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 mm_iommu_put(container->mm, mem);
207 list_add(&tcemem->next, &container->prereg_list);
209 container->enabled = true;
214 static bool tce_page_is_contained(struct page *page, unsigned page_shift)
217 * Check that the TCE table granularity is not bigger than the size of
218 * a page we just found. Otherwise the hardware can get access to
219 * a bigger memory chunk that it should.
221 return (PAGE_SHIFT + compound_order(compound_head(page))) >= page_shift;
224 static inline bool tce_groups_attached(struct tce_container *container)
226 return !list_empty(&container->group_list);
229 static long tce_iommu_find_table(struct tce_container *container,
230 phys_addr_t ioba, struct iommu_table **ptbl)
234 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
235 struct iommu_table *tbl = container->tables[i];
238 unsigned long entry = ioba >> tbl->it_page_shift;
239 unsigned long start = tbl->it_offset;
240 unsigned long end = start + tbl->it_size;
242 if ((start <= entry) && (entry < end)) {
252 static int tce_iommu_find_free_table(struct tce_container *container)
256 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
257 if (!container->tables[i])
264 static int tce_iommu_enable(struct tce_container *container)
267 unsigned long locked;
268 struct iommu_table_group *table_group;
269 struct tce_iommu_group *tcegrp;
271 if (container->enabled)
275 * When userspace pages are mapped into the IOMMU, they are effectively
276 * locked memory, so, theoretically, we need to update the accounting
277 * of locked pages on each map and unmap. For powerpc, the map unmap
278 * paths can be very hot, though, and the accounting would kill
279 * performance, especially since it would be difficult to impossible
280 * to handle the accounting in real mode only.
282 * To address that, rather than precisely accounting every page, we
283 * instead account for a worst case on locked memory when the iommu is
284 * enabled and disabled. The worst case upper bound on locked memory
285 * is the size of the whole iommu window, which is usually relatively
286 * small (compared to total memory sizes) on POWER hardware.
288 * Also we don't have a nice way to fail on H_PUT_TCE due to ulimits,
289 * that would effectively kill the guest at random points, much better
290 * enforcing the limit based on the max that the guest can map.
292 * Unfortunately at the moment it counts whole tables, no matter how
293 * much memory the guest has. I.e. for 4GB guest and 4 IOMMU groups
294 * each with 2GB DMA window, 8GB will be counted here. The reason for
295 * this is that we cannot tell here the amount of RAM used by the guest
296 * as this information is only available from KVM and VFIO is
299 * So we do not allow enabling a container without a group attached
300 * as there is no way to know how much we should increment
301 * the locked_vm counter.
303 if (!tce_groups_attached(container))
306 tcegrp = list_first_entry(&container->group_list,
307 struct tce_iommu_group, next);
308 table_group = iommu_group_get_iommudata(tcegrp->grp);
312 if (!table_group->tce32_size)
315 ret = tce_iommu_mm_set(container);
319 locked = table_group->tce32_size >> PAGE_SHIFT;
320 ret = try_increment_locked_vm(container->mm, locked);
324 container->locked_pages = locked;
326 container->enabled = true;
331 static void tce_iommu_disable(struct tce_container *container)
333 if (!container->enabled)
336 container->enabled = false;
338 BUG_ON(!container->mm);
339 decrement_locked_vm(container->mm, container->locked_pages);
342 static void *tce_iommu_open(unsigned long arg)
344 struct tce_container *container;
346 if ((arg != VFIO_SPAPR_TCE_IOMMU) && (arg != VFIO_SPAPR_TCE_v2_IOMMU)) {
347 pr_err("tce_vfio: Wrong IOMMU type\n");
348 return ERR_PTR(-EINVAL);
351 container = kzalloc(sizeof(*container), GFP_KERNEL);
353 return ERR_PTR(-ENOMEM);
355 mutex_init(&container->lock);
356 INIT_LIST_HEAD_RCU(&container->group_list);
357 INIT_LIST_HEAD_RCU(&container->prereg_list);
359 container->v2 = arg == VFIO_SPAPR_TCE_v2_IOMMU;
364 static int tce_iommu_clear(struct tce_container *container,
365 struct iommu_table *tbl,
366 unsigned long entry, unsigned long pages);
367 static void tce_iommu_free_table(struct tce_container *container,
368 struct iommu_table *tbl);
370 static void tce_iommu_release(void *iommu_data)
372 struct tce_container *container = iommu_data;
373 struct tce_iommu_group *tcegrp;
376 while (tce_groups_attached(container)) {
377 tcegrp = list_first_entry(&container->group_list,
378 struct tce_iommu_group, next);
379 tce_iommu_detach_group(iommu_data, tcegrp->grp);
383 * If VFIO created a table, it was not disposed
384 * by tce_iommu_detach_group() so do it now.
386 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
387 struct iommu_table *tbl = container->tables[i];
392 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
393 tce_iommu_free_table(container, tbl);
396 while (!list_empty(&container->prereg_list)) {
397 struct tce_iommu_prereg *tcemem;
399 tcemem = list_first_entry(&container->prereg_list,
400 struct tce_iommu_prereg, next);
401 WARN_ON_ONCE(tce_iommu_prereg_free(container, tcemem));
404 tce_iommu_disable(container);
406 mmdrop(container->mm);
407 mutex_destroy(&container->lock);
412 static void tce_iommu_unuse_page(struct tce_container *container,
417 page = pfn_to_page(hpa >> PAGE_SHIFT);
421 static int tce_iommu_prereg_ua_to_hpa(struct tce_container *container,
422 unsigned long tce, unsigned long shift,
423 unsigned long *phpa, struct mm_iommu_table_group_mem_t **pmem)
426 struct mm_iommu_table_group_mem_t *mem;
428 mem = mm_iommu_lookup(container->mm, tce, 1ULL << shift);
432 ret = mm_iommu_ua_to_hpa(mem, tce, shift, phpa);
441 static void tce_iommu_unuse_page_v2(struct tce_container *container,
442 struct iommu_table *tbl, unsigned long entry)
444 struct mm_iommu_table_group_mem_t *mem = NULL;
446 unsigned long hpa = 0;
447 __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry);
452 ret = tce_iommu_prereg_ua_to_hpa(container, be64_to_cpu(*pua),
453 tbl->it_page_shift, &hpa, &mem);
455 pr_debug("%s: tce %llx at #%lx was not cached, ret=%d\n",
456 __func__, be64_to_cpu(*pua), entry, ret);
458 mm_iommu_mapped_dec(mem);
460 *pua = cpu_to_be64(0);
463 static int tce_iommu_clear(struct tce_container *container,
464 struct iommu_table *tbl,
465 unsigned long entry, unsigned long pages)
467 unsigned long oldhpa;
469 enum dma_data_direction direction;
471 for ( ; pages; --pages, ++entry) {
474 direction = DMA_NONE;
476 ret = iommu_tce_xchg(tbl, entry, &oldhpa, &direction);
480 if (direction == DMA_NONE)
484 tce_iommu_unuse_page_v2(container, tbl, entry);
488 tce_iommu_unuse_page(container, oldhpa);
494 static int tce_iommu_use_page(unsigned long tce, unsigned long *hpa)
496 struct page *page = NULL;
497 enum dma_data_direction direction = iommu_tce_direction(tce);
499 if (get_user_pages_fast(tce & PAGE_MASK, 1,
500 direction != DMA_TO_DEVICE, &page) != 1)
503 *hpa = __pa((unsigned long) page_address(page));
508 static long tce_iommu_build(struct tce_container *container,
509 struct iommu_table *tbl,
510 unsigned long entry, unsigned long tce, unsigned long pages,
511 enum dma_data_direction direction)
516 enum dma_data_direction dirtmp;
518 for (i = 0; i < pages; ++i) {
519 unsigned long offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
521 ret = tce_iommu_use_page(tce, &hpa);
525 page = pfn_to_page(hpa >> PAGE_SHIFT);
526 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
533 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
535 tce_iommu_unuse_page(container, hpa);
536 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
537 __func__, entry << tbl->it_page_shift,
542 if (dirtmp != DMA_NONE)
543 tce_iommu_unuse_page(container, hpa);
545 tce += IOMMU_PAGE_SIZE(tbl);
549 tce_iommu_clear(container, tbl, entry, i);
554 static long tce_iommu_build_v2(struct tce_container *container,
555 struct iommu_table *tbl,
556 unsigned long entry, unsigned long tce, unsigned long pages,
557 enum dma_data_direction direction)
562 enum dma_data_direction dirtmp;
564 for (i = 0; i < pages; ++i) {
565 struct mm_iommu_table_group_mem_t *mem = NULL;
566 __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry + i);
568 ret = tce_iommu_prereg_ua_to_hpa(container,
569 tce, tbl->it_page_shift, &hpa, &mem);
573 page = pfn_to_page(hpa >> PAGE_SHIFT);
574 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
579 /* Preserve offset within IOMMU page */
580 hpa |= tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
583 /* The registered region is being unregistered */
584 if (mm_iommu_mapped_inc(mem))
587 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
589 /* dirtmp cannot be DMA_NONE here */
590 tce_iommu_unuse_page_v2(container, tbl, entry + i);
591 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
592 __func__, entry << tbl->it_page_shift,
597 if (dirtmp != DMA_NONE)
598 tce_iommu_unuse_page_v2(container, tbl, entry + i);
600 *pua = cpu_to_be64(tce);
602 tce += IOMMU_PAGE_SIZE(tbl);
606 tce_iommu_clear(container, tbl, entry, i);
611 static long tce_iommu_create_table(struct tce_container *container,
612 struct iommu_table_group *table_group,
617 struct iommu_table **ptbl)
619 long ret, table_size;
621 table_size = table_group->ops->get_table_size(page_shift, window_size,
626 ret = try_increment_locked_vm(container->mm, table_size >> PAGE_SHIFT);
630 ret = table_group->ops->create_table(table_group, num,
631 page_shift, window_size, levels, ptbl);
633 WARN_ON(!ret && !(*ptbl)->it_ops->free);
634 WARN_ON(!ret && ((*ptbl)->it_allocated_size > table_size));
639 static void tce_iommu_free_table(struct tce_container *container,
640 struct iommu_table *tbl)
642 unsigned long pages = tbl->it_allocated_size >> PAGE_SHIFT;
644 iommu_tce_table_put(tbl);
645 decrement_locked_vm(container->mm, pages);
648 static long tce_iommu_create_window(struct tce_container *container,
649 __u32 page_shift, __u64 window_size, __u32 levels,
652 struct tce_iommu_group *tcegrp;
653 struct iommu_table_group *table_group;
654 struct iommu_table *tbl = NULL;
657 num = tce_iommu_find_free_table(container);
661 /* Get the first group for ops::create_table */
662 tcegrp = list_first_entry(&container->group_list,
663 struct tce_iommu_group, next);
664 table_group = iommu_group_get_iommudata(tcegrp->grp);
668 if (!(table_group->pgsizes & (1ULL << page_shift)))
671 if (!table_group->ops->set_window || !table_group->ops->unset_window ||
672 !table_group->ops->get_table_size ||
673 !table_group->ops->create_table)
676 /* Create TCE table */
677 ret = tce_iommu_create_table(container, table_group, num,
678 page_shift, window_size, levels, &tbl);
682 BUG_ON(!tbl->it_ops->free);
685 * Program the table to every group.
686 * Groups have been tested for compatibility at the attach time.
688 list_for_each_entry(tcegrp, &container->group_list, next) {
689 table_group = iommu_group_get_iommudata(tcegrp->grp);
691 ret = table_group->ops->set_window(table_group, num, tbl);
696 container->tables[num] = tbl;
698 /* Return start address assigned by platform in create_table() */
699 *start_addr = tbl->it_offset << tbl->it_page_shift;
704 list_for_each_entry(tcegrp, &container->group_list, next) {
705 table_group = iommu_group_get_iommudata(tcegrp->grp);
706 table_group->ops->unset_window(table_group, num);
708 tce_iommu_free_table(container, tbl);
713 static long tce_iommu_remove_window(struct tce_container *container,
716 struct iommu_table_group *table_group = NULL;
717 struct iommu_table *tbl;
718 struct tce_iommu_group *tcegrp;
721 num = tce_iommu_find_table(container, start_addr, &tbl);
725 BUG_ON(!tbl->it_size);
727 /* Detach groups from IOMMUs */
728 list_for_each_entry(tcegrp, &container->group_list, next) {
729 table_group = iommu_group_get_iommudata(tcegrp->grp);
732 * SPAPR TCE IOMMU exposes the default DMA window to
733 * the guest via dma32_window_start/size of
734 * VFIO_IOMMU_SPAPR_TCE_GET_INFO. Some platforms allow
735 * the userspace to remove this window, some do not so
736 * here we check for the platform capability.
738 if (!table_group->ops || !table_group->ops->unset_window)
741 table_group->ops->unset_window(table_group, num);
745 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
746 tce_iommu_free_table(container, tbl);
747 container->tables[num] = NULL;
752 static long tce_iommu_create_default_window(struct tce_container *container)
755 __u64 start_addr = 0;
756 struct tce_iommu_group *tcegrp;
757 struct iommu_table_group *table_group;
759 if (!container->def_window_pending)
762 if (!tce_groups_attached(container))
765 tcegrp = list_first_entry(&container->group_list,
766 struct tce_iommu_group, next);
767 table_group = iommu_group_get_iommudata(tcegrp->grp);
771 ret = tce_iommu_create_window(container, IOMMU_PAGE_SHIFT_4K,
772 table_group->tce32_size, 1, &start_addr);
773 WARN_ON_ONCE(!ret && start_addr);
776 container->def_window_pending = false;
781 static long tce_iommu_ioctl(void *iommu_data,
782 unsigned int cmd, unsigned long arg)
784 struct tce_container *container = iommu_data;
785 unsigned long minsz, ddwsz;
789 case VFIO_CHECK_EXTENSION:
791 case VFIO_SPAPR_TCE_IOMMU:
792 case VFIO_SPAPR_TCE_v2_IOMMU:
796 ret = vfio_spapr_iommu_eeh_ioctl(NULL, cmd, arg);
800 return (ret < 0) ? 0 : ret;
804 * Sanity check to prevent one userspace from manipulating
805 * another userspace mm.
808 if (container->mm && container->mm != current->mm)
812 case VFIO_IOMMU_SPAPR_TCE_GET_INFO: {
813 struct vfio_iommu_spapr_tce_info info;
814 struct tce_iommu_group *tcegrp;
815 struct iommu_table_group *table_group;
817 if (!tce_groups_attached(container))
820 tcegrp = list_first_entry(&container->group_list,
821 struct tce_iommu_group, next);
822 table_group = iommu_group_get_iommudata(tcegrp->grp);
827 minsz = offsetofend(struct vfio_iommu_spapr_tce_info,
830 if (copy_from_user(&info, (void __user *)arg, minsz))
833 if (info.argsz < minsz)
836 info.dma32_window_start = table_group->tce32_start;
837 info.dma32_window_size = table_group->tce32_size;
839 memset(&info.ddw, 0, sizeof(info.ddw));
841 if (table_group->max_dynamic_windows_supported &&
843 info.flags |= VFIO_IOMMU_SPAPR_INFO_DDW;
844 info.ddw.pgsizes = table_group->pgsizes;
845 info.ddw.max_dynamic_windows_supported =
846 table_group->max_dynamic_windows_supported;
847 info.ddw.levels = table_group->max_levels;
850 ddwsz = offsetofend(struct vfio_iommu_spapr_tce_info, ddw);
852 if (info.argsz >= ddwsz)
855 if (copy_to_user((void __user *)arg, &info, minsz))
860 case VFIO_IOMMU_MAP_DMA: {
861 struct vfio_iommu_type1_dma_map param;
862 struct iommu_table *tbl = NULL;
864 enum dma_data_direction direction;
866 if (!container->enabled)
869 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
871 if (copy_from_user(¶m, (void __user *)arg, minsz))
874 if (param.argsz < minsz)
877 if (param.flags & ~(VFIO_DMA_MAP_FLAG_READ |
878 VFIO_DMA_MAP_FLAG_WRITE))
881 ret = tce_iommu_create_default_window(container);
885 num = tce_iommu_find_table(container, param.iova, &tbl);
889 if ((param.size & ~IOMMU_PAGE_MASK(tbl)) ||
890 (param.vaddr & ~IOMMU_PAGE_MASK(tbl)))
893 /* iova is checked by the IOMMU API */
894 if (param.flags & VFIO_DMA_MAP_FLAG_READ) {
895 if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
896 direction = DMA_BIDIRECTIONAL;
898 direction = DMA_TO_DEVICE;
900 if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
901 direction = DMA_FROM_DEVICE;
906 ret = iommu_tce_put_param_check(tbl, param.iova, param.vaddr);
911 ret = tce_iommu_build_v2(container, tbl,
912 param.iova >> tbl->it_page_shift,
914 param.size >> tbl->it_page_shift,
917 ret = tce_iommu_build(container, tbl,
918 param.iova >> tbl->it_page_shift,
920 param.size >> tbl->it_page_shift,
923 iommu_flush_tce(tbl);
927 case VFIO_IOMMU_UNMAP_DMA: {
928 struct vfio_iommu_type1_dma_unmap param;
929 struct iommu_table *tbl = NULL;
932 if (!container->enabled)
935 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap,
938 if (copy_from_user(¶m, (void __user *)arg, minsz))
941 if (param.argsz < minsz)
944 /* No flag is supported now */
948 ret = tce_iommu_create_default_window(container);
952 num = tce_iommu_find_table(container, param.iova, &tbl);
956 if (param.size & ~IOMMU_PAGE_MASK(tbl))
959 ret = iommu_tce_clear_param_check(tbl, param.iova, 0,
960 param.size >> tbl->it_page_shift);
964 ret = tce_iommu_clear(container, tbl,
965 param.iova >> tbl->it_page_shift,
966 param.size >> tbl->it_page_shift);
967 iommu_flush_tce(tbl);
971 case VFIO_IOMMU_SPAPR_REGISTER_MEMORY: {
972 struct vfio_iommu_spapr_register_memory param;
977 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
980 ret = tce_iommu_mm_set(container);
984 if (copy_from_user(¶m, (void __user *)arg, minsz))
987 if (param.argsz < minsz)
990 /* No flag is supported now */
994 mutex_lock(&container->lock);
995 ret = tce_iommu_register_pages(container, param.vaddr,
997 mutex_unlock(&container->lock);
1001 case VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY: {
1002 struct vfio_iommu_spapr_register_memory param;
1010 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1013 if (copy_from_user(¶m, (void __user *)arg, minsz))
1016 if (param.argsz < minsz)
1019 /* No flag is supported now */
1023 mutex_lock(&container->lock);
1024 ret = tce_iommu_unregister_pages(container, param.vaddr,
1026 mutex_unlock(&container->lock);
1030 case VFIO_IOMMU_ENABLE:
1034 mutex_lock(&container->lock);
1035 ret = tce_iommu_enable(container);
1036 mutex_unlock(&container->lock);
1040 case VFIO_IOMMU_DISABLE:
1044 mutex_lock(&container->lock);
1045 tce_iommu_disable(container);
1046 mutex_unlock(&container->lock);
1049 case VFIO_EEH_PE_OP: {
1050 struct tce_iommu_group *tcegrp;
1053 list_for_each_entry(tcegrp, &container->group_list, next) {
1054 ret = vfio_spapr_iommu_eeh_ioctl(tcegrp->grp,
1062 case VFIO_IOMMU_SPAPR_TCE_CREATE: {
1063 struct vfio_iommu_spapr_tce_create create;
1068 ret = tce_iommu_mm_set(container);
1072 if (!tce_groups_attached(container))
1075 minsz = offsetofend(struct vfio_iommu_spapr_tce_create,
1078 if (copy_from_user(&create, (void __user *)arg, minsz))
1081 if (create.argsz < minsz)
1087 mutex_lock(&container->lock);
1089 ret = tce_iommu_create_default_window(container);
1091 ret = tce_iommu_create_window(container,
1093 create.window_size, create.levels,
1094 &create.start_addr);
1096 mutex_unlock(&container->lock);
1098 if (!ret && copy_to_user((void __user *)arg, &create, minsz))
1103 case VFIO_IOMMU_SPAPR_TCE_REMOVE: {
1104 struct vfio_iommu_spapr_tce_remove remove;
1109 ret = tce_iommu_mm_set(container);
1113 if (!tce_groups_attached(container))
1116 minsz = offsetofend(struct vfio_iommu_spapr_tce_remove,
1119 if (copy_from_user(&remove, (void __user *)arg, minsz))
1122 if (remove.argsz < minsz)
1128 if (container->def_window_pending && !remove.start_addr) {
1129 container->def_window_pending = false;
1133 mutex_lock(&container->lock);
1135 ret = tce_iommu_remove_window(container, remove.start_addr);
1137 mutex_unlock(&container->lock);
1146 static void tce_iommu_release_ownership(struct tce_container *container,
1147 struct iommu_table_group *table_group)
1151 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1152 struct iommu_table *tbl = container->tables[i];
1157 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
1159 iommu_release_ownership(tbl);
1161 container->tables[i] = NULL;
1165 static int tce_iommu_take_ownership(struct tce_container *container,
1166 struct iommu_table_group *table_group)
1170 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1171 struct iommu_table *tbl = table_group->tables[i];
1173 if (!tbl || !tbl->it_map)
1176 rc = iommu_take_ownership(tbl);
1178 for (j = 0; j < i; ++j)
1179 iommu_release_ownership(
1180 table_group->tables[j]);
1186 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1187 container->tables[i] = table_group->tables[i];
1192 static void tce_iommu_release_ownership_ddw(struct tce_container *container,
1193 struct iommu_table_group *table_group)
1197 if (!table_group->ops->unset_window) {
1202 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1203 table_group->ops->unset_window(table_group, i);
1205 table_group->ops->release_ownership(table_group);
1208 static long tce_iommu_take_ownership_ddw(struct tce_container *container,
1209 struct iommu_table_group *table_group)
1213 if (!table_group->ops->create_table || !table_group->ops->set_window ||
1214 !table_group->ops->release_ownership) {
1219 table_group->ops->take_ownership(table_group);
1221 /* Set all windows to the new group */
1222 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1223 struct iommu_table *tbl = container->tables[i];
1228 ret = table_group->ops->set_window(table_group, i, tbl);
1236 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1237 table_group->ops->unset_window(table_group, i);
1239 table_group->ops->release_ownership(table_group);
1244 static int tce_iommu_attach_group(void *iommu_data,
1245 struct iommu_group *iommu_group)
1248 struct tce_container *container = iommu_data;
1249 struct iommu_table_group *table_group;
1250 struct tce_iommu_group *tcegrp = NULL;
1252 mutex_lock(&container->lock);
1254 /* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
1255 iommu_group_id(iommu_group), iommu_group); */
1256 table_group = iommu_group_get_iommudata(iommu_group);
1262 if (tce_groups_attached(container) && (!table_group->ops ||
1263 !table_group->ops->take_ownership ||
1264 !table_group->ops->release_ownership)) {
1269 /* Check if new group has the same iommu_ops (i.e. compatible) */
1270 list_for_each_entry(tcegrp, &container->group_list, next) {
1271 struct iommu_table_group *table_group_tmp;
1273 if (tcegrp->grp == iommu_group) {
1274 pr_warn("tce_vfio: Group %d is already attached\n",
1275 iommu_group_id(iommu_group));
1279 table_group_tmp = iommu_group_get_iommudata(tcegrp->grp);
1280 if (table_group_tmp->ops->create_table !=
1281 table_group->ops->create_table) {
1282 pr_warn("tce_vfio: Group %d is incompatible with group %d\n",
1283 iommu_group_id(iommu_group),
1284 iommu_group_id(tcegrp->grp));
1290 tcegrp = kzalloc(sizeof(*tcegrp), GFP_KERNEL);
1296 if (!table_group->ops || !table_group->ops->take_ownership ||
1297 !table_group->ops->release_ownership) {
1298 if (container->v2) {
1302 ret = tce_iommu_take_ownership(container, table_group);
1304 if (!container->v2) {
1308 ret = tce_iommu_take_ownership_ddw(container, table_group);
1309 if (!tce_groups_attached(container) && !container->tables[0])
1310 container->def_window_pending = true;
1314 tcegrp->grp = iommu_group;
1315 list_add(&tcegrp->next, &container->group_list);
1322 mutex_unlock(&container->lock);
1327 static void tce_iommu_detach_group(void *iommu_data,
1328 struct iommu_group *iommu_group)
1330 struct tce_container *container = iommu_data;
1331 struct iommu_table_group *table_group;
1333 struct tce_iommu_group *tcegrp;
1335 mutex_lock(&container->lock);
1337 list_for_each_entry(tcegrp, &container->group_list, next) {
1338 if (tcegrp->grp == iommu_group) {
1345 pr_warn("tce_vfio: detaching unattached group #%u\n",
1346 iommu_group_id(iommu_group));
1350 list_del(&tcegrp->next);
1353 table_group = iommu_group_get_iommudata(iommu_group);
1354 BUG_ON(!table_group);
1356 if (!table_group->ops || !table_group->ops->release_ownership)
1357 tce_iommu_release_ownership(container, table_group);
1359 tce_iommu_release_ownership_ddw(container, table_group);
1362 mutex_unlock(&container->lock);
1365 const struct vfio_iommu_driver_ops tce_iommu_driver_ops = {
1366 .name = "iommu-vfio-powerpc",
1367 .owner = THIS_MODULE,
1368 .open = tce_iommu_open,
1369 .release = tce_iommu_release,
1370 .ioctl = tce_iommu_ioctl,
1371 .attach_group = tce_iommu_attach_group,
1372 .detach_group = tce_iommu_detach_group,
1375 static int __init tce_iommu_init(void)
1377 return vfio_register_iommu_driver(&tce_iommu_driver_ops);
1380 static void __exit tce_iommu_cleanup(void)
1382 vfio_unregister_iommu_driver(&tce_iommu_driver_ops);
1385 module_init(tce_iommu_init);
1386 module_exit(tce_iommu_cleanup);
1388 MODULE_VERSION(DRIVER_VERSION);
1389 MODULE_LICENSE("GPL v2");
1390 MODULE_AUTHOR(DRIVER_AUTHOR);
1391 MODULE_DESCRIPTION(DRIVER_DESC);