1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  *  IOMMU helpers in MMU context.
4  *
5  *  Copyright (C) 2015 IBM Corp. <aik@ozlabs.ru>
6  */
7 
8 #include <linux/sched/signal.h>
9 #include <linux/slab.h>
10 #include <linux/rculist.h>
11 #include <linux/vmalloc.h>
12 #include <linux/mutex.h>
13 #include <linux/migrate.h>
14 #include <linux/hugetlb.h>
15 #include <linux/swap.h>
16 #include <linux/sizes.h>
17 #include <linux/mm.h>
18 #include <asm/mmu_context.h>
19 #include <asm/pte-walk.h>
20 #include <linux/mm_inline.h>
21 
22 static DEFINE_MUTEX(mem_list_mutex);
23 
24 #define MM_IOMMU_TABLE_GROUP_PAGE_DIRTY	0x1
25 #define MM_IOMMU_TABLE_GROUP_PAGE_MASK	~(SZ_4K - 1)
26 
27 struct mm_iommu_table_group_mem_t {
28 	struct list_head next;
29 	struct rcu_head rcu;
30 	unsigned long used;
31 	atomic64_t mapped;
32 	unsigned int pageshift;
33 	u64 ua;			/* userspace address */
34 	u64 entries;		/* number of entries in hpas/hpages[] */
35 	/*
36 	 * in mm_iommu_get we temporarily use this to store
37 	 * struct page address.
38 	 *
39 	 * We need to convert ua to hpa in real mode. Make it
40 	 * simpler by storing physical address.
41 	 */
42 	union {
43 		struct page **hpages;	/* vmalloc'ed */
44 		phys_addr_t *hpas;
45 	};
46 #define MM_IOMMU_TABLE_INVALID_HPA	((uint64_t)-1)
47 	u64 dev_hpa;		/* Device memory base address */
48 };
49 
50 bool mm_iommu_preregistered(struct mm_struct *mm)
51 {
52 	return !list_empty(&mm->context.iommu_group_mem_list);
53 }
54 EXPORT_SYMBOL_GPL(mm_iommu_preregistered);
55 
56 static long mm_iommu_do_alloc(struct mm_struct *mm, unsigned long ua,
57 			      unsigned long entries, unsigned long dev_hpa,
58 			      struct mm_iommu_table_group_mem_t **pmem)
59 {
60 	struct mm_iommu_table_group_mem_t *mem, *mem2;
61 	long i, ret, locked_entries = 0, pinned = 0;
62 	unsigned int pageshift;
63 	unsigned long entry, chunk;
64 
65 	if (dev_hpa == MM_IOMMU_TABLE_INVALID_HPA) {
66 		ret = account_locked_vm(mm, entries, true);
67 		if (ret)
68 			return ret;
69 
70 		locked_entries = entries;
71 	}
72 
73 	mem = kzalloc(sizeof(*mem), GFP_KERNEL);
74 	if (!mem) {
75 		ret = -ENOMEM;
76 		goto unlock_exit;
77 	}
78 
79 	if (dev_hpa != MM_IOMMU_TABLE_INVALID_HPA) {
80 		mem->pageshift = __ffs(dev_hpa | (entries << PAGE_SHIFT));
81 		mem->dev_hpa = dev_hpa;
82 		goto good_exit;
83 	}
84 	mem->dev_hpa = MM_IOMMU_TABLE_INVALID_HPA;
85 
86 	/*
87 	 * For a starting point for a maximum page size calculation
88 	 * we use @ua and @entries natural alignment to allow IOMMU pages
89 	 * smaller than huge pages but still bigger than PAGE_SIZE.
90 	 */
91 	mem->pageshift = __ffs(ua | (entries << PAGE_SHIFT));
92 	mem->hpas = vzalloc(array_size(entries, sizeof(mem->hpas[0])));
93 	if (!mem->hpas) {
94 		kfree(mem);
95 		ret = -ENOMEM;
96 		goto unlock_exit;
97 	}
98 
99 	down_read(&mm->mmap_sem);
100 	chunk = (1UL << (PAGE_SHIFT + MAX_ORDER - 1)) /
101 			sizeof(struct vm_area_struct *);
102 	chunk = min(chunk, entries);
103 	for (entry = 0; entry < entries; entry += chunk) {
104 		unsigned long n = min(entries - entry, chunk);
105 
106 		ret = get_user_pages(ua + (entry << PAGE_SHIFT), n,
107 				FOLL_WRITE | FOLL_LONGTERM,
108 				mem->hpages + entry, NULL);
109 		if (ret == n) {
110 			pinned += n;
111 			continue;
112 		}
113 		if (ret > 0)
114 			pinned += ret;
115 		break;
116 	}
117 	up_read(&mm->mmap_sem);
118 	if (pinned != entries) {
119 		if (!ret)
120 			ret = -EFAULT;
121 		goto free_exit;
122 	}
123 
124 	pageshift = PAGE_SHIFT;
125 	for (i = 0; i < entries; ++i) {
126 		struct page *page = mem->hpages[i];
127 
128 		/*
129 		 * Allow to use larger than 64k IOMMU pages. Only do that
130 		 * if we are backed by hugetlb.
131 		 */
132 		if ((mem->pageshift > PAGE_SHIFT) && PageHuge(page))
133 			pageshift = page_shift(compound_head(page));
134 		mem->pageshift = min(mem->pageshift, pageshift);
135 		/*
136 		 * We don't need struct page reference any more, switch
137 		 * to physical address.
138 		 */
139 		mem->hpas[i] = page_to_pfn(page) << PAGE_SHIFT;
140 	}
141 
142 good_exit:
143 	atomic64_set(&mem->mapped, 1);
144 	mem->used = 1;
145 	mem->ua = ua;
146 	mem->entries = entries;
147 
148 	mutex_lock(&mem_list_mutex);
149 
150 	list_for_each_entry_rcu(mem2, &mm->context.iommu_group_mem_list, next) {
151 		/* Overlap? */
152 		if ((mem2->ua < (ua + (entries << PAGE_SHIFT))) &&
153 				(ua < (mem2->ua +
154 				       (mem2->entries << PAGE_SHIFT)))) {
155 			ret = -EINVAL;
156 			mutex_unlock(&mem_list_mutex);
157 			goto free_exit;
158 		}
159 	}
160 
161 	list_add_rcu(&mem->next, &mm->context.iommu_group_mem_list);
162 
163 	mutex_unlock(&mem_list_mutex);
164 
165 	*pmem = mem;
166 
167 	return 0;
168 
169 free_exit:
170 	/* free the reference taken */
171 	for (i = 0; i < pinned; i++)
172 		put_page(mem->hpages[i]);
173 
174 	vfree(mem->hpas);
175 	kfree(mem);
176 
177 unlock_exit:
178 	account_locked_vm(mm, locked_entries, false);
179 
180 	return ret;
181 }
182 
183 long mm_iommu_new(struct mm_struct *mm, unsigned long ua, unsigned long entries,
184 		struct mm_iommu_table_group_mem_t **pmem)
185 {
186 	return mm_iommu_do_alloc(mm, ua, entries, MM_IOMMU_TABLE_INVALID_HPA,
187 			pmem);
188 }
189 EXPORT_SYMBOL_GPL(mm_iommu_new);
190 
191 long mm_iommu_newdev(struct mm_struct *mm, unsigned long ua,
192 		unsigned long entries, unsigned long dev_hpa,
193 		struct mm_iommu_table_group_mem_t **pmem)
194 {
195 	return mm_iommu_do_alloc(mm, ua, entries, dev_hpa, pmem);
196 }
197 EXPORT_SYMBOL_GPL(mm_iommu_newdev);
198 
199 static void mm_iommu_unpin(struct mm_iommu_table_group_mem_t *mem)
200 {
201 	long i;
202 	struct page *page = NULL;
203 
204 	if (!mem->hpas)
205 		return;
206 
207 	for (i = 0; i < mem->entries; ++i) {
208 		if (!mem->hpas[i])
209 			continue;
210 
211 		page = pfn_to_page(mem->hpas[i] >> PAGE_SHIFT);
212 		if (!page)
213 			continue;
214 
215 		if (mem->hpas[i] & MM_IOMMU_TABLE_GROUP_PAGE_DIRTY)
216 			SetPageDirty(page);
217 
218 		put_page(page);
219 		mem->hpas[i] = 0;
220 	}
221 }
222 
223 static void mm_iommu_do_free(struct mm_iommu_table_group_mem_t *mem)
224 {
225 
226 	mm_iommu_unpin(mem);
227 	vfree(mem->hpas);
228 	kfree(mem);
229 }
230 
231 static void mm_iommu_free(struct rcu_head *head)
232 {
233 	struct mm_iommu_table_group_mem_t *mem = container_of(head,
234 			struct mm_iommu_table_group_mem_t, rcu);
235 
236 	mm_iommu_do_free(mem);
237 }
238 
239 static void mm_iommu_release(struct mm_iommu_table_group_mem_t *mem)
240 {
241 	list_del_rcu(&mem->next);
242 	call_rcu(&mem->rcu, mm_iommu_free);
243 }
244 
245 long mm_iommu_put(struct mm_struct *mm, struct mm_iommu_table_group_mem_t *mem)
246 {
247 	long ret = 0;
248 	unsigned long unlock_entries = 0;
249 
250 	mutex_lock(&mem_list_mutex);
251 
252 	if (mem->used == 0) {
253 		ret = -ENOENT;
254 		goto unlock_exit;
255 	}
256 
257 	--mem->used;
258 	/* There are still users, exit */
259 	if (mem->used)
260 		goto unlock_exit;
261 
262 	/* Are there still mappings? */
263 	if (atomic_cmpxchg(&mem->mapped, 1, 0) != 1) {
264 		++mem->used;
265 		ret = -EBUSY;
266 		goto unlock_exit;
267 	}
268 
269 	if (mem->dev_hpa == MM_IOMMU_TABLE_INVALID_HPA)
270 		unlock_entries = mem->entries;
271 
272 	/* @mapped became 0 so now mappings are disabled, release the region */
273 	mm_iommu_release(mem);
274 
275 unlock_exit:
276 	mutex_unlock(&mem_list_mutex);
277 
278 	account_locked_vm(mm, unlock_entries, false);
279 
280 	return ret;
281 }
282 EXPORT_SYMBOL_GPL(mm_iommu_put);
283 
284 struct mm_iommu_table_group_mem_t *mm_iommu_lookup(struct mm_struct *mm,
285 		unsigned long ua, unsigned long size)
286 {
287 	struct mm_iommu_table_group_mem_t *mem, *ret = NULL;
288 
289 	list_for_each_entry_rcu(mem, &mm->context.iommu_group_mem_list, next) {
290 		if ((mem->ua <= ua) &&
291 				(ua + size <= mem->ua +
292 				 (mem->entries << PAGE_SHIFT))) {
293 			ret = mem;
294 			break;
295 		}
296 	}
297 
298 	return ret;
299 }
300 EXPORT_SYMBOL_GPL(mm_iommu_lookup);
301 
302 struct mm_iommu_table_group_mem_t *mm_iommu_lookup_rm(struct mm_struct *mm,
303 		unsigned long ua, unsigned long size)
304 {
305 	struct mm_iommu_table_group_mem_t *mem, *ret = NULL;
306 
307 	list_for_each_entry_lockless(mem, &mm->context.iommu_group_mem_list,
308 			next) {
309 		if ((mem->ua <= ua) &&
310 				(ua + size <= mem->ua +
311 				 (mem->entries << PAGE_SHIFT))) {
312 			ret = mem;
313 			break;
314 		}
315 	}
316 
317 	return ret;
318 }
319 
320 struct mm_iommu_table_group_mem_t *mm_iommu_get(struct mm_struct *mm,
321 		unsigned long ua, unsigned long entries)
322 {
323 	struct mm_iommu_table_group_mem_t *mem, *ret = NULL;
324 
325 	mutex_lock(&mem_list_mutex);
326 
327 	list_for_each_entry_rcu(mem, &mm->context.iommu_group_mem_list, next) {
328 		if ((mem->ua == ua) && (mem->entries == entries)) {
329 			ret = mem;
330 			++mem->used;
331 			break;
332 		}
333 	}
334 
335 	mutex_unlock(&mem_list_mutex);
336 
337 	return ret;
338 }
339 EXPORT_SYMBOL_GPL(mm_iommu_get);
340 
341 long mm_iommu_ua_to_hpa(struct mm_iommu_table_group_mem_t *mem,
342 		unsigned long ua, unsigned int pageshift, unsigned long *hpa)
343 {
344 	const long entry = (ua - mem->ua) >> PAGE_SHIFT;
345 	u64 *va;
346 
347 	if (entry >= mem->entries)
348 		return -EFAULT;
349 
350 	if (pageshift > mem->pageshift)
351 		return -EFAULT;
352 
353 	if (!mem->hpas) {
354 		*hpa = mem->dev_hpa + (ua - mem->ua);
355 		return 0;
356 	}
357 
358 	va = &mem->hpas[entry];
359 	*hpa = (*va & MM_IOMMU_TABLE_GROUP_PAGE_MASK) | (ua & ~PAGE_MASK);
360 
361 	return 0;
362 }
363 EXPORT_SYMBOL_GPL(mm_iommu_ua_to_hpa);
364 
365 long mm_iommu_ua_to_hpa_rm(struct mm_iommu_table_group_mem_t *mem,
366 		unsigned long ua, unsigned int pageshift, unsigned long *hpa)
367 {
368 	const long entry = (ua - mem->ua) >> PAGE_SHIFT;
369 	unsigned long *pa;
370 
371 	if (entry >= mem->entries)
372 		return -EFAULT;
373 
374 	if (pageshift > mem->pageshift)
375 		return -EFAULT;
376 
377 	if (!mem->hpas) {
378 		*hpa = mem->dev_hpa + (ua - mem->ua);
379 		return 0;
380 	}
381 
382 	pa = (void *) vmalloc_to_phys(&mem->hpas[entry]);
383 	if (!pa)
384 		return -EFAULT;
385 
386 	*hpa = (*pa & MM_IOMMU_TABLE_GROUP_PAGE_MASK) | (ua & ~PAGE_MASK);
387 
388 	return 0;
389 }
390 
391 extern void mm_iommu_ua_mark_dirty_rm(struct mm_struct *mm, unsigned long ua)
392 {
393 	struct mm_iommu_table_group_mem_t *mem;
394 	long entry;
395 	void *va;
396 	unsigned long *pa;
397 
398 	mem = mm_iommu_lookup_rm(mm, ua, PAGE_SIZE);
399 	if (!mem)
400 		return;
401 
402 	if (mem->dev_hpa != MM_IOMMU_TABLE_INVALID_HPA)
403 		return;
404 
405 	entry = (ua - mem->ua) >> PAGE_SHIFT;
406 	va = &mem->hpas[entry];
407 
408 	pa = (void *) vmalloc_to_phys(va);
409 	if (!pa)
410 		return;
411 
412 	*pa |= MM_IOMMU_TABLE_GROUP_PAGE_DIRTY;
413 }
414 
415 bool mm_iommu_is_devmem(struct mm_struct *mm, unsigned long hpa,
416 		unsigned int pageshift, unsigned long *size)
417 {
418 	struct mm_iommu_table_group_mem_t *mem;
419 	unsigned long end;
420 
421 	list_for_each_entry_rcu(mem, &mm->context.iommu_group_mem_list, next) {
422 		if (mem->dev_hpa == MM_IOMMU_TABLE_INVALID_HPA)
423 			continue;
424 
425 		end = mem->dev_hpa + (mem->entries << PAGE_SHIFT);
426 		if ((mem->dev_hpa <= hpa) && (hpa < end)) {
427 			/*
428 			 * Since the IOMMU page size might be bigger than
429 			 * PAGE_SIZE, the amount of preregistered memory
430 			 * starting from @hpa might be smaller than 1<<pageshift
431 			 * and the caller needs to distinguish this situation.
432 			 */
433 			*size = min(1UL << pageshift, end - hpa);
434 			return true;
435 		}
436 	}
437 
438 	return false;
439 }
440 EXPORT_SYMBOL_GPL(mm_iommu_is_devmem);
441 
442 long mm_iommu_mapped_inc(struct mm_iommu_table_group_mem_t *mem)
443 {
444 	if (atomic64_inc_not_zero(&mem->mapped))
445 		return 0;
446 
447 	/* Last mm_iommu_put() has been called, no more mappings allowed() */
448 	return -ENXIO;
449 }
450 EXPORT_SYMBOL_GPL(mm_iommu_mapped_inc);
451 
452 void mm_iommu_mapped_dec(struct mm_iommu_table_group_mem_t *mem)
453 {
454 	atomic64_add_unless(&mem->mapped, -1, 1);
455 }
456 EXPORT_SYMBOL_GPL(mm_iommu_mapped_dec);
457 
458 void mm_iommu_init(struct mm_struct *mm)
459 {
460 	INIT_LIST_HEAD_RCU(&mm->context.iommu_group_mem_list);
461 }
462