1 // SPDX-License-Identifier: GPL-2.0-only
3 * VFIO: IOMMU DMA mapping support for TCE on POWER
5 * Copyright (C) 2013 IBM Corp. All rights reserved.
6 * Author: Alexey Kardashevskiy <aik@ozlabs.ru>
8 * Derived from original vfio_iommu_type1.c:
9 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
10 * Author: Alex Williamson <alex.williamson@redhat.com>
13 #include <linux/module.h>
14 #include <linux/pci.h>
15 #include <linux/slab.h>
16 #include <linux/uaccess.h>
17 #include <linux/err.h>
18 #include <linux/vfio.h>
19 #include <linux/vmalloc.h>
20 #include <linux/sched/mm.h>
21 #include <linux/sched/signal.h>
25 #include <asm/iommu.h>
27 #include <asm/mmu_context.h>
29 #define DRIVER_VERSION "0.1"
30 #define DRIVER_AUTHOR "aik@ozlabs.ru"
31 #define DRIVER_DESC "VFIO IOMMU SPAPR TCE"
33 static void tce_iommu_detach_group(void *iommu_data,
34 struct iommu_group *iommu_group);
37 * VFIO IOMMU fd for SPAPR_TCE IOMMU implementation
39 * This code handles mapping and unmapping of user data buffers
40 * into DMA'ble space using the IOMMU
43 struct tce_iommu_group {
44 struct list_head next;
45 struct iommu_group *grp;
49 * A container needs to remember which preregistered region it has
50 * referenced to do proper cleanup at the userspace process exit.
52 struct tce_iommu_prereg {
53 struct list_head next;
54 struct mm_iommu_table_group_mem_t *mem;
58 * The container descriptor supports only a single group per container.
59 * Required by the API as the container is not supplied with the IOMMU group
60 * at the moment of initialization.
62 struct tce_container {
66 bool def_window_pending;
67 unsigned long locked_pages;
69 struct iommu_table *tables[IOMMU_TABLE_GROUP_MAX_TABLES];
70 struct list_head group_list;
71 struct list_head prereg_list;
74 static long tce_iommu_mm_set(struct tce_container *container)
77 if (container->mm == current->mm)
82 container->mm = current->mm;
83 mmgrab(container->mm);
88 static long tce_iommu_prereg_free(struct tce_container *container,
89 struct tce_iommu_prereg *tcemem)
93 ret = mm_iommu_put(container->mm, tcemem->mem);
97 list_del(&tcemem->next);
103 static long tce_iommu_unregister_pages(struct tce_container *container,
104 __u64 vaddr, __u64 size)
106 struct mm_iommu_table_group_mem_t *mem;
107 struct tce_iommu_prereg *tcemem;
111 if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK))
114 mem = mm_iommu_get(container->mm, vaddr, size >> PAGE_SHIFT);
118 list_for_each_entry(tcemem, &container->prereg_list, next) {
119 if (tcemem->mem == mem) {
128 ret = tce_iommu_prereg_free(container, tcemem);
130 mm_iommu_put(container->mm, mem);
135 static long tce_iommu_register_pages(struct tce_container *container,
136 __u64 vaddr, __u64 size)
139 struct mm_iommu_table_group_mem_t *mem = NULL;
140 struct tce_iommu_prereg *tcemem;
141 unsigned long entries = size >> PAGE_SHIFT;
143 if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK) ||
144 ((vaddr + size) < vaddr))
147 mem = mm_iommu_get(container->mm, vaddr, entries);
149 list_for_each_entry(tcemem, &container->prereg_list, next) {
150 if (tcemem->mem == mem) {
156 ret = mm_iommu_new(container->mm, vaddr, entries, &mem);
161 tcemem = kzalloc(sizeof(*tcemem), GFP_KERNEL);
168 list_add(&tcemem->next, &container->prereg_list);
170 container->enabled = true;
175 mm_iommu_put(container->mm, mem);
179 static bool tce_page_is_contained(struct mm_struct *mm, unsigned long hpa,
180 unsigned int it_page_shift)
183 unsigned long size = 0;
185 if (mm_iommu_is_devmem(mm, hpa, it_page_shift, &size))
186 return size == (1UL << it_page_shift);
188 page = pfn_to_page(hpa >> PAGE_SHIFT);
190 * Check that the TCE table granularity is not bigger than the size of
191 * a page we just found. Otherwise the hardware can get access to
192 * a bigger memory chunk that it should.
194 return page_shift(compound_head(page)) >= it_page_shift;
197 static inline bool tce_groups_attached(struct tce_container *container)
199 return !list_empty(&container->group_list);
202 static long tce_iommu_find_table(struct tce_container *container,
203 phys_addr_t ioba, struct iommu_table **ptbl)
207 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
208 struct iommu_table *tbl = container->tables[i];
211 unsigned long entry = ioba >> tbl->it_page_shift;
212 unsigned long start = tbl->it_offset;
213 unsigned long end = start + tbl->it_size;
215 if ((start <= entry) && (entry < end)) {
225 static int tce_iommu_find_free_table(struct tce_container *container)
229 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
230 if (!container->tables[i])
237 static int tce_iommu_enable(struct tce_container *container)
240 unsigned long locked;
241 struct iommu_table_group *table_group;
242 struct tce_iommu_group *tcegrp;
244 if (container->enabled)
248 * When userspace pages are mapped into the IOMMU, they are effectively
249 * locked memory, so, theoretically, we need to update the accounting
250 * of locked pages on each map and unmap. For powerpc, the map unmap
251 * paths can be very hot, though, and the accounting would kill
252 * performance, especially since it would be difficult to impossible
253 * to handle the accounting in real mode only.
255 * To address that, rather than precisely accounting every page, we
256 * instead account for a worst case on locked memory when the iommu is
257 * enabled and disabled. The worst case upper bound on locked memory
258 * is the size of the whole iommu window, which is usually relatively
259 * small (compared to total memory sizes) on POWER hardware.
261 * Also we don't have a nice way to fail on H_PUT_TCE due to ulimits,
262 * that would effectively kill the guest at random points, much better
263 * enforcing the limit based on the max that the guest can map.
265 * Unfortunately at the moment it counts whole tables, no matter how
266 * much memory the guest has. I.e. for 4GB guest and 4 IOMMU groups
267 * each with 2GB DMA window, 8GB will be counted here. The reason for
268 * this is that we cannot tell here the amount of RAM used by the guest
269 * as this information is only available from KVM and VFIO is
272 * So we do not allow enabling a container without a group attached
273 * as there is no way to know how much we should increment
274 * the locked_vm counter.
276 if (!tce_groups_attached(container))
279 tcegrp = list_first_entry(&container->group_list,
280 struct tce_iommu_group, next);
281 table_group = iommu_group_get_iommudata(tcegrp->grp);
285 if (!table_group->tce32_size)
288 ret = tce_iommu_mm_set(container);
292 locked = table_group->tce32_size >> PAGE_SHIFT;
293 ret = account_locked_vm(container->mm, locked, true);
297 container->locked_pages = locked;
299 container->enabled = true;
304 static void tce_iommu_disable(struct tce_container *container)
306 if (!container->enabled)
309 container->enabled = false;
311 BUG_ON(!container->mm);
312 account_locked_vm(container->mm, container->locked_pages, false);
315 static void *tce_iommu_open(unsigned long arg)
317 struct tce_container *container;
319 if ((arg != VFIO_SPAPR_TCE_IOMMU) && (arg != VFIO_SPAPR_TCE_v2_IOMMU)) {
320 pr_err("tce_vfio: Wrong IOMMU type\n");
321 return ERR_PTR(-EINVAL);
324 container = kzalloc(sizeof(*container), GFP_KERNEL);
326 return ERR_PTR(-ENOMEM);
328 mutex_init(&container->lock);
329 INIT_LIST_HEAD_RCU(&container->group_list);
330 INIT_LIST_HEAD_RCU(&container->prereg_list);
332 container->v2 = arg == VFIO_SPAPR_TCE_v2_IOMMU;
337 static int tce_iommu_clear(struct tce_container *container,
338 struct iommu_table *tbl,
339 unsigned long entry, unsigned long pages);
340 static void tce_iommu_free_table(struct tce_container *container,
341 struct iommu_table *tbl);
343 static void tce_iommu_release(void *iommu_data)
345 struct tce_container *container = iommu_data;
346 struct tce_iommu_group *tcegrp;
347 struct tce_iommu_prereg *tcemem, *tmtmp;
350 while (tce_groups_attached(container)) {
351 tcegrp = list_first_entry(&container->group_list,
352 struct tce_iommu_group, next);
353 tce_iommu_detach_group(iommu_data, tcegrp->grp);
357 * If VFIO created a table, it was not disposed
358 * by tce_iommu_detach_group() so do it now.
360 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
361 struct iommu_table *tbl = container->tables[i];
366 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
367 tce_iommu_free_table(container, tbl);
370 list_for_each_entry_safe(tcemem, tmtmp, &container->prereg_list, next)
371 WARN_ON(tce_iommu_prereg_free(container, tcemem));
373 tce_iommu_disable(container);
375 mmdrop(container->mm);
376 mutex_destroy(&container->lock);
381 static void tce_iommu_unuse_page(unsigned long hpa)
385 page = pfn_to_page(hpa >> PAGE_SHIFT);
386 unpin_user_page(page);
389 static int tce_iommu_prereg_ua_to_hpa(struct tce_container *container,
390 unsigned long tce, unsigned long shift,
391 unsigned long *phpa, struct mm_iommu_table_group_mem_t **pmem)
394 struct mm_iommu_table_group_mem_t *mem;
396 mem = mm_iommu_lookup(container->mm, tce, 1ULL << shift);
400 ret = mm_iommu_ua_to_hpa(mem, tce, shift, phpa);
409 static void tce_iommu_unuse_page_v2(struct tce_container *container,
410 struct iommu_table *tbl, unsigned long entry)
412 struct mm_iommu_table_group_mem_t *mem = NULL;
414 unsigned long hpa = 0;
415 __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY_RO(tbl, entry);
420 ret = tce_iommu_prereg_ua_to_hpa(container, be64_to_cpu(*pua),
421 tbl->it_page_shift, &hpa, &mem);
423 pr_debug("%s: tce %llx at #%lx was not cached, ret=%d\n",
424 __func__, be64_to_cpu(*pua), entry, ret);
426 mm_iommu_mapped_dec(mem);
428 *pua = cpu_to_be64(0);
431 static int tce_iommu_clear(struct tce_container *container,
432 struct iommu_table *tbl,
433 unsigned long entry, unsigned long pages)
435 unsigned long oldhpa;
437 enum dma_data_direction direction;
438 unsigned long lastentry = entry + pages, firstentry = entry;
440 for ( ; entry < lastentry; ++entry) {
441 if (tbl->it_indirect_levels && tbl->it_userspace) {
443 * For multilevel tables, we can take a shortcut here
444 * and skip some TCEs as we know that the userspace
445 * addresses cache is a mirror of the real TCE table
446 * and if it is missing some indirect levels, then
447 * the hardware table does not have them allocated
448 * either and therefore does not require updating.
450 __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY_RO(tbl,
453 /* align to level_size which is power of two */
454 entry |= tbl->it_level_size - 1;
461 direction = DMA_NONE;
463 ret = iommu_tce_xchg_no_kill(container->mm, tbl, entry, &oldhpa,
468 if (direction == DMA_NONE)
472 tce_iommu_unuse_page_v2(container, tbl, entry);
476 tce_iommu_unuse_page(oldhpa);
479 iommu_tce_kill(tbl, firstentry, pages);
484 static int tce_iommu_use_page(unsigned long tce, unsigned long *hpa)
486 struct page *page = NULL;
487 enum dma_data_direction direction = iommu_tce_direction(tce);
489 if (pin_user_pages_fast(tce & PAGE_MASK, 1,
490 direction != DMA_TO_DEVICE ? FOLL_WRITE : 0,
494 *hpa = __pa((unsigned long) page_address(page));
499 static long tce_iommu_build(struct tce_container *container,
500 struct iommu_table *tbl,
501 unsigned long entry, unsigned long tce, unsigned long pages,
502 enum dma_data_direction direction)
506 enum dma_data_direction dirtmp;
508 for (i = 0; i < pages; ++i) {
509 unsigned long offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
511 ret = tce_iommu_use_page(tce, &hpa);
515 if (!tce_page_is_contained(container->mm, hpa,
516 tbl->it_page_shift)) {
523 ret = iommu_tce_xchg_no_kill(container->mm, tbl, entry + i,
526 tce_iommu_unuse_page(hpa);
527 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
528 __func__, entry << tbl->it_page_shift,
533 if (dirtmp != DMA_NONE)
534 tce_iommu_unuse_page(hpa);
536 tce += IOMMU_PAGE_SIZE(tbl);
540 tce_iommu_clear(container, tbl, entry, i);
542 iommu_tce_kill(tbl, entry, pages);
547 static long tce_iommu_build_v2(struct tce_container *container,
548 struct iommu_table *tbl,
549 unsigned long entry, unsigned long tce, unsigned long pages,
550 enum dma_data_direction direction)
554 enum dma_data_direction dirtmp;
556 for (i = 0; i < pages; ++i) {
557 struct mm_iommu_table_group_mem_t *mem = NULL;
558 __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry + i);
560 ret = tce_iommu_prereg_ua_to_hpa(container,
561 tce, tbl->it_page_shift, &hpa, &mem);
565 if (!tce_page_is_contained(container->mm, hpa,
566 tbl->it_page_shift)) {
571 /* Preserve offset within IOMMU page */
572 hpa |= tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
575 /* The registered region is being unregistered */
576 if (mm_iommu_mapped_inc(mem))
579 ret = iommu_tce_xchg_no_kill(container->mm, tbl, entry + i,
582 /* dirtmp cannot be DMA_NONE here */
583 tce_iommu_unuse_page_v2(container, tbl, entry + i);
584 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
585 __func__, entry << tbl->it_page_shift,
590 if (dirtmp != DMA_NONE)
591 tce_iommu_unuse_page_v2(container, tbl, entry + i);
593 *pua = cpu_to_be64(tce);
595 tce += IOMMU_PAGE_SIZE(tbl);
599 tce_iommu_clear(container, tbl, entry, i);
601 iommu_tce_kill(tbl, entry, pages);
606 static long tce_iommu_create_table(struct tce_container *container,
607 struct iommu_table_group *table_group,
612 struct iommu_table **ptbl)
614 long ret, table_size;
616 table_size = table_group->ops->get_table_size(page_shift, window_size,
621 ret = account_locked_vm(container->mm, table_size >> PAGE_SHIFT, true);
625 ret = table_group->ops->create_table(table_group, num,
626 page_shift, window_size, levels, ptbl);
628 WARN_ON(!ret && !(*ptbl)->it_ops->free);
629 WARN_ON(!ret && ((*ptbl)->it_allocated_size > table_size));
634 static void tce_iommu_free_table(struct tce_container *container,
635 struct iommu_table *tbl)
637 unsigned long pages = tbl->it_allocated_size >> PAGE_SHIFT;
639 iommu_tce_table_put(tbl);
640 account_locked_vm(container->mm, pages, false);
643 static long tce_iommu_create_window(struct tce_container *container,
644 __u32 page_shift, __u64 window_size, __u32 levels,
647 struct tce_iommu_group *tcegrp;
648 struct iommu_table_group *table_group;
649 struct iommu_table *tbl = NULL;
652 num = tce_iommu_find_free_table(container);
656 /* Get the first group for ops::create_table */
657 tcegrp = list_first_entry(&container->group_list,
658 struct tce_iommu_group, next);
659 table_group = iommu_group_get_iommudata(tcegrp->grp);
663 if (!(table_group->pgsizes & (1ULL << page_shift)))
666 if (!table_group->ops->set_window || !table_group->ops->unset_window ||
667 !table_group->ops->get_table_size ||
668 !table_group->ops->create_table)
671 /* Create TCE table */
672 ret = tce_iommu_create_table(container, table_group, num,
673 page_shift, window_size, levels, &tbl);
677 BUG_ON(!tbl->it_ops->free);
680 * Program the table to every group.
681 * Groups have been tested for compatibility at the attach time.
683 list_for_each_entry(tcegrp, &container->group_list, next) {
684 table_group = iommu_group_get_iommudata(tcegrp->grp);
686 ret = table_group->ops->set_window(table_group, num, tbl);
691 container->tables[num] = tbl;
693 /* Return start address assigned by platform in create_table() */
694 *start_addr = tbl->it_offset << tbl->it_page_shift;
699 list_for_each_entry(tcegrp, &container->group_list, next) {
700 table_group = iommu_group_get_iommudata(tcegrp->grp);
701 table_group->ops->unset_window(table_group, num);
703 tce_iommu_free_table(container, tbl);
708 static long tce_iommu_remove_window(struct tce_container *container,
711 struct iommu_table_group *table_group = NULL;
712 struct iommu_table *tbl;
713 struct tce_iommu_group *tcegrp;
716 num = tce_iommu_find_table(container, start_addr, &tbl);
720 BUG_ON(!tbl->it_size);
722 /* Detach groups from IOMMUs */
723 list_for_each_entry(tcegrp, &container->group_list, next) {
724 table_group = iommu_group_get_iommudata(tcegrp->grp);
727 * SPAPR TCE IOMMU exposes the default DMA window to
728 * the guest via dma32_window_start/size of
729 * VFIO_IOMMU_SPAPR_TCE_GET_INFO. Some platforms allow
730 * the userspace to remove this window, some do not so
731 * here we check for the platform capability.
733 if (!table_group->ops || !table_group->ops->unset_window)
736 table_group->ops->unset_window(table_group, num);
740 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
741 tce_iommu_free_table(container, tbl);
742 container->tables[num] = NULL;
747 static long tce_iommu_create_default_window(struct tce_container *container)
750 __u64 start_addr = 0;
751 struct tce_iommu_group *tcegrp;
752 struct iommu_table_group *table_group;
754 if (!container->def_window_pending)
757 if (!tce_groups_attached(container))
760 tcegrp = list_first_entry(&container->group_list,
761 struct tce_iommu_group, next);
762 table_group = iommu_group_get_iommudata(tcegrp->grp);
766 ret = tce_iommu_create_window(container, IOMMU_PAGE_SHIFT_4K,
767 table_group->tce32_size, 1, &start_addr);
768 WARN_ON_ONCE(!ret && start_addr);
771 container->def_window_pending = false;
776 static long tce_iommu_ioctl(void *iommu_data,
777 unsigned int cmd, unsigned long arg)
779 struct tce_container *container = iommu_data;
780 unsigned long minsz, ddwsz;
784 case VFIO_CHECK_EXTENSION:
786 case VFIO_SPAPR_TCE_IOMMU:
787 case VFIO_SPAPR_TCE_v2_IOMMU:
791 ret = vfio_spapr_iommu_eeh_ioctl(NULL, cmd, arg);
795 return (ret < 0) ? 0 : ret;
799 * Sanity check to prevent one userspace from manipulating
800 * another userspace mm.
803 if (container->mm && container->mm != current->mm)
807 case VFIO_IOMMU_SPAPR_TCE_GET_INFO: {
808 struct vfio_iommu_spapr_tce_info info;
809 struct tce_iommu_group *tcegrp;
810 struct iommu_table_group *table_group;
812 if (!tce_groups_attached(container))
815 tcegrp = list_first_entry(&container->group_list,
816 struct tce_iommu_group, next);
817 table_group = iommu_group_get_iommudata(tcegrp->grp);
822 minsz = offsetofend(struct vfio_iommu_spapr_tce_info,
825 if (copy_from_user(&info, (void __user *)arg, minsz))
828 if (info.argsz < minsz)
831 info.dma32_window_start = table_group->tce32_start;
832 info.dma32_window_size = table_group->tce32_size;
834 memset(&info.ddw, 0, sizeof(info.ddw));
836 if (table_group->max_dynamic_windows_supported &&
838 info.flags |= VFIO_IOMMU_SPAPR_INFO_DDW;
839 info.ddw.pgsizes = table_group->pgsizes;
840 info.ddw.max_dynamic_windows_supported =
841 table_group->max_dynamic_windows_supported;
842 info.ddw.levels = table_group->max_levels;
845 ddwsz = offsetofend(struct vfio_iommu_spapr_tce_info, ddw);
847 if (info.argsz >= ddwsz)
850 if (copy_to_user((void __user *)arg, &info, minsz))
855 case VFIO_IOMMU_MAP_DMA: {
856 struct vfio_iommu_type1_dma_map param;
857 struct iommu_table *tbl = NULL;
859 enum dma_data_direction direction;
861 if (!container->enabled)
864 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
866 if (copy_from_user(¶m, (void __user *)arg, minsz))
869 if (param.argsz < minsz)
872 if (param.flags & ~(VFIO_DMA_MAP_FLAG_READ |
873 VFIO_DMA_MAP_FLAG_WRITE))
876 ret = tce_iommu_create_default_window(container);
880 num = tce_iommu_find_table(container, param.iova, &tbl);
884 if ((param.size & ~IOMMU_PAGE_MASK(tbl)) ||
885 (param.vaddr & ~IOMMU_PAGE_MASK(tbl)))
888 /* iova is checked by the IOMMU API */
889 if (param.flags & VFIO_DMA_MAP_FLAG_READ) {
890 if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
891 direction = DMA_BIDIRECTIONAL;
893 direction = DMA_TO_DEVICE;
895 if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
896 direction = DMA_FROM_DEVICE;
901 ret = iommu_tce_put_param_check(tbl, param.iova, param.vaddr);
906 ret = tce_iommu_build_v2(container, tbl,
907 param.iova >> tbl->it_page_shift,
909 param.size >> tbl->it_page_shift,
912 ret = tce_iommu_build(container, tbl,
913 param.iova >> tbl->it_page_shift,
915 param.size >> tbl->it_page_shift,
918 iommu_flush_tce(tbl);
922 case VFIO_IOMMU_UNMAP_DMA: {
923 struct vfio_iommu_type1_dma_unmap param;
924 struct iommu_table *tbl = NULL;
927 if (!container->enabled)
930 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap,
933 if (copy_from_user(¶m, (void __user *)arg, minsz))
936 if (param.argsz < minsz)
939 /* No flag is supported now */
943 ret = tce_iommu_create_default_window(container);
947 num = tce_iommu_find_table(container, param.iova, &tbl);
951 if (param.size & ~IOMMU_PAGE_MASK(tbl))
954 ret = iommu_tce_clear_param_check(tbl, param.iova, 0,
955 param.size >> tbl->it_page_shift);
959 ret = tce_iommu_clear(container, tbl,
960 param.iova >> tbl->it_page_shift,
961 param.size >> tbl->it_page_shift);
962 iommu_flush_tce(tbl);
966 case VFIO_IOMMU_SPAPR_REGISTER_MEMORY: {
967 struct vfio_iommu_spapr_register_memory param;
972 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
975 ret = tce_iommu_mm_set(container);
979 if (copy_from_user(¶m, (void __user *)arg, minsz))
982 if (param.argsz < minsz)
985 /* No flag is supported now */
989 mutex_lock(&container->lock);
990 ret = tce_iommu_register_pages(container, param.vaddr,
992 mutex_unlock(&container->lock);
996 case VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY: {
997 struct vfio_iommu_spapr_register_memory param;
1005 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1008 if (copy_from_user(¶m, (void __user *)arg, minsz))
1011 if (param.argsz < minsz)
1014 /* No flag is supported now */
1018 mutex_lock(&container->lock);
1019 ret = tce_iommu_unregister_pages(container, param.vaddr,
1021 mutex_unlock(&container->lock);
1025 case VFIO_IOMMU_ENABLE:
1029 mutex_lock(&container->lock);
1030 ret = tce_iommu_enable(container);
1031 mutex_unlock(&container->lock);
1035 case VFIO_IOMMU_DISABLE:
1039 mutex_lock(&container->lock);
1040 tce_iommu_disable(container);
1041 mutex_unlock(&container->lock);
1044 case VFIO_EEH_PE_OP: {
1045 struct tce_iommu_group *tcegrp;
1048 list_for_each_entry(tcegrp, &container->group_list, next) {
1049 ret = vfio_spapr_iommu_eeh_ioctl(tcegrp->grp,
1057 case VFIO_IOMMU_SPAPR_TCE_CREATE: {
1058 struct vfio_iommu_spapr_tce_create create;
1063 ret = tce_iommu_mm_set(container);
1067 if (!tce_groups_attached(container))
1070 minsz = offsetofend(struct vfio_iommu_spapr_tce_create,
1073 if (copy_from_user(&create, (void __user *)arg, minsz))
1076 if (create.argsz < minsz)
1082 mutex_lock(&container->lock);
1084 ret = tce_iommu_create_default_window(container);
1086 ret = tce_iommu_create_window(container,
1088 create.window_size, create.levels,
1089 &create.start_addr);
1091 mutex_unlock(&container->lock);
1093 if (!ret && copy_to_user((void __user *)arg, &create, minsz))
1098 case VFIO_IOMMU_SPAPR_TCE_REMOVE: {
1099 struct vfio_iommu_spapr_tce_remove remove;
1104 ret = tce_iommu_mm_set(container);
1108 if (!tce_groups_attached(container))
1111 minsz = offsetofend(struct vfio_iommu_spapr_tce_remove,
1114 if (copy_from_user(&remove, (void __user *)arg, minsz))
1117 if (remove.argsz < minsz)
1123 if (container->def_window_pending && !remove.start_addr) {
1124 container->def_window_pending = false;
1128 mutex_lock(&container->lock);
1130 ret = tce_iommu_remove_window(container, remove.start_addr);
1132 mutex_unlock(&container->lock);
1141 static void tce_iommu_release_ownership(struct tce_container *container,
1142 struct iommu_table_group *table_group)
1146 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1147 struct iommu_table *tbl = container->tables[i];
1152 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
1154 iommu_release_ownership(tbl);
1156 container->tables[i] = NULL;
1160 static int tce_iommu_take_ownership(struct tce_container *container,
1161 struct iommu_table_group *table_group)
1165 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1166 struct iommu_table *tbl = table_group->tables[i];
1168 if (!tbl || !tbl->it_map)
1171 rc = iommu_take_ownership(tbl);
1173 for (j = 0; j < i; ++j)
1174 iommu_release_ownership(
1175 table_group->tables[j]);
1181 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1182 container->tables[i] = table_group->tables[i];
1187 static void tce_iommu_release_ownership_ddw(struct tce_container *container,
1188 struct iommu_table_group *table_group)
1192 if (!table_group->ops->unset_window) {
1197 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1198 if (container->tables[i])
1199 table_group->ops->unset_window(table_group, i);
1201 table_group->ops->release_ownership(table_group);
1204 static long tce_iommu_take_ownership_ddw(struct tce_container *container,
1205 struct iommu_table_group *table_group)
1209 if (!table_group->ops->create_table || !table_group->ops->set_window ||
1210 !table_group->ops->release_ownership) {
1215 table_group->ops->take_ownership(table_group);
1217 /* Set all windows to the new group */
1218 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1219 struct iommu_table *tbl = container->tables[i];
1224 ret = table_group->ops->set_window(table_group, i, tbl);
1232 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1233 table_group->ops->unset_window(table_group, i);
1235 table_group->ops->release_ownership(table_group);
1240 static int tce_iommu_attach_group(void *iommu_data,
1241 struct iommu_group *iommu_group, enum vfio_group_type type)
1244 struct tce_container *container = iommu_data;
1245 struct iommu_table_group *table_group;
1246 struct tce_iommu_group *tcegrp = NULL;
1248 if (type == VFIO_EMULATED_IOMMU)
1251 mutex_lock(&container->lock);
1253 /* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
1254 iommu_group_id(iommu_group), iommu_group); */
1255 table_group = iommu_group_get_iommudata(iommu_group);
1261 if (tce_groups_attached(container) && (!table_group->ops ||
1262 !table_group->ops->take_ownership ||
1263 !table_group->ops->release_ownership)) {
1269 * Check if new group has the same iommu_table_group_ops
1272 list_for_each_entry(tcegrp, &container->group_list, next) {
1273 struct iommu_table_group *table_group_tmp;
1275 if (tcegrp->grp == iommu_group) {
1276 pr_warn("tce_vfio: Group %d is already attached\n",
1277 iommu_group_id(iommu_group));
1281 table_group_tmp = iommu_group_get_iommudata(tcegrp->grp);
1282 if (table_group_tmp->ops->create_table !=
1283 table_group->ops->create_table) {
1284 pr_warn("tce_vfio: Group %d is incompatible with group %d\n",
1285 iommu_group_id(iommu_group),
1286 iommu_group_id(tcegrp->grp));
1292 tcegrp = kzalloc(sizeof(*tcegrp), GFP_KERNEL);
1298 if (!table_group->ops || !table_group->ops->take_ownership ||
1299 !table_group->ops->release_ownership) {
1300 if (container->v2) {
1304 ret = tce_iommu_take_ownership(container, table_group);
1306 if (!container->v2) {
1310 ret = tce_iommu_take_ownership_ddw(container, table_group);
1311 if (!tce_groups_attached(container) && !container->tables[0])
1312 container->def_window_pending = true;
1316 tcegrp->grp = iommu_group;
1317 list_add(&tcegrp->next, &container->group_list);
1325 mutex_unlock(&container->lock);
1330 static void tce_iommu_detach_group(void *iommu_data,
1331 struct iommu_group *iommu_group)
1333 struct tce_container *container = iommu_data;
1334 struct iommu_table_group *table_group;
1336 struct tce_iommu_group *tcegrp;
1338 mutex_lock(&container->lock);
1340 list_for_each_entry(tcegrp, &container->group_list, next) {
1341 if (tcegrp->grp == iommu_group) {
1348 pr_warn("tce_vfio: detaching unattached group #%u\n",
1349 iommu_group_id(iommu_group));
1353 list_del(&tcegrp->next);
1356 table_group = iommu_group_get_iommudata(iommu_group);
1357 BUG_ON(!table_group);
1359 if (!table_group->ops || !table_group->ops->release_ownership)
1360 tce_iommu_release_ownership(container, table_group);
1362 tce_iommu_release_ownership_ddw(container, table_group);
1365 mutex_unlock(&container->lock);
1368 static const struct vfio_iommu_driver_ops tce_iommu_driver_ops = {
1369 .name = "iommu-vfio-powerpc",
1370 .owner = THIS_MODULE,
1371 .open = tce_iommu_open,
1372 .release = tce_iommu_release,
1373 .ioctl = tce_iommu_ioctl,
1374 .attach_group = tce_iommu_attach_group,
1375 .detach_group = tce_iommu_detach_group,
1378 static int __init tce_iommu_init(void)
1380 return vfio_register_iommu_driver(&tce_iommu_driver_ops);
1383 static void __exit tce_iommu_cleanup(void)
1385 vfio_unregister_iommu_driver(&tce_iommu_driver_ops);
1388 module_init(tce_iommu_init);
1389 module_exit(tce_iommu_cleanup);
1391 MODULE_VERSION(DRIVER_VERSION);
1392 MODULE_LICENSE("GPL v2");
1393 MODULE_AUTHOR(DRIVER_AUTHOR);
1394 MODULE_DESCRIPTION(DRIVER_DESC);