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 tce_iommu_group *tcegrp;
337 while (tce_groups_attached(container)) {
338 tcegrp = list_first_entry(&container->group_list,
339 struct tce_iommu_group, next);
340 tce_iommu_detach_group(iommu_data, tcegrp->grp);
344 * If VFIO created a table, it was not disposed
345 * by tce_iommu_detach_group() so do it now.
347 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
348 struct iommu_table *tbl = container->tables[i];
353 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
354 tce_iommu_free_table(tbl);
357 tce_iommu_disable(container);
358 mutex_destroy(&container->lock);
363 static void tce_iommu_unuse_page(struct tce_container *container,
368 page = pfn_to_page(hpa >> PAGE_SHIFT);
372 static int tce_iommu_prereg_ua_to_hpa(unsigned long tce, unsigned long size,
373 unsigned long *phpa, struct mm_iommu_table_group_mem_t **pmem)
376 struct mm_iommu_table_group_mem_t *mem;
378 mem = mm_iommu_lookup(tce, size);
382 ret = mm_iommu_ua_to_hpa(mem, tce, phpa);
391 static void tce_iommu_unuse_page_v2(struct iommu_table *tbl,
394 struct mm_iommu_table_group_mem_t *mem = NULL;
396 unsigned long hpa = 0;
397 unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry);
399 if (!pua || !current || !current->mm)
402 ret = tce_iommu_prereg_ua_to_hpa(*pua, IOMMU_PAGE_SIZE(tbl),
405 pr_debug("%s: tce %lx at #%lx was not cached, ret=%d\n",
406 __func__, *pua, entry, ret);
408 mm_iommu_mapped_dec(mem);
413 static int tce_iommu_clear(struct tce_container *container,
414 struct iommu_table *tbl,
415 unsigned long entry, unsigned long pages)
417 unsigned long oldhpa;
419 enum dma_data_direction direction;
421 for ( ; pages; --pages, ++entry) {
422 direction = DMA_NONE;
424 ret = iommu_tce_xchg(tbl, entry, &oldhpa, &direction);
428 if (direction == DMA_NONE)
432 tce_iommu_unuse_page_v2(tbl, entry);
436 tce_iommu_unuse_page(container, oldhpa);
442 static int tce_iommu_use_page(unsigned long tce, unsigned long *hpa)
444 struct page *page = NULL;
445 enum dma_data_direction direction = iommu_tce_direction(tce);
447 if (get_user_pages_fast(tce & PAGE_MASK, 1,
448 direction != DMA_TO_DEVICE, &page) != 1)
451 *hpa = __pa((unsigned long) page_address(page));
456 static long tce_iommu_build(struct tce_container *container,
457 struct iommu_table *tbl,
458 unsigned long entry, unsigned long tce, unsigned long pages,
459 enum dma_data_direction direction)
464 enum dma_data_direction dirtmp;
466 for (i = 0; i < pages; ++i) {
467 unsigned long offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
469 ret = tce_iommu_use_page(tce, &hpa);
473 page = pfn_to_page(hpa >> PAGE_SHIFT);
474 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
481 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
483 tce_iommu_unuse_page(container, hpa);
484 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
485 __func__, entry << tbl->it_page_shift,
490 if (dirtmp != DMA_NONE)
491 tce_iommu_unuse_page(container, hpa);
493 tce += IOMMU_PAGE_SIZE(tbl);
497 tce_iommu_clear(container, tbl, entry, i);
502 static long tce_iommu_build_v2(struct tce_container *container,
503 struct iommu_table *tbl,
504 unsigned long entry, unsigned long tce, unsigned long pages,
505 enum dma_data_direction direction)
510 enum dma_data_direction dirtmp;
512 for (i = 0; i < pages; ++i) {
513 struct mm_iommu_table_group_mem_t *mem = NULL;
514 unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl,
517 ret = tce_iommu_prereg_ua_to_hpa(tce, IOMMU_PAGE_SIZE(tbl),
522 page = pfn_to_page(hpa >> PAGE_SHIFT);
523 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
528 /* Preserve offset within IOMMU page */
529 hpa |= tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
532 /* The registered region is being unregistered */
533 if (mm_iommu_mapped_inc(mem))
536 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
538 /* dirtmp cannot be DMA_NONE here */
539 tce_iommu_unuse_page_v2(tbl, entry + i);
540 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
541 __func__, entry << tbl->it_page_shift,
546 if (dirtmp != DMA_NONE)
547 tce_iommu_unuse_page_v2(tbl, entry + i);
551 tce += IOMMU_PAGE_SIZE(tbl);
555 tce_iommu_clear(container, tbl, entry, i);
560 static long tce_iommu_create_table(struct tce_container *container,
561 struct iommu_table_group *table_group,
566 struct iommu_table **ptbl)
568 long ret, table_size;
570 table_size = table_group->ops->get_table_size(page_shift, window_size,
575 ret = try_increment_locked_vm(table_size >> PAGE_SHIFT);
579 ret = table_group->ops->create_table(table_group, num,
580 page_shift, window_size, levels, ptbl);
582 WARN_ON(!ret && !(*ptbl)->it_ops->free);
583 WARN_ON(!ret && ((*ptbl)->it_allocated_size != table_size));
585 if (!ret && container->v2) {
586 ret = tce_iommu_userspace_view_alloc(*ptbl);
588 (*ptbl)->it_ops->free(*ptbl);
592 decrement_locked_vm(table_size >> PAGE_SHIFT);
597 static void tce_iommu_free_table(struct iommu_table *tbl)
599 unsigned long pages = tbl->it_allocated_size >> PAGE_SHIFT;
601 tce_iommu_userspace_view_free(tbl);
602 tbl->it_ops->free(tbl);
603 decrement_locked_vm(pages);
606 static long tce_iommu_create_window(struct tce_container *container,
607 __u32 page_shift, __u64 window_size, __u32 levels,
610 struct tce_iommu_group *tcegrp;
611 struct iommu_table_group *table_group;
612 struct iommu_table *tbl = NULL;
615 num = tce_iommu_find_free_table(container);
619 /* Get the first group for ops::create_table */
620 tcegrp = list_first_entry(&container->group_list,
621 struct tce_iommu_group, next);
622 table_group = iommu_group_get_iommudata(tcegrp->grp);
626 if (!(table_group->pgsizes & (1ULL << page_shift)))
629 if (!table_group->ops->set_window || !table_group->ops->unset_window ||
630 !table_group->ops->get_table_size ||
631 !table_group->ops->create_table)
634 /* Create TCE table */
635 ret = tce_iommu_create_table(container, table_group, num,
636 page_shift, window_size, levels, &tbl);
640 BUG_ON(!tbl->it_ops->free);
643 * Program the table to every group.
644 * Groups have been tested for compatibility at the attach time.
646 list_for_each_entry(tcegrp, &container->group_list, next) {
647 table_group = iommu_group_get_iommudata(tcegrp->grp);
649 ret = table_group->ops->set_window(table_group, num, tbl);
654 container->tables[num] = tbl;
656 /* Return start address assigned by platform in create_table() */
657 *start_addr = tbl->it_offset << tbl->it_page_shift;
662 list_for_each_entry(tcegrp, &container->group_list, next) {
663 table_group = iommu_group_get_iommudata(tcegrp->grp);
664 table_group->ops->unset_window(table_group, num);
666 tce_iommu_free_table(tbl);
671 static long tce_iommu_remove_window(struct tce_container *container,
674 struct iommu_table_group *table_group = NULL;
675 struct iommu_table *tbl;
676 struct tce_iommu_group *tcegrp;
679 num = tce_iommu_find_table(container, start_addr, &tbl);
683 BUG_ON(!tbl->it_size);
685 /* Detach groups from IOMMUs */
686 list_for_each_entry(tcegrp, &container->group_list, next) {
687 table_group = iommu_group_get_iommudata(tcegrp->grp);
690 * SPAPR TCE IOMMU exposes the default DMA window to
691 * the guest via dma32_window_start/size of
692 * VFIO_IOMMU_SPAPR_TCE_GET_INFO. Some platforms allow
693 * the userspace to remove this window, some do not so
694 * here we check for the platform capability.
696 if (!table_group->ops || !table_group->ops->unset_window)
699 table_group->ops->unset_window(table_group, num);
703 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
704 tce_iommu_free_table(tbl);
705 container->tables[num] = NULL;
710 static long tce_iommu_ioctl(void *iommu_data,
711 unsigned int cmd, unsigned long arg)
713 struct tce_container *container = iommu_data;
714 unsigned long minsz, ddwsz;
718 case VFIO_CHECK_EXTENSION:
720 case VFIO_SPAPR_TCE_IOMMU:
721 case VFIO_SPAPR_TCE_v2_IOMMU:
725 ret = vfio_spapr_iommu_eeh_ioctl(NULL, cmd, arg);
729 return (ret < 0) ? 0 : ret;
731 case VFIO_IOMMU_SPAPR_TCE_GET_INFO: {
732 struct vfio_iommu_spapr_tce_info info;
733 struct tce_iommu_group *tcegrp;
734 struct iommu_table_group *table_group;
736 if (!tce_groups_attached(container))
739 tcegrp = list_first_entry(&container->group_list,
740 struct tce_iommu_group, next);
741 table_group = iommu_group_get_iommudata(tcegrp->grp);
746 minsz = offsetofend(struct vfio_iommu_spapr_tce_info,
749 if (copy_from_user(&info, (void __user *)arg, minsz))
752 if (info.argsz < minsz)
755 info.dma32_window_start = table_group->tce32_start;
756 info.dma32_window_size = table_group->tce32_size;
758 memset(&info.ddw, 0, sizeof(info.ddw));
760 if (table_group->max_dynamic_windows_supported &&
762 info.flags |= VFIO_IOMMU_SPAPR_INFO_DDW;
763 info.ddw.pgsizes = table_group->pgsizes;
764 info.ddw.max_dynamic_windows_supported =
765 table_group->max_dynamic_windows_supported;
766 info.ddw.levels = table_group->max_levels;
769 ddwsz = offsetofend(struct vfio_iommu_spapr_tce_info, ddw);
771 if (info.argsz >= ddwsz)
774 if (copy_to_user((void __user *)arg, &info, minsz))
779 case VFIO_IOMMU_MAP_DMA: {
780 struct vfio_iommu_type1_dma_map param;
781 struct iommu_table *tbl = NULL;
783 enum dma_data_direction direction;
785 if (!container->enabled)
788 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
790 if (copy_from_user(¶m, (void __user *)arg, minsz))
793 if (param.argsz < minsz)
796 if (param.flags & ~(VFIO_DMA_MAP_FLAG_READ |
797 VFIO_DMA_MAP_FLAG_WRITE))
800 num = tce_iommu_find_table(container, param.iova, &tbl);
804 if ((param.size & ~IOMMU_PAGE_MASK(tbl)) ||
805 (param.vaddr & ~IOMMU_PAGE_MASK(tbl)))
808 /* iova is checked by the IOMMU API */
809 if (param.flags & VFIO_DMA_MAP_FLAG_READ) {
810 if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
811 direction = DMA_BIDIRECTIONAL;
813 direction = DMA_TO_DEVICE;
815 if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
816 direction = DMA_FROM_DEVICE;
821 ret = iommu_tce_put_param_check(tbl, param.iova, param.vaddr);
826 ret = tce_iommu_build_v2(container, tbl,
827 param.iova >> tbl->it_page_shift,
829 param.size >> tbl->it_page_shift,
832 ret = tce_iommu_build(container, tbl,
833 param.iova >> tbl->it_page_shift,
835 param.size >> tbl->it_page_shift,
838 iommu_flush_tce(tbl);
842 case VFIO_IOMMU_UNMAP_DMA: {
843 struct vfio_iommu_type1_dma_unmap param;
844 struct iommu_table *tbl = NULL;
847 if (!container->enabled)
850 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap,
853 if (copy_from_user(¶m, (void __user *)arg, minsz))
856 if (param.argsz < minsz)
859 /* No flag is supported now */
863 num = tce_iommu_find_table(container, param.iova, &tbl);
867 if (param.size & ~IOMMU_PAGE_MASK(tbl))
870 ret = iommu_tce_clear_param_check(tbl, param.iova, 0,
871 param.size >> tbl->it_page_shift);
875 ret = tce_iommu_clear(container, tbl,
876 param.iova >> tbl->it_page_shift,
877 param.size >> tbl->it_page_shift);
878 iommu_flush_tce(tbl);
882 case VFIO_IOMMU_SPAPR_REGISTER_MEMORY: {
883 struct vfio_iommu_spapr_register_memory param;
888 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
891 if (copy_from_user(¶m, (void __user *)arg, minsz))
894 if (param.argsz < minsz)
897 /* No flag is supported now */
901 mutex_lock(&container->lock);
902 ret = tce_iommu_register_pages(container, param.vaddr,
904 mutex_unlock(&container->lock);
908 case VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY: {
909 struct vfio_iommu_spapr_register_memory param;
914 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
917 if (copy_from_user(¶m, (void __user *)arg, minsz))
920 if (param.argsz < minsz)
923 /* No flag is supported now */
927 mutex_lock(&container->lock);
928 ret = tce_iommu_unregister_pages(container, param.vaddr,
930 mutex_unlock(&container->lock);
934 case VFIO_IOMMU_ENABLE:
938 mutex_lock(&container->lock);
939 ret = tce_iommu_enable(container);
940 mutex_unlock(&container->lock);
944 case VFIO_IOMMU_DISABLE:
948 mutex_lock(&container->lock);
949 tce_iommu_disable(container);
950 mutex_unlock(&container->lock);
953 case VFIO_EEH_PE_OP: {
954 struct tce_iommu_group *tcegrp;
957 list_for_each_entry(tcegrp, &container->group_list, next) {
958 ret = vfio_spapr_iommu_eeh_ioctl(tcegrp->grp,
966 case VFIO_IOMMU_SPAPR_TCE_CREATE: {
967 struct vfio_iommu_spapr_tce_create create;
972 if (!tce_groups_attached(container))
975 minsz = offsetofend(struct vfio_iommu_spapr_tce_create,
978 if (copy_from_user(&create, (void __user *)arg, minsz))
981 if (create.argsz < minsz)
987 mutex_lock(&container->lock);
989 ret = tce_iommu_create_window(container, create.page_shift,
990 create.window_size, create.levels,
993 mutex_unlock(&container->lock);
995 if (!ret && copy_to_user((void __user *)arg, &create, minsz))
1000 case VFIO_IOMMU_SPAPR_TCE_REMOVE: {
1001 struct vfio_iommu_spapr_tce_remove remove;
1006 if (!tce_groups_attached(container))
1009 minsz = offsetofend(struct vfio_iommu_spapr_tce_remove,
1012 if (copy_from_user(&remove, (void __user *)arg, minsz))
1015 if (remove.argsz < minsz)
1021 mutex_lock(&container->lock);
1023 ret = tce_iommu_remove_window(container, remove.start_addr);
1025 mutex_unlock(&container->lock);
1034 static void tce_iommu_release_ownership(struct tce_container *container,
1035 struct iommu_table_group *table_group)
1039 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1040 struct iommu_table *tbl = container->tables[i];
1045 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
1046 tce_iommu_userspace_view_free(tbl);
1048 iommu_release_ownership(tbl);
1050 container->tables[i] = NULL;
1054 static int tce_iommu_take_ownership(struct tce_container *container,
1055 struct iommu_table_group *table_group)
1059 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1060 struct iommu_table *tbl = table_group->tables[i];
1062 if (!tbl || !tbl->it_map)
1065 rc = tce_iommu_userspace_view_alloc(tbl);
1067 rc = iommu_take_ownership(tbl);
1070 for (j = 0; j < i; ++j)
1071 iommu_release_ownership(
1072 table_group->tables[j]);
1078 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1079 container->tables[i] = table_group->tables[i];
1084 static void tce_iommu_release_ownership_ddw(struct tce_container *container,
1085 struct iommu_table_group *table_group)
1089 if (!table_group->ops->unset_window) {
1094 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1095 table_group->ops->unset_window(table_group, i);
1097 table_group->ops->release_ownership(table_group);
1100 static long tce_iommu_take_ownership_ddw(struct tce_container *container,
1101 struct iommu_table_group *table_group)
1104 struct iommu_table *tbl = NULL;
1106 if (!table_group->ops->create_table || !table_group->ops->set_window ||
1107 !table_group->ops->release_ownership) {
1112 table_group->ops->take_ownership(table_group);
1115 * If it the first group attached, check if there is
1116 * a default DMA window and create one if none as
1117 * the userspace expects it to exist.
1119 if (!tce_groups_attached(container) && !container->tables[0]) {
1120 ret = tce_iommu_create_table(container,
1122 0, /* window number */
1123 IOMMU_PAGE_SHIFT_4K,
1124 table_group->tce32_size,
1125 1, /* default levels */
1130 container->tables[0] = tbl;
1133 /* Set all windows to the new group */
1134 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1135 tbl = container->tables[i];
1140 /* Set the default window to a new group */
1141 ret = table_group->ops->set_window(table_group, i, tbl);
1149 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1150 table_group->ops->unset_window(table_group, i);
1152 table_group->ops->release_ownership(table_group);
1157 static int tce_iommu_attach_group(void *iommu_data,
1158 struct iommu_group *iommu_group)
1161 struct tce_container *container = iommu_data;
1162 struct iommu_table_group *table_group;
1163 struct tce_iommu_group *tcegrp = NULL;
1165 mutex_lock(&container->lock);
1167 /* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
1168 iommu_group_id(iommu_group), iommu_group); */
1169 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->create_table !=
1190 table_group->ops->create_table) {
1191 pr_warn("tce_vfio: Group %d is incompatible with group %d\n",
1192 iommu_group_id(iommu_group),
1193 iommu_group_id(tcegrp->grp));
1199 tcegrp = kzalloc(sizeof(*tcegrp), GFP_KERNEL);
1205 if (!table_group->ops || !table_group->ops->take_ownership ||
1206 !table_group->ops->release_ownership)
1207 ret = tce_iommu_take_ownership(container, table_group);
1209 ret = tce_iommu_take_ownership_ddw(container, table_group);
1212 tcegrp->grp = iommu_group;
1213 list_add(&tcegrp->next, &container->group_list);
1220 mutex_unlock(&container->lock);
1225 static void tce_iommu_detach_group(void *iommu_data,
1226 struct iommu_group *iommu_group)
1228 struct tce_container *container = iommu_data;
1229 struct iommu_table_group *table_group;
1231 struct tce_iommu_group *tcegrp;
1233 mutex_lock(&container->lock);
1235 list_for_each_entry(tcegrp, &container->group_list, next) {
1236 if (tcegrp->grp == iommu_group) {
1243 pr_warn("tce_vfio: detaching unattached group #%u\n",
1244 iommu_group_id(iommu_group));
1248 list_del(&tcegrp->next);
1251 table_group = iommu_group_get_iommudata(iommu_group);
1252 BUG_ON(!table_group);
1254 if (!table_group->ops || !table_group->ops->release_ownership)
1255 tce_iommu_release_ownership(container, table_group);
1257 tce_iommu_release_ownership_ddw(container, table_group);
1260 mutex_unlock(&container->lock);
1263 const struct vfio_iommu_driver_ops tce_iommu_driver_ops = {
1264 .name = "iommu-vfio-powerpc",
1265 .owner = THIS_MODULE,
1266 .open = tce_iommu_open,
1267 .release = tce_iommu_release,
1268 .ioctl = tce_iommu_ioctl,
1269 .attach_group = tce_iommu_attach_group,
1270 .detach_group = tce_iommu_detach_group,
1273 static int __init tce_iommu_init(void)
1275 return vfio_register_iommu_driver(&tce_iommu_driver_ops);
1278 static void __exit tce_iommu_cleanup(void)
1280 vfio_unregister_iommu_driver(&tce_iommu_driver_ops);
1283 module_init(tce_iommu_init);
1284 module_exit(tce_iommu_cleanup);
1286 MODULE_VERSION(DRIVER_VERSION);
1287 MODULE_LICENSE("GPL v2");
1288 MODULE_AUTHOR(DRIVER_AUTHOR);
1289 MODULE_DESCRIPTION(DRIVER_DESC);