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(struct mm_struct *mm, long npages)
36 long ret = 0, locked, lock_limit;
38 if (WARN_ON_ONCE(!mm))
44 down_write(&mm->mmap_sem);
45 locked = mm->locked_vm + npages;
46 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
47 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
50 mm->locked_vm += npages;
52 pr_debug("[%d] RLIMIT_MEMLOCK +%ld %ld/%ld%s\n", current->pid,
54 mm->locked_vm << PAGE_SHIFT,
55 rlimit(RLIMIT_MEMLOCK),
56 ret ? " - exceeded" : "");
58 up_write(&mm->mmap_sem);
63 static void decrement_locked_vm(struct mm_struct *mm, long npages)
68 down_write(&mm->mmap_sem);
69 if (WARN_ON_ONCE(npages > mm->locked_vm))
70 npages = mm->locked_vm;
71 mm->locked_vm -= npages;
72 pr_debug("[%d] RLIMIT_MEMLOCK -%ld %ld/%ld\n", current->pid,
74 mm->locked_vm << PAGE_SHIFT,
75 rlimit(RLIMIT_MEMLOCK));
76 up_write(&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 * A container needs to remember which preregistered region it has
93 * referenced to do proper cleanup at the userspace process exit.
95 struct tce_iommu_prereg {
96 struct list_head next;
97 struct mm_iommu_table_group_mem_t *mem;
101 * The container descriptor supports only a single group per container.
102 * Required by the API as the container is not supplied with the IOMMU group
103 * at the moment of initialization.
105 struct tce_container {
109 bool def_window_pending;
110 unsigned long locked_pages;
111 struct mm_struct *mm;
112 struct iommu_table *tables[IOMMU_TABLE_GROUP_MAX_TABLES];
113 struct list_head group_list;
114 struct list_head prereg_list;
117 static long tce_iommu_mm_set(struct tce_container *container)
120 if (container->mm == current->mm)
124 BUG_ON(!current->mm);
125 container->mm = current->mm;
126 atomic_inc(&container->mm->mm_count);
131 static long tce_iommu_prereg_free(struct tce_container *container,
132 struct tce_iommu_prereg *tcemem)
136 ret = mm_iommu_put(container->mm, tcemem->mem);
140 list_del(&tcemem->next);
146 static long tce_iommu_unregister_pages(struct tce_container *container,
147 __u64 vaddr, __u64 size)
149 struct mm_iommu_table_group_mem_t *mem;
150 struct tce_iommu_prereg *tcemem;
153 if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK))
156 mem = mm_iommu_find(container->mm, vaddr, size >> PAGE_SHIFT);
160 list_for_each_entry(tcemem, &container->prereg_list, next) {
161 if (tcemem->mem == mem) {
170 return tce_iommu_prereg_free(container, tcemem);
173 static long tce_iommu_register_pages(struct tce_container *container,
174 __u64 vaddr, __u64 size)
177 struct mm_iommu_table_group_mem_t *mem = NULL;
178 struct tce_iommu_prereg *tcemem;
179 unsigned long entries = size >> PAGE_SHIFT;
181 if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK) ||
182 ((vaddr + size) < vaddr))
185 mem = mm_iommu_find(container->mm, vaddr, entries);
187 list_for_each_entry(tcemem, &container->prereg_list, next) {
188 if (tcemem->mem == mem)
193 ret = mm_iommu_get(container->mm, vaddr, entries, &mem);
197 tcemem = kzalloc(sizeof(*tcemem), GFP_KERNEL);
199 list_add(&tcemem->next, &container->prereg_list);
201 container->enabled = true;
206 static long tce_iommu_userspace_view_alloc(struct iommu_table *tbl,
207 struct mm_struct *mm)
209 unsigned long cb = _ALIGN_UP(sizeof(tbl->it_userspace[0]) *
210 tbl->it_size, PAGE_SIZE);
214 BUG_ON(tbl->it_userspace);
216 ret = try_increment_locked_vm(mm, cb >> PAGE_SHIFT);
222 decrement_locked_vm(mm, cb >> PAGE_SHIFT);
225 tbl->it_userspace = uas;
230 static void tce_iommu_userspace_view_free(struct iommu_table *tbl,
231 struct mm_struct *mm)
233 unsigned long cb = _ALIGN_UP(sizeof(tbl->it_userspace[0]) *
234 tbl->it_size, PAGE_SIZE);
236 if (!tbl->it_userspace)
239 vfree(tbl->it_userspace);
240 tbl->it_userspace = NULL;
241 decrement_locked_vm(mm, cb >> PAGE_SHIFT);
244 static bool tce_page_is_contained(struct page *page, unsigned page_shift)
247 * Check that the TCE table granularity is not bigger than the size of
248 * a page we just found. Otherwise the hardware can get access to
249 * a bigger memory chunk that it should.
251 return (PAGE_SHIFT + compound_order(compound_head(page))) >= page_shift;
254 static inline bool tce_groups_attached(struct tce_container *container)
256 return !list_empty(&container->group_list);
259 static long tce_iommu_find_table(struct tce_container *container,
260 phys_addr_t ioba, struct iommu_table **ptbl)
264 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
265 struct iommu_table *tbl = container->tables[i];
268 unsigned long entry = ioba >> tbl->it_page_shift;
269 unsigned long start = tbl->it_offset;
270 unsigned long end = start + tbl->it_size;
272 if ((start <= entry) && (entry < end)) {
282 static int tce_iommu_find_free_table(struct tce_container *container)
286 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
287 if (!container->tables[i])
294 static int tce_iommu_enable(struct tce_container *container)
297 unsigned long locked;
298 struct iommu_table_group *table_group;
299 struct tce_iommu_group *tcegrp;
301 if (container->enabled)
305 * When userspace pages are mapped into the IOMMU, they are effectively
306 * locked memory, so, theoretically, we need to update the accounting
307 * of locked pages on each map and unmap. For powerpc, the map unmap
308 * paths can be very hot, though, and the accounting would kill
309 * performance, especially since it would be difficult to impossible
310 * to handle the accounting in real mode only.
312 * To address that, rather than precisely accounting every page, we
313 * instead account for a worst case on locked memory when the iommu is
314 * enabled and disabled. The worst case upper bound on locked memory
315 * is the size of the whole iommu window, which is usually relatively
316 * small (compared to total memory sizes) on POWER hardware.
318 * Also we don't have a nice way to fail on H_PUT_TCE due to ulimits,
319 * that would effectively kill the guest at random points, much better
320 * enforcing the limit based on the max that the guest can map.
322 * Unfortunately at the moment it counts whole tables, no matter how
323 * much memory the guest has. I.e. for 4GB guest and 4 IOMMU groups
324 * each with 2GB DMA window, 8GB will be counted here. The reason for
325 * this is that we cannot tell here the amount of RAM used by the guest
326 * as this information is only available from KVM and VFIO is
329 * So we do not allow enabling a container without a group attached
330 * as there is no way to know how much we should increment
331 * the locked_vm counter.
333 if (!tce_groups_attached(container))
336 tcegrp = list_first_entry(&container->group_list,
337 struct tce_iommu_group, next);
338 table_group = iommu_group_get_iommudata(tcegrp->grp);
342 if (!table_group->tce32_size)
345 ret = tce_iommu_mm_set(container);
349 locked = table_group->tce32_size >> PAGE_SHIFT;
350 ret = try_increment_locked_vm(container->mm, locked);
354 container->locked_pages = locked;
356 container->enabled = true;
361 static void tce_iommu_disable(struct tce_container *container)
363 if (!container->enabled)
366 container->enabled = false;
368 BUG_ON(!container->mm);
369 decrement_locked_vm(container->mm, container->locked_pages);
372 static void *tce_iommu_open(unsigned long arg)
374 struct tce_container *container;
376 if ((arg != VFIO_SPAPR_TCE_IOMMU) && (arg != VFIO_SPAPR_TCE_v2_IOMMU)) {
377 pr_err("tce_vfio: Wrong IOMMU type\n");
378 return ERR_PTR(-EINVAL);
381 container = kzalloc(sizeof(*container), GFP_KERNEL);
383 return ERR_PTR(-ENOMEM);
385 mutex_init(&container->lock);
386 INIT_LIST_HEAD_RCU(&container->group_list);
387 INIT_LIST_HEAD_RCU(&container->prereg_list);
389 container->v2 = arg == VFIO_SPAPR_TCE_v2_IOMMU;
394 static int tce_iommu_clear(struct tce_container *container,
395 struct iommu_table *tbl,
396 unsigned long entry, unsigned long pages);
397 static void tce_iommu_free_table(struct tce_container *container,
398 struct iommu_table *tbl);
400 static void tce_iommu_release(void *iommu_data)
402 struct tce_container *container = iommu_data;
403 struct tce_iommu_group *tcegrp;
406 while (tce_groups_attached(container)) {
407 tcegrp = list_first_entry(&container->group_list,
408 struct tce_iommu_group, next);
409 tce_iommu_detach_group(iommu_data, tcegrp->grp);
413 * If VFIO created a table, it was not disposed
414 * by tce_iommu_detach_group() so do it now.
416 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
417 struct iommu_table *tbl = container->tables[i];
422 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
423 tce_iommu_free_table(container, tbl);
426 while (!list_empty(&container->prereg_list)) {
427 struct tce_iommu_prereg *tcemem;
429 tcemem = list_first_entry(&container->prereg_list,
430 struct tce_iommu_prereg, next);
431 WARN_ON_ONCE(tce_iommu_prereg_free(container, tcemem));
434 tce_iommu_disable(container);
436 mmdrop(container->mm);
437 mutex_destroy(&container->lock);
442 static void tce_iommu_unuse_page(struct tce_container *container,
447 page = pfn_to_page(hpa >> PAGE_SHIFT);
451 static int tce_iommu_prereg_ua_to_hpa(struct tce_container *container,
452 unsigned long tce, unsigned long size,
453 unsigned long *phpa, struct mm_iommu_table_group_mem_t **pmem)
456 struct mm_iommu_table_group_mem_t *mem;
458 mem = mm_iommu_lookup(container->mm, tce, size);
462 ret = mm_iommu_ua_to_hpa(mem, tce, phpa);
471 static void tce_iommu_unuse_page_v2(struct tce_container *container,
472 struct iommu_table *tbl, unsigned long entry)
474 struct mm_iommu_table_group_mem_t *mem = NULL;
476 unsigned long hpa = 0;
477 unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry);
482 ret = tce_iommu_prereg_ua_to_hpa(container, *pua, IOMMU_PAGE_SIZE(tbl),
485 pr_debug("%s: tce %lx at #%lx was not cached, ret=%d\n",
486 __func__, *pua, entry, ret);
488 mm_iommu_mapped_dec(mem);
493 static int tce_iommu_clear(struct tce_container *container,
494 struct iommu_table *tbl,
495 unsigned long entry, unsigned long pages)
497 unsigned long oldhpa;
499 enum dma_data_direction direction;
501 for ( ; pages; --pages, ++entry) {
502 direction = DMA_NONE;
504 ret = iommu_tce_xchg(tbl, entry, &oldhpa, &direction);
508 if (direction == DMA_NONE)
512 tce_iommu_unuse_page_v2(container, tbl, entry);
516 tce_iommu_unuse_page(container, oldhpa);
522 static int tce_iommu_use_page(unsigned long tce, unsigned long *hpa)
524 struct page *page = NULL;
525 enum dma_data_direction direction = iommu_tce_direction(tce);
527 if (get_user_pages_fast(tce & PAGE_MASK, 1,
528 direction != DMA_TO_DEVICE, &page) != 1)
531 *hpa = __pa((unsigned long) page_address(page));
536 static long tce_iommu_build(struct tce_container *container,
537 struct iommu_table *tbl,
538 unsigned long entry, unsigned long tce, unsigned long pages,
539 enum dma_data_direction direction)
544 enum dma_data_direction dirtmp;
546 for (i = 0; i < pages; ++i) {
547 unsigned long offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
549 ret = tce_iommu_use_page(tce, &hpa);
553 page = pfn_to_page(hpa >> PAGE_SHIFT);
554 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
561 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
563 tce_iommu_unuse_page(container, hpa);
564 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
565 __func__, entry << tbl->it_page_shift,
570 if (dirtmp != DMA_NONE)
571 tce_iommu_unuse_page(container, hpa);
573 tce += IOMMU_PAGE_SIZE(tbl);
577 tce_iommu_clear(container, tbl, entry, i);
582 static long tce_iommu_build_v2(struct tce_container *container,
583 struct iommu_table *tbl,
584 unsigned long entry, unsigned long tce, unsigned long pages,
585 enum dma_data_direction direction)
590 enum dma_data_direction dirtmp;
592 if (!tbl->it_userspace) {
593 ret = tce_iommu_userspace_view_alloc(tbl, container->mm);
598 for (i = 0; i < pages; ++i) {
599 struct mm_iommu_table_group_mem_t *mem = NULL;
600 unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl,
603 ret = tce_iommu_prereg_ua_to_hpa(container,
604 tce, IOMMU_PAGE_SIZE(tbl), &hpa, &mem);
608 page = pfn_to_page(hpa >> PAGE_SHIFT);
609 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
614 /* Preserve offset within IOMMU page */
615 hpa |= tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
618 /* The registered region is being unregistered */
619 if (mm_iommu_mapped_inc(mem))
622 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
624 /* dirtmp cannot be DMA_NONE here */
625 tce_iommu_unuse_page_v2(container, tbl, entry + i);
626 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
627 __func__, entry << tbl->it_page_shift,
632 if (dirtmp != DMA_NONE)
633 tce_iommu_unuse_page_v2(container, tbl, entry + i);
637 tce += IOMMU_PAGE_SIZE(tbl);
641 tce_iommu_clear(container, tbl, entry, i);
646 static long tce_iommu_create_table(struct tce_container *container,
647 struct iommu_table_group *table_group,
652 struct iommu_table **ptbl)
654 long ret, table_size;
656 table_size = table_group->ops->get_table_size(page_shift, window_size,
661 ret = try_increment_locked_vm(container->mm, table_size >> PAGE_SHIFT);
665 ret = table_group->ops->create_table(table_group, num,
666 page_shift, window_size, levels, ptbl);
668 WARN_ON(!ret && !(*ptbl)->it_ops->free);
669 WARN_ON(!ret && ((*ptbl)->it_allocated_size != table_size));
674 static void tce_iommu_free_table(struct tce_container *container,
675 struct iommu_table *tbl)
677 unsigned long pages = tbl->it_allocated_size >> PAGE_SHIFT;
679 tce_iommu_userspace_view_free(tbl, container->mm);
680 tbl->it_ops->free(tbl);
681 decrement_locked_vm(container->mm, pages);
684 static long tce_iommu_create_window(struct tce_container *container,
685 __u32 page_shift, __u64 window_size, __u32 levels,
688 struct tce_iommu_group *tcegrp;
689 struct iommu_table_group *table_group;
690 struct iommu_table *tbl = NULL;
693 num = tce_iommu_find_free_table(container);
697 /* Get the first group for ops::create_table */
698 tcegrp = list_first_entry(&container->group_list,
699 struct tce_iommu_group, next);
700 table_group = iommu_group_get_iommudata(tcegrp->grp);
704 if (!(table_group->pgsizes & (1ULL << page_shift)))
707 if (!table_group->ops->set_window || !table_group->ops->unset_window ||
708 !table_group->ops->get_table_size ||
709 !table_group->ops->create_table)
712 /* Create TCE table */
713 ret = tce_iommu_create_table(container, table_group, num,
714 page_shift, window_size, levels, &tbl);
718 BUG_ON(!tbl->it_ops->free);
721 * Program the table to every group.
722 * Groups have been tested for compatibility at the attach time.
724 list_for_each_entry(tcegrp, &container->group_list, next) {
725 table_group = iommu_group_get_iommudata(tcegrp->grp);
727 ret = table_group->ops->set_window(table_group, num, tbl);
732 container->tables[num] = tbl;
734 /* Return start address assigned by platform in create_table() */
735 *start_addr = tbl->it_offset << tbl->it_page_shift;
740 list_for_each_entry(tcegrp, &container->group_list, next) {
741 table_group = iommu_group_get_iommudata(tcegrp->grp);
742 table_group->ops->unset_window(table_group, num);
744 tce_iommu_free_table(container, tbl);
749 static long tce_iommu_remove_window(struct tce_container *container,
752 struct iommu_table_group *table_group = NULL;
753 struct iommu_table *tbl;
754 struct tce_iommu_group *tcegrp;
757 num = tce_iommu_find_table(container, start_addr, &tbl);
761 BUG_ON(!tbl->it_size);
763 /* Detach groups from IOMMUs */
764 list_for_each_entry(tcegrp, &container->group_list, next) {
765 table_group = iommu_group_get_iommudata(tcegrp->grp);
768 * SPAPR TCE IOMMU exposes the default DMA window to
769 * the guest via dma32_window_start/size of
770 * VFIO_IOMMU_SPAPR_TCE_GET_INFO. Some platforms allow
771 * the userspace to remove this window, some do not so
772 * here we check for the platform capability.
774 if (!table_group->ops || !table_group->ops->unset_window)
777 table_group->ops->unset_window(table_group, num);
781 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
782 tce_iommu_free_table(container, tbl);
783 container->tables[num] = NULL;
788 static long tce_iommu_create_default_window(struct tce_container *container)
791 __u64 start_addr = 0;
792 struct tce_iommu_group *tcegrp;
793 struct iommu_table_group *table_group;
795 if (!container->def_window_pending)
798 if (!tce_groups_attached(container))
801 tcegrp = list_first_entry(&container->group_list,
802 struct tce_iommu_group, next);
803 table_group = iommu_group_get_iommudata(tcegrp->grp);
807 ret = tce_iommu_create_window(container, IOMMU_PAGE_SHIFT_4K,
808 table_group->tce32_size, 1, &start_addr);
809 WARN_ON_ONCE(!ret && start_addr);
812 container->def_window_pending = false;
817 static long tce_iommu_ioctl(void *iommu_data,
818 unsigned int cmd, unsigned long arg)
820 struct tce_container *container = iommu_data;
821 unsigned long minsz, ddwsz;
825 case VFIO_CHECK_EXTENSION:
827 case VFIO_SPAPR_TCE_IOMMU:
828 case VFIO_SPAPR_TCE_v2_IOMMU:
832 ret = vfio_spapr_iommu_eeh_ioctl(NULL, cmd, arg);
836 return (ret < 0) ? 0 : ret;
840 * Sanity check to prevent one userspace from manipulating
841 * another userspace mm.
844 if (container->mm && container->mm != current->mm)
848 case VFIO_IOMMU_SPAPR_TCE_GET_INFO: {
849 struct vfio_iommu_spapr_tce_info info;
850 struct tce_iommu_group *tcegrp;
851 struct iommu_table_group *table_group;
853 if (!tce_groups_attached(container))
856 tcegrp = list_first_entry(&container->group_list,
857 struct tce_iommu_group, next);
858 table_group = iommu_group_get_iommudata(tcegrp->grp);
863 minsz = offsetofend(struct vfio_iommu_spapr_tce_info,
866 if (copy_from_user(&info, (void __user *)arg, minsz))
869 if (info.argsz < minsz)
872 info.dma32_window_start = table_group->tce32_start;
873 info.dma32_window_size = table_group->tce32_size;
875 memset(&info.ddw, 0, sizeof(info.ddw));
877 if (table_group->max_dynamic_windows_supported &&
879 info.flags |= VFIO_IOMMU_SPAPR_INFO_DDW;
880 info.ddw.pgsizes = table_group->pgsizes;
881 info.ddw.max_dynamic_windows_supported =
882 table_group->max_dynamic_windows_supported;
883 info.ddw.levels = table_group->max_levels;
886 ddwsz = offsetofend(struct vfio_iommu_spapr_tce_info, ddw);
888 if (info.argsz >= ddwsz)
891 if (copy_to_user((void __user *)arg, &info, minsz))
896 case VFIO_IOMMU_MAP_DMA: {
897 struct vfio_iommu_type1_dma_map param;
898 struct iommu_table *tbl = NULL;
900 enum dma_data_direction direction;
902 if (!container->enabled)
905 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
907 if (copy_from_user(¶m, (void __user *)arg, minsz))
910 if (param.argsz < minsz)
913 if (param.flags & ~(VFIO_DMA_MAP_FLAG_READ |
914 VFIO_DMA_MAP_FLAG_WRITE))
917 ret = tce_iommu_create_default_window(container);
921 num = tce_iommu_find_table(container, param.iova, &tbl);
925 if ((param.size & ~IOMMU_PAGE_MASK(tbl)) ||
926 (param.vaddr & ~IOMMU_PAGE_MASK(tbl)))
929 /* iova is checked by the IOMMU API */
930 if (param.flags & VFIO_DMA_MAP_FLAG_READ) {
931 if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
932 direction = DMA_BIDIRECTIONAL;
934 direction = DMA_TO_DEVICE;
936 if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
937 direction = DMA_FROM_DEVICE;
942 ret = iommu_tce_put_param_check(tbl, param.iova, param.vaddr);
947 ret = tce_iommu_build_v2(container, tbl,
948 param.iova >> tbl->it_page_shift,
950 param.size >> tbl->it_page_shift,
953 ret = tce_iommu_build(container, tbl,
954 param.iova >> tbl->it_page_shift,
956 param.size >> tbl->it_page_shift,
959 iommu_flush_tce(tbl);
963 case VFIO_IOMMU_UNMAP_DMA: {
964 struct vfio_iommu_type1_dma_unmap param;
965 struct iommu_table *tbl = NULL;
968 if (!container->enabled)
971 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap,
974 if (copy_from_user(¶m, (void __user *)arg, minsz))
977 if (param.argsz < minsz)
980 /* No flag is supported now */
984 ret = tce_iommu_create_default_window(container);
988 num = tce_iommu_find_table(container, param.iova, &tbl);
992 if (param.size & ~IOMMU_PAGE_MASK(tbl))
995 ret = iommu_tce_clear_param_check(tbl, param.iova, 0,
996 param.size >> tbl->it_page_shift);
1000 ret = tce_iommu_clear(container, tbl,
1001 param.iova >> tbl->it_page_shift,
1002 param.size >> tbl->it_page_shift);
1003 iommu_flush_tce(tbl);
1007 case VFIO_IOMMU_SPAPR_REGISTER_MEMORY: {
1008 struct vfio_iommu_spapr_register_memory param;
1013 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1016 ret = tce_iommu_mm_set(container);
1020 if (copy_from_user(¶m, (void __user *)arg, minsz))
1023 if (param.argsz < minsz)
1026 /* No flag is supported now */
1030 mutex_lock(&container->lock);
1031 ret = tce_iommu_register_pages(container, param.vaddr,
1033 mutex_unlock(&container->lock);
1037 case VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY: {
1038 struct vfio_iommu_spapr_register_memory param;
1046 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1049 if (copy_from_user(¶m, (void __user *)arg, minsz))
1052 if (param.argsz < minsz)
1055 /* No flag is supported now */
1059 mutex_lock(&container->lock);
1060 ret = tce_iommu_unregister_pages(container, param.vaddr,
1062 mutex_unlock(&container->lock);
1066 case VFIO_IOMMU_ENABLE:
1070 mutex_lock(&container->lock);
1071 ret = tce_iommu_enable(container);
1072 mutex_unlock(&container->lock);
1076 case VFIO_IOMMU_DISABLE:
1080 mutex_lock(&container->lock);
1081 tce_iommu_disable(container);
1082 mutex_unlock(&container->lock);
1085 case VFIO_EEH_PE_OP: {
1086 struct tce_iommu_group *tcegrp;
1089 list_for_each_entry(tcegrp, &container->group_list, next) {
1090 ret = vfio_spapr_iommu_eeh_ioctl(tcegrp->grp,
1098 case VFIO_IOMMU_SPAPR_TCE_CREATE: {
1099 struct vfio_iommu_spapr_tce_create create;
1104 ret = tce_iommu_mm_set(container);
1108 if (!tce_groups_attached(container))
1111 minsz = offsetofend(struct vfio_iommu_spapr_tce_create,
1114 if (copy_from_user(&create, (void __user *)arg, minsz))
1117 if (create.argsz < minsz)
1123 mutex_lock(&container->lock);
1125 ret = tce_iommu_create_default_window(container);
1129 ret = tce_iommu_create_window(container, create.page_shift,
1130 create.window_size, create.levels,
1131 &create.start_addr);
1133 mutex_unlock(&container->lock);
1135 if (!ret && copy_to_user((void __user *)arg, &create, minsz))
1140 case VFIO_IOMMU_SPAPR_TCE_REMOVE: {
1141 struct vfio_iommu_spapr_tce_remove remove;
1146 ret = tce_iommu_mm_set(container);
1150 if (!tce_groups_attached(container))
1153 minsz = offsetofend(struct vfio_iommu_spapr_tce_remove,
1156 if (copy_from_user(&remove, (void __user *)arg, minsz))
1159 if (remove.argsz < minsz)
1165 if (container->def_window_pending && !remove.start_addr) {
1166 container->def_window_pending = false;
1170 mutex_lock(&container->lock);
1172 ret = tce_iommu_remove_window(container, remove.start_addr);
1174 mutex_unlock(&container->lock);
1183 static void tce_iommu_release_ownership(struct tce_container *container,
1184 struct iommu_table_group *table_group)
1188 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1189 struct iommu_table *tbl = container->tables[i];
1194 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
1195 tce_iommu_userspace_view_free(tbl, container->mm);
1197 iommu_release_ownership(tbl);
1199 container->tables[i] = NULL;
1203 static int tce_iommu_take_ownership(struct tce_container *container,
1204 struct iommu_table_group *table_group)
1208 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1209 struct iommu_table *tbl = table_group->tables[i];
1211 if (!tbl || !tbl->it_map)
1214 rc = iommu_take_ownership(tbl);
1216 for (j = 0; j < i; ++j)
1217 iommu_release_ownership(
1218 table_group->tables[j]);
1224 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1225 container->tables[i] = table_group->tables[i];
1230 static void tce_iommu_release_ownership_ddw(struct tce_container *container,
1231 struct iommu_table_group *table_group)
1235 if (!table_group->ops->unset_window) {
1240 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1241 table_group->ops->unset_window(table_group, i);
1243 table_group->ops->release_ownership(table_group);
1246 static long tce_iommu_take_ownership_ddw(struct tce_container *container,
1247 struct iommu_table_group *table_group)
1249 if (!table_group->ops->create_table || !table_group->ops->set_window ||
1250 !table_group->ops->release_ownership) {
1255 table_group->ops->take_ownership(table_group);
1260 static int tce_iommu_attach_group(void *iommu_data,
1261 struct iommu_group *iommu_group)
1264 struct tce_container *container = iommu_data;
1265 struct iommu_table_group *table_group;
1266 struct tce_iommu_group *tcegrp = NULL;
1268 mutex_lock(&container->lock);
1270 /* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
1271 iommu_group_id(iommu_group), iommu_group); */
1272 table_group = iommu_group_get_iommudata(iommu_group);
1274 if (tce_groups_attached(container) && (!table_group->ops ||
1275 !table_group->ops->take_ownership ||
1276 !table_group->ops->release_ownership)) {
1281 /* Check if new group has the same iommu_ops (i.e. compatible) */
1282 list_for_each_entry(tcegrp, &container->group_list, next) {
1283 struct iommu_table_group *table_group_tmp;
1285 if (tcegrp->grp == iommu_group) {
1286 pr_warn("tce_vfio: Group %d is already attached\n",
1287 iommu_group_id(iommu_group));
1291 table_group_tmp = iommu_group_get_iommudata(tcegrp->grp);
1292 if (table_group_tmp->ops->create_table !=
1293 table_group->ops->create_table) {
1294 pr_warn("tce_vfio: Group %d is incompatible with group %d\n",
1295 iommu_group_id(iommu_group),
1296 iommu_group_id(tcegrp->grp));
1302 tcegrp = kzalloc(sizeof(*tcegrp), GFP_KERNEL);
1308 if (!table_group->ops || !table_group->ops->take_ownership ||
1309 !table_group->ops->release_ownership) {
1310 ret = tce_iommu_take_ownership(container, table_group);
1312 ret = tce_iommu_take_ownership_ddw(container, table_group);
1313 if (!tce_groups_attached(container) && !container->tables[0])
1314 container->def_window_pending = true;
1318 tcegrp->grp = iommu_group;
1319 list_add(&tcegrp->next, &container->group_list);
1326 mutex_unlock(&container->lock);
1331 static void tce_iommu_detach_group(void *iommu_data,
1332 struct iommu_group *iommu_group)
1334 struct tce_container *container = iommu_data;
1335 struct iommu_table_group *table_group;
1337 struct tce_iommu_group *tcegrp;
1339 mutex_lock(&container->lock);
1341 list_for_each_entry(tcegrp, &container->group_list, next) {
1342 if (tcegrp->grp == iommu_group) {
1349 pr_warn("tce_vfio: detaching unattached group #%u\n",
1350 iommu_group_id(iommu_group));
1354 list_del(&tcegrp->next);
1357 table_group = iommu_group_get_iommudata(iommu_group);
1358 BUG_ON(!table_group);
1360 if (!table_group->ops || !table_group->ops->release_ownership)
1361 tce_iommu_release_ownership(container, table_group);
1363 tce_iommu_release_ownership_ddw(container, table_group);
1366 mutex_unlock(&container->lock);
1369 const struct vfio_iommu_driver_ops tce_iommu_driver_ops = {
1370 .name = "iommu-vfio-powerpc",
1371 .owner = THIS_MODULE,
1372 .open = tce_iommu_open,
1373 .release = tce_iommu_release,
1374 .ioctl = tce_iommu_ioctl,
1375 .attach_group = tce_iommu_attach_group,
1376 .detach_group = tce_iommu_detach_group,
1379 static int __init tce_iommu_init(void)
1381 return vfio_register_iommu_driver(&tce_iommu_driver_ops);
1384 static void __exit tce_iommu_cleanup(void)
1386 vfio_unregister_iommu_driver(&tce_iommu_driver_ops);
1389 module_init(tce_iommu_init);
1390 module_exit(tce_iommu_cleanup);
1392 MODULE_VERSION(DRIVER_VERSION);
1393 MODULE_LICENSE("GPL v2");
1394 MODULE_AUTHOR(DRIVER_AUTHOR);
1395 MODULE_DESCRIPTION(DRIVER_DESC);