2 * VFIO: IOMMU DMA mapping support for TCE on POWER
4 * Copyright (C) 2013 IBM Corp. All rights reserved.
5 * Author: Alexey Kardashevskiy <aik@ozlabs.ru>
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
11 * Derived from original vfio_iommu_type1.c:
12 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
13 * Author: Alex Williamson <alex.williamson@redhat.com>
16 #include <linux/module.h>
17 #include <linux/pci.h>
18 #include <linux/slab.h>
19 #include <linux/uaccess.h>
20 #include <linux/err.h>
21 #include <linux/vfio.h>
22 #include <linux/vmalloc.h>
23 #include <linux/sched/mm.h>
24 #include <linux/sched/signal.h>
26 #include <asm/iommu.h>
28 #include <asm/mmu_context.h>
30 #define DRIVER_VERSION "0.1"
31 #define DRIVER_AUTHOR "aik@ozlabs.ru"
32 #define DRIVER_DESC "VFIO IOMMU SPAPR TCE"
34 static void tce_iommu_detach_group(void *iommu_data,
35 struct iommu_group *iommu_group);
37 static long try_increment_locked_vm(struct mm_struct *mm, long npages)
39 long ret = 0, locked, lock_limit;
41 if (WARN_ON_ONCE(!mm))
47 down_write(&mm->mmap_sem);
48 locked = mm->locked_vm + npages;
49 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
50 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
53 mm->locked_vm += npages;
55 pr_debug("[%d] RLIMIT_MEMLOCK +%ld %ld/%ld%s\n", current->pid,
57 mm->locked_vm << PAGE_SHIFT,
58 rlimit(RLIMIT_MEMLOCK),
59 ret ? " - exceeded" : "");
61 up_write(&mm->mmap_sem);
66 static void decrement_locked_vm(struct mm_struct *mm, long npages)
71 down_write(&mm->mmap_sem);
72 if (WARN_ON_ONCE(npages > mm->locked_vm))
73 npages = mm->locked_vm;
74 mm->locked_vm -= npages;
75 pr_debug("[%d] RLIMIT_MEMLOCK -%ld %ld/%ld\n", current->pid,
77 mm->locked_vm << PAGE_SHIFT,
78 rlimit(RLIMIT_MEMLOCK));
79 up_write(&mm->mmap_sem);
83 * VFIO IOMMU fd for SPAPR_TCE IOMMU implementation
85 * This code handles mapping and unmapping of user data buffers
86 * into DMA'ble space using the IOMMU
89 struct tce_iommu_group {
90 struct list_head next;
91 struct iommu_group *grp;
95 * A container needs to remember which preregistered region it has
96 * referenced to do proper cleanup at the userspace process exit.
98 struct tce_iommu_prereg {
99 struct list_head next;
100 struct mm_iommu_table_group_mem_t *mem;
104 * The container descriptor supports only a single group per container.
105 * Required by the API as the container is not supplied with the IOMMU group
106 * at the moment of initialization.
108 struct tce_container {
112 bool def_window_pending;
113 unsigned long locked_pages;
114 struct mm_struct *mm;
115 struct iommu_table *tables[IOMMU_TABLE_GROUP_MAX_TABLES];
116 struct list_head group_list;
117 struct list_head prereg_list;
120 static long tce_iommu_mm_set(struct tce_container *container)
123 if (container->mm == current->mm)
127 BUG_ON(!current->mm);
128 container->mm = current->mm;
129 atomic_inc(&container->mm->mm_count);
134 static long tce_iommu_prereg_free(struct tce_container *container,
135 struct tce_iommu_prereg *tcemem)
139 ret = mm_iommu_put(container->mm, tcemem->mem);
143 list_del(&tcemem->next);
149 static long tce_iommu_unregister_pages(struct tce_container *container,
150 __u64 vaddr, __u64 size)
152 struct mm_iommu_table_group_mem_t *mem;
153 struct tce_iommu_prereg *tcemem;
156 if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK))
159 mem = mm_iommu_find(container->mm, vaddr, size >> PAGE_SHIFT);
163 list_for_each_entry(tcemem, &container->prereg_list, next) {
164 if (tcemem->mem == mem) {
173 return tce_iommu_prereg_free(container, tcemem);
176 static long tce_iommu_register_pages(struct tce_container *container,
177 __u64 vaddr, __u64 size)
180 struct mm_iommu_table_group_mem_t *mem = NULL;
181 struct tce_iommu_prereg *tcemem;
182 unsigned long entries = size >> PAGE_SHIFT;
184 if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK) ||
185 ((vaddr + size) < vaddr))
188 mem = mm_iommu_find(container->mm, vaddr, entries);
190 list_for_each_entry(tcemem, &container->prereg_list, next) {
191 if (tcemem->mem == mem)
196 ret = mm_iommu_get(container->mm, vaddr, entries, &mem);
200 tcemem = kzalloc(sizeof(*tcemem), GFP_KERNEL);
202 list_add(&tcemem->next, &container->prereg_list);
204 container->enabled = true;
209 static long tce_iommu_userspace_view_alloc(struct iommu_table *tbl,
210 struct mm_struct *mm)
212 unsigned long cb = _ALIGN_UP(sizeof(tbl->it_userspace[0]) *
213 tbl->it_size, PAGE_SIZE);
217 BUG_ON(tbl->it_userspace);
219 ret = try_increment_locked_vm(mm, cb >> PAGE_SHIFT);
225 decrement_locked_vm(mm, cb >> PAGE_SHIFT);
228 tbl->it_userspace = uas;
233 static void tce_iommu_userspace_view_free(struct iommu_table *tbl,
234 struct mm_struct *mm)
236 unsigned long cb = _ALIGN_UP(sizeof(tbl->it_userspace[0]) *
237 tbl->it_size, PAGE_SIZE);
239 if (!tbl->it_userspace)
242 vfree(tbl->it_userspace);
243 tbl->it_userspace = NULL;
244 decrement_locked_vm(mm, cb >> PAGE_SHIFT);
247 static bool tce_page_is_contained(struct page *page, unsigned page_shift)
250 * Check that the TCE table granularity is not bigger than the size of
251 * a page we just found. Otherwise the hardware can get access to
252 * a bigger memory chunk that it should.
254 return (PAGE_SHIFT + compound_order(compound_head(page))) >= page_shift;
257 static inline bool tce_groups_attached(struct tce_container *container)
259 return !list_empty(&container->group_list);
262 static long tce_iommu_find_table(struct tce_container *container,
263 phys_addr_t ioba, struct iommu_table **ptbl)
267 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
268 struct iommu_table *tbl = container->tables[i];
271 unsigned long entry = ioba >> tbl->it_page_shift;
272 unsigned long start = tbl->it_offset;
273 unsigned long end = start + tbl->it_size;
275 if ((start <= entry) && (entry < end)) {
285 static int tce_iommu_find_free_table(struct tce_container *container)
289 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
290 if (!container->tables[i])
297 static int tce_iommu_enable(struct tce_container *container)
300 unsigned long locked;
301 struct iommu_table_group *table_group;
302 struct tce_iommu_group *tcegrp;
304 if (container->enabled)
308 * When userspace pages are mapped into the IOMMU, they are effectively
309 * locked memory, so, theoretically, we need to update the accounting
310 * of locked pages on each map and unmap. For powerpc, the map unmap
311 * paths can be very hot, though, and the accounting would kill
312 * performance, especially since it would be difficult to impossible
313 * to handle the accounting in real mode only.
315 * To address that, rather than precisely accounting every page, we
316 * instead account for a worst case on locked memory when the iommu is
317 * enabled and disabled. The worst case upper bound on locked memory
318 * is the size of the whole iommu window, which is usually relatively
319 * small (compared to total memory sizes) on POWER hardware.
321 * Also we don't have a nice way to fail on H_PUT_TCE due to ulimits,
322 * that would effectively kill the guest at random points, much better
323 * enforcing the limit based on the max that the guest can map.
325 * Unfortunately at the moment it counts whole tables, no matter how
326 * much memory the guest has. I.e. for 4GB guest and 4 IOMMU groups
327 * each with 2GB DMA window, 8GB will be counted here. The reason for
328 * this is that we cannot tell here the amount of RAM used by the guest
329 * as this information is only available from KVM and VFIO is
332 * So we do not allow enabling a container without a group attached
333 * as there is no way to know how much we should increment
334 * the locked_vm counter.
336 if (!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);
345 if (!table_group->tce32_size)
348 ret = tce_iommu_mm_set(container);
352 locked = table_group->tce32_size >> PAGE_SHIFT;
353 ret = try_increment_locked_vm(container->mm, locked);
357 container->locked_pages = locked;
359 container->enabled = true;
364 static void tce_iommu_disable(struct tce_container *container)
366 if (!container->enabled)
369 container->enabled = false;
371 BUG_ON(!container->mm);
372 decrement_locked_vm(container->mm, container->locked_pages);
375 static void *tce_iommu_open(unsigned long arg)
377 struct tce_container *container;
379 if ((arg != VFIO_SPAPR_TCE_IOMMU) && (arg != VFIO_SPAPR_TCE_v2_IOMMU)) {
380 pr_err("tce_vfio: Wrong IOMMU type\n");
381 return ERR_PTR(-EINVAL);
384 container = kzalloc(sizeof(*container), GFP_KERNEL);
386 return ERR_PTR(-ENOMEM);
388 mutex_init(&container->lock);
389 INIT_LIST_HEAD_RCU(&container->group_list);
390 INIT_LIST_HEAD_RCU(&container->prereg_list);
392 container->v2 = arg == VFIO_SPAPR_TCE_v2_IOMMU;
397 static int tce_iommu_clear(struct tce_container *container,
398 struct iommu_table *tbl,
399 unsigned long entry, unsigned long pages);
400 static void tce_iommu_free_table(struct tce_container *container,
401 struct iommu_table *tbl);
403 static void tce_iommu_release(void *iommu_data)
405 struct tce_container *container = iommu_data;
406 struct tce_iommu_group *tcegrp;
409 while (tce_groups_attached(container)) {
410 tcegrp = list_first_entry(&container->group_list,
411 struct tce_iommu_group, next);
412 tce_iommu_detach_group(iommu_data, tcegrp->grp);
416 * If VFIO created a table, it was not disposed
417 * by tce_iommu_detach_group() so do it now.
419 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
420 struct iommu_table *tbl = container->tables[i];
425 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
426 tce_iommu_free_table(container, tbl);
429 while (!list_empty(&container->prereg_list)) {
430 struct tce_iommu_prereg *tcemem;
432 tcemem = list_first_entry(&container->prereg_list,
433 struct tce_iommu_prereg, next);
434 WARN_ON_ONCE(tce_iommu_prereg_free(container, tcemem));
437 tce_iommu_disable(container);
439 mmdrop(container->mm);
440 mutex_destroy(&container->lock);
445 static void tce_iommu_unuse_page(struct tce_container *container,
450 page = pfn_to_page(hpa >> PAGE_SHIFT);
454 static int tce_iommu_prereg_ua_to_hpa(struct tce_container *container,
455 unsigned long tce, unsigned long size,
456 unsigned long *phpa, struct mm_iommu_table_group_mem_t **pmem)
459 struct mm_iommu_table_group_mem_t *mem;
461 mem = mm_iommu_lookup(container->mm, tce, size);
465 ret = mm_iommu_ua_to_hpa(mem, tce, phpa);
474 static void tce_iommu_unuse_page_v2(struct tce_container *container,
475 struct iommu_table *tbl, unsigned long entry)
477 struct mm_iommu_table_group_mem_t *mem = NULL;
479 unsigned long hpa = 0;
480 unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry);
485 ret = tce_iommu_prereg_ua_to_hpa(container, *pua, IOMMU_PAGE_SIZE(tbl),
488 pr_debug("%s: tce %lx at #%lx was not cached, ret=%d\n",
489 __func__, *pua, entry, ret);
491 mm_iommu_mapped_dec(mem);
496 static int tce_iommu_clear(struct tce_container *container,
497 struct iommu_table *tbl,
498 unsigned long entry, unsigned long pages)
500 unsigned long oldhpa;
502 enum dma_data_direction direction;
504 for ( ; pages; --pages, ++entry) {
505 direction = DMA_NONE;
507 ret = iommu_tce_xchg(tbl, entry, &oldhpa, &direction);
511 if (direction == DMA_NONE)
515 tce_iommu_unuse_page_v2(container, tbl, entry);
519 tce_iommu_unuse_page(container, oldhpa);
525 static int tce_iommu_use_page(unsigned long tce, unsigned long *hpa)
527 struct page *page = NULL;
528 enum dma_data_direction direction = iommu_tce_direction(tce);
530 if (get_user_pages_fast(tce & PAGE_MASK, 1,
531 direction != DMA_TO_DEVICE, &page) != 1)
534 *hpa = __pa((unsigned long) page_address(page));
539 static long tce_iommu_build(struct tce_container *container,
540 struct iommu_table *tbl,
541 unsigned long entry, unsigned long tce, unsigned long pages,
542 enum dma_data_direction direction)
547 enum dma_data_direction dirtmp;
549 for (i = 0; i < pages; ++i) {
550 unsigned long offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
552 ret = tce_iommu_use_page(tce, &hpa);
556 page = pfn_to_page(hpa >> PAGE_SHIFT);
557 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
564 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
566 tce_iommu_unuse_page(container, hpa);
567 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
568 __func__, entry << tbl->it_page_shift,
573 if (dirtmp != DMA_NONE)
574 tce_iommu_unuse_page(container, hpa);
576 tce += IOMMU_PAGE_SIZE(tbl);
580 tce_iommu_clear(container, tbl, entry, i);
585 static long tce_iommu_build_v2(struct tce_container *container,
586 struct iommu_table *tbl,
587 unsigned long entry, unsigned long tce, unsigned long pages,
588 enum dma_data_direction direction)
593 enum dma_data_direction dirtmp;
595 if (!tbl->it_userspace) {
596 ret = tce_iommu_userspace_view_alloc(tbl, container->mm);
601 for (i = 0; i < pages; ++i) {
602 struct mm_iommu_table_group_mem_t *mem = NULL;
603 unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl,
606 ret = tce_iommu_prereg_ua_to_hpa(container,
607 tce, IOMMU_PAGE_SIZE(tbl), &hpa, &mem);
611 page = pfn_to_page(hpa >> PAGE_SHIFT);
612 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
617 /* Preserve offset within IOMMU page */
618 hpa |= tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
621 /* The registered region is being unregistered */
622 if (mm_iommu_mapped_inc(mem))
625 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
627 /* dirtmp cannot be DMA_NONE here */
628 tce_iommu_unuse_page_v2(container, tbl, entry + i);
629 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
630 __func__, entry << tbl->it_page_shift,
635 if (dirtmp != DMA_NONE)
636 tce_iommu_unuse_page_v2(container, tbl, entry + i);
640 tce += IOMMU_PAGE_SIZE(tbl);
644 tce_iommu_clear(container, tbl, entry, i);
649 static long tce_iommu_create_table(struct tce_container *container,
650 struct iommu_table_group *table_group,
655 struct iommu_table **ptbl)
657 long ret, table_size;
659 table_size = table_group->ops->get_table_size(page_shift, window_size,
664 ret = try_increment_locked_vm(container->mm, table_size >> PAGE_SHIFT);
668 ret = table_group->ops->create_table(table_group, num,
669 page_shift, window_size, levels, ptbl);
671 WARN_ON(!ret && !(*ptbl)->it_ops->free);
672 WARN_ON(!ret && ((*ptbl)->it_allocated_size != table_size));
677 static void tce_iommu_free_table(struct tce_container *container,
678 struct iommu_table *tbl)
680 unsigned long pages = tbl->it_allocated_size >> PAGE_SHIFT;
682 tce_iommu_userspace_view_free(tbl, container->mm);
683 tbl->it_ops->free(tbl);
684 decrement_locked_vm(container->mm, pages);
687 static long tce_iommu_create_window(struct tce_container *container,
688 __u32 page_shift, __u64 window_size, __u32 levels,
691 struct tce_iommu_group *tcegrp;
692 struct iommu_table_group *table_group;
693 struct iommu_table *tbl = NULL;
696 num = tce_iommu_find_free_table(container);
700 /* Get the first group for ops::create_table */
701 tcegrp = list_first_entry(&container->group_list,
702 struct tce_iommu_group, next);
703 table_group = iommu_group_get_iommudata(tcegrp->grp);
707 if (!(table_group->pgsizes & (1ULL << page_shift)))
710 if (!table_group->ops->set_window || !table_group->ops->unset_window ||
711 !table_group->ops->get_table_size ||
712 !table_group->ops->create_table)
715 /* Create TCE table */
716 ret = tce_iommu_create_table(container, table_group, num,
717 page_shift, window_size, levels, &tbl);
721 BUG_ON(!tbl->it_ops->free);
724 * Program the table to every group.
725 * Groups have been tested for compatibility at the attach time.
727 list_for_each_entry(tcegrp, &container->group_list, next) {
728 table_group = iommu_group_get_iommudata(tcegrp->grp);
730 ret = table_group->ops->set_window(table_group, num, tbl);
735 container->tables[num] = tbl;
737 /* Return start address assigned by platform in create_table() */
738 *start_addr = tbl->it_offset << tbl->it_page_shift;
743 list_for_each_entry(tcegrp, &container->group_list, next) {
744 table_group = iommu_group_get_iommudata(tcegrp->grp);
745 table_group->ops->unset_window(table_group, num);
747 tce_iommu_free_table(container, tbl);
752 static long tce_iommu_remove_window(struct tce_container *container,
755 struct iommu_table_group *table_group = NULL;
756 struct iommu_table *tbl;
757 struct tce_iommu_group *tcegrp;
760 num = tce_iommu_find_table(container, start_addr, &tbl);
764 BUG_ON(!tbl->it_size);
766 /* Detach groups from IOMMUs */
767 list_for_each_entry(tcegrp, &container->group_list, next) {
768 table_group = iommu_group_get_iommudata(tcegrp->grp);
771 * SPAPR TCE IOMMU exposes the default DMA window to
772 * the guest via dma32_window_start/size of
773 * VFIO_IOMMU_SPAPR_TCE_GET_INFO. Some platforms allow
774 * the userspace to remove this window, some do not so
775 * here we check for the platform capability.
777 if (!table_group->ops || !table_group->ops->unset_window)
780 table_group->ops->unset_window(table_group, num);
784 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
785 tce_iommu_free_table(container, tbl);
786 container->tables[num] = NULL;
791 static long tce_iommu_create_default_window(struct tce_container *container)
794 __u64 start_addr = 0;
795 struct tce_iommu_group *tcegrp;
796 struct iommu_table_group *table_group;
798 if (!container->def_window_pending)
801 if (!tce_groups_attached(container))
804 tcegrp = list_first_entry(&container->group_list,
805 struct tce_iommu_group, next);
806 table_group = iommu_group_get_iommudata(tcegrp->grp);
810 ret = tce_iommu_create_window(container, IOMMU_PAGE_SHIFT_4K,
811 table_group->tce32_size, 1, &start_addr);
812 WARN_ON_ONCE(!ret && start_addr);
815 container->def_window_pending = false;
820 static long tce_iommu_ioctl(void *iommu_data,
821 unsigned int cmd, unsigned long arg)
823 struct tce_container *container = iommu_data;
824 unsigned long minsz, ddwsz;
828 case VFIO_CHECK_EXTENSION:
830 case VFIO_SPAPR_TCE_IOMMU:
831 case VFIO_SPAPR_TCE_v2_IOMMU:
835 ret = vfio_spapr_iommu_eeh_ioctl(NULL, cmd, arg);
839 return (ret < 0) ? 0 : ret;
843 * Sanity check to prevent one userspace from manipulating
844 * another userspace mm.
847 if (container->mm && container->mm != current->mm)
851 case VFIO_IOMMU_SPAPR_TCE_GET_INFO: {
852 struct vfio_iommu_spapr_tce_info info;
853 struct tce_iommu_group *tcegrp;
854 struct iommu_table_group *table_group;
856 if (!tce_groups_attached(container))
859 tcegrp = list_first_entry(&container->group_list,
860 struct tce_iommu_group, next);
861 table_group = iommu_group_get_iommudata(tcegrp->grp);
866 minsz = offsetofend(struct vfio_iommu_spapr_tce_info,
869 if (copy_from_user(&info, (void __user *)arg, minsz))
872 if (info.argsz < minsz)
875 info.dma32_window_start = table_group->tce32_start;
876 info.dma32_window_size = table_group->tce32_size;
878 memset(&info.ddw, 0, sizeof(info.ddw));
880 if (table_group->max_dynamic_windows_supported &&
882 info.flags |= VFIO_IOMMU_SPAPR_INFO_DDW;
883 info.ddw.pgsizes = table_group->pgsizes;
884 info.ddw.max_dynamic_windows_supported =
885 table_group->max_dynamic_windows_supported;
886 info.ddw.levels = table_group->max_levels;
889 ddwsz = offsetofend(struct vfio_iommu_spapr_tce_info, ddw);
891 if (info.argsz >= ddwsz)
894 if (copy_to_user((void __user *)arg, &info, minsz))
899 case VFIO_IOMMU_MAP_DMA: {
900 struct vfio_iommu_type1_dma_map param;
901 struct iommu_table *tbl = NULL;
903 enum dma_data_direction direction;
905 if (!container->enabled)
908 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
910 if (copy_from_user(¶m, (void __user *)arg, minsz))
913 if (param.argsz < minsz)
916 if (param.flags & ~(VFIO_DMA_MAP_FLAG_READ |
917 VFIO_DMA_MAP_FLAG_WRITE))
920 ret = tce_iommu_create_default_window(container);
924 num = tce_iommu_find_table(container, param.iova, &tbl);
928 if ((param.size & ~IOMMU_PAGE_MASK(tbl)) ||
929 (param.vaddr & ~IOMMU_PAGE_MASK(tbl)))
932 /* iova is checked by the IOMMU API */
933 if (param.flags & VFIO_DMA_MAP_FLAG_READ) {
934 if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
935 direction = DMA_BIDIRECTIONAL;
937 direction = DMA_TO_DEVICE;
939 if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
940 direction = DMA_FROM_DEVICE;
945 ret = iommu_tce_put_param_check(tbl, param.iova, param.vaddr);
950 ret = tce_iommu_build_v2(container, tbl,
951 param.iova >> tbl->it_page_shift,
953 param.size >> tbl->it_page_shift,
956 ret = tce_iommu_build(container, tbl,
957 param.iova >> tbl->it_page_shift,
959 param.size >> tbl->it_page_shift,
962 iommu_flush_tce(tbl);
966 case VFIO_IOMMU_UNMAP_DMA: {
967 struct vfio_iommu_type1_dma_unmap param;
968 struct iommu_table *tbl = NULL;
971 if (!container->enabled)
974 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap,
977 if (copy_from_user(¶m, (void __user *)arg, minsz))
980 if (param.argsz < minsz)
983 /* No flag is supported now */
987 ret = tce_iommu_create_default_window(container);
991 num = tce_iommu_find_table(container, param.iova, &tbl);
995 if (param.size & ~IOMMU_PAGE_MASK(tbl))
998 ret = iommu_tce_clear_param_check(tbl, param.iova, 0,
999 param.size >> tbl->it_page_shift);
1003 ret = tce_iommu_clear(container, tbl,
1004 param.iova >> tbl->it_page_shift,
1005 param.size >> tbl->it_page_shift);
1006 iommu_flush_tce(tbl);
1010 case VFIO_IOMMU_SPAPR_REGISTER_MEMORY: {
1011 struct vfio_iommu_spapr_register_memory param;
1016 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1019 ret = tce_iommu_mm_set(container);
1023 if (copy_from_user(¶m, (void __user *)arg, minsz))
1026 if (param.argsz < minsz)
1029 /* No flag is supported now */
1033 mutex_lock(&container->lock);
1034 ret = tce_iommu_register_pages(container, param.vaddr,
1036 mutex_unlock(&container->lock);
1040 case VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY: {
1041 struct vfio_iommu_spapr_register_memory param;
1049 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1052 if (copy_from_user(¶m, (void __user *)arg, minsz))
1055 if (param.argsz < minsz)
1058 /* No flag is supported now */
1062 mutex_lock(&container->lock);
1063 ret = tce_iommu_unregister_pages(container, param.vaddr,
1065 mutex_unlock(&container->lock);
1069 case VFIO_IOMMU_ENABLE:
1073 mutex_lock(&container->lock);
1074 ret = tce_iommu_enable(container);
1075 mutex_unlock(&container->lock);
1079 case VFIO_IOMMU_DISABLE:
1083 mutex_lock(&container->lock);
1084 tce_iommu_disable(container);
1085 mutex_unlock(&container->lock);
1088 case VFIO_EEH_PE_OP: {
1089 struct tce_iommu_group *tcegrp;
1092 list_for_each_entry(tcegrp, &container->group_list, next) {
1093 ret = vfio_spapr_iommu_eeh_ioctl(tcegrp->grp,
1101 case VFIO_IOMMU_SPAPR_TCE_CREATE: {
1102 struct vfio_iommu_spapr_tce_create create;
1107 ret = tce_iommu_mm_set(container);
1111 if (!tce_groups_attached(container))
1114 minsz = offsetofend(struct vfio_iommu_spapr_tce_create,
1117 if (copy_from_user(&create, (void __user *)arg, minsz))
1120 if (create.argsz < minsz)
1126 mutex_lock(&container->lock);
1128 ret = tce_iommu_create_default_window(container);
1130 ret = tce_iommu_create_window(container,
1132 create.window_size, create.levels,
1133 &create.start_addr);
1135 mutex_unlock(&container->lock);
1137 if (!ret && copy_to_user((void __user *)arg, &create, minsz))
1142 case VFIO_IOMMU_SPAPR_TCE_REMOVE: {
1143 struct vfio_iommu_spapr_tce_remove remove;
1148 ret = tce_iommu_mm_set(container);
1152 if (!tce_groups_attached(container))
1155 minsz = offsetofend(struct vfio_iommu_spapr_tce_remove,
1158 if (copy_from_user(&remove, (void __user *)arg, minsz))
1161 if (remove.argsz < minsz)
1167 if (container->def_window_pending && !remove.start_addr) {
1168 container->def_window_pending = false;
1172 mutex_lock(&container->lock);
1174 ret = tce_iommu_remove_window(container, remove.start_addr);
1176 mutex_unlock(&container->lock);
1185 static void tce_iommu_release_ownership(struct tce_container *container,
1186 struct iommu_table_group *table_group)
1190 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1191 struct iommu_table *tbl = container->tables[i];
1196 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
1197 tce_iommu_userspace_view_free(tbl, container->mm);
1199 iommu_release_ownership(tbl);
1201 container->tables[i] = NULL;
1205 static int tce_iommu_take_ownership(struct tce_container *container,
1206 struct iommu_table_group *table_group)
1210 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1211 struct iommu_table *tbl = table_group->tables[i];
1213 if (!tbl || !tbl->it_map)
1216 rc = iommu_take_ownership(tbl);
1218 for (j = 0; j < i; ++j)
1219 iommu_release_ownership(
1220 table_group->tables[j]);
1226 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1227 container->tables[i] = table_group->tables[i];
1232 static void tce_iommu_release_ownership_ddw(struct tce_container *container,
1233 struct iommu_table_group *table_group)
1237 if (!table_group->ops->unset_window) {
1242 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1243 table_group->ops->unset_window(table_group, i);
1245 table_group->ops->release_ownership(table_group);
1248 static long tce_iommu_take_ownership_ddw(struct tce_container *container,
1249 struct iommu_table_group *table_group)
1253 if (!table_group->ops->create_table || !table_group->ops->set_window ||
1254 !table_group->ops->release_ownership) {
1259 table_group->ops->take_ownership(table_group);
1261 /* Set all windows to the new group */
1262 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1263 struct iommu_table *tbl = container->tables[i];
1268 ret = table_group->ops->set_window(table_group, i, tbl);
1276 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1277 table_group->ops->unset_window(table_group, i);
1279 table_group->ops->release_ownership(table_group);
1284 static int tce_iommu_attach_group(void *iommu_data,
1285 struct iommu_group *iommu_group)
1288 struct tce_container *container = iommu_data;
1289 struct iommu_table_group *table_group;
1290 struct tce_iommu_group *tcegrp = NULL;
1292 mutex_lock(&container->lock);
1294 /* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
1295 iommu_group_id(iommu_group), iommu_group); */
1296 table_group = iommu_group_get_iommudata(iommu_group);
1302 if (tce_groups_attached(container) && (!table_group->ops ||
1303 !table_group->ops->take_ownership ||
1304 !table_group->ops->release_ownership)) {
1309 /* Check if new group has the same iommu_ops (i.e. compatible) */
1310 list_for_each_entry(tcegrp, &container->group_list, next) {
1311 struct iommu_table_group *table_group_tmp;
1313 if (tcegrp->grp == iommu_group) {
1314 pr_warn("tce_vfio: Group %d is already attached\n",
1315 iommu_group_id(iommu_group));
1319 table_group_tmp = iommu_group_get_iommudata(tcegrp->grp);
1320 if (table_group_tmp->ops->create_table !=
1321 table_group->ops->create_table) {
1322 pr_warn("tce_vfio: Group %d is incompatible with group %d\n",
1323 iommu_group_id(iommu_group),
1324 iommu_group_id(tcegrp->grp));
1330 tcegrp = kzalloc(sizeof(*tcegrp), GFP_KERNEL);
1336 if (!table_group->ops || !table_group->ops->take_ownership ||
1337 !table_group->ops->release_ownership) {
1338 ret = tce_iommu_take_ownership(container, table_group);
1340 ret = tce_iommu_take_ownership_ddw(container, table_group);
1341 if (!tce_groups_attached(container) && !container->tables[0])
1342 container->def_window_pending = true;
1346 tcegrp->grp = iommu_group;
1347 list_add(&tcegrp->next, &container->group_list);
1354 mutex_unlock(&container->lock);
1359 static void tce_iommu_detach_group(void *iommu_data,
1360 struct iommu_group *iommu_group)
1362 struct tce_container *container = iommu_data;
1363 struct iommu_table_group *table_group;
1365 struct tce_iommu_group *tcegrp;
1367 mutex_lock(&container->lock);
1369 list_for_each_entry(tcegrp, &container->group_list, next) {
1370 if (tcegrp->grp == iommu_group) {
1377 pr_warn("tce_vfio: detaching unattached group #%u\n",
1378 iommu_group_id(iommu_group));
1382 list_del(&tcegrp->next);
1385 table_group = iommu_group_get_iommudata(iommu_group);
1386 BUG_ON(!table_group);
1388 if (!table_group->ops || !table_group->ops->release_ownership)
1389 tce_iommu_release_ownership(container, table_group);
1391 tce_iommu_release_ownership_ddw(container, table_group);
1394 mutex_unlock(&container->lock);
1397 const struct vfio_iommu_driver_ops tce_iommu_driver_ops = {
1398 .name = "iommu-vfio-powerpc",
1399 .owner = THIS_MODULE,
1400 .open = tce_iommu_open,
1401 .release = tce_iommu_release,
1402 .ioctl = tce_iommu_ioctl,
1403 .attach_group = tce_iommu_attach_group,
1404 .detach_group = tce_iommu_detach_group,
1407 static int __init tce_iommu_init(void)
1409 return vfio_register_iommu_driver(&tce_iommu_driver_ops);
1412 static void __exit tce_iommu_cleanup(void)
1414 vfio_unregister_iommu_driver(&tce_iommu_driver_ops);
1417 module_init(tce_iommu_init);
1418 module_exit(tce_iommu_cleanup);
1420 MODULE_VERSION(DRIVER_VERSION);
1421 MODULE_LICENSE("GPL v2");
1422 MODULE_AUTHOR(DRIVER_AUTHOR);
1423 MODULE_DESCRIPTION(DRIVER_DESC);