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 <asm/iommu.h>
25 #include <asm/mmu_context.h>
27 #define DRIVER_VERSION "0.1"
28 #define DRIVER_AUTHOR "aik@ozlabs.ru"
29 #define DRIVER_DESC "VFIO IOMMU SPAPR TCE"
31 static void tce_iommu_detach_group(void *iommu_data,
32 struct iommu_group *iommu_group);
34 static long try_increment_locked_vm(long npages)
36 long ret = 0, locked, lock_limit;
38 if (!current || !current->mm)
39 return -ESRCH; /* process exited */
44 down_write(¤t->mm->mmap_sem);
45 locked = current->mm->locked_vm + npages;
46 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
47 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
50 current->mm->locked_vm += npages;
52 pr_debug("[%d] RLIMIT_MEMLOCK +%ld %ld/%ld%s\n", current->pid,
54 current->mm->locked_vm << PAGE_SHIFT,
55 rlimit(RLIMIT_MEMLOCK),
56 ret ? " - exceeded" : "");
58 up_write(¤t->mm->mmap_sem);
63 static void decrement_locked_vm(long npages)
65 if (!current || !current->mm || !npages)
66 return; /* process exited */
68 down_write(¤t->mm->mmap_sem);
69 if (WARN_ON_ONCE(npages > current->mm->locked_vm))
70 npages = current->mm->locked_vm;
71 current->mm->locked_vm -= npages;
72 pr_debug("[%d] RLIMIT_MEMLOCK -%ld %ld/%ld\n", current->pid,
74 current->mm->locked_vm << PAGE_SHIFT,
75 rlimit(RLIMIT_MEMLOCK));
76 up_write(¤t->mm->mmap_sem);
80 * VFIO IOMMU fd for SPAPR_TCE IOMMU implementation
82 * This code handles mapping and unmapping of user data buffers
83 * into DMA'ble space using the IOMMU
86 struct tce_iommu_group {
87 struct list_head next;
88 struct iommu_group *grp;
92 * The container descriptor supports only a single group per container.
93 * Required by the API as the container is not supplied with the IOMMU group
94 * at the moment of initialization.
96 struct tce_container {
100 unsigned long locked_pages;
101 struct iommu_table *tables[IOMMU_TABLE_GROUP_MAX_TABLES];
102 struct list_head group_list;
105 static long tce_iommu_unregister_pages(struct tce_container *container,
106 __u64 vaddr, __u64 size)
108 struct mm_iommu_table_group_mem_t *mem;
110 if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK))
113 mem = mm_iommu_find(vaddr, size >> PAGE_SHIFT);
117 return mm_iommu_put(mem);
120 static long tce_iommu_register_pages(struct tce_container *container,
121 __u64 vaddr, __u64 size)
124 struct mm_iommu_table_group_mem_t *mem = NULL;
125 unsigned long entries = size >> PAGE_SHIFT;
127 if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK) ||
128 ((vaddr + size) < vaddr))
131 ret = mm_iommu_get(vaddr, entries, &mem);
135 container->enabled = true;
140 static long tce_iommu_userspace_view_alloc(struct iommu_table *tbl)
142 unsigned long cb = _ALIGN_UP(sizeof(tbl->it_userspace[0]) *
143 tbl->it_size, PAGE_SIZE);
147 BUG_ON(tbl->it_userspace);
149 ret = try_increment_locked_vm(cb >> PAGE_SHIFT);
155 decrement_locked_vm(cb >> PAGE_SHIFT);
158 tbl->it_userspace = uas;
163 static void tce_iommu_userspace_view_free(struct iommu_table *tbl)
165 unsigned long cb = _ALIGN_UP(sizeof(tbl->it_userspace[0]) *
166 tbl->it_size, PAGE_SIZE);
168 if (!tbl->it_userspace)
171 vfree(tbl->it_userspace);
172 tbl->it_userspace = NULL;
173 decrement_locked_vm(cb >> PAGE_SHIFT);
176 static bool tce_page_is_contained(struct page *page, unsigned page_shift)
179 * Check that the TCE table granularity is not bigger than the size of
180 * a page we just found. Otherwise the hardware can get access to
181 * a bigger memory chunk that it should.
183 return (PAGE_SHIFT + compound_order(compound_head(page))) >= page_shift;
186 static inline bool tce_groups_attached(struct tce_container *container)
188 return !list_empty(&container->group_list);
191 static long tce_iommu_find_table(struct tce_container *container,
192 phys_addr_t ioba, struct iommu_table **ptbl)
196 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
197 struct iommu_table *tbl = container->tables[i];
200 unsigned long entry = ioba >> tbl->it_page_shift;
201 unsigned long start = tbl->it_offset;
202 unsigned long end = start + tbl->it_size;
204 if ((start <= entry) && (entry < end)) {
214 static int tce_iommu_find_free_table(struct tce_container *container)
218 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
219 if (!container->tables[i])
226 static int tce_iommu_enable(struct tce_container *container)
229 unsigned long locked;
230 struct iommu_table_group *table_group;
231 struct tce_iommu_group *tcegrp;
234 return -ESRCH; /* process exited */
236 if (container->enabled)
240 * When userspace pages are mapped into the IOMMU, they are effectively
241 * locked memory, so, theoretically, we need to update the accounting
242 * of locked pages on each map and unmap. For powerpc, the map unmap
243 * paths can be very hot, though, and the accounting would kill
244 * performance, especially since it would be difficult to impossible
245 * to handle the accounting in real mode only.
247 * To address that, rather than precisely accounting every page, we
248 * instead account for a worst case on locked memory when the iommu is
249 * enabled and disabled. The worst case upper bound on locked memory
250 * is the size of the whole iommu window, which is usually relatively
251 * small (compared to total memory sizes) on POWER hardware.
253 * Also we don't have a nice way to fail on H_PUT_TCE due to ulimits,
254 * that would effectively kill the guest at random points, much better
255 * enforcing the limit based on the max that the guest can map.
257 * Unfortunately at the moment it counts whole tables, no matter how
258 * much memory the guest has. I.e. for 4GB guest and 4 IOMMU groups
259 * each with 2GB DMA window, 8GB will be counted here. The reason for
260 * this is that we cannot tell here the amount of RAM used by the guest
261 * as this information is only available from KVM and VFIO is
264 * So we do not allow enabling a container without a group attached
265 * as there is no way to know how much we should increment
266 * the locked_vm counter.
268 if (!tce_groups_attached(container))
271 tcegrp = list_first_entry(&container->group_list,
272 struct tce_iommu_group, next);
273 table_group = iommu_group_get_iommudata(tcegrp->grp);
277 if (!table_group->tce32_size)
280 locked = table_group->tce32_size >> PAGE_SHIFT;
281 ret = try_increment_locked_vm(locked);
285 container->locked_pages = locked;
287 container->enabled = true;
292 static void tce_iommu_disable(struct tce_container *container)
294 if (!container->enabled)
297 container->enabled = false;
302 decrement_locked_vm(container->locked_pages);
305 static void *tce_iommu_open(unsigned long arg)
307 struct tce_container *container;
309 if ((arg != VFIO_SPAPR_TCE_IOMMU) && (arg != VFIO_SPAPR_TCE_v2_IOMMU)) {
310 pr_err("tce_vfio: Wrong IOMMU type\n");
311 return ERR_PTR(-EINVAL);
314 container = kzalloc(sizeof(*container), GFP_KERNEL);
316 return ERR_PTR(-ENOMEM);
318 mutex_init(&container->lock);
319 INIT_LIST_HEAD_RCU(&container->group_list);
321 container->v2 = arg == VFIO_SPAPR_TCE_v2_IOMMU;
326 static int tce_iommu_clear(struct tce_container *container,
327 struct iommu_table *tbl,
328 unsigned long entry, unsigned long pages);
329 static void tce_iommu_free_table(struct iommu_table *tbl);
331 static void tce_iommu_release(void *iommu_data)
333 struct tce_container *container = iommu_data;
334 struct iommu_table_group *table_group;
335 struct tce_iommu_group *tcegrp;
338 while (tce_groups_attached(container)) {
339 tcegrp = list_first_entry(&container->group_list,
340 struct tce_iommu_group, next);
341 table_group = iommu_group_get_iommudata(tcegrp->grp);
342 tce_iommu_detach_group(iommu_data, tcegrp->grp);
346 * If VFIO created a table, it was not disposed
347 * by tce_iommu_detach_group() so do it now.
349 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
350 struct iommu_table *tbl = container->tables[i];
355 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
356 tce_iommu_free_table(tbl);
359 tce_iommu_disable(container);
360 mutex_destroy(&container->lock);
365 static void tce_iommu_unuse_page(struct tce_container *container,
370 page = pfn_to_page(hpa >> PAGE_SHIFT);
374 static int tce_iommu_prereg_ua_to_hpa(unsigned long tce, unsigned long size,
375 unsigned long *phpa, struct mm_iommu_table_group_mem_t **pmem)
378 struct mm_iommu_table_group_mem_t *mem;
380 mem = mm_iommu_lookup(tce, size);
384 ret = mm_iommu_ua_to_hpa(mem, tce, phpa);
393 static void tce_iommu_unuse_page_v2(struct iommu_table *tbl,
396 struct mm_iommu_table_group_mem_t *mem = NULL;
398 unsigned long hpa = 0;
399 unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry);
401 if (!pua || !current || !current->mm)
404 ret = tce_iommu_prereg_ua_to_hpa(*pua, IOMMU_PAGE_SIZE(tbl),
407 pr_debug("%s: tce %lx at #%lx was not cached, ret=%d\n",
408 __func__, *pua, entry, ret);
410 mm_iommu_mapped_dec(mem);
415 static int tce_iommu_clear(struct tce_container *container,
416 struct iommu_table *tbl,
417 unsigned long entry, unsigned long pages)
419 unsigned long oldhpa;
421 enum dma_data_direction direction;
423 for ( ; pages; --pages, ++entry) {
424 direction = DMA_NONE;
426 ret = iommu_tce_xchg(tbl, entry, &oldhpa, &direction);
430 if (direction == DMA_NONE)
434 tce_iommu_unuse_page_v2(tbl, entry);
438 tce_iommu_unuse_page(container, oldhpa);
444 static int tce_iommu_use_page(unsigned long tce, unsigned long *hpa)
446 struct page *page = NULL;
447 enum dma_data_direction direction = iommu_tce_direction(tce);
449 if (get_user_pages_fast(tce & PAGE_MASK, 1,
450 direction != DMA_TO_DEVICE, &page) != 1)
453 *hpa = __pa((unsigned long) page_address(page));
458 static long tce_iommu_build(struct tce_container *container,
459 struct iommu_table *tbl,
460 unsigned long entry, unsigned long tce, unsigned long pages,
461 enum dma_data_direction direction)
466 enum dma_data_direction dirtmp;
468 for (i = 0; i < pages; ++i) {
469 unsigned long offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
471 ret = tce_iommu_use_page(tce, &hpa);
475 page = pfn_to_page(hpa >> PAGE_SHIFT);
476 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
483 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
485 tce_iommu_unuse_page(container, hpa);
486 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
487 __func__, entry << tbl->it_page_shift,
492 if (dirtmp != DMA_NONE)
493 tce_iommu_unuse_page(container, hpa);
495 tce += IOMMU_PAGE_SIZE(tbl);
499 tce_iommu_clear(container, tbl, entry, i);
504 static long tce_iommu_build_v2(struct tce_container *container,
505 struct iommu_table *tbl,
506 unsigned long entry, unsigned long tce, unsigned long pages,
507 enum dma_data_direction direction)
512 enum dma_data_direction dirtmp;
514 if (!tbl->it_userspace) {
515 ret = tce_iommu_userspace_view_alloc(tbl);
520 for (i = 0; i < pages; ++i) {
521 struct mm_iommu_table_group_mem_t *mem = NULL;
522 unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl,
525 ret = tce_iommu_prereg_ua_to_hpa(tce, IOMMU_PAGE_SIZE(tbl),
530 page = pfn_to_page(hpa >> PAGE_SHIFT);
531 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
536 /* Preserve offset within IOMMU page */
537 hpa |= tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
540 /* The registered region is being unregistered */
541 if (mm_iommu_mapped_inc(mem))
544 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
546 /* dirtmp cannot be DMA_NONE here */
547 tce_iommu_unuse_page_v2(tbl, entry + i);
548 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
549 __func__, entry << tbl->it_page_shift,
554 if (dirtmp != DMA_NONE)
555 tce_iommu_unuse_page_v2(tbl, entry + i);
559 tce += IOMMU_PAGE_SIZE(tbl);
563 tce_iommu_clear(container, tbl, entry, i);
568 static long tce_iommu_create_table(struct tce_container *container,
569 struct iommu_table_group *table_group,
574 struct iommu_table **ptbl)
576 long ret, table_size;
578 table_size = table_group->ops->get_table_size(page_shift, window_size,
583 ret = try_increment_locked_vm(table_size >> PAGE_SHIFT);
587 ret = table_group->ops->create_table(table_group, num,
588 page_shift, window_size, levels, ptbl);
590 WARN_ON(!ret && !(*ptbl)->it_ops->free);
591 WARN_ON(!ret && ((*ptbl)->it_allocated_size != table_size));
596 static void tce_iommu_free_table(struct iommu_table *tbl)
598 unsigned long pages = tbl->it_allocated_size >> PAGE_SHIFT;
600 tce_iommu_userspace_view_free(tbl);
601 tbl->it_ops->free(tbl);
602 decrement_locked_vm(pages);
605 static long tce_iommu_create_window(struct tce_container *container,
606 __u32 page_shift, __u64 window_size, __u32 levels,
609 struct tce_iommu_group *tcegrp;
610 struct iommu_table_group *table_group;
611 struct iommu_table *tbl = NULL;
614 num = tce_iommu_find_free_table(container);
618 /* Get the first group for ops::create_table */
619 tcegrp = list_first_entry(&container->group_list,
620 struct tce_iommu_group, next);
621 table_group = iommu_group_get_iommudata(tcegrp->grp);
625 if (!(table_group->pgsizes & (1ULL << page_shift)))
628 if (!table_group->ops->set_window || !table_group->ops->unset_window ||
629 !table_group->ops->get_table_size ||
630 !table_group->ops->create_table)
633 /* Create TCE table */
634 ret = tce_iommu_create_table(container, table_group, num,
635 page_shift, window_size, levels, &tbl);
639 BUG_ON(!tbl->it_ops->free);
642 * Program the table to every group.
643 * Groups have been tested for compatibility at the attach time.
645 list_for_each_entry(tcegrp, &container->group_list, next) {
646 table_group = iommu_group_get_iommudata(tcegrp->grp);
648 ret = table_group->ops->set_window(table_group, num, tbl);
653 container->tables[num] = tbl;
655 /* Return start address assigned by platform in create_table() */
656 *start_addr = tbl->it_offset << tbl->it_page_shift;
661 list_for_each_entry(tcegrp, &container->group_list, next) {
662 table_group = iommu_group_get_iommudata(tcegrp->grp);
663 table_group->ops->unset_window(table_group, num);
665 tce_iommu_free_table(tbl);
670 static long tce_iommu_remove_window(struct tce_container *container,
673 struct iommu_table_group *table_group = NULL;
674 struct iommu_table *tbl;
675 struct tce_iommu_group *tcegrp;
678 num = tce_iommu_find_table(container, start_addr, &tbl);
682 BUG_ON(!tbl->it_size);
684 /* Detach groups from IOMMUs */
685 list_for_each_entry(tcegrp, &container->group_list, next) {
686 table_group = iommu_group_get_iommudata(tcegrp->grp);
689 * SPAPR TCE IOMMU exposes the default DMA window to
690 * the guest via dma32_window_start/size of
691 * VFIO_IOMMU_SPAPR_TCE_GET_INFO. Some platforms allow
692 * the userspace to remove this window, some do not so
693 * here we check for the platform capability.
695 if (!table_group->ops || !table_group->ops->unset_window)
698 table_group->ops->unset_window(table_group, num);
702 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
703 tce_iommu_free_table(tbl);
704 container->tables[num] = NULL;
709 static long tce_iommu_ioctl(void *iommu_data,
710 unsigned int cmd, unsigned long arg)
712 struct tce_container *container = iommu_data;
713 unsigned long minsz, ddwsz;
717 case VFIO_CHECK_EXTENSION:
719 case VFIO_SPAPR_TCE_IOMMU:
720 case VFIO_SPAPR_TCE_v2_IOMMU:
724 ret = vfio_spapr_iommu_eeh_ioctl(NULL, cmd, arg);
728 return (ret < 0) ? 0 : ret;
730 case VFIO_IOMMU_SPAPR_TCE_GET_INFO: {
731 struct vfio_iommu_spapr_tce_info info;
732 struct tce_iommu_group *tcegrp;
733 struct iommu_table_group *table_group;
735 if (!tce_groups_attached(container))
738 tcegrp = list_first_entry(&container->group_list,
739 struct tce_iommu_group, next);
740 table_group = iommu_group_get_iommudata(tcegrp->grp);
745 minsz = offsetofend(struct vfio_iommu_spapr_tce_info,
748 if (copy_from_user(&info, (void __user *)arg, minsz))
751 if (info.argsz < minsz)
754 info.dma32_window_start = table_group->tce32_start;
755 info.dma32_window_size = table_group->tce32_size;
757 memset(&info.ddw, 0, sizeof(info.ddw));
759 if (table_group->max_dynamic_windows_supported &&
761 info.flags |= VFIO_IOMMU_SPAPR_INFO_DDW;
762 info.ddw.pgsizes = table_group->pgsizes;
763 info.ddw.max_dynamic_windows_supported =
764 table_group->max_dynamic_windows_supported;
765 info.ddw.levels = table_group->max_levels;
768 ddwsz = offsetofend(struct vfio_iommu_spapr_tce_info, ddw);
770 if (info.argsz >= ddwsz)
773 if (copy_to_user((void __user *)arg, &info, minsz))
778 case VFIO_IOMMU_MAP_DMA: {
779 struct vfio_iommu_type1_dma_map param;
780 struct iommu_table *tbl = NULL;
782 enum dma_data_direction direction;
784 if (!container->enabled)
787 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
789 if (copy_from_user(¶m, (void __user *)arg, minsz))
792 if (param.argsz < minsz)
795 if (param.flags & ~(VFIO_DMA_MAP_FLAG_READ |
796 VFIO_DMA_MAP_FLAG_WRITE))
799 num = tce_iommu_find_table(container, param.iova, &tbl);
803 if ((param.size & ~IOMMU_PAGE_MASK(tbl)) ||
804 (param.vaddr & ~IOMMU_PAGE_MASK(tbl)))
807 /* iova is checked by the IOMMU API */
808 if (param.flags & VFIO_DMA_MAP_FLAG_READ) {
809 if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
810 direction = DMA_BIDIRECTIONAL;
812 direction = DMA_TO_DEVICE;
814 if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
815 direction = DMA_FROM_DEVICE;
820 ret = iommu_tce_put_param_check(tbl, param.iova, param.vaddr);
825 ret = tce_iommu_build_v2(container, tbl,
826 param.iova >> tbl->it_page_shift,
828 param.size >> tbl->it_page_shift,
831 ret = tce_iommu_build(container, tbl,
832 param.iova >> tbl->it_page_shift,
834 param.size >> tbl->it_page_shift,
837 iommu_flush_tce(tbl);
841 case VFIO_IOMMU_UNMAP_DMA: {
842 struct vfio_iommu_type1_dma_unmap param;
843 struct iommu_table *tbl = NULL;
846 if (!container->enabled)
849 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap,
852 if (copy_from_user(¶m, (void __user *)arg, minsz))
855 if (param.argsz < minsz)
858 /* No flag is supported now */
862 num = tce_iommu_find_table(container, param.iova, &tbl);
866 if (param.size & ~IOMMU_PAGE_MASK(tbl))
869 ret = iommu_tce_clear_param_check(tbl, param.iova, 0,
870 param.size >> tbl->it_page_shift);
874 ret = tce_iommu_clear(container, tbl,
875 param.iova >> tbl->it_page_shift,
876 param.size >> tbl->it_page_shift);
877 iommu_flush_tce(tbl);
881 case VFIO_IOMMU_SPAPR_REGISTER_MEMORY: {
882 struct vfio_iommu_spapr_register_memory param;
887 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
890 if (copy_from_user(¶m, (void __user *)arg, minsz))
893 if (param.argsz < minsz)
896 /* No flag is supported now */
900 mutex_lock(&container->lock);
901 ret = tce_iommu_register_pages(container, param.vaddr,
903 mutex_unlock(&container->lock);
907 case VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY: {
908 struct vfio_iommu_spapr_register_memory param;
913 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
916 if (copy_from_user(¶m, (void __user *)arg, minsz))
919 if (param.argsz < minsz)
922 /* No flag is supported now */
926 mutex_lock(&container->lock);
927 ret = tce_iommu_unregister_pages(container, param.vaddr,
929 mutex_unlock(&container->lock);
933 case VFIO_IOMMU_ENABLE:
937 mutex_lock(&container->lock);
938 ret = tce_iommu_enable(container);
939 mutex_unlock(&container->lock);
943 case VFIO_IOMMU_DISABLE:
947 mutex_lock(&container->lock);
948 tce_iommu_disable(container);
949 mutex_unlock(&container->lock);
952 case VFIO_EEH_PE_OP: {
953 struct tce_iommu_group *tcegrp;
956 list_for_each_entry(tcegrp, &container->group_list, next) {
957 ret = vfio_spapr_iommu_eeh_ioctl(tcegrp->grp,
965 case VFIO_IOMMU_SPAPR_TCE_CREATE: {
966 struct vfio_iommu_spapr_tce_create create;
971 if (!tce_groups_attached(container))
974 minsz = offsetofend(struct vfio_iommu_spapr_tce_create,
977 if (copy_from_user(&create, (void __user *)arg, minsz))
980 if (create.argsz < minsz)
986 mutex_lock(&container->lock);
988 ret = tce_iommu_create_window(container, create.page_shift,
989 create.window_size, create.levels,
992 mutex_unlock(&container->lock);
994 if (!ret && copy_to_user((void __user *)arg, &create, minsz))
999 case VFIO_IOMMU_SPAPR_TCE_REMOVE: {
1000 struct vfio_iommu_spapr_tce_remove remove;
1005 if (!tce_groups_attached(container))
1008 minsz = offsetofend(struct vfio_iommu_spapr_tce_remove,
1011 if (copy_from_user(&remove, (void __user *)arg, minsz))
1014 if (remove.argsz < minsz)
1020 mutex_lock(&container->lock);
1022 ret = tce_iommu_remove_window(container, remove.start_addr);
1024 mutex_unlock(&container->lock);
1033 static void tce_iommu_release_ownership(struct tce_container *container,
1034 struct iommu_table_group *table_group)
1038 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1039 struct iommu_table *tbl = container->tables[i];
1044 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
1045 tce_iommu_userspace_view_free(tbl);
1047 iommu_release_ownership(tbl);
1049 container->tables[i] = NULL;
1053 static int tce_iommu_take_ownership(struct tce_container *container,
1054 struct iommu_table_group *table_group)
1058 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1059 struct iommu_table *tbl = table_group->tables[i];
1061 if (!tbl || !tbl->it_map)
1064 rc = iommu_take_ownership(tbl);
1066 for (j = 0; j < i; ++j)
1067 iommu_release_ownership(
1068 table_group->tables[j]);
1074 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1075 container->tables[i] = table_group->tables[i];
1080 static void tce_iommu_release_ownership_ddw(struct tce_container *container,
1081 struct iommu_table_group *table_group)
1085 if (!table_group->ops->unset_window) {
1090 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1091 table_group->ops->unset_window(table_group, i);
1093 table_group->ops->release_ownership(table_group);
1096 static long tce_iommu_take_ownership_ddw(struct tce_container *container,
1097 struct iommu_table_group *table_group)
1100 struct iommu_table *tbl = NULL;
1102 if (!table_group->ops->create_table || !table_group->ops->set_window ||
1103 !table_group->ops->release_ownership) {
1108 table_group->ops->take_ownership(table_group);
1111 * If it the first group attached, check if there is
1112 * a default DMA window and create one if none as
1113 * the userspace expects it to exist.
1115 if (!tce_groups_attached(container) && !container->tables[0]) {
1116 ret = tce_iommu_create_table(container,
1118 0, /* window number */
1119 IOMMU_PAGE_SHIFT_4K,
1120 table_group->tce32_size,
1121 1, /* default levels */
1126 container->tables[0] = tbl;
1129 /* Set all windows to the new group */
1130 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1131 tbl = container->tables[i];
1136 /* Set the default window to a new group */
1137 ret = table_group->ops->set_window(table_group, i, tbl);
1145 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1146 table_group->ops->unset_window(table_group, i);
1148 table_group->ops->release_ownership(table_group);
1153 static int tce_iommu_attach_group(void *iommu_data,
1154 struct iommu_group *iommu_group)
1157 struct tce_container *container = iommu_data;
1158 struct iommu_table_group *table_group;
1159 struct tce_iommu_group *tcegrp = NULL;
1161 mutex_lock(&container->lock);
1163 /* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
1164 iommu_group_id(iommu_group), iommu_group); */
1165 table_group = iommu_group_get_iommudata(iommu_group);
1171 if (tce_groups_attached(container) && (!table_group->ops ||
1172 !table_group->ops->take_ownership ||
1173 !table_group->ops->release_ownership)) {
1178 /* Check if new group has the same iommu_ops (i.e. compatible) */
1179 list_for_each_entry(tcegrp, &container->group_list, next) {
1180 struct iommu_table_group *table_group_tmp;
1182 if (tcegrp->grp == iommu_group) {
1183 pr_warn("tce_vfio: Group %d is already attached\n",
1184 iommu_group_id(iommu_group));
1188 table_group_tmp = iommu_group_get_iommudata(tcegrp->grp);
1189 if (table_group_tmp->ops != table_group->ops) {
1190 pr_warn("tce_vfio: Group %d is incompatible with group %d\n",
1191 iommu_group_id(iommu_group),
1192 iommu_group_id(tcegrp->grp));
1198 tcegrp = kzalloc(sizeof(*tcegrp), GFP_KERNEL);
1204 if (!table_group->ops || !table_group->ops->take_ownership ||
1205 !table_group->ops->release_ownership)
1206 ret = tce_iommu_take_ownership(container, table_group);
1208 ret = tce_iommu_take_ownership_ddw(container, table_group);
1211 tcegrp->grp = iommu_group;
1212 list_add(&tcegrp->next, &container->group_list);
1219 mutex_unlock(&container->lock);
1224 static void tce_iommu_detach_group(void *iommu_data,
1225 struct iommu_group *iommu_group)
1227 struct tce_container *container = iommu_data;
1228 struct iommu_table_group *table_group;
1230 struct tce_iommu_group *tcegrp;
1232 mutex_lock(&container->lock);
1234 list_for_each_entry(tcegrp, &container->group_list, next) {
1235 if (tcegrp->grp == iommu_group) {
1242 pr_warn("tce_vfio: detaching unattached group #%u\n",
1243 iommu_group_id(iommu_group));
1247 list_del(&tcegrp->next);
1250 table_group = iommu_group_get_iommudata(iommu_group);
1251 BUG_ON(!table_group);
1253 if (!table_group->ops || !table_group->ops->release_ownership)
1254 tce_iommu_release_ownership(container, table_group);
1256 tce_iommu_release_ownership_ddw(container, table_group);
1259 mutex_unlock(&container->lock);
1262 const struct vfio_iommu_driver_ops tce_iommu_driver_ops = {
1263 .name = "iommu-vfio-powerpc",
1264 .owner = THIS_MODULE,
1265 .open = tce_iommu_open,
1266 .release = tce_iommu_release,
1267 .ioctl = tce_iommu_ioctl,
1268 .attach_group = tce_iommu_attach_group,
1269 .detach_group = tce_iommu_detach_group,
1272 static int __init tce_iommu_init(void)
1274 return vfio_register_iommu_driver(&tce_iommu_driver_ops);
1277 static void __exit tce_iommu_cleanup(void)
1279 vfio_unregister_iommu_driver(&tce_iommu_driver_ops);
1282 module_init(tce_iommu_init);
1283 module_exit(tce_iommu_cleanup);
1285 MODULE_VERSION(DRIVER_VERSION);
1286 MODULE_LICENSE("GPL v2");
1287 MODULE_AUTHOR(DRIVER_AUTHOR);
1288 MODULE_DESCRIPTION(DRIVER_DESC);