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_RO(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;
470 unsigned long lastentry = entry + pages;
472 for ( ; entry < lastentry; ++entry) {
473 if (tbl->it_indirect_levels && tbl->it_userspace) {
475 * For multilevel tables, we can take a shortcut here
476 * and skip some TCEs as we know that the userspace
477 * addresses cache is a mirror of the real TCE table
478 * and if it is missing some indirect levels, then
479 * the hardware table does not have them allocated
480 * either and therefore does not require updating.
482 __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY_RO(tbl,
485 /* align to level_size which is power of two */
486 entry |= tbl->it_level_size - 1;
493 direction = DMA_NONE;
495 ret = iommu_tce_xchg(tbl, entry, &oldhpa, &direction);
499 if (direction == DMA_NONE)
503 tce_iommu_unuse_page_v2(container, tbl, entry);
507 tce_iommu_unuse_page(container, oldhpa);
513 static int tce_iommu_use_page(unsigned long tce, unsigned long *hpa)
515 struct page *page = NULL;
516 enum dma_data_direction direction = iommu_tce_direction(tce);
518 if (get_user_pages_fast(tce & PAGE_MASK, 1,
519 direction != DMA_TO_DEVICE, &page) != 1)
522 *hpa = __pa((unsigned long) page_address(page));
527 static long tce_iommu_build(struct tce_container *container,
528 struct iommu_table *tbl,
529 unsigned long entry, unsigned long tce, unsigned long pages,
530 enum dma_data_direction direction)
535 enum dma_data_direction dirtmp;
537 for (i = 0; i < pages; ++i) {
538 unsigned long offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
540 ret = tce_iommu_use_page(tce, &hpa);
544 page = pfn_to_page(hpa >> PAGE_SHIFT);
545 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
552 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
554 tce_iommu_unuse_page(container, hpa);
555 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
556 __func__, entry << tbl->it_page_shift,
561 if (dirtmp != DMA_NONE)
562 tce_iommu_unuse_page(container, hpa);
564 tce += IOMMU_PAGE_SIZE(tbl);
568 tce_iommu_clear(container, tbl, entry, i);
573 static long tce_iommu_build_v2(struct tce_container *container,
574 struct iommu_table *tbl,
575 unsigned long entry, unsigned long tce, unsigned long pages,
576 enum dma_data_direction direction)
581 enum dma_data_direction dirtmp;
583 for (i = 0; i < pages; ++i) {
584 struct mm_iommu_table_group_mem_t *mem = NULL;
585 __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry + i);
587 ret = tce_iommu_prereg_ua_to_hpa(container,
588 tce, tbl->it_page_shift, &hpa, &mem);
592 page = pfn_to_page(hpa >> PAGE_SHIFT);
593 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
598 /* Preserve offset within IOMMU page */
599 hpa |= tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
602 /* The registered region is being unregistered */
603 if (mm_iommu_mapped_inc(mem))
606 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
608 /* dirtmp cannot be DMA_NONE here */
609 tce_iommu_unuse_page_v2(container, tbl, entry + i);
610 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
611 __func__, entry << tbl->it_page_shift,
616 if (dirtmp != DMA_NONE)
617 tce_iommu_unuse_page_v2(container, tbl, entry + i);
619 *pua = cpu_to_be64(tce);
621 tce += IOMMU_PAGE_SIZE(tbl);
625 tce_iommu_clear(container, tbl, entry, i);
630 static long tce_iommu_create_table(struct tce_container *container,
631 struct iommu_table_group *table_group,
636 struct iommu_table **ptbl)
638 long ret, table_size;
640 table_size = table_group->ops->get_table_size(page_shift, window_size,
645 ret = try_increment_locked_vm(container->mm, table_size >> PAGE_SHIFT);
649 ret = table_group->ops->create_table(table_group, num,
650 page_shift, window_size, levels, ptbl);
652 WARN_ON(!ret && !(*ptbl)->it_ops->free);
653 WARN_ON(!ret && ((*ptbl)->it_allocated_size > table_size));
658 static void tce_iommu_free_table(struct tce_container *container,
659 struct iommu_table *tbl)
661 unsigned long pages = tbl->it_allocated_size >> PAGE_SHIFT;
663 iommu_tce_table_put(tbl);
664 decrement_locked_vm(container->mm, pages);
667 static long tce_iommu_create_window(struct tce_container *container,
668 __u32 page_shift, __u64 window_size, __u32 levels,
671 struct tce_iommu_group *tcegrp;
672 struct iommu_table_group *table_group;
673 struct iommu_table *tbl = NULL;
676 num = tce_iommu_find_free_table(container);
680 /* Get the first group for ops::create_table */
681 tcegrp = list_first_entry(&container->group_list,
682 struct tce_iommu_group, next);
683 table_group = iommu_group_get_iommudata(tcegrp->grp);
687 if (!(table_group->pgsizes & (1ULL << page_shift)))
690 if (!table_group->ops->set_window || !table_group->ops->unset_window ||
691 !table_group->ops->get_table_size ||
692 !table_group->ops->create_table)
695 /* Create TCE table */
696 ret = tce_iommu_create_table(container, table_group, num,
697 page_shift, window_size, levels, &tbl);
701 BUG_ON(!tbl->it_ops->free);
704 * Program the table to every group.
705 * Groups have been tested for compatibility at the attach time.
707 list_for_each_entry(tcegrp, &container->group_list, next) {
708 table_group = iommu_group_get_iommudata(tcegrp->grp);
710 ret = table_group->ops->set_window(table_group, num, tbl);
715 container->tables[num] = tbl;
717 /* Return start address assigned by platform in create_table() */
718 *start_addr = tbl->it_offset << tbl->it_page_shift;
723 list_for_each_entry(tcegrp, &container->group_list, next) {
724 table_group = iommu_group_get_iommudata(tcegrp->grp);
725 table_group->ops->unset_window(table_group, num);
727 tce_iommu_free_table(container, tbl);
732 static long tce_iommu_remove_window(struct tce_container *container,
735 struct iommu_table_group *table_group = NULL;
736 struct iommu_table *tbl;
737 struct tce_iommu_group *tcegrp;
740 num = tce_iommu_find_table(container, start_addr, &tbl);
744 BUG_ON(!tbl->it_size);
746 /* Detach groups from IOMMUs */
747 list_for_each_entry(tcegrp, &container->group_list, next) {
748 table_group = iommu_group_get_iommudata(tcegrp->grp);
751 * SPAPR TCE IOMMU exposes the default DMA window to
752 * the guest via dma32_window_start/size of
753 * VFIO_IOMMU_SPAPR_TCE_GET_INFO. Some platforms allow
754 * the userspace to remove this window, some do not so
755 * here we check for the platform capability.
757 if (!table_group->ops || !table_group->ops->unset_window)
760 table_group->ops->unset_window(table_group, num);
764 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
765 tce_iommu_free_table(container, tbl);
766 container->tables[num] = NULL;
771 static long tce_iommu_create_default_window(struct tce_container *container)
774 __u64 start_addr = 0;
775 struct tce_iommu_group *tcegrp;
776 struct iommu_table_group *table_group;
778 if (!container->def_window_pending)
781 if (!tce_groups_attached(container))
784 tcegrp = list_first_entry(&container->group_list,
785 struct tce_iommu_group, next);
786 table_group = iommu_group_get_iommudata(tcegrp->grp);
790 ret = tce_iommu_create_window(container, IOMMU_PAGE_SHIFT_4K,
791 table_group->tce32_size, 1, &start_addr);
792 WARN_ON_ONCE(!ret && start_addr);
795 container->def_window_pending = false;
800 static long tce_iommu_ioctl(void *iommu_data,
801 unsigned int cmd, unsigned long arg)
803 struct tce_container *container = iommu_data;
804 unsigned long minsz, ddwsz;
808 case VFIO_CHECK_EXTENSION:
810 case VFIO_SPAPR_TCE_IOMMU:
811 case VFIO_SPAPR_TCE_v2_IOMMU:
815 ret = vfio_spapr_iommu_eeh_ioctl(NULL, cmd, arg);
819 return (ret < 0) ? 0 : ret;
823 * Sanity check to prevent one userspace from manipulating
824 * another userspace mm.
827 if (container->mm && container->mm != current->mm)
831 case VFIO_IOMMU_SPAPR_TCE_GET_INFO: {
832 struct vfio_iommu_spapr_tce_info info;
833 struct tce_iommu_group *tcegrp;
834 struct iommu_table_group *table_group;
836 if (!tce_groups_attached(container))
839 tcegrp = list_first_entry(&container->group_list,
840 struct tce_iommu_group, next);
841 table_group = iommu_group_get_iommudata(tcegrp->grp);
846 minsz = offsetofend(struct vfio_iommu_spapr_tce_info,
849 if (copy_from_user(&info, (void __user *)arg, minsz))
852 if (info.argsz < minsz)
855 info.dma32_window_start = table_group->tce32_start;
856 info.dma32_window_size = table_group->tce32_size;
858 memset(&info.ddw, 0, sizeof(info.ddw));
860 if (table_group->max_dynamic_windows_supported &&
862 info.flags |= VFIO_IOMMU_SPAPR_INFO_DDW;
863 info.ddw.pgsizes = table_group->pgsizes;
864 info.ddw.max_dynamic_windows_supported =
865 table_group->max_dynamic_windows_supported;
866 info.ddw.levels = table_group->max_levels;
869 ddwsz = offsetofend(struct vfio_iommu_spapr_tce_info, ddw);
871 if (info.argsz >= ddwsz)
874 if (copy_to_user((void __user *)arg, &info, minsz))
879 case VFIO_IOMMU_MAP_DMA: {
880 struct vfio_iommu_type1_dma_map param;
881 struct iommu_table *tbl = NULL;
883 enum dma_data_direction direction;
885 if (!container->enabled)
888 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
890 if (copy_from_user(¶m, (void __user *)arg, minsz))
893 if (param.argsz < minsz)
896 if (param.flags & ~(VFIO_DMA_MAP_FLAG_READ |
897 VFIO_DMA_MAP_FLAG_WRITE))
900 ret = tce_iommu_create_default_window(container);
904 num = tce_iommu_find_table(container, param.iova, &tbl);
908 if ((param.size & ~IOMMU_PAGE_MASK(tbl)) ||
909 (param.vaddr & ~IOMMU_PAGE_MASK(tbl)))
912 /* iova is checked by the IOMMU API */
913 if (param.flags & VFIO_DMA_MAP_FLAG_READ) {
914 if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
915 direction = DMA_BIDIRECTIONAL;
917 direction = DMA_TO_DEVICE;
919 if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
920 direction = DMA_FROM_DEVICE;
925 ret = iommu_tce_put_param_check(tbl, param.iova, param.vaddr);
930 ret = tce_iommu_build_v2(container, tbl,
931 param.iova >> tbl->it_page_shift,
933 param.size >> tbl->it_page_shift,
936 ret = tce_iommu_build(container, tbl,
937 param.iova >> tbl->it_page_shift,
939 param.size >> tbl->it_page_shift,
942 iommu_flush_tce(tbl);
946 case VFIO_IOMMU_UNMAP_DMA: {
947 struct vfio_iommu_type1_dma_unmap param;
948 struct iommu_table *tbl = NULL;
951 if (!container->enabled)
954 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap,
957 if (copy_from_user(¶m, (void __user *)arg, minsz))
960 if (param.argsz < minsz)
963 /* No flag is supported now */
967 ret = tce_iommu_create_default_window(container);
971 num = tce_iommu_find_table(container, param.iova, &tbl);
975 if (param.size & ~IOMMU_PAGE_MASK(tbl))
978 ret = iommu_tce_clear_param_check(tbl, param.iova, 0,
979 param.size >> tbl->it_page_shift);
983 ret = tce_iommu_clear(container, tbl,
984 param.iova >> tbl->it_page_shift,
985 param.size >> tbl->it_page_shift);
986 iommu_flush_tce(tbl);
990 case VFIO_IOMMU_SPAPR_REGISTER_MEMORY: {
991 struct vfio_iommu_spapr_register_memory param;
996 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
999 ret = tce_iommu_mm_set(container);
1003 if (copy_from_user(¶m, (void __user *)arg, minsz))
1006 if (param.argsz < minsz)
1009 /* No flag is supported now */
1013 mutex_lock(&container->lock);
1014 ret = tce_iommu_register_pages(container, param.vaddr,
1016 mutex_unlock(&container->lock);
1020 case VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY: {
1021 struct vfio_iommu_spapr_register_memory param;
1029 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1032 if (copy_from_user(¶m, (void __user *)arg, minsz))
1035 if (param.argsz < minsz)
1038 /* No flag is supported now */
1042 mutex_lock(&container->lock);
1043 ret = tce_iommu_unregister_pages(container, param.vaddr,
1045 mutex_unlock(&container->lock);
1049 case VFIO_IOMMU_ENABLE:
1053 mutex_lock(&container->lock);
1054 ret = tce_iommu_enable(container);
1055 mutex_unlock(&container->lock);
1059 case VFIO_IOMMU_DISABLE:
1063 mutex_lock(&container->lock);
1064 tce_iommu_disable(container);
1065 mutex_unlock(&container->lock);
1068 case VFIO_EEH_PE_OP: {
1069 struct tce_iommu_group *tcegrp;
1072 list_for_each_entry(tcegrp, &container->group_list, next) {
1073 ret = vfio_spapr_iommu_eeh_ioctl(tcegrp->grp,
1081 case VFIO_IOMMU_SPAPR_TCE_CREATE: {
1082 struct vfio_iommu_spapr_tce_create create;
1087 ret = tce_iommu_mm_set(container);
1091 if (!tce_groups_attached(container))
1094 minsz = offsetofend(struct vfio_iommu_spapr_tce_create,
1097 if (copy_from_user(&create, (void __user *)arg, minsz))
1100 if (create.argsz < minsz)
1106 mutex_lock(&container->lock);
1108 ret = tce_iommu_create_default_window(container);
1110 ret = tce_iommu_create_window(container,
1112 create.window_size, create.levels,
1113 &create.start_addr);
1115 mutex_unlock(&container->lock);
1117 if (!ret && copy_to_user((void __user *)arg, &create, minsz))
1122 case VFIO_IOMMU_SPAPR_TCE_REMOVE: {
1123 struct vfio_iommu_spapr_tce_remove remove;
1128 ret = tce_iommu_mm_set(container);
1132 if (!tce_groups_attached(container))
1135 minsz = offsetofend(struct vfio_iommu_spapr_tce_remove,
1138 if (copy_from_user(&remove, (void __user *)arg, minsz))
1141 if (remove.argsz < minsz)
1147 if (container->def_window_pending && !remove.start_addr) {
1148 container->def_window_pending = false;
1152 mutex_lock(&container->lock);
1154 ret = tce_iommu_remove_window(container, remove.start_addr);
1156 mutex_unlock(&container->lock);
1165 static void tce_iommu_release_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 = container->tables[i];
1176 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
1178 iommu_release_ownership(tbl);
1180 container->tables[i] = NULL;
1184 static int tce_iommu_take_ownership(struct tce_container *container,
1185 struct iommu_table_group *table_group)
1189 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1190 struct iommu_table *tbl = table_group->tables[i];
1192 if (!tbl || !tbl->it_map)
1195 rc = iommu_take_ownership(tbl);
1197 for (j = 0; j < i; ++j)
1198 iommu_release_ownership(
1199 table_group->tables[j]);
1205 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1206 container->tables[i] = table_group->tables[i];
1211 static void tce_iommu_release_ownership_ddw(struct tce_container *container,
1212 struct iommu_table_group *table_group)
1216 if (!table_group->ops->unset_window) {
1221 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1222 table_group->ops->unset_window(table_group, i);
1224 table_group->ops->release_ownership(table_group);
1227 static long tce_iommu_take_ownership_ddw(struct tce_container *container,
1228 struct iommu_table_group *table_group)
1232 if (!table_group->ops->create_table || !table_group->ops->set_window ||
1233 !table_group->ops->release_ownership) {
1238 table_group->ops->take_ownership(table_group);
1240 /* Set all windows to the new group */
1241 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1242 struct iommu_table *tbl = container->tables[i];
1247 ret = table_group->ops->set_window(table_group, i, tbl);
1255 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1256 table_group->ops->unset_window(table_group, i);
1258 table_group->ops->release_ownership(table_group);
1263 static int tce_iommu_attach_group(void *iommu_data,
1264 struct iommu_group *iommu_group)
1267 struct tce_container *container = iommu_data;
1268 struct iommu_table_group *table_group;
1269 struct tce_iommu_group *tcegrp = NULL;
1271 mutex_lock(&container->lock);
1273 /* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
1274 iommu_group_id(iommu_group), iommu_group); */
1275 table_group = iommu_group_get_iommudata(iommu_group);
1281 if (tce_groups_attached(container) && (!table_group->ops ||
1282 !table_group->ops->take_ownership ||
1283 !table_group->ops->release_ownership)) {
1288 /* Check if new group has the same iommu_ops (i.e. compatible) */
1289 list_for_each_entry(tcegrp, &container->group_list, next) {
1290 struct iommu_table_group *table_group_tmp;
1292 if (tcegrp->grp == iommu_group) {
1293 pr_warn("tce_vfio: Group %d is already attached\n",
1294 iommu_group_id(iommu_group));
1298 table_group_tmp = iommu_group_get_iommudata(tcegrp->grp);
1299 if (table_group_tmp->ops->create_table !=
1300 table_group->ops->create_table) {
1301 pr_warn("tce_vfio: Group %d is incompatible with group %d\n",
1302 iommu_group_id(iommu_group),
1303 iommu_group_id(tcegrp->grp));
1309 tcegrp = kzalloc(sizeof(*tcegrp), GFP_KERNEL);
1315 if (!table_group->ops || !table_group->ops->take_ownership ||
1316 !table_group->ops->release_ownership) {
1317 if (container->v2) {
1321 ret = tce_iommu_take_ownership(container, table_group);
1323 if (!container->v2) {
1327 ret = tce_iommu_take_ownership_ddw(container, table_group);
1328 if (!tce_groups_attached(container) && !container->tables[0])
1329 container->def_window_pending = true;
1333 tcegrp->grp = iommu_group;
1334 list_add(&tcegrp->next, &container->group_list);
1341 mutex_unlock(&container->lock);
1346 static void tce_iommu_detach_group(void *iommu_data,
1347 struct iommu_group *iommu_group)
1349 struct tce_container *container = iommu_data;
1350 struct iommu_table_group *table_group;
1352 struct tce_iommu_group *tcegrp;
1354 mutex_lock(&container->lock);
1356 list_for_each_entry(tcegrp, &container->group_list, next) {
1357 if (tcegrp->grp == iommu_group) {
1364 pr_warn("tce_vfio: detaching unattached group #%u\n",
1365 iommu_group_id(iommu_group));
1369 list_del(&tcegrp->next);
1372 table_group = iommu_group_get_iommudata(iommu_group);
1373 BUG_ON(!table_group);
1375 if (!table_group->ops || !table_group->ops->release_ownership)
1376 tce_iommu_release_ownership(container, table_group);
1378 tce_iommu_release_ownership_ddw(container, table_group);
1381 mutex_unlock(&container->lock);
1384 const struct vfio_iommu_driver_ops tce_iommu_driver_ops = {
1385 .name = "iommu-vfio-powerpc",
1386 .owner = THIS_MODULE,
1387 .open = tce_iommu_open,
1388 .release = tce_iommu_release,
1389 .ioctl = tce_iommu_ioctl,
1390 .attach_group = tce_iommu_attach_group,
1391 .detach_group = tce_iommu_detach_group,
1394 static int __init tce_iommu_init(void)
1396 return vfio_register_iommu_driver(&tce_iommu_driver_ops);
1399 static void __exit tce_iommu_cleanup(void)
1401 vfio_unregister_iommu_driver(&tce_iommu_driver_ops);
1404 module_init(tce_iommu_init);
1405 module_exit(tce_iommu_cleanup);
1407 MODULE_VERSION(DRIVER_VERSION);
1408 MODULE_LICENSE("GPL v2");
1409 MODULE_AUTHOR(DRIVER_AUTHOR);
1410 MODULE_DESCRIPTION(DRIVER_DESC);