2 * VFIO: IOMMU DMA mapping support for TCE on POWER
4 * Copyright (C) 2013 IBM Corp. All rights reserved.
5 * Author: Alexey Kardashevskiy <aik@ozlabs.ru>
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_iommu_type1.c:
12 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
13 * Author: Alex Williamson <alex.williamson@redhat.com>
16 #include <linux/module.h>
17 #include <linux/pci.h>
18 #include <linux/slab.h>
19 #include <linux/uaccess.h>
20 #include <linux/err.h>
21 #include <linux/vfio.h>
22 #include <linux/vmalloc.h>
23 #include <linux/sched/mm.h>
24 #include <linux/sched/signal.h>
26 #include <asm/iommu.h>
28 #include <asm/mmu_context.h>
30 #define DRIVER_VERSION "0.1"
31 #define DRIVER_AUTHOR "aik@ozlabs.ru"
32 #define DRIVER_DESC "VFIO IOMMU SPAPR TCE"
34 static void tce_iommu_detach_group(void *iommu_data,
35 struct iommu_group *iommu_group);
37 static long try_increment_locked_vm(struct mm_struct *mm, long npages)
39 long ret = 0, locked, lock_limit;
41 if (WARN_ON_ONCE(!mm))
47 down_write(&mm->mmap_sem);
48 locked = mm->locked_vm + npages;
49 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
50 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
53 mm->locked_vm += npages;
55 pr_debug("[%d] RLIMIT_MEMLOCK +%ld %ld/%ld%s\n", current->pid,
57 mm->locked_vm << PAGE_SHIFT,
58 rlimit(RLIMIT_MEMLOCK),
59 ret ? " - exceeded" : "");
61 up_write(&mm->mmap_sem);
66 static void decrement_locked_vm(struct mm_struct *mm, long npages)
71 down_write(&mm->mmap_sem);
72 if (WARN_ON_ONCE(npages > mm->locked_vm))
73 npages = mm->locked_vm;
74 mm->locked_vm -= npages;
75 pr_debug("[%d] RLIMIT_MEMLOCK -%ld %ld/%ld\n", current->pid,
77 mm->locked_vm << PAGE_SHIFT,
78 rlimit(RLIMIT_MEMLOCK));
79 up_write(&mm->mmap_sem);
83 * VFIO IOMMU fd for SPAPR_TCE IOMMU implementation
85 * This code handles mapping and unmapping of user data buffers
86 * into DMA'ble space using the IOMMU
89 struct tce_iommu_group {
90 struct list_head next;
91 struct iommu_group *grp;
95 * A container needs to remember which preregistered region it has
96 * referenced to do proper cleanup at the userspace process exit.
98 struct tce_iommu_prereg {
99 struct list_head next;
100 struct mm_iommu_table_group_mem_t *mem;
104 * The container descriptor supports only a single group per container.
105 * Required by the API as the container is not supplied with the IOMMU group
106 * at the moment of initialization.
108 struct tce_container {
112 bool def_window_pending;
113 unsigned long locked_pages;
114 struct mm_struct *mm;
115 struct iommu_table *tables[IOMMU_TABLE_GROUP_MAX_TABLES];
116 struct list_head group_list;
117 struct list_head prereg_list;
120 static long tce_iommu_mm_set(struct tce_container *container)
123 if (container->mm == current->mm)
127 BUG_ON(!current->mm);
128 container->mm = current->mm;
129 atomic_inc(&container->mm->mm_count);
134 static long tce_iommu_prereg_free(struct tce_container *container,
135 struct tce_iommu_prereg *tcemem)
139 ret = mm_iommu_put(container->mm, tcemem->mem);
143 list_del(&tcemem->next);
149 static long tce_iommu_unregister_pages(struct tce_container *container,
150 __u64 vaddr, __u64 size)
152 struct mm_iommu_table_group_mem_t *mem;
153 struct tce_iommu_prereg *tcemem;
156 if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK))
159 mem = mm_iommu_find(container->mm, vaddr, size >> PAGE_SHIFT);
163 list_for_each_entry(tcemem, &container->prereg_list, next) {
164 if (tcemem->mem == mem) {
173 return tce_iommu_prereg_free(container, tcemem);
176 static long tce_iommu_register_pages(struct tce_container *container,
177 __u64 vaddr, __u64 size)
180 struct mm_iommu_table_group_mem_t *mem = NULL;
181 struct tce_iommu_prereg *tcemem;
182 unsigned long entries = size >> PAGE_SHIFT;
184 if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK) ||
185 ((vaddr + size) < vaddr))
188 mem = mm_iommu_find(container->mm, vaddr, entries);
190 list_for_each_entry(tcemem, &container->prereg_list, next) {
191 if (tcemem->mem == mem)
196 ret = mm_iommu_get(container->mm, vaddr, entries, &mem);
200 tcemem = kzalloc(sizeof(*tcemem), GFP_KERNEL);
202 mm_iommu_put(container->mm, mem);
207 list_add(&tcemem->next, &container->prereg_list);
209 container->enabled = true;
214 static long tce_iommu_userspace_view_alloc(struct iommu_table *tbl,
215 struct mm_struct *mm)
217 unsigned long cb = _ALIGN_UP(sizeof(tbl->it_userspace[0]) *
218 tbl->it_size, PAGE_SIZE);
222 BUG_ON(tbl->it_userspace);
224 ret = try_increment_locked_vm(mm, cb >> PAGE_SHIFT);
230 decrement_locked_vm(mm, cb >> PAGE_SHIFT);
233 tbl->it_userspace = uas;
238 static void tce_iommu_userspace_view_free(struct iommu_table *tbl,
239 struct mm_struct *mm)
241 unsigned long cb = _ALIGN_UP(sizeof(tbl->it_userspace[0]) *
242 tbl->it_size, PAGE_SIZE);
244 if (!tbl->it_userspace)
247 vfree(tbl->it_userspace);
248 tbl->it_userspace = NULL;
249 decrement_locked_vm(mm, cb >> PAGE_SHIFT);
252 static bool tce_page_is_contained(struct page *page, unsigned page_shift)
255 * Check that the TCE table granularity is not bigger than the size of
256 * a page we just found. Otherwise the hardware can get access to
257 * a bigger memory chunk that it should.
259 return (PAGE_SHIFT + compound_order(compound_head(page))) >= page_shift;
262 static inline bool tce_groups_attached(struct tce_container *container)
264 return !list_empty(&container->group_list);
267 static long tce_iommu_find_table(struct tce_container *container,
268 phys_addr_t ioba, struct iommu_table **ptbl)
272 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
273 struct iommu_table *tbl = container->tables[i];
276 unsigned long entry = ioba >> tbl->it_page_shift;
277 unsigned long start = tbl->it_offset;
278 unsigned long end = start + tbl->it_size;
280 if ((start <= entry) && (entry < end)) {
290 static int tce_iommu_find_free_table(struct tce_container *container)
294 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
295 if (!container->tables[i])
302 static int tce_iommu_enable(struct tce_container *container)
305 unsigned long locked;
306 struct iommu_table_group *table_group;
307 struct tce_iommu_group *tcegrp;
309 if (container->enabled)
313 * When userspace pages are mapped into the IOMMU, they are effectively
314 * locked memory, so, theoretically, we need to update the accounting
315 * of locked pages on each map and unmap. For powerpc, the map unmap
316 * paths can be very hot, though, and the accounting would kill
317 * performance, especially since it would be difficult to impossible
318 * to handle the accounting in real mode only.
320 * To address that, rather than precisely accounting every page, we
321 * instead account for a worst case on locked memory when the iommu is
322 * enabled and disabled. The worst case upper bound on locked memory
323 * is the size of the whole iommu window, which is usually relatively
324 * small (compared to total memory sizes) on POWER hardware.
326 * Also we don't have a nice way to fail on H_PUT_TCE due to ulimits,
327 * that would effectively kill the guest at random points, much better
328 * enforcing the limit based on the max that the guest can map.
330 * Unfortunately at the moment it counts whole tables, no matter how
331 * much memory the guest has. I.e. for 4GB guest and 4 IOMMU groups
332 * each with 2GB DMA window, 8GB will be counted here. The reason for
333 * this is that we cannot tell here the amount of RAM used by the guest
334 * as this information is only available from KVM and VFIO is
337 * So we do not allow enabling a container without a group attached
338 * as there is no way to know how much we should increment
339 * the locked_vm counter.
341 if (!tce_groups_attached(container))
344 tcegrp = list_first_entry(&container->group_list,
345 struct tce_iommu_group, next);
346 table_group = iommu_group_get_iommudata(tcegrp->grp);
350 if (!table_group->tce32_size)
353 ret = tce_iommu_mm_set(container);
357 locked = table_group->tce32_size >> PAGE_SHIFT;
358 ret = try_increment_locked_vm(container->mm, locked);
362 container->locked_pages = locked;
364 container->enabled = true;
369 static void tce_iommu_disable(struct tce_container *container)
371 if (!container->enabled)
374 container->enabled = false;
376 BUG_ON(!container->mm);
377 decrement_locked_vm(container->mm, container->locked_pages);
380 static void *tce_iommu_open(unsigned long arg)
382 struct tce_container *container;
384 if ((arg != VFIO_SPAPR_TCE_IOMMU) && (arg != VFIO_SPAPR_TCE_v2_IOMMU)) {
385 pr_err("tce_vfio: Wrong IOMMU type\n");
386 return ERR_PTR(-EINVAL);
389 container = kzalloc(sizeof(*container), GFP_KERNEL);
391 return ERR_PTR(-ENOMEM);
393 mutex_init(&container->lock);
394 INIT_LIST_HEAD_RCU(&container->group_list);
395 INIT_LIST_HEAD_RCU(&container->prereg_list);
397 container->v2 = arg == VFIO_SPAPR_TCE_v2_IOMMU;
402 static int tce_iommu_clear(struct tce_container *container,
403 struct iommu_table *tbl,
404 unsigned long entry, unsigned long pages);
405 static void tce_iommu_free_table(struct tce_container *container,
406 struct iommu_table *tbl);
408 static void tce_iommu_release(void *iommu_data)
410 struct tce_container *container = iommu_data;
411 struct tce_iommu_group *tcegrp;
414 while (tce_groups_attached(container)) {
415 tcegrp = list_first_entry(&container->group_list,
416 struct tce_iommu_group, next);
417 tce_iommu_detach_group(iommu_data, tcegrp->grp);
421 * If VFIO created a table, it was not disposed
422 * by tce_iommu_detach_group() so do it now.
424 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
425 struct iommu_table *tbl = container->tables[i];
430 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
431 tce_iommu_free_table(container, tbl);
434 while (!list_empty(&container->prereg_list)) {
435 struct tce_iommu_prereg *tcemem;
437 tcemem = list_first_entry(&container->prereg_list,
438 struct tce_iommu_prereg, next);
439 WARN_ON_ONCE(tce_iommu_prereg_free(container, tcemem));
442 tce_iommu_disable(container);
444 mmdrop(container->mm);
445 mutex_destroy(&container->lock);
450 static void tce_iommu_unuse_page(struct tce_container *container,
455 page = pfn_to_page(hpa >> PAGE_SHIFT);
459 static int tce_iommu_prereg_ua_to_hpa(struct tce_container *container,
460 unsigned long tce, unsigned long size,
461 unsigned long *phpa, struct mm_iommu_table_group_mem_t **pmem)
464 struct mm_iommu_table_group_mem_t *mem;
466 mem = mm_iommu_lookup(container->mm, tce, size);
470 ret = mm_iommu_ua_to_hpa(mem, tce, phpa);
479 static void tce_iommu_unuse_page_v2(struct tce_container *container,
480 struct iommu_table *tbl, unsigned long entry)
482 struct mm_iommu_table_group_mem_t *mem = NULL;
484 unsigned long hpa = 0;
485 unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry);
490 ret = tce_iommu_prereg_ua_to_hpa(container, *pua, IOMMU_PAGE_SIZE(tbl),
493 pr_debug("%s: tce %lx at #%lx was not cached, ret=%d\n",
494 __func__, *pua, entry, ret);
496 mm_iommu_mapped_dec(mem);
501 static int tce_iommu_clear(struct tce_container *container,
502 struct iommu_table *tbl,
503 unsigned long entry, unsigned long pages)
505 unsigned long oldhpa;
507 enum dma_data_direction direction;
509 for ( ; pages; --pages, ++entry) {
510 direction = DMA_NONE;
512 ret = iommu_tce_xchg(tbl, entry, &oldhpa, &direction);
516 if (direction == DMA_NONE)
520 tce_iommu_unuse_page_v2(container, tbl, entry);
524 tce_iommu_unuse_page(container, oldhpa);
530 static int tce_iommu_use_page(unsigned long tce, unsigned long *hpa)
532 struct page *page = NULL;
533 enum dma_data_direction direction = iommu_tce_direction(tce);
535 if (get_user_pages_fast(tce & PAGE_MASK, 1,
536 direction != DMA_TO_DEVICE, &page) != 1)
539 *hpa = __pa((unsigned long) page_address(page));
544 static long tce_iommu_build(struct tce_container *container,
545 struct iommu_table *tbl,
546 unsigned long entry, unsigned long tce, unsigned long pages,
547 enum dma_data_direction direction)
552 enum dma_data_direction dirtmp;
554 for (i = 0; i < pages; ++i) {
555 unsigned long offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
557 ret = tce_iommu_use_page(tce, &hpa);
561 page = pfn_to_page(hpa >> PAGE_SHIFT);
562 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
569 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
571 tce_iommu_unuse_page(container, hpa);
572 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
573 __func__, entry << tbl->it_page_shift,
578 if (dirtmp != DMA_NONE)
579 tce_iommu_unuse_page(container, hpa);
581 tce += IOMMU_PAGE_SIZE(tbl);
585 tce_iommu_clear(container, tbl, entry, i);
590 static long tce_iommu_build_v2(struct tce_container *container,
591 struct iommu_table *tbl,
592 unsigned long entry, unsigned long tce, unsigned long pages,
593 enum dma_data_direction direction)
598 enum dma_data_direction dirtmp;
600 if (!tbl->it_userspace) {
601 ret = tce_iommu_userspace_view_alloc(tbl, container->mm);
606 for (i = 0; i < pages; ++i) {
607 struct mm_iommu_table_group_mem_t *mem = NULL;
608 unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl,
611 ret = tce_iommu_prereg_ua_to_hpa(container,
612 tce, IOMMU_PAGE_SIZE(tbl), &hpa, &mem);
616 page = pfn_to_page(hpa >> PAGE_SHIFT);
617 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
622 /* Preserve offset within IOMMU page */
623 hpa |= tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
626 /* The registered region is being unregistered */
627 if (mm_iommu_mapped_inc(mem))
630 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
632 /* dirtmp cannot be DMA_NONE here */
633 tce_iommu_unuse_page_v2(container, tbl, entry + i);
634 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
635 __func__, entry << tbl->it_page_shift,
640 if (dirtmp != DMA_NONE)
641 tce_iommu_unuse_page_v2(container, tbl, entry + i);
645 tce += IOMMU_PAGE_SIZE(tbl);
649 tce_iommu_clear(container, tbl, entry, i);
654 static long tce_iommu_create_table(struct tce_container *container,
655 struct iommu_table_group *table_group,
660 struct iommu_table **ptbl)
662 long ret, table_size;
664 table_size = table_group->ops->get_table_size(page_shift, window_size,
669 ret = try_increment_locked_vm(container->mm, table_size >> PAGE_SHIFT);
673 ret = table_group->ops->create_table(table_group, num,
674 page_shift, window_size, levels, ptbl);
676 WARN_ON(!ret && !(*ptbl)->it_ops->free);
677 WARN_ON(!ret && ((*ptbl)->it_allocated_size != table_size));
682 static void tce_iommu_free_table(struct tce_container *container,
683 struct iommu_table *tbl)
685 unsigned long pages = tbl->it_allocated_size >> PAGE_SHIFT;
687 tce_iommu_userspace_view_free(tbl, container->mm);
688 tbl->it_ops->free(tbl);
689 decrement_locked_vm(container->mm, pages);
692 static long tce_iommu_create_window(struct tce_container *container,
693 __u32 page_shift, __u64 window_size, __u32 levels,
696 struct tce_iommu_group *tcegrp;
697 struct iommu_table_group *table_group;
698 struct iommu_table *tbl = NULL;
701 num = tce_iommu_find_free_table(container);
705 /* Get the first group for ops::create_table */
706 tcegrp = list_first_entry(&container->group_list,
707 struct tce_iommu_group, next);
708 table_group = iommu_group_get_iommudata(tcegrp->grp);
712 if (!(table_group->pgsizes & (1ULL << page_shift)))
715 if (!table_group->ops->set_window || !table_group->ops->unset_window ||
716 !table_group->ops->get_table_size ||
717 !table_group->ops->create_table)
720 /* Create TCE table */
721 ret = tce_iommu_create_table(container, table_group, num,
722 page_shift, window_size, levels, &tbl);
726 BUG_ON(!tbl->it_ops->free);
729 * Program the table to every group.
730 * Groups have been tested for compatibility at the attach time.
732 list_for_each_entry(tcegrp, &container->group_list, next) {
733 table_group = iommu_group_get_iommudata(tcegrp->grp);
735 ret = table_group->ops->set_window(table_group, num, tbl);
740 container->tables[num] = tbl;
742 /* Return start address assigned by platform in create_table() */
743 *start_addr = tbl->it_offset << tbl->it_page_shift;
748 list_for_each_entry(tcegrp, &container->group_list, next) {
749 table_group = iommu_group_get_iommudata(tcegrp->grp);
750 table_group->ops->unset_window(table_group, num);
752 tce_iommu_free_table(container, tbl);
757 static long tce_iommu_remove_window(struct tce_container *container,
760 struct iommu_table_group *table_group = NULL;
761 struct iommu_table *tbl;
762 struct tce_iommu_group *tcegrp;
765 num = tce_iommu_find_table(container, start_addr, &tbl);
769 BUG_ON(!tbl->it_size);
771 /* Detach groups from IOMMUs */
772 list_for_each_entry(tcegrp, &container->group_list, next) {
773 table_group = iommu_group_get_iommudata(tcegrp->grp);
776 * SPAPR TCE IOMMU exposes the default DMA window to
777 * the guest via dma32_window_start/size of
778 * VFIO_IOMMU_SPAPR_TCE_GET_INFO. Some platforms allow
779 * the userspace to remove this window, some do not so
780 * here we check for the platform capability.
782 if (!table_group->ops || !table_group->ops->unset_window)
785 table_group->ops->unset_window(table_group, num);
789 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
790 tce_iommu_free_table(container, tbl);
791 container->tables[num] = NULL;
796 static long tce_iommu_create_default_window(struct tce_container *container)
799 __u64 start_addr = 0;
800 struct tce_iommu_group *tcegrp;
801 struct iommu_table_group *table_group;
803 if (!container->def_window_pending)
806 if (!tce_groups_attached(container))
809 tcegrp = list_first_entry(&container->group_list,
810 struct tce_iommu_group, next);
811 table_group = iommu_group_get_iommudata(tcegrp->grp);
815 ret = tce_iommu_create_window(container, IOMMU_PAGE_SHIFT_4K,
816 table_group->tce32_size, 1, &start_addr);
817 WARN_ON_ONCE(!ret && start_addr);
820 container->def_window_pending = false;
825 static long tce_iommu_ioctl(void *iommu_data,
826 unsigned int cmd, unsigned long arg)
828 struct tce_container *container = iommu_data;
829 unsigned long minsz, ddwsz;
833 case VFIO_CHECK_EXTENSION:
835 case VFIO_SPAPR_TCE_IOMMU:
836 case VFIO_SPAPR_TCE_v2_IOMMU:
840 ret = vfio_spapr_iommu_eeh_ioctl(NULL, cmd, arg);
844 return (ret < 0) ? 0 : ret;
848 * Sanity check to prevent one userspace from manipulating
849 * another userspace mm.
852 if (container->mm && container->mm != current->mm)
856 case VFIO_IOMMU_SPAPR_TCE_GET_INFO: {
857 struct vfio_iommu_spapr_tce_info info;
858 struct tce_iommu_group *tcegrp;
859 struct iommu_table_group *table_group;
861 if (!tce_groups_attached(container))
864 tcegrp = list_first_entry(&container->group_list,
865 struct tce_iommu_group, next);
866 table_group = iommu_group_get_iommudata(tcegrp->grp);
871 minsz = offsetofend(struct vfio_iommu_spapr_tce_info,
874 if (copy_from_user(&info, (void __user *)arg, minsz))
877 if (info.argsz < minsz)
880 info.dma32_window_start = table_group->tce32_start;
881 info.dma32_window_size = table_group->tce32_size;
883 memset(&info.ddw, 0, sizeof(info.ddw));
885 if (table_group->max_dynamic_windows_supported &&
887 info.flags |= VFIO_IOMMU_SPAPR_INFO_DDW;
888 info.ddw.pgsizes = table_group->pgsizes;
889 info.ddw.max_dynamic_windows_supported =
890 table_group->max_dynamic_windows_supported;
891 info.ddw.levels = table_group->max_levels;
894 ddwsz = offsetofend(struct vfio_iommu_spapr_tce_info, ddw);
896 if (info.argsz >= ddwsz)
899 if (copy_to_user((void __user *)arg, &info, minsz))
904 case VFIO_IOMMU_MAP_DMA: {
905 struct vfio_iommu_type1_dma_map param;
906 struct iommu_table *tbl = NULL;
908 enum dma_data_direction direction;
910 if (!container->enabled)
913 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
915 if (copy_from_user(¶m, (void __user *)arg, minsz))
918 if (param.argsz < minsz)
921 if (param.flags & ~(VFIO_DMA_MAP_FLAG_READ |
922 VFIO_DMA_MAP_FLAG_WRITE))
925 ret = tce_iommu_create_default_window(container);
929 num = tce_iommu_find_table(container, param.iova, &tbl);
933 if ((param.size & ~IOMMU_PAGE_MASK(tbl)) ||
934 (param.vaddr & ~IOMMU_PAGE_MASK(tbl)))
937 /* iova is checked by the IOMMU API */
938 if (param.flags & VFIO_DMA_MAP_FLAG_READ) {
939 if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
940 direction = DMA_BIDIRECTIONAL;
942 direction = DMA_TO_DEVICE;
944 if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
945 direction = DMA_FROM_DEVICE;
950 ret = iommu_tce_put_param_check(tbl, param.iova, param.vaddr);
955 ret = tce_iommu_build_v2(container, tbl,
956 param.iova >> tbl->it_page_shift,
958 param.size >> tbl->it_page_shift,
961 ret = tce_iommu_build(container, tbl,
962 param.iova >> tbl->it_page_shift,
964 param.size >> tbl->it_page_shift,
967 iommu_flush_tce(tbl);
971 case VFIO_IOMMU_UNMAP_DMA: {
972 struct vfio_iommu_type1_dma_unmap param;
973 struct iommu_table *tbl = NULL;
976 if (!container->enabled)
979 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap,
982 if (copy_from_user(¶m, (void __user *)arg, minsz))
985 if (param.argsz < minsz)
988 /* No flag is supported now */
992 ret = tce_iommu_create_default_window(container);
996 num = tce_iommu_find_table(container, param.iova, &tbl);
1000 if (param.size & ~IOMMU_PAGE_MASK(tbl))
1003 ret = iommu_tce_clear_param_check(tbl, param.iova, 0,
1004 param.size >> tbl->it_page_shift);
1008 ret = tce_iommu_clear(container, tbl,
1009 param.iova >> tbl->it_page_shift,
1010 param.size >> tbl->it_page_shift);
1011 iommu_flush_tce(tbl);
1015 case VFIO_IOMMU_SPAPR_REGISTER_MEMORY: {
1016 struct vfio_iommu_spapr_register_memory param;
1021 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1024 ret = tce_iommu_mm_set(container);
1028 if (copy_from_user(¶m, (void __user *)arg, minsz))
1031 if (param.argsz < minsz)
1034 /* No flag is supported now */
1038 mutex_lock(&container->lock);
1039 ret = tce_iommu_register_pages(container, param.vaddr,
1041 mutex_unlock(&container->lock);
1045 case VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY: {
1046 struct vfio_iommu_spapr_register_memory param;
1054 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1057 if (copy_from_user(¶m, (void __user *)arg, minsz))
1060 if (param.argsz < minsz)
1063 /* No flag is supported now */
1067 mutex_lock(&container->lock);
1068 ret = tce_iommu_unregister_pages(container, param.vaddr,
1070 mutex_unlock(&container->lock);
1074 case VFIO_IOMMU_ENABLE:
1078 mutex_lock(&container->lock);
1079 ret = tce_iommu_enable(container);
1080 mutex_unlock(&container->lock);
1084 case VFIO_IOMMU_DISABLE:
1088 mutex_lock(&container->lock);
1089 tce_iommu_disable(container);
1090 mutex_unlock(&container->lock);
1093 case VFIO_EEH_PE_OP: {
1094 struct tce_iommu_group *tcegrp;
1097 list_for_each_entry(tcegrp, &container->group_list, next) {
1098 ret = vfio_spapr_iommu_eeh_ioctl(tcegrp->grp,
1106 case VFIO_IOMMU_SPAPR_TCE_CREATE: {
1107 struct vfio_iommu_spapr_tce_create create;
1112 ret = tce_iommu_mm_set(container);
1116 if (!tce_groups_attached(container))
1119 minsz = offsetofend(struct vfio_iommu_spapr_tce_create,
1122 if (copy_from_user(&create, (void __user *)arg, minsz))
1125 if (create.argsz < minsz)
1131 mutex_lock(&container->lock);
1133 ret = tce_iommu_create_default_window(container);
1135 ret = tce_iommu_create_window(container,
1137 create.window_size, create.levels,
1138 &create.start_addr);
1140 mutex_unlock(&container->lock);
1142 if (!ret && copy_to_user((void __user *)arg, &create, minsz))
1147 case VFIO_IOMMU_SPAPR_TCE_REMOVE: {
1148 struct vfio_iommu_spapr_tce_remove remove;
1153 ret = tce_iommu_mm_set(container);
1157 if (!tce_groups_attached(container))
1160 minsz = offsetofend(struct vfio_iommu_spapr_tce_remove,
1163 if (copy_from_user(&remove, (void __user *)arg, minsz))
1166 if (remove.argsz < minsz)
1172 if (container->def_window_pending && !remove.start_addr) {
1173 container->def_window_pending = false;
1177 mutex_lock(&container->lock);
1179 ret = tce_iommu_remove_window(container, remove.start_addr);
1181 mutex_unlock(&container->lock);
1190 static void tce_iommu_release_ownership(struct tce_container *container,
1191 struct iommu_table_group *table_group)
1195 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1196 struct iommu_table *tbl = container->tables[i];
1201 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
1202 tce_iommu_userspace_view_free(tbl, container->mm);
1204 iommu_release_ownership(tbl);
1206 container->tables[i] = NULL;
1210 static int tce_iommu_take_ownership(struct tce_container *container,
1211 struct iommu_table_group *table_group)
1215 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1216 struct iommu_table *tbl = table_group->tables[i];
1218 if (!tbl || !tbl->it_map)
1221 rc = iommu_take_ownership(tbl);
1223 for (j = 0; j < i; ++j)
1224 iommu_release_ownership(
1225 table_group->tables[j]);
1231 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1232 container->tables[i] = table_group->tables[i];
1237 static void tce_iommu_release_ownership_ddw(struct tce_container *container,
1238 struct iommu_table_group *table_group)
1242 if (!table_group->ops->unset_window) {
1247 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1248 table_group->ops->unset_window(table_group, i);
1250 table_group->ops->release_ownership(table_group);
1253 static long tce_iommu_take_ownership_ddw(struct tce_container *container,
1254 struct iommu_table_group *table_group)
1258 if (!table_group->ops->create_table || !table_group->ops->set_window ||
1259 !table_group->ops->release_ownership) {
1264 table_group->ops->take_ownership(table_group);
1266 /* Set all windows to the new group */
1267 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1268 struct iommu_table *tbl = container->tables[i];
1273 ret = table_group->ops->set_window(table_group, i, tbl);
1281 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1282 table_group->ops->unset_window(table_group, i);
1284 table_group->ops->release_ownership(table_group);
1289 static int tce_iommu_attach_group(void *iommu_data,
1290 struct iommu_group *iommu_group)
1293 struct tce_container *container = iommu_data;
1294 struct iommu_table_group *table_group;
1295 struct tce_iommu_group *tcegrp = NULL;
1297 mutex_lock(&container->lock);
1299 /* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
1300 iommu_group_id(iommu_group), iommu_group); */
1301 table_group = iommu_group_get_iommudata(iommu_group);
1307 if (tce_groups_attached(container) && (!table_group->ops ||
1308 !table_group->ops->take_ownership ||
1309 !table_group->ops->release_ownership)) {
1314 /* Check if new group has the same iommu_ops (i.e. compatible) */
1315 list_for_each_entry(tcegrp, &container->group_list, next) {
1316 struct iommu_table_group *table_group_tmp;
1318 if (tcegrp->grp == iommu_group) {
1319 pr_warn("tce_vfio: Group %d is already attached\n",
1320 iommu_group_id(iommu_group));
1324 table_group_tmp = iommu_group_get_iommudata(tcegrp->grp);
1325 if (table_group_tmp->ops->create_table !=
1326 table_group->ops->create_table) {
1327 pr_warn("tce_vfio: Group %d is incompatible with group %d\n",
1328 iommu_group_id(iommu_group),
1329 iommu_group_id(tcegrp->grp));
1335 tcegrp = kzalloc(sizeof(*tcegrp), GFP_KERNEL);
1341 if (!table_group->ops || !table_group->ops->take_ownership ||
1342 !table_group->ops->release_ownership) {
1343 if (container->v2) {
1347 ret = tce_iommu_take_ownership(container, table_group);
1349 if (!container->v2) {
1353 ret = tce_iommu_take_ownership_ddw(container, table_group);
1354 if (!tce_groups_attached(container) && !container->tables[0])
1355 container->def_window_pending = true;
1359 tcegrp->grp = iommu_group;
1360 list_add(&tcegrp->next, &container->group_list);
1367 mutex_unlock(&container->lock);
1372 static void tce_iommu_detach_group(void *iommu_data,
1373 struct iommu_group *iommu_group)
1375 struct tce_container *container = iommu_data;
1376 struct iommu_table_group *table_group;
1378 struct tce_iommu_group *tcegrp;
1380 mutex_lock(&container->lock);
1382 list_for_each_entry(tcegrp, &container->group_list, next) {
1383 if (tcegrp->grp == iommu_group) {
1390 pr_warn("tce_vfio: detaching unattached group #%u\n",
1391 iommu_group_id(iommu_group));
1395 list_del(&tcegrp->next);
1398 table_group = iommu_group_get_iommudata(iommu_group);
1399 BUG_ON(!table_group);
1401 if (!table_group->ops || !table_group->ops->release_ownership)
1402 tce_iommu_release_ownership(container, table_group);
1404 tce_iommu_release_ownership_ddw(container, table_group);
1407 mutex_unlock(&container->lock);
1410 const struct vfio_iommu_driver_ops tce_iommu_driver_ops = {
1411 .name = "iommu-vfio-powerpc",
1412 .owner = THIS_MODULE,
1413 .open = tce_iommu_open,
1414 .release = tce_iommu_release,
1415 .ioctl = tce_iommu_ioctl,
1416 .attach_group = tce_iommu_attach_group,
1417 .detach_group = tce_iommu_detach_group,
1420 static int __init tce_iommu_init(void)
1422 return vfio_register_iommu_driver(&tce_iommu_driver_ops);
1425 static void __exit tce_iommu_cleanup(void)
1427 vfio_unregister_iommu_driver(&tce_iommu_driver_ops);
1430 module_init(tce_iommu_init);
1431 module_exit(tce_iommu_cleanup);
1433 MODULE_VERSION(DRIVER_VERSION);
1434 MODULE_LICENSE("GPL v2");
1435 MODULE_AUTHOR(DRIVER_AUTHOR);
1436 MODULE_DESCRIPTION(DRIVER_DESC);