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
execmem_vmalloc(struct execmem_range * range,size_t size,pgprot_t pgprot,unsigned long vm_flags)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 if (vm_flags & VM_ALLOW_HUGE_VMAP)
42 align = PMD_SIZE;
43
44 p = __vmalloc_node_range(size, align, start, end, gfp_flags,
45 pgprot, vm_flags, NUMA_NO_NODE,
46 __builtin_return_address(0));
47 if (!p && range->fallback_start) {
48 start = range->fallback_start;
49 end = range->fallback_end;
50 p = __vmalloc_node_range(size, align, start, end, gfp_flags,
51 pgprot, vm_flags, NUMA_NO_NODE,
52 __builtin_return_address(0));
53 }
54
55 if (!p) {
56 pr_warn_ratelimited("unable to allocate memory\n");
57 return NULL;
58 }
59
60 if (kasan && (kasan_alloc_module_shadow(p, size, GFP_KERNEL) < 0)) {
61 vfree(p);
62 return NULL;
63 }
64
65 return p;
66 }
67
execmem_vmap(size_t size)68 struct vm_struct *execmem_vmap(size_t size)
69 {
70 struct execmem_range *range = &execmem_info->ranges[EXECMEM_MODULE_DATA];
71 struct vm_struct *area;
72
73 area = __get_vm_area_node(size, range->alignment, PAGE_SHIFT, VM_ALLOC,
74 range->start, range->end, NUMA_NO_NODE,
75 GFP_KERNEL, __builtin_return_address(0));
76 if (!area && range->fallback_start)
77 area = __get_vm_area_node(size, range->alignment, PAGE_SHIFT, VM_ALLOC,
78 range->fallback_start, range->fallback_end,
79 NUMA_NO_NODE, GFP_KERNEL, __builtin_return_address(0));
80
81 return area;
82 }
83 #else
execmem_vmalloc(struct execmem_range * range,size_t size,pgprot_t pgprot,unsigned long vm_flags)84 static void *execmem_vmalloc(struct execmem_range *range, size_t size,
85 pgprot_t pgprot, unsigned long vm_flags)
86 {
87 return vmalloc(size);
88 }
89 #endif /* CONFIG_MMU */
90
91 #ifdef CONFIG_ARCH_HAS_EXECMEM_ROX
92 struct execmem_cache {
93 struct mutex mutex;
94 struct maple_tree busy_areas;
95 struct maple_tree free_areas;
96 };
97
98 static struct execmem_cache execmem_cache = {
99 .mutex = __MUTEX_INITIALIZER(execmem_cache.mutex),
100 .busy_areas = MTREE_INIT_EXT(busy_areas, MT_FLAGS_LOCK_EXTERN,
101 execmem_cache.mutex),
102 .free_areas = MTREE_INIT_EXT(free_areas, MT_FLAGS_LOCK_EXTERN,
103 execmem_cache.mutex),
104 };
105
mas_range_len(struct ma_state * mas)106 static inline unsigned long mas_range_len(struct ma_state *mas)
107 {
108 return mas->last - mas->index + 1;
109 }
110
execmem_set_direct_map_valid(struct vm_struct * vm,bool valid)111 static int execmem_set_direct_map_valid(struct vm_struct *vm, bool valid)
112 {
113 unsigned int nr = (1 << get_vm_area_page_order(vm));
114 unsigned int updated = 0;
115 int err = 0;
116
117 for (int i = 0; i < vm->nr_pages; i += nr) {
118 err = set_direct_map_valid_noflush(vm->pages[i], nr, valid);
119 if (err)
120 goto err_restore;
121 updated += nr;
122 }
123
124 return 0;
125
126 err_restore:
127 for (int i = 0; i < updated; i += nr)
128 set_direct_map_valid_noflush(vm->pages[i], nr, !valid);
129
130 return err;
131 }
132
execmem_cache_clean(struct work_struct * work)133 static void execmem_cache_clean(struct work_struct *work)
134 {
135 struct maple_tree *free_areas = &execmem_cache.free_areas;
136 struct mutex *mutex = &execmem_cache.mutex;
137 MA_STATE(mas, free_areas, 0, ULONG_MAX);
138 void *area;
139
140 mutex_lock(mutex);
141 mas_for_each(&mas, area, ULONG_MAX) {
142 size_t size = mas_range_len(&mas);
143
144 if (IS_ALIGNED(size, PMD_SIZE) &&
145 IS_ALIGNED(mas.index, PMD_SIZE)) {
146 struct vm_struct *vm = find_vm_area(area);
147
148 execmem_set_direct_map_valid(vm, true);
149 mas_store_gfp(&mas, NULL, GFP_KERNEL);
150 vfree(area);
151 }
152 }
153 mutex_unlock(mutex);
154 }
155
156 static DECLARE_WORK(execmem_cache_clean_work, execmem_cache_clean);
157
execmem_cache_add(void * ptr,size_t size)158 static int execmem_cache_add(void *ptr, size_t size)
159 {
160 struct maple_tree *free_areas = &execmem_cache.free_areas;
161 struct mutex *mutex = &execmem_cache.mutex;
162 unsigned long addr = (unsigned long)ptr;
163 MA_STATE(mas, free_areas, addr - 1, addr + 1);
164 unsigned long lower, upper;
165 void *area = NULL;
166 int err;
167
168 lower = addr;
169 upper = addr + size - 1;
170
171 mutex_lock(mutex);
172 area = mas_walk(&mas);
173 if (area && mas.last == addr - 1)
174 lower = mas.index;
175
176 area = mas_next(&mas, ULONG_MAX);
177 if (area && mas.index == addr + size)
178 upper = mas.last;
179
180 mas_set_range(&mas, lower, upper);
181 err = mas_store_gfp(&mas, (void *)lower, GFP_KERNEL);
182 mutex_unlock(mutex);
183 if (err)
184 return err;
185
186 return 0;
187 }
188
within_range(struct execmem_range * range,struct ma_state * mas,size_t size)189 static bool within_range(struct execmem_range *range, struct ma_state *mas,
190 size_t size)
191 {
192 unsigned long addr = mas->index;
193
194 if (addr >= range->start && addr + size < range->end)
195 return true;
196
197 if (range->fallback_start &&
198 addr >= range->fallback_start && addr + size < range->fallback_end)
199 return true;
200
201 return false;
202 }
203
__execmem_cache_alloc(struct execmem_range * range,size_t size)204 static void *__execmem_cache_alloc(struct execmem_range *range, size_t size)
205 {
206 struct maple_tree *free_areas = &execmem_cache.free_areas;
207 struct maple_tree *busy_areas = &execmem_cache.busy_areas;
208 MA_STATE(mas_free, free_areas, 0, ULONG_MAX);
209 MA_STATE(mas_busy, busy_areas, 0, ULONG_MAX);
210 struct mutex *mutex = &execmem_cache.mutex;
211 unsigned long addr, last, area_size = 0;
212 void *area, *ptr = NULL;
213 int err;
214
215 mutex_lock(mutex);
216 mas_for_each(&mas_free, area, ULONG_MAX) {
217 area_size = mas_range_len(&mas_free);
218
219 if (area_size >= size && within_range(range, &mas_free, size))
220 break;
221 }
222
223 if (area_size < size)
224 goto out_unlock;
225
226 addr = mas_free.index;
227 last = mas_free.last;
228
229 /* insert allocated size to busy_areas at range [addr, addr + size) */
230 mas_set_range(&mas_busy, addr, addr + size - 1);
231 err = mas_store_gfp(&mas_busy, (void *)addr, GFP_KERNEL);
232 if (err)
233 goto out_unlock;
234
235 mas_store_gfp(&mas_free, NULL, GFP_KERNEL);
236 if (area_size > size) {
237 void *ptr = (void *)(addr + size);
238
239 /*
240 * re-insert remaining free size to free_areas at range
241 * [addr + size, last]
242 */
243 mas_set_range(&mas_free, addr + size, last);
244 err = mas_store_gfp(&mas_free, ptr, GFP_KERNEL);
245 if (err) {
246 mas_store_gfp(&mas_busy, NULL, GFP_KERNEL);
247 goto out_unlock;
248 }
249 }
250 ptr = (void *)addr;
251
252 out_unlock:
253 mutex_unlock(mutex);
254 return ptr;
255 }
256
257 static bool execmem_cache_rox = false;
258
execmem_cache_make_ro(void)259 void execmem_cache_make_ro(void)
260 {
261 struct maple_tree *free_areas = &execmem_cache.free_areas;
262 struct maple_tree *busy_areas = &execmem_cache.busy_areas;
263 MA_STATE(mas_free, free_areas, 0, ULONG_MAX);
264 MA_STATE(mas_busy, busy_areas, 0, ULONG_MAX);
265 struct mutex *mutex = &execmem_cache.mutex;
266 void *area;
267
268 execmem_cache_rox = true;
269
270 mutex_lock(mutex);
271
272 mas_for_each(&mas_free, area, ULONG_MAX) {
273 unsigned long pages = mas_range_len(&mas_free) >> PAGE_SHIFT;
274 set_memory_ro(mas_free.index, pages);
275 }
276
277 mas_for_each(&mas_busy, area, ULONG_MAX) {
278 unsigned long pages = mas_range_len(&mas_busy) >> PAGE_SHIFT;
279 set_memory_ro(mas_busy.index, pages);
280 }
281
282 mutex_unlock(mutex);
283 }
284
execmem_cache_populate(struct execmem_range * range,size_t size)285 static int execmem_cache_populate(struct execmem_range *range, size_t size)
286 {
287 unsigned long vm_flags = VM_ALLOW_HUGE_VMAP;
288 struct vm_struct *vm;
289 size_t alloc_size;
290 int err = -ENOMEM;
291 void *p;
292
293 alloc_size = round_up(size, PMD_SIZE);
294 p = execmem_vmalloc(range, alloc_size, PAGE_KERNEL, vm_flags);
295 if (!p)
296 return err;
297
298 vm = find_vm_area(p);
299 if (!vm)
300 goto err_free_mem;
301
302 /* fill memory with instructions that will trap */
303 execmem_fill_trapping_insns(p, alloc_size, /* writable = */ true);
304
305 if (execmem_cache_rox) {
306 err = set_memory_rox((unsigned long)p, vm->nr_pages);
307 if (err)
308 goto err_free_mem;
309 } else {
310 err = set_memory_x((unsigned long)p, vm->nr_pages);
311 if (err)
312 goto err_free_mem;
313 }
314
315 err = execmem_cache_add(p, alloc_size);
316 if (err)
317 goto err_reset_direct_map;
318
319 return 0;
320
321 err_reset_direct_map:
322 execmem_set_direct_map_valid(vm, true);
323 err_free_mem:
324 vfree(p);
325 return err;
326 }
327
execmem_cache_alloc(struct execmem_range * range,size_t size)328 static void *execmem_cache_alloc(struct execmem_range *range, size_t size)
329 {
330 void *p;
331 int err;
332
333 p = __execmem_cache_alloc(range, size);
334 if (p)
335 return p;
336
337 err = execmem_cache_populate(range, size);
338 if (err)
339 return NULL;
340
341 return __execmem_cache_alloc(range, size);
342 }
343
execmem_cache_free(void * ptr)344 static bool execmem_cache_free(void *ptr)
345 {
346 struct maple_tree *busy_areas = &execmem_cache.busy_areas;
347 struct mutex *mutex = &execmem_cache.mutex;
348 unsigned long addr = (unsigned long)ptr;
349 MA_STATE(mas, busy_areas, addr, addr);
350 size_t size;
351 void *area;
352
353 mutex_lock(mutex);
354 area = mas_walk(&mas);
355 if (!area) {
356 mutex_unlock(mutex);
357 return false;
358 }
359 size = mas_range_len(&mas);
360
361 mas_store_gfp(&mas, NULL, GFP_KERNEL);
362 mutex_unlock(mutex);
363
364 execmem_fill_trapping_insns(ptr, size, /* writable = */ false);
365
366 execmem_cache_add(ptr, size);
367
368 schedule_work(&execmem_cache_clean_work);
369
370 return true;
371 }
372
execmem_make_temp_rw(void * ptr,size_t size)373 int execmem_make_temp_rw(void *ptr, size_t size)
374 {
375 unsigned int nr = PAGE_ALIGN(size) >> PAGE_SHIFT;
376 unsigned long addr = (unsigned long)ptr;
377 int ret;
378
379 ret = set_memory_nx(addr, nr);
380 if (ret)
381 return ret;
382
383 return set_memory_rw(addr, nr);
384 }
385
execmem_restore_rox(void * ptr,size_t size)386 int execmem_restore_rox(void *ptr, size_t size)
387 {
388 unsigned int nr = PAGE_ALIGN(size) >> PAGE_SHIFT;
389 unsigned long addr = (unsigned long)ptr;
390
391 return set_memory_rox(addr, nr);
392 }
393
394 #else /* CONFIG_ARCH_HAS_EXECMEM_ROX */
execmem_cache_alloc(struct execmem_range * range,size_t size)395 static void *execmem_cache_alloc(struct execmem_range *range, size_t size)
396 {
397 return NULL;
398 }
399
execmem_cache_free(void * ptr)400 static bool execmem_cache_free(void *ptr)
401 {
402 return false;
403 }
404 #endif /* CONFIG_ARCH_HAS_EXECMEM_ROX */
405
execmem_alloc(enum execmem_type type,size_t size)406 void *execmem_alloc(enum execmem_type type, size_t size)
407 {
408 struct execmem_range *range = &execmem_info->ranges[type];
409 bool use_cache = range->flags & EXECMEM_ROX_CACHE;
410 unsigned long vm_flags = VM_FLUSH_RESET_PERMS;
411 pgprot_t pgprot = range->pgprot;
412 void *p;
413
414 if (use_cache)
415 p = execmem_cache_alloc(range, size);
416 else
417 p = execmem_vmalloc(range, size, pgprot, vm_flags);
418
419 return kasan_reset_tag(p);
420 }
421
execmem_free(void * ptr)422 void execmem_free(void *ptr)
423 {
424 /*
425 * This memory may be RO, and freeing RO memory in an interrupt is not
426 * supported by vmalloc.
427 */
428 WARN_ON(in_interrupt());
429
430 if (!execmem_cache_free(ptr))
431 vfree(ptr);
432 }
433
execmem_update_copy(void * dst,const void * src,size_t size)434 void *execmem_update_copy(void *dst, const void *src, size_t size)
435 {
436 return text_poke_copy(dst, src, size);
437 }
438
execmem_is_rox(enum execmem_type type)439 bool execmem_is_rox(enum execmem_type type)
440 {
441 return !!(execmem_info->ranges[type].flags & EXECMEM_ROX_CACHE);
442 }
443
execmem_validate(struct execmem_info * info)444 static bool execmem_validate(struct execmem_info *info)
445 {
446 struct execmem_range *r = &info->ranges[EXECMEM_DEFAULT];
447
448 if (!r->alignment || !r->start || !r->end || !pgprot_val(r->pgprot)) {
449 pr_crit("Invalid parameters for execmem allocator, module loading will fail");
450 return false;
451 }
452
453 if (!IS_ENABLED(CONFIG_ARCH_HAS_EXECMEM_ROX)) {
454 for (int i = EXECMEM_DEFAULT; i < EXECMEM_TYPE_MAX; i++) {
455 r = &info->ranges[i];
456
457 if (r->flags & EXECMEM_ROX_CACHE) {
458 pr_warn_once("ROX cache is not supported\n");
459 r->flags &= ~EXECMEM_ROX_CACHE;
460 }
461 }
462 }
463
464 return true;
465 }
466
execmem_init_missing(struct execmem_info * info)467 static void execmem_init_missing(struct execmem_info *info)
468 {
469 struct execmem_range *default_range = &info->ranges[EXECMEM_DEFAULT];
470
471 for (int i = EXECMEM_DEFAULT + 1; i < EXECMEM_TYPE_MAX; i++) {
472 struct execmem_range *r = &info->ranges[i];
473
474 if (!r->start) {
475 if (i == EXECMEM_MODULE_DATA)
476 r->pgprot = PAGE_KERNEL;
477 else
478 r->pgprot = default_range->pgprot;
479 r->alignment = default_range->alignment;
480 r->start = default_range->start;
481 r->end = default_range->end;
482 r->flags = default_range->flags;
483 r->fallback_start = default_range->fallback_start;
484 r->fallback_end = default_range->fallback_end;
485 }
486 }
487 }
488
execmem_arch_setup(void)489 struct execmem_info * __weak execmem_arch_setup(void)
490 {
491 return NULL;
492 }
493
__execmem_init(void)494 static void __init __execmem_init(void)
495 {
496 struct execmem_info *info = execmem_arch_setup();
497
498 if (!info) {
499 info = execmem_info = &default_execmem_info;
500 info->ranges[EXECMEM_DEFAULT].start = VMALLOC_START;
501 info->ranges[EXECMEM_DEFAULT].end = VMALLOC_END;
502 info->ranges[EXECMEM_DEFAULT].pgprot = PAGE_KERNEL_EXEC;
503 info->ranges[EXECMEM_DEFAULT].alignment = 1;
504 }
505
506 if (!execmem_validate(info))
507 return;
508
509 execmem_init_missing(info);
510
511 execmem_info = info;
512 }
513
514 #ifdef CONFIG_ARCH_WANTS_EXECMEM_LATE
execmem_late_init(void)515 static int __init execmem_late_init(void)
516 {
517 __execmem_init();
518 return 0;
519 }
520 core_initcall(execmem_late_init);
521 #else
execmem_init(void)522 void __init execmem_init(void)
523 {
524 __execmem_init();
525 }
526 #endif
527