2 * IOMMU helpers in MMU context.
4 * Copyright (C) 2015 IBM Corp. <aik@ozlabs.ru>
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * as published by the Free Software Foundation; either version
9 * 2 of the License, or (at your option) any later version.
13 #include <linux/sched/signal.h>
14 #include <linux/slab.h>
15 #include <linux/rculist.h>
16 #include <linux/vmalloc.h>
17 #include <linux/mutex.h>
18 #include <linux/migrate.h>
19 #include <linux/hugetlb.h>
20 #include <linux/swap.h>
21 #include <linux/sizes.h>
22 #include <asm/mmu_context.h>
23 #include <asm/pte-walk.h>
25 static DEFINE_MUTEX(mem_list_mutex);
27 #define MM_IOMMU_TABLE_GROUP_PAGE_DIRTY 0x1
28 #define MM_IOMMU_TABLE_GROUP_PAGE_MASK ~(SZ_4K - 1)
30 struct mm_iommu_table_group_mem_t {
31 struct list_head next;
35 unsigned int pageshift;
36 u64 ua; /* userspace address */
37 u64 entries; /* number of entries in hpas[] */
38 u64 *hpas; /* vmalloc'ed */
41 static long mm_iommu_adjust_locked_vm(struct mm_struct *mm,
42 unsigned long npages, bool incr)
44 long ret = 0, locked, lock_limit;
49 down_write(&mm->mmap_sem);
52 locked = mm->locked_vm + npages;
53 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
54 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
57 mm->locked_vm += npages;
59 if (WARN_ON_ONCE(npages > mm->locked_vm))
60 npages = mm->locked_vm;
61 mm->locked_vm -= npages;
64 pr_debug("[%d] RLIMIT_MEMLOCK HASH64 %c%ld %ld/%ld\n",
65 current ? current->pid : 0,
68 mm->locked_vm << PAGE_SHIFT,
69 rlimit(RLIMIT_MEMLOCK));
70 up_write(&mm->mmap_sem);
75 bool mm_iommu_preregistered(struct mm_struct *mm)
77 return !list_empty(&mm->context.iommu_group_mem_list);
79 EXPORT_SYMBOL_GPL(mm_iommu_preregistered);
82 * Taken from alloc_migrate_target with changes to remove CMA allocations
84 struct page *new_iommu_non_cma_page(struct page *page, unsigned long private)
86 gfp_t gfp_mask = GFP_USER;
87 struct page *new_page;
89 if (PageCompound(page))
92 if (PageHighMem(page))
93 gfp_mask |= __GFP_HIGHMEM;
96 * We don't want the allocation to force an OOM if possibe
98 new_page = alloc_page(gfp_mask | __GFP_NORETRY | __GFP_NOWARN);
102 static int mm_iommu_move_page_from_cma(struct page *page)
105 LIST_HEAD(cma_migrate_pages);
107 /* Ignore huge pages for now */
108 if (PageCompound(page))
112 ret = isolate_lru_page(page);
116 list_add(&page->lru, &cma_migrate_pages);
117 put_page(page); /* Drop the gup reference */
119 ret = migrate_pages(&cma_migrate_pages, new_iommu_non_cma_page,
120 NULL, 0, MIGRATE_SYNC, MR_CONTIG_RANGE);
122 if (!list_empty(&cma_migrate_pages))
123 putback_movable_pages(&cma_migrate_pages);
129 long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries,
130 struct mm_iommu_table_group_mem_t **pmem)
132 struct mm_iommu_table_group_mem_t *mem;
133 long i, j, ret = 0, locked_entries = 0;
134 unsigned int pageshift;
136 unsigned long cur_ua;
137 struct page *page = NULL;
139 mutex_lock(&mem_list_mutex);
141 list_for_each_entry_rcu(mem, &mm->context.iommu_group_mem_list,
143 if ((mem->ua == ua) && (mem->entries == entries)) {
150 if ((mem->ua < (ua + (entries << PAGE_SHIFT))) &&
152 (mem->entries << PAGE_SHIFT)))) {
159 ret = mm_iommu_adjust_locked_vm(mm, entries, true);
163 locked_entries = entries;
165 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
172 * For a starting point for a maximum page size calculation
173 * we use @ua and @entries natural alignment to allow IOMMU pages
174 * smaller than huge pages but still bigger than PAGE_SIZE.
176 mem->pageshift = __ffs(ua | (entries << PAGE_SHIFT));
177 mem->hpas = vzalloc(array_size(entries, sizeof(mem->hpas[0])));
184 for (i = 0; i < entries; ++i) {
185 cur_ua = ua + (i << PAGE_SHIFT);
186 if (1 != get_user_pages_fast(cur_ua,
187 1/* pages */, 1/* iswrite */, &page)) {
189 for (j = 0; j < i; ++j)
190 put_page(pfn_to_page(mem->hpas[j] >>
197 * If we get a page from the CMA zone, since we are going to
198 * be pinning these entries, we might as well move them out
199 * of the CMA zone if possible. NOTE: faulting in + migration
200 * can be expensive. Batching can be considered later
202 if (is_migrate_cma_page(page)) {
203 if (mm_iommu_move_page_from_cma(page))
205 if (1 != get_user_pages_fast(cur_ua,
206 1/* pages */, 1/* iswrite */,
209 for (j = 0; j < i; ++j)
210 put_page(pfn_to_page(mem->hpas[j] >>
218 pageshift = PAGE_SHIFT;
219 if (mem->pageshift > PAGE_SHIFT && PageCompound(page)) {
221 struct page *head = compound_head(page);
222 unsigned int compshift = compound_order(head);
223 unsigned int pteshift;
225 local_irq_save(flags); /* disables as well */
226 pte = find_linux_pte(mm->pgd, cur_ua, NULL, &pteshift);
228 /* Double check it is still the same pinned page */
229 if (pte && pte_page(*pte) == head &&
230 pteshift == compshift + PAGE_SHIFT)
231 pageshift = max_t(unsigned int, pteshift,
233 local_irq_restore(flags);
235 mem->pageshift = min(mem->pageshift, pageshift);
236 mem->hpas[i] = page_to_pfn(page) << PAGE_SHIFT;
239 atomic64_set(&mem->mapped, 1);
242 mem->entries = entries;
245 list_add_rcu(&mem->next, &mm->context.iommu_group_mem_list);
248 if (locked_entries && ret)
249 mm_iommu_adjust_locked_vm(mm, locked_entries, false);
251 mutex_unlock(&mem_list_mutex);
255 EXPORT_SYMBOL_GPL(mm_iommu_get);
257 static void mm_iommu_unpin(struct mm_iommu_table_group_mem_t *mem)
260 struct page *page = NULL;
262 for (i = 0; i < mem->entries; ++i) {
266 page = pfn_to_page(mem->hpas[i] >> PAGE_SHIFT);
270 if (mem->hpas[i] & MM_IOMMU_TABLE_GROUP_PAGE_DIRTY)
278 static void mm_iommu_do_free(struct mm_iommu_table_group_mem_t *mem)
286 static void mm_iommu_free(struct rcu_head *head)
288 struct mm_iommu_table_group_mem_t *mem = container_of(head,
289 struct mm_iommu_table_group_mem_t, rcu);
291 mm_iommu_do_free(mem);
294 static void mm_iommu_release(struct mm_iommu_table_group_mem_t *mem)
296 list_del_rcu(&mem->next);
297 call_rcu(&mem->rcu, mm_iommu_free);
300 long mm_iommu_put(struct mm_struct *mm, struct mm_iommu_table_group_mem_t *mem)
304 mutex_lock(&mem_list_mutex);
306 if (mem->used == 0) {
312 /* There are still users, exit */
316 /* Are there still mappings? */
317 if (atomic_cmpxchg(&mem->mapped, 1, 0) != 1) {
323 /* @mapped became 0 so now mappings are disabled, release the region */
324 mm_iommu_release(mem);
326 mm_iommu_adjust_locked_vm(mm, mem->entries, false);
329 mutex_unlock(&mem_list_mutex);
333 EXPORT_SYMBOL_GPL(mm_iommu_put);
335 struct mm_iommu_table_group_mem_t *mm_iommu_lookup(struct mm_struct *mm,
336 unsigned long ua, unsigned long size)
338 struct mm_iommu_table_group_mem_t *mem, *ret = NULL;
340 list_for_each_entry_rcu(mem, &mm->context.iommu_group_mem_list, next) {
341 if ((mem->ua <= ua) &&
342 (ua + size <= mem->ua +
343 (mem->entries << PAGE_SHIFT))) {
351 EXPORT_SYMBOL_GPL(mm_iommu_lookup);
353 struct mm_iommu_table_group_mem_t *mm_iommu_lookup_rm(struct mm_struct *mm,
354 unsigned long ua, unsigned long size)
356 struct mm_iommu_table_group_mem_t *mem, *ret = NULL;
358 list_for_each_entry_lockless(mem, &mm->context.iommu_group_mem_list,
360 if ((mem->ua <= ua) &&
361 (ua + size <= mem->ua +
362 (mem->entries << PAGE_SHIFT))) {
371 struct mm_iommu_table_group_mem_t *mm_iommu_find(struct mm_struct *mm,
372 unsigned long ua, unsigned long entries)
374 struct mm_iommu_table_group_mem_t *mem, *ret = NULL;
376 list_for_each_entry_rcu(mem, &mm->context.iommu_group_mem_list, next) {
377 if ((mem->ua == ua) && (mem->entries == entries)) {
385 EXPORT_SYMBOL_GPL(mm_iommu_find);
387 long mm_iommu_ua_to_hpa(struct mm_iommu_table_group_mem_t *mem,
388 unsigned long ua, unsigned int pageshift, unsigned long *hpa)
390 const long entry = (ua - mem->ua) >> PAGE_SHIFT;
391 u64 *va = &mem->hpas[entry];
393 if (entry >= mem->entries)
396 if (pageshift > mem->pageshift)
399 *hpa = (*va & MM_IOMMU_TABLE_GROUP_PAGE_MASK) | (ua & ~PAGE_MASK);
403 EXPORT_SYMBOL_GPL(mm_iommu_ua_to_hpa);
405 long mm_iommu_ua_to_hpa_rm(struct mm_iommu_table_group_mem_t *mem,
406 unsigned long ua, unsigned int pageshift, unsigned long *hpa)
408 const long entry = (ua - mem->ua) >> PAGE_SHIFT;
409 void *va = &mem->hpas[entry];
412 if (entry >= mem->entries)
415 if (pageshift > mem->pageshift)
418 pa = (void *) vmalloc_to_phys(va);
422 *hpa = (*pa & MM_IOMMU_TABLE_GROUP_PAGE_MASK) | (ua & ~PAGE_MASK);
427 extern void mm_iommu_ua_mark_dirty_rm(struct mm_struct *mm, unsigned long ua)
429 struct mm_iommu_table_group_mem_t *mem;
434 mem = mm_iommu_lookup_rm(mm, ua, PAGE_SIZE);
438 entry = (ua - mem->ua) >> PAGE_SHIFT;
439 va = &mem->hpas[entry];
441 pa = (void *) vmalloc_to_phys(va);
445 *pa |= MM_IOMMU_TABLE_GROUP_PAGE_DIRTY;
448 long mm_iommu_mapped_inc(struct mm_iommu_table_group_mem_t *mem)
450 if (atomic64_inc_not_zero(&mem->mapped))
453 /* Last mm_iommu_put() has been called, no more mappings allowed() */
456 EXPORT_SYMBOL_GPL(mm_iommu_mapped_inc);
458 void mm_iommu_mapped_dec(struct mm_iommu_table_group_mem_t *mem)
460 atomic64_add_unless(&mem->mapped, -1, 1);
462 EXPORT_SYMBOL_GPL(mm_iommu_mapped_dec);
464 void mm_iommu_init(struct mm_struct *mm)
466 INIT_LIST_HEAD_RCU(&mm->context.iommu_group_mem_list);