xref: /linux/mm/execmem.c (revision 1871d548fc4feb007644efb6d669c93a4e191254) 
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2002 Richard Henderson
4  * Copyright (C) 2001 Rusty Russell, 2002, 2010 Rusty Russell IBM.
5  * Copyright (C) 2023 Luis Chamberlain <mcgrof@kernel.org>
6  * Copyright (C) 2024 Mike Rapoport IBM.
7  */
8 
9 #define pr_fmt(fmt) "execmem: " fmt
10 
11 #include <linux/mm.h>
12 #include <linux/mutex.h>
13 #include <linux/vmalloc.h>
14 #include <linux/execmem.h>
15 #include <linux/maple_tree.h>
16 #include <linux/set_memory.h>
17 #include <linux/moduleloader.h>
18 #include <linux/text-patching.h>
19 
20 #include <asm/tlbflush.h>
21 
22 #include "internal.h"
23 
24 static struct execmem_info *execmem_info __ro_after_init;
25 static struct execmem_info default_execmem_info __ro_after_init;
26 
27 #ifdef CONFIG_MMU
28 static void *execmem_vmalloc(struct execmem_range *range, size_t size,
29 			     pgprot_t pgprot, unsigned long vm_flags)
30 {
31 	bool kasan = range->flags & EXECMEM_KASAN_SHADOW;
32 	gfp_t gfp_flags = GFP_KERNEL | __GFP_NOWARN;
33 	unsigned int align = range->alignment;
34 	unsigned long start = range->start;
35 	unsigned long end = range->end;
36 	void *p;
37 
38 	if (kasan)
39 		vm_flags |= VM_DEFER_KMEMLEAK;
40 
41 	p = __vmalloc_node_range(size, align, start, end, gfp_flags,
42 				 pgprot, vm_flags, NUMA_NO_NODE,
43 				 __builtin_return_address(0));
44 	if (!p && range->fallback_start) {
45 		start = range->fallback_start;
46 		end = range->fallback_end;
47 		p = __vmalloc_node_range(size, align, start, end, gfp_flags,
48 					 pgprot, vm_flags, NUMA_NO_NODE,
49 					 __builtin_return_address(0));
50 	}
51 
52 	if (!p) {
53 		pr_warn_ratelimited("unable to allocate memory\n");
54 		return NULL;
55 	}
56 
57 	if (kasan && (kasan_alloc_module_shadow(p, size, GFP_KERNEL) < 0)) {
58 		vfree(p);
59 		return NULL;
60 	}
61 
62 	return p;
63 }
64 
65 struct vm_struct *execmem_vmap(size_t size)
66 {
67 	struct execmem_range *range = &execmem_info->ranges[EXECMEM_MODULE_DATA];
68 	struct vm_struct *area;
69 
70 	area = __get_vm_area_node(size, range->alignment, PAGE_SHIFT, VM_ALLOC,
71 				  range->start, range->end, NUMA_NO_NODE,
72 				  GFP_KERNEL, __builtin_return_address(0));
73 	if (!area && range->fallback_start)
74 		area = __get_vm_area_node(size, range->alignment, PAGE_SHIFT, VM_ALLOC,
75 					  range->fallback_start, range->fallback_end,
76 					  NUMA_NO_NODE, GFP_KERNEL, __builtin_return_address(0));
77 
78 	return area;
79 }
80 #else
81 static void *execmem_vmalloc(struct execmem_range *range, size_t size,
82 			     pgprot_t pgprot, unsigned long vm_flags)
83 {
84 	return vmalloc(size);
85 }
86 #endif /* CONFIG_MMU */
87 
88 #ifdef CONFIG_ARCH_HAS_EXECMEM_ROX
89 struct execmem_cache {
90 	struct mutex mutex;
91 	struct maple_tree busy_areas;
92 	struct maple_tree free_areas;
93 	unsigned int pending_free_cnt;	/* protected by mutex */
94 };
95 
96 /* delay to schedule asynchronous free if fast path free fails */
97 #define FREE_DELAY	(msecs_to_jiffies(10))
98 
99 /* mark entries in busy_areas that should be freed asynchronously */
100 #define PENDING_FREE_MASK	(1 << (PAGE_SHIFT - 1))
101 
102 static struct execmem_cache execmem_cache = {
103 	.mutex = __MUTEX_INITIALIZER(execmem_cache.mutex),
104 	.busy_areas = MTREE_INIT_EXT(busy_areas, MT_FLAGS_LOCK_EXTERN,
105 				     execmem_cache.mutex),
106 	.free_areas = MTREE_INIT_EXT(free_areas, MT_FLAGS_LOCK_EXTERN,
107 				     execmem_cache.mutex),
108 };
109 
110 static inline unsigned long mas_range_len(struct ma_state *mas)
111 {
112 	return mas->last - mas->index + 1;
113 }
114 
115 static int execmem_set_direct_map_valid(struct vm_struct *vm, bool valid)
116 {
117 	unsigned int nr = (1 << get_vm_area_page_order(vm));
118 	unsigned int updated = 0;
119 	int err = 0;
120 
121 	for (int i = 0; i < vm->nr_pages; i += nr) {
122 		err = set_direct_map_valid_noflush(vm->pages[i], nr, valid);
123 		if (err)
124 			goto err_restore;
125 		updated += nr;
126 	}
127 
128 	return 0;
129 
130 err_restore:
131 	for (int i = 0; i < updated; i += nr)
132 		set_direct_map_valid_noflush(vm->pages[i], nr, !valid);
133 
134 	return err;
135 }
136 
137 static int execmem_force_rw(void *ptr, size_t size)
138 {
139 	unsigned int nr = PAGE_ALIGN(size) >> PAGE_SHIFT;
140 	unsigned long addr = (unsigned long)ptr;
141 	int ret;
142 
143 	ret = set_memory_nx(addr, nr);
144 	if (ret)
145 		return ret;
146 
147 	return set_memory_rw(addr, nr);
148 }
149 
150 int execmem_restore_rox(void *ptr, size_t size)
151 {
152 	unsigned int nr = PAGE_ALIGN(size) >> PAGE_SHIFT;
153 	unsigned long addr = (unsigned long)ptr;
154 
155 	return set_memory_rox(addr, nr);
156 }
157 
158 static void execmem_cache_clean(struct work_struct *work)
159 {
160 	struct maple_tree *free_areas = &execmem_cache.free_areas;
161 	struct mutex *mutex = &execmem_cache.mutex;
162 	MA_STATE(mas, free_areas, 0, ULONG_MAX);
163 	void *area;
164 
165 	mutex_lock(mutex);
166 	mas_for_each(&mas, area, ULONG_MAX) {
167 		size_t size = mas_range_len(&mas);
168 
169 		if (IS_ALIGNED(size, PMD_SIZE) &&
170 		    IS_ALIGNED(mas.index, PMD_SIZE)) {
171 			struct vm_struct *vm = find_vm_area(area);
172 
173 			execmem_set_direct_map_valid(vm, true);
174 			mas_store_gfp(&mas, NULL, GFP_KERNEL);
175 			vfree(area);
176 		}
177 	}
178 	mutex_unlock(mutex);
179 }
180 
181 static DECLARE_WORK(execmem_cache_clean_work, execmem_cache_clean);
182 
183 static int execmem_cache_add_locked(void *ptr, size_t size, gfp_t gfp_mask)
184 {
185 	struct maple_tree *free_areas = &execmem_cache.free_areas;
186 	unsigned long addr = (unsigned long)ptr;
187 	MA_STATE(mas, free_areas, addr - 1, addr + 1);
188 	unsigned long lower, upper;
189 	void *area = NULL;
190 
191 	lower = addr;
192 	upper = addr + size - 1;
193 
194 	area = mas_walk(&mas);
195 	if (area && mas.last == addr - 1)
196 		lower = mas.index;
197 
198 	area = mas_next(&mas, ULONG_MAX);
199 	if (area && mas.index == addr + size)
200 		upper = mas.last;
201 
202 	mas_set_range(&mas, lower, upper);
203 	return mas_store_gfp(&mas, (void *)lower, gfp_mask);
204 }
205 
206 static bool within_range(struct execmem_range *range, struct ma_state *mas,
207 			 size_t size)
208 {
209 	unsigned long addr = mas->index;
210 
211 	if (addr >= range->start && addr + size < range->end)
212 		return true;
213 
214 	if (range->fallback_start &&
215 	    addr >= range->fallback_start && addr + size < range->fallback_end)
216 		return true;
217 
218 	return false;
219 }
220 
221 static void *execmem_cache_alloc_locked(struct execmem_range *range, size_t size)
222 {
223 	struct maple_tree *free_areas = &execmem_cache.free_areas;
224 	struct maple_tree *busy_areas = &execmem_cache.busy_areas;
225 	MA_STATE(mas_free, free_areas, 0, ULONG_MAX);
226 	MA_STATE(mas_busy, busy_areas, 0, ULONG_MAX);
227 	unsigned long addr, last, area_size = 0;
228 	void *area, *ptr = NULL;
229 	int err;
230 
231 	mas_for_each(&mas_free, area, ULONG_MAX) {
232 		area_size = mas_range_len(&mas_free);
233 
234 		if (area_size >= size && within_range(range, &mas_free, size))
235 			break;
236 	}
237 
238 	if (area_size < size)
239 		return NULL;
240 
241 	addr = mas_free.index;
242 	last = mas_free.last;
243 
244 	/* insert allocated size to busy_areas at range [addr, addr + size) */
245 	mas_set_range(&mas_busy, addr, addr + size - 1);
246 	err = mas_store_gfp(&mas_busy, (void *)addr, GFP_KERNEL);
247 	if (err)
248 		return NULL;
249 
250 	mas_store_gfp(&mas_free, NULL, GFP_KERNEL);
251 	if (area_size > size) {
252 		void *ptr = (void *)(addr + size);
253 
254 		/*
255 		 * re-insert remaining free size to free_areas at range
256 		 * [addr + size, last]
257 		 */
258 		mas_set_range(&mas_free, addr + size, last);
259 		err = mas_store_gfp(&mas_free, ptr, GFP_KERNEL);
260 		if (err) {
261 			mas_store_gfp(&mas_busy, NULL, GFP_KERNEL);
262 			return NULL;
263 		}
264 	}
265 	ptr = (void *)addr;
266 
267 	return ptr;
268 }
269 
270 static void *__execmem_cache_alloc(struct execmem_range *range, size_t size)
271 {
272 	guard(mutex)(&execmem_cache.mutex);
273 
274 	return execmem_cache_alloc_locked(range, size);
275 }
276 
277 static void *execmem_cache_populate_alloc(struct execmem_range *range, size_t size)
278 {
279 	unsigned long vm_flags = VM_ALLOW_HUGE_VMAP;
280 	struct mutex *mutex = &execmem_cache.mutex;
281 	struct vm_struct *vm;
282 	size_t alloc_size;
283 	int err = -ENOMEM;
284 	void *p;
285 
286 	alloc_size = round_up(size, PMD_SIZE);
287 	p = execmem_vmalloc(range, alloc_size, PAGE_KERNEL, vm_flags);
288 	if (!p) {
289 		alloc_size = size;
290 		p = execmem_vmalloc(range, alloc_size, PAGE_KERNEL, vm_flags);
291 	}
292 
293 	if (!p)
294 		return NULL;
295 
296 	vm = find_vm_area(p);
297 	if (!vm)
298 		goto err_free_mem;
299 
300 	/* fill memory with instructions that will trap */
301 	execmem_fill_trapping_insns(p, alloc_size);
302 
303 	err = set_memory_rox((unsigned long)p, vm->nr_pages);
304 	if (err)
305 		goto err_free_mem;
306 
307 	/*
308 	 * New memory blocks must be allocated and added to the cache
309 	 * as an atomic operation, otherwise they may be consumed
310 	 * by a parallel call to the execmem_cache_alloc function.
311 	 */
312 	mutex_lock(mutex);
313 	err = execmem_cache_add_locked(p, alloc_size, GFP_KERNEL);
314 	if (err)
315 		goto err_reset_direct_map;
316 
317 	p = execmem_cache_alloc_locked(range, size);
318 
319 	mutex_unlock(mutex);
320 
321 	return p;
322 
323 err_reset_direct_map:
324 	mutex_unlock(mutex);
325 	execmem_set_direct_map_valid(vm, true);
326 err_free_mem:
327 	vfree(p);
328 	return NULL;
329 }
330 
331 static void *execmem_cache_alloc(struct execmem_range *range, size_t size)
332 {
333 	void *p;
334 
335 	p = __execmem_cache_alloc(range, size);
336 	if (p)
337 		return p;
338 
339 	return execmem_cache_populate_alloc(range, size);
340 }
341 
342 static inline bool is_pending_free(void *ptr)
343 {
344 	return ((unsigned long)ptr & PENDING_FREE_MASK);
345 }
346 
347 static inline void *pending_free_set(void *ptr)
348 {
349 	return (void *)((unsigned long)ptr | PENDING_FREE_MASK);
350 }
351 
352 static inline void *pending_free_clear(void *ptr)
353 {
354 	return (void *)((unsigned long)ptr & ~PENDING_FREE_MASK);
355 }
356 
357 static int __execmem_cache_free(struct ma_state *mas, void *ptr, gfp_t gfp_mask)
358 {
359 	size_t size = mas_range_len(mas);
360 	int err;
361 
362 	err = execmem_force_rw(ptr, size);
363 	if (err)
364 		return err;
365 
366 	execmem_fill_trapping_insns(ptr, size);
367 	execmem_restore_rox(ptr, size);
368 
369 	err = execmem_cache_add_locked(ptr, size, gfp_mask);
370 	if (err)
371 		return err;
372 
373 	mas_store_gfp(mas, NULL, gfp_mask);
374 	return 0;
375 }
376 
377 static void execmem_cache_free_slow(struct work_struct *work);
378 static DECLARE_DELAYED_WORK(execmem_cache_free_work, execmem_cache_free_slow);
379 
380 static void execmem_cache_free_slow(struct work_struct *work)
381 {
382 	struct maple_tree *busy_areas = &execmem_cache.busy_areas;
383 	MA_STATE(mas, busy_areas, 0, ULONG_MAX);
384 	void *area;
385 
386 	guard(mutex)(&execmem_cache.mutex);
387 
388 	if (!execmem_cache.pending_free_cnt)
389 		return;
390 
391 	mas_for_each(&mas, area, ULONG_MAX) {
392 		if (!is_pending_free(area))
393 			continue;
394 
395 		area = pending_free_clear(area);
396 		if (__execmem_cache_free(&mas, area, GFP_KERNEL))
397 			continue;
398 
399 		execmem_cache.pending_free_cnt--;
400 	}
401 
402 	if (execmem_cache.pending_free_cnt)
403 		schedule_delayed_work(&execmem_cache_free_work, FREE_DELAY);
404 	else
405 		schedule_work(&execmem_cache_clean_work);
406 }
407 
408 static bool execmem_cache_free(void *ptr)
409 {
410 	struct maple_tree *busy_areas = &execmem_cache.busy_areas;
411 	unsigned long addr = (unsigned long)ptr;
412 	MA_STATE(mas, busy_areas, addr, addr);
413 	void *area;
414 	int err;
415 
416 	guard(mutex)(&execmem_cache.mutex);
417 
418 	area = mas_walk(&mas);
419 	if (!area)
420 		return false;
421 
422 	err = __execmem_cache_free(&mas, area, GFP_KERNEL | __GFP_NORETRY);
423 	if (err) {
424 		/*
425 		 * mas points to exact slot we've got the area from, nothing
426 		 * else can modify the tree because of the mutex, so there
427 		 * won't be any allocations in mas_store_gfp() and it will just
428 		 * change the pointer.
429 		 */
430 		area = pending_free_set(area);
431 		mas_store_gfp(&mas, area, GFP_KERNEL);
432 		execmem_cache.pending_free_cnt++;
433 		schedule_delayed_work(&execmem_cache_free_work, FREE_DELAY);
434 		return true;
435 	}
436 
437 	schedule_work(&execmem_cache_clean_work);
438 
439 	return true;
440 }
441 
442 #else /* CONFIG_ARCH_HAS_EXECMEM_ROX */
443 /*
444  * when ROX cache is not used the permissions defined by architectures for
445  * execmem ranges that are updated before use (e.g. EXECMEM_MODULE_TEXT) must
446  * be writable anyway
447  */
448 static inline int execmem_force_rw(void *ptr, size_t size)
449 {
450 	return 0;
451 }
452 
453 static void *execmem_cache_alloc(struct execmem_range *range, size_t size)
454 {
455 	return NULL;
456 }
457 
458 static bool execmem_cache_free(void *ptr)
459 {
460 	return false;
461 }
462 #endif /* CONFIG_ARCH_HAS_EXECMEM_ROX */
463 
464 void *execmem_alloc(enum execmem_type type, size_t size)
465 {
466 	struct execmem_range *range = &execmem_info->ranges[type];
467 	bool use_cache = range->flags & EXECMEM_ROX_CACHE;
468 	unsigned long vm_flags = VM_FLUSH_RESET_PERMS;
469 	pgprot_t pgprot = range->pgprot;
470 	void *p = NULL;
471 
472 	size = PAGE_ALIGN(size);
473 
474 	if (use_cache)
475 		p = execmem_cache_alloc(range, size);
476 	else
477 		p = execmem_vmalloc(range, size, pgprot, vm_flags);
478 
479 	return kasan_reset_tag(p);
480 }
481 
482 void *execmem_alloc_rw(enum execmem_type type, size_t size)
483 {
484 	void *p __free(execmem) = execmem_alloc(type, size);
485 	int err;
486 
487 	if (!p)
488 		return NULL;
489 
490 	err = execmem_force_rw(p, size);
491 	if (err)
492 		return NULL;
493 
494 	return no_free_ptr(p);
495 }
496 
497 void execmem_free(void *ptr)
498 {
499 	/*
500 	 * This memory may be RO, and freeing RO memory in an interrupt is not
501 	 * supported by vmalloc.
502 	 */
503 	WARN_ON(in_interrupt());
504 
505 	if (!execmem_cache_free(ptr))
506 		vfree(ptr);
507 }
508 
509 bool execmem_is_rox(enum execmem_type type)
510 {
511 	return !!(execmem_info->ranges[type].flags & EXECMEM_ROX_CACHE);
512 }
513 
514 static bool execmem_validate(struct execmem_info *info)
515 {
516 	struct execmem_range *r = &info->ranges[EXECMEM_DEFAULT];
517 
518 	if (!r->alignment || !r->start || !r->end || !pgprot_val(r->pgprot)) {
519 		pr_crit("Invalid parameters for execmem allocator, module loading will fail");
520 		return false;
521 	}
522 
523 	if (!IS_ENABLED(CONFIG_ARCH_HAS_EXECMEM_ROX)) {
524 		for (int i = EXECMEM_DEFAULT; i < EXECMEM_TYPE_MAX; i++) {
525 			r = &info->ranges[i];
526 
527 			if (r->flags & EXECMEM_ROX_CACHE) {
528 				pr_warn_once("ROX cache is not supported\n");
529 				r->flags &= ~EXECMEM_ROX_CACHE;
530 			}
531 		}
532 	}
533 
534 	return true;
535 }
536 
537 static void execmem_init_missing(struct execmem_info *info)
538 {
539 	struct execmem_range *default_range = &info->ranges[EXECMEM_DEFAULT];
540 
541 	for (int i = EXECMEM_DEFAULT + 1; i < EXECMEM_TYPE_MAX; i++) {
542 		struct execmem_range *r = &info->ranges[i];
543 
544 		if (!r->start) {
545 			if (i == EXECMEM_MODULE_DATA)
546 				r->pgprot = PAGE_KERNEL;
547 			else
548 				r->pgprot = default_range->pgprot;
549 			r->alignment = default_range->alignment;
550 			r->start = default_range->start;
551 			r->end = default_range->end;
552 			r->flags = default_range->flags;
553 			r->fallback_start = default_range->fallback_start;
554 			r->fallback_end = default_range->fallback_end;
555 		}
556 	}
557 }
558 
559 struct execmem_info * __weak execmem_arch_setup(void)
560 {
561 	return NULL;
562 }
563 
564 static void __init __execmem_init(void)
565 {
566 	struct execmem_info *info = execmem_arch_setup();
567 
568 	if (!info) {
569 		info = execmem_info = &default_execmem_info;
570 		info->ranges[EXECMEM_DEFAULT].start = VMALLOC_START;
571 		info->ranges[EXECMEM_DEFAULT].end = VMALLOC_END;
572 		info->ranges[EXECMEM_DEFAULT].pgprot = PAGE_KERNEL_EXEC;
573 		info->ranges[EXECMEM_DEFAULT].alignment = 1;
574 	}
575 
576 	if (!execmem_validate(info))
577 		return;
578 
579 	execmem_init_missing(info);
580 
581 	execmem_info = info;
582 }
583 
584 #ifdef CONFIG_ARCH_WANTS_EXECMEM_LATE
585 static int __init execmem_late_init(void)
586 {
587 	__execmem_init();
588 	return 0;
589 }
590 core_initcall(execmem_late_init);
591 #else
592 void __init execmem_init(void)
593 {
594 	__execmem_init();
595 }
596 #endif
597