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) {
512 direction = DMA_NONE;
514 ret = iommu_tce_xchg(tbl, entry, &oldhpa, &direction);
518 if (direction == DMA_NONE)
522 tce_iommu_unuse_page_v2(container, tbl, entry);
526 tce_iommu_unuse_page(container, oldhpa);
532 static int tce_iommu_use_page(unsigned long tce, unsigned long *hpa)
534 struct page *page = NULL;
535 enum dma_data_direction direction = iommu_tce_direction(tce);
537 if (get_user_pages_fast(tce & PAGE_MASK, 1,
538 direction != DMA_TO_DEVICE, &page) != 1)
541 *hpa = __pa((unsigned long) page_address(page));
546 static long tce_iommu_build(struct tce_container *container,
547 struct iommu_table *tbl,
548 unsigned long entry, unsigned long tce, unsigned long pages,
549 enum dma_data_direction direction)
554 enum dma_data_direction dirtmp;
556 for (i = 0; i < pages; ++i) {
557 unsigned long offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
559 ret = tce_iommu_use_page(tce, &hpa);
563 page = pfn_to_page(hpa >> PAGE_SHIFT);
564 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
571 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
573 tce_iommu_unuse_page(container, hpa);
574 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
575 __func__, entry << tbl->it_page_shift,
580 if (dirtmp != DMA_NONE)
581 tce_iommu_unuse_page(container, hpa);
583 tce += IOMMU_PAGE_SIZE(tbl);
587 tce_iommu_clear(container, tbl, entry, i);
592 static long tce_iommu_build_v2(struct tce_container *container,
593 struct iommu_table *tbl,
594 unsigned long entry, unsigned long tce, unsigned long pages,
595 enum dma_data_direction direction)
600 enum dma_data_direction dirtmp;
602 if (!tbl->it_userspace) {
603 ret = tce_iommu_userspace_view_alloc(tbl, container->mm);
608 for (i = 0; i < pages; ++i) {
609 struct mm_iommu_table_group_mem_t *mem = NULL;
610 unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl,
613 ret = tce_iommu_prereg_ua_to_hpa(container,
614 tce, IOMMU_PAGE_SIZE(tbl), &hpa, &mem);
618 page = pfn_to_page(hpa >> PAGE_SHIFT);
619 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
624 /* Preserve offset within IOMMU page */
625 hpa |= tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
628 /* The registered region is being unregistered */
629 if (mm_iommu_mapped_inc(mem))
632 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
634 /* dirtmp cannot be DMA_NONE here */
635 tce_iommu_unuse_page_v2(container, tbl, entry + i);
636 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
637 __func__, entry << tbl->it_page_shift,
642 if (dirtmp != DMA_NONE)
643 tce_iommu_unuse_page_v2(container, tbl, entry + i);
647 tce += IOMMU_PAGE_SIZE(tbl);
651 tce_iommu_clear(container, tbl, entry, i);
656 static long tce_iommu_create_table(struct tce_container *container,
657 struct iommu_table_group *table_group,
662 struct iommu_table **ptbl)
664 long ret, table_size;
666 table_size = table_group->ops->get_table_size(page_shift, window_size,
671 ret = try_increment_locked_vm(container->mm, table_size >> PAGE_SHIFT);
675 ret = table_group->ops->create_table(table_group, num,
676 page_shift, window_size, levels, ptbl);
678 WARN_ON(!ret && !(*ptbl)->it_ops->free);
679 WARN_ON(!ret && ((*ptbl)->it_allocated_size != table_size));
684 static void tce_iommu_free_table(struct tce_container *container,
685 struct iommu_table *tbl)
687 unsigned long pages = tbl->it_allocated_size >> PAGE_SHIFT;
689 tce_iommu_userspace_view_free(tbl, container->mm);
690 iommu_tce_table_put(tbl);
691 decrement_locked_vm(container->mm, pages);
694 static long tce_iommu_create_window(struct tce_container *container,
695 __u32 page_shift, __u64 window_size, __u32 levels,
698 struct tce_iommu_group *tcegrp;
699 struct iommu_table_group *table_group;
700 struct iommu_table *tbl = NULL;
703 num = tce_iommu_find_free_table(container);
707 /* Get the first group for ops::create_table */
708 tcegrp = list_first_entry(&container->group_list,
709 struct tce_iommu_group, next);
710 table_group = iommu_group_get_iommudata(tcegrp->grp);
714 if (!(table_group->pgsizes & (1ULL << page_shift)))
717 if (!table_group->ops->set_window || !table_group->ops->unset_window ||
718 !table_group->ops->get_table_size ||
719 !table_group->ops->create_table)
722 /* Create TCE table */
723 ret = tce_iommu_create_table(container, table_group, num,
724 page_shift, window_size, levels, &tbl);
728 BUG_ON(!tbl->it_ops->free);
731 * Program the table to every group.
732 * Groups have been tested for compatibility at the attach time.
734 list_for_each_entry(tcegrp, &container->group_list, next) {
735 table_group = iommu_group_get_iommudata(tcegrp->grp);
737 ret = table_group->ops->set_window(table_group, num, tbl);
742 container->tables[num] = tbl;
744 /* Return start address assigned by platform in create_table() */
745 *start_addr = tbl->it_offset << tbl->it_page_shift;
750 list_for_each_entry(tcegrp, &container->group_list, next) {
751 table_group = iommu_group_get_iommudata(tcegrp->grp);
752 table_group->ops->unset_window(table_group, num);
754 tce_iommu_free_table(container, tbl);
759 static long tce_iommu_remove_window(struct tce_container *container,
762 struct iommu_table_group *table_group = NULL;
763 struct iommu_table *tbl;
764 struct tce_iommu_group *tcegrp;
767 num = tce_iommu_find_table(container, start_addr, &tbl);
771 BUG_ON(!tbl->it_size);
773 /* Detach groups from IOMMUs */
774 list_for_each_entry(tcegrp, &container->group_list, next) {
775 table_group = iommu_group_get_iommudata(tcegrp->grp);
778 * SPAPR TCE IOMMU exposes the default DMA window to
779 * the guest via dma32_window_start/size of
780 * VFIO_IOMMU_SPAPR_TCE_GET_INFO. Some platforms allow
781 * the userspace to remove this window, some do not so
782 * here we check for the platform capability.
784 if (!table_group->ops || !table_group->ops->unset_window)
787 table_group->ops->unset_window(table_group, num);
791 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
792 tce_iommu_free_table(container, tbl);
793 container->tables[num] = NULL;
798 static long tce_iommu_create_default_window(struct tce_container *container)
801 __u64 start_addr = 0;
802 struct tce_iommu_group *tcegrp;
803 struct iommu_table_group *table_group;
805 if (!container->def_window_pending)
808 if (!tce_groups_attached(container))
811 tcegrp = list_first_entry(&container->group_list,
812 struct tce_iommu_group, next);
813 table_group = iommu_group_get_iommudata(tcegrp->grp);
817 ret = tce_iommu_create_window(container, IOMMU_PAGE_SHIFT_4K,
818 table_group->tce32_size, 1, &start_addr);
819 WARN_ON_ONCE(!ret && start_addr);
822 container->def_window_pending = false;
827 static long tce_iommu_ioctl(void *iommu_data,
828 unsigned int cmd, unsigned long arg)
830 struct tce_container *container = iommu_data;
831 unsigned long minsz, ddwsz;
835 case VFIO_CHECK_EXTENSION:
837 case VFIO_SPAPR_TCE_IOMMU:
838 case VFIO_SPAPR_TCE_v2_IOMMU:
842 ret = vfio_spapr_iommu_eeh_ioctl(NULL, cmd, arg);
846 return (ret < 0) ? 0 : ret;
850 * Sanity check to prevent one userspace from manipulating
851 * another userspace mm.
854 if (container->mm && container->mm != current->mm)
858 case VFIO_IOMMU_SPAPR_TCE_GET_INFO: {
859 struct vfio_iommu_spapr_tce_info info;
860 struct tce_iommu_group *tcegrp;
861 struct iommu_table_group *table_group;
863 if (!tce_groups_attached(container))
866 tcegrp = list_first_entry(&container->group_list,
867 struct tce_iommu_group, next);
868 table_group = iommu_group_get_iommudata(tcegrp->grp);
873 minsz = offsetofend(struct vfio_iommu_spapr_tce_info,
876 if (copy_from_user(&info, (void __user *)arg, minsz))
879 if (info.argsz < minsz)
882 info.dma32_window_start = table_group->tce32_start;
883 info.dma32_window_size = table_group->tce32_size;
885 memset(&info.ddw, 0, sizeof(info.ddw));
887 if (table_group->max_dynamic_windows_supported &&
889 info.flags |= VFIO_IOMMU_SPAPR_INFO_DDW;
890 info.ddw.pgsizes = table_group->pgsizes;
891 info.ddw.max_dynamic_windows_supported =
892 table_group->max_dynamic_windows_supported;
893 info.ddw.levels = table_group->max_levels;
896 ddwsz = offsetofend(struct vfio_iommu_spapr_tce_info, ddw);
898 if (info.argsz >= ddwsz)
901 if (copy_to_user((void __user *)arg, &info, minsz))
906 case VFIO_IOMMU_MAP_DMA: {
907 struct vfio_iommu_type1_dma_map param;
908 struct iommu_table *tbl = NULL;
910 enum dma_data_direction direction;
912 if (!container->enabled)
915 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
917 if (copy_from_user(¶m, (void __user *)arg, minsz))
920 if (param.argsz < minsz)
923 if (param.flags & ~(VFIO_DMA_MAP_FLAG_READ |
924 VFIO_DMA_MAP_FLAG_WRITE))
927 ret = tce_iommu_create_default_window(container);
931 num = tce_iommu_find_table(container, param.iova, &tbl);
935 if ((param.size & ~IOMMU_PAGE_MASK(tbl)) ||
936 (param.vaddr & ~IOMMU_PAGE_MASK(tbl)))
939 /* iova is checked by the IOMMU API */
940 if (param.flags & VFIO_DMA_MAP_FLAG_READ) {
941 if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
942 direction = DMA_BIDIRECTIONAL;
944 direction = DMA_TO_DEVICE;
946 if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
947 direction = DMA_FROM_DEVICE;
952 ret = iommu_tce_put_param_check(tbl, param.iova, param.vaddr);
957 ret = tce_iommu_build_v2(container, tbl,
958 param.iova >> tbl->it_page_shift,
960 param.size >> tbl->it_page_shift,
963 ret = tce_iommu_build(container, tbl,
964 param.iova >> tbl->it_page_shift,
966 param.size >> tbl->it_page_shift,
969 iommu_flush_tce(tbl);
973 case VFIO_IOMMU_UNMAP_DMA: {
974 struct vfio_iommu_type1_dma_unmap param;
975 struct iommu_table *tbl = NULL;
978 if (!container->enabled)
981 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap,
984 if (copy_from_user(¶m, (void __user *)arg, minsz))
987 if (param.argsz < minsz)
990 /* No flag is supported now */
994 ret = tce_iommu_create_default_window(container);
998 num = tce_iommu_find_table(container, param.iova, &tbl);
1002 if (param.size & ~IOMMU_PAGE_MASK(tbl))
1005 ret = iommu_tce_clear_param_check(tbl, param.iova, 0,
1006 param.size >> tbl->it_page_shift);
1010 ret = tce_iommu_clear(container, tbl,
1011 param.iova >> tbl->it_page_shift,
1012 param.size >> tbl->it_page_shift);
1013 iommu_flush_tce(tbl);
1017 case VFIO_IOMMU_SPAPR_REGISTER_MEMORY: {
1018 struct vfio_iommu_spapr_register_memory param;
1023 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1026 ret = tce_iommu_mm_set(container);
1030 if (copy_from_user(¶m, (void __user *)arg, minsz))
1033 if (param.argsz < minsz)
1036 /* No flag is supported now */
1040 mutex_lock(&container->lock);
1041 ret = tce_iommu_register_pages(container, param.vaddr,
1043 mutex_unlock(&container->lock);
1047 case VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY: {
1048 struct vfio_iommu_spapr_register_memory param;
1056 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1059 if (copy_from_user(¶m, (void __user *)arg, minsz))
1062 if (param.argsz < minsz)
1065 /* No flag is supported now */
1069 mutex_lock(&container->lock);
1070 ret = tce_iommu_unregister_pages(container, param.vaddr,
1072 mutex_unlock(&container->lock);
1076 case VFIO_IOMMU_ENABLE:
1080 mutex_lock(&container->lock);
1081 ret = tce_iommu_enable(container);
1082 mutex_unlock(&container->lock);
1086 case VFIO_IOMMU_DISABLE:
1090 mutex_lock(&container->lock);
1091 tce_iommu_disable(container);
1092 mutex_unlock(&container->lock);
1095 case VFIO_EEH_PE_OP: {
1096 struct tce_iommu_group *tcegrp;
1099 list_for_each_entry(tcegrp, &container->group_list, next) {
1100 ret = vfio_spapr_iommu_eeh_ioctl(tcegrp->grp,
1108 case VFIO_IOMMU_SPAPR_TCE_CREATE: {
1109 struct vfio_iommu_spapr_tce_create create;
1114 ret = tce_iommu_mm_set(container);
1118 if (!tce_groups_attached(container))
1121 minsz = offsetofend(struct vfio_iommu_spapr_tce_create,
1124 if (copy_from_user(&create, (void __user *)arg, minsz))
1127 if (create.argsz < minsz)
1133 mutex_lock(&container->lock);
1135 ret = tce_iommu_create_default_window(container);
1137 ret = tce_iommu_create_window(container,
1139 create.window_size, create.levels,
1140 &create.start_addr);
1142 mutex_unlock(&container->lock);
1144 if (!ret && copy_to_user((void __user *)arg, &create, minsz))
1149 case VFIO_IOMMU_SPAPR_TCE_REMOVE: {
1150 struct vfio_iommu_spapr_tce_remove remove;
1155 ret = tce_iommu_mm_set(container);
1159 if (!tce_groups_attached(container))
1162 minsz = offsetofend(struct vfio_iommu_spapr_tce_remove,
1165 if (copy_from_user(&remove, (void __user *)arg, minsz))
1168 if (remove.argsz < minsz)
1174 if (container->def_window_pending && !remove.start_addr) {
1175 container->def_window_pending = false;
1179 mutex_lock(&container->lock);
1181 ret = tce_iommu_remove_window(container, remove.start_addr);
1183 mutex_unlock(&container->lock);
1192 static void tce_iommu_release_ownership(struct tce_container *container,
1193 struct iommu_table_group *table_group)
1197 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1198 struct iommu_table *tbl = container->tables[i];
1203 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
1204 tce_iommu_userspace_view_free(tbl, container->mm);
1206 iommu_release_ownership(tbl);
1208 container->tables[i] = NULL;
1212 static int tce_iommu_take_ownership(struct tce_container *container,
1213 struct iommu_table_group *table_group)
1217 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1218 struct iommu_table *tbl = table_group->tables[i];
1220 if (!tbl || !tbl->it_map)
1223 rc = iommu_take_ownership(tbl);
1225 for (j = 0; j < i; ++j)
1226 iommu_release_ownership(
1227 table_group->tables[j]);
1233 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1234 container->tables[i] = table_group->tables[i];
1239 static void tce_iommu_release_ownership_ddw(struct tce_container *container,
1240 struct iommu_table_group *table_group)
1244 if (!table_group->ops->unset_window) {
1249 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1250 table_group->ops->unset_window(table_group, i);
1252 table_group->ops->release_ownership(table_group);
1255 static long tce_iommu_take_ownership_ddw(struct tce_container *container,
1256 struct iommu_table_group *table_group)
1260 if (!table_group->ops->create_table || !table_group->ops->set_window ||
1261 !table_group->ops->release_ownership) {
1266 table_group->ops->take_ownership(table_group);
1268 /* Set all windows to the new group */
1269 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1270 struct iommu_table *tbl = container->tables[i];
1275 ret = table_group->ops->set_window(table_group, i, tbl);
1283 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1284 table_group->ops->unset_window(table_group, i);
1286 table_group->ops->release_ownership(table_group);
1291 static int tce_iommu_attach_group(void *iommu_data,
1292 struct iommu_group *iommu_group)
1295 struct tce_container *container = iommu_data;
1296 struct iommu_table_group *table_group;
1297 struct tce_iommu_group *tcegrp = NULL;
1299 mutex_lock(&container->lock);
1301 /* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
1302 iommu_group_id(iommu_group), iommu_group); */
1303 table_group = iommu_group_get_iommudata(iommu_group);
1309 if (tce_groups_attached(container) && (!table_group->ops ||
1310 !table_group->ops->take_ownership ||
1311 !table_group->ops->release_ownership)) {
1316 /* Check if new group has the same iommu_ops (i.e. compatible) */
1317 list_for_each_entry(tcegrp, &container->group_list, next) {
1318 struct iommu_table_group *table_group_tmp;
1320 if (tcegrp->grp == iommu_group) {
1321 pr_warn("tce_vfio: Group %d is already attached\n",
1322 iommu_group_id(iommu_group));
1326 table_group_tmp = iommu_group_get_iommudata(tcegrp->grp);
1327 if (table_group_tmp->ops->create_table !=
1328 table_group->ops->create_table) {
1329 pr_warn("tce_vfio: Group %d is incompatible with group %d\n",
1330 iommu_group_id(iommu_group),
1331 iommu_group_id(tcegrp->grp));
1337 tcegrp = kzalloc(sizeof(*tcegrp), GFP_KERNEL);
1343 if (!table_group->ops || !table_group->ops->take_ownership ||
1344 !table_group->ops->release_ownership) {
1345 if (container->v2) {
1349 ret = tce_iommu_take_ownership(container, table_group);
1351 if (!container->v2) {
1355 ret = tce_iommu_take_ownership_ddw(container, table_group);
1356 if (!tce_groups_attached(container) && !container->tables[0])
1357 container->def_window_pending = true;
1361 tcegrp->grp = iommu_group;
1362 list_add(&tcegrp->next, &container->group_list);
1369 mutex_unlock(&container->lock);
1374 static void tce_iommu_detach_group(void *iommu_data,
1375 struct iommu_group *iommu_group)
1377 struct tce_container *container = iommu_data;
1378 struct iommu_table_group *table_group;
1380 struct tce_iommu_group *tcegrp;
1382 mutex_lock(&container->lock);
1384 list_for_each_entry(tcegrp, &container->group_list, next) {
1385 if (tcegrp->grp == iommu_group) {
1392 pr_warn("tce_vfio: detaching unattached group #%u\n",
1393 iommu_group_id(iommu_group));
1397 list_del(&tcegrp->next);
1400 table_group = iommu_group_get_iommudata(iommu_group);
1401 BUG_ON(!table_group);
1403 if (!table_group->ops || !table_group->ops->release_ownership)
1404 tce_iommu_release_ownership(container, table_group);
1406 tce_iommu_release_ownership_ddw(container, table_group);
1409 mutex_unlock(&container->lock);
1412 const struct vfio_iommu_driver_ops tce_iommu_driver_ops = {
1413 .name = "iommu-vfio-powerpc",
1414 .owner = THIS_MODULE,
1415 .open = tce_iommu_open,
1416 .release = tce_iommu_release,
1417 .ioctl = tce_iommu_ioctl,
1418 .attach_group = tce_iommu_attach_group,
1419 .detach_group = tce_iommu_detach_group,
1422 static int __init tce_iommu_init(void)
1424 return vfio_register_iommu_driver(&tce_iommu_driver_ops);
1427 static void __exit tce_iommu_cleanup(void)
1429 vfio_unregister_iommu_driver(&tce_iommu_driver_ops);
1432 module_init(tce_iommu_init);
1433 module_exit(tce_iommu_cleanup);
1435 MODULE_VERSION(DRIVER_VERSION);
1436 MODULE_LICENSE("GPL v2");
1437 MODULE_AUTHOR(DRIVER_AUTHOR);
1438 MODULE_DESCRIPTION(DRIVER_DESC);