_INFO) {
1535 struct vfio_iommu_type1_info info;
1537 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
1539 if (copy_from_user(&info, (void __user *)arg, minsz))
1542 if (info.argsz < minsz)
1545 info.flags = VFIO_IOMMU_INFO_PGSIZES;
1547 info.iova_pgsizes = vfio_pgsize_bitmap(iommu);
1549 return copy_to_user((void __user *)arg, &info, minsz) ?
1552 } else if (cmd == VFIO_IOMMU_MAP_DMA) {
1553 struct vfio_iommu_type1_dma_map map;
1554 uint32_t mask = VFIO_DMA_MAP_FLAG_READ |
1555 VFIO_DMA_MAP_FLAG_WRITE;
1557 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
1559 if (copy_from_user(&map, (void __user *)arg, minsz))
1562 if (map.argsz < minsz || map.flags & ~mask)
1565 return vfio_dma_do_map(iommu, &map);
1567 } else if (cmd == VFIO_IOMMU_UNMAP_DMA) {
1568 struct vfio_iommu_type1_dma_unmap unmap;
1571 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
1573 if (copy_from_user(&unmap, (void __user *)arg, minsz))
1576 if (unmap.argsz < minsz || unmap.flags)
1579 ret = vfio_dma_do_unmap(iommu, &unmap);
1583 return copy_to_user((void __user *)arg, &unmap, minsz) ?
1590 static int vfio_iommu_type1_register_notifier(void *iommu_data,
1591 unsigned long *events,
1592 struct notifier_block *nb)
1594 struct vfio_iommu *iommu = iommu_data;
1596 /* clear known events */
1597 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
1599 /* refuse to register if still events remaining */
1603 return blocking_notifier_chain_register(&iommu->notifier, nb);
1606 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
1607 struct notifier_block *nb)
1609 struct vfio_iommu *iommu = iommu_data;
1611 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
1614 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
1615 .name = "vfio-iommu-type1",
1616 .owner = THIS_MODULE,
1617 .open = vfio_iommu_type1_open,
1618 .release = vfio_iommu_type1_release,
1619 .ioctl = vfio_iommu_type1_ioctl,
1620 .attach_group = vfio_iommu_type1_attach_group,
1621 .detach_group = vfio_iommu_type1_detach_group,
1622 .pin_pages = vfio_iommu_type1_pin_pages,
1623 .unpin_pages = vfio_iommu_type1_unpin_pages,
1624 .register_notifier = vfio_iommu_type1_register_notifier,
1625 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
1628 static int __init vfio_iommu_type1_init(void)
1630 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
1633 static void __exit vfio_iommu_type1_cleanup(void)
1635 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
1638 module_init(vfio_iommu_type1_init);
1639 module_exit(vfio_iommu_type1_cleanup);
1641 MODULE_VERSION(DRIVER_VERSION);
1642 MODULE_LICENSE("GPL v2");
1643 MODULE_AUTHOR(DRIVER_AUTHOR);
1644 MODULE_DESCRIPTION(DRIVER_DESC);