1// SPDX-License-Identifier: GPL-2.0
2/*
3 * This file contains KASAN runtime code that manages shadow memory for
4 * generic and software tag-based KASAN modes.
5 *
6 * Copyright (c) 2014 Samsung Electronics Co., Ltd.
7 * Author: Andrey Ryabinin <ryabinin.a.a@gmail.com>
8 *
9 * Some code borrowed from https://github.com/xairy/kasan-prototype by
10 * Andrey Konovalov <andreyknvl@gmail.com>
11 */
12
13#include <linux/init.h>
14#include <linux/kasan.h>
15#include <linux/kernel.h>
16#include <linux/kfence.h>
17#include <linux/kmemleak.h>
18#include <linux/memory.h>
19#include <linux/mm.h>
20#include <linux/string.h>
21#include <linux/types.h>
22#include <linux/vmalloc.h>
23
24#include <asm/cacheflush.h>
25#include <asm/tlbflush.h>
26
27#include "kasan.h"
28
29bool __kasan_check_read(const volatile void *p, unsigned int size)
30{
31 return kasan_check_range((void *)p, size, false, _RET_IP_);
32}
33EXPORT_SYMBOL(__kasan_check_read);
34
35bool __kasan_check_write(const volatile void *p, unsigned int size)
36{
37 return kasan_check_range((void *)p, size, true, _RET_IP_);
38}
39EXPORT_SYMBOL(__kasan_check_write);
40
41#if !defined(CONFIG_CC_HAS_KASAN_MEMINTRINSIC_PREFIX) && !defined(CONFIG_GENERIC_ENTRY)
42/*
43 * CONFIG_GENERIC_ENTRY relies on compiler emitted mem*() calls to not be
44 * instrumented. KASAN enabled toolchains should emit __asan_mem*() functions
45 * for the sites they want to instrument.
46 *
47 * If we have a compiler that can instrument meminstrinsics, never override
48 * these, so that non-instrumented files can safely consider them as builtins.
49 */
50#undef memset
51void *memset(void *addr, int c, size_t len)
52{
53 if (!kasan_check_range(addr, len, true, _RET_IP_))
54 return NULL;
55
56 return __memset(addr, c, len);
57}
58
59#ifdef __HAVE_ARCH_MEMMOVE
60#undef memmove
61void *memmove(void *dest, const void *src, size_t len)
62{
63 if (!kasan_check_range(src, len, false, _RET_IP_) ||
64 !kasan_check_range(dest, len, true, _RET_IP_))
65 return NULL;
66
67 return __memmove(dest, src, len);
68}
69#endif
70
71#undef memcpy
72void *memcpy(void *dest, const void *src, size_t len)
73{
74 if (!kasan_check_range(src, len, false, _RET_IP_) ||
75 !kasan_check_range(dest, len, true, _RET_IP_))
76 return NULL;
77
78 return __memcpy(dest, src, len);
79}
80#endif
81
82void *__asan_memset(void *addr, int c, ssize_t len)
83{
84 if (!kasan_check_range(addr, len, true, _RET_IP_))
85 return NULL;
86
87 return __memset(s: addr, c, n: len);
88}
89EXPORT_SYMBOL(__asan_memset);
90
91#ifdef __HAVE_ARCH_MEMMOVE
92void *__asan_memmove(void *dest, const void *src, ssize_t len)
93{
94 if (!kasan_check_range(src, len, false, _RET_IP_) ||
95 !kasan_check_range(dest, len, true, _RET_IP_))
96 return NULL;
97
98 return __memmove(dest, src, count: len);
99}
100EXPORT_SYMBOL(__asan_memmove);
101#endif
102
103void *__asan_memcpy(void *dest, const void *src, ssize_t len)
104{
105 if (!kasan_check_range(src, len, false, _RET_IP_) ||
106 !kasan_check_range(dest, len, true, _RET_IP_))
107 return NULL;
108
109 return __memcpy(to: dest, from: src, len);
110}
111EXPORT_SYMBOL(__asan_memcpy);
112
113#ifdef CONFIG_KASAN_SW_TAGS
114void *__hwasan_memset(void *addr, int c, ssize_t len) __alias(__asan_memset);
115EXPORT_SYMBOL(__hwasan_memset);
116#ifdef __HAVE_ARCH_MEMMOVE
117void *__hwasan_memmove(void *dest, const void *src, ssize_t len) __alias(__asan_memmove);
118EXPORT_SYMBOL(__hwasan_memmove);
119#endif
120void *__hwasan_memcpy(void *dest, const void *src, ssize_t len) __alias(__asan_memcpy);
121EXPORT_SYMBOL(__hwasan_memcpy);
122#endif
123
124void kasan_poison(const void *addr, size_t size, u8 value, bool init)
125{
126 void *shadow_start, *shadow_end;
127
128 if (!kasan_arch_is_ready())
129 return;
130
131 /*
132 * Perform shadow offset calculation based on untagged address, as
133 * some of the callers (e.g. kasan_poison_new_object) pass tagged
134 * addresses to this function.
135 */
136 addr = kasan_reset_tag(addr);
137
138 if (WARN_ON((unsigned long)addr & KASAN_GRANULE_MASK))
139 return;
140 if (WARN_ON(size & KASAN_GRANULE_MASK))
141 return;
142
143 shadow_start = kasan_mem_to_shadow(addr);
144 shadow_end = kasan_mem_to_shadow(addr + size);
145
146 __memset(s: shadow_start, c: value, n: shadow_end - shadow_start);
147}
148EXPORT_SYMBOL_GPL(kasan_poison);
149
150#ifdef CONFIG_KASAN_GENERIC
151void kasan_poison_last_granule(const void *addr, size_t size)
152{
153 if (!kasan_arch_is_ready())
154 return;
155
156 if (size & KASAN_GRANULE_MASK) {
157 u8 *shadow = (u8 *)kasan_mem_to_shadow(addr + size);
158 *shadow = size & KASAN_GRANULE_MASK;
159 }
160}
161#endif
162
163void kasan_unpoison(const void *addr, size_t size, bool init)
164{
165 u8 tag = get_tag(addr);
166
167 /*
168 * Perform shadow offset calculation based on untagged address, as
169 * some of the callers (e.g. kasan_unpoison_new_object) pass tagged
170 * addresses to this function.
171 */
172 addr = kasan_reset_tag(addr);
173
174 if (WARN_ON((unsigned long)addr & KASAN_GRANULE_MASK))
175 return;
176
177 /* Unpoison all granules that cover the object. */
178 kasan_poison(addr, round_up(size, KASAN_GRANULE_SIZE), tag, false);
179
180 /* Partially poison the last granule for the generic mode. */
181 if (IS_ENABLED(CONFIG_KASAN_GENERIC))
182 kasan_poison_last_granule(address: addr, size);
183}
184
185#ifdef CONFIG_MEMORY_HOTPLUG
186static bool shadow_mapped(unsigned long addr)
187{
188 pgd_t *pgd = pgd_offset_k(addr);
189 p4d_t *p4d;
190 pud_t *pud;
191 pmd_t *pmd;
192 pte_t *pte;
193
194 if (pgd_none(pgd: *pgd))
195 return false;
196 p4d = p4d_offset(pgd, address: addr);
197 if (p4d_none(p4d: *p4d))
198 return false;
199 pud = pud_offset(p4d, address: addr);
200 if (pud_none(pud: *pud))
201 return false;
202 if (pud_leaf(pud: *pud))
203 return true;
204 pmd = pmd_offset(pud, address: addr);
205 if (pmd_none(pmd: *pmd))
206 return false;
207 if (pmd_leaf(pte: *pmd))
208 return true;
209 pte = pte_offset_kernel(pmd, address: addr);
210 return !pte_none(pte: ptep_get(ptep: pte));
211}
212
213static int __meminit kasan_mem_notifier(struct notifier_block *nb,
214 unsigned long action, void *data)
215{
216 struct memory_notify *mem_data = data;
217 unsigned long nr_shadow_pages, start_kaddr, shadow_start;
218 unsigned long shadow_end, shadow_size;
219
220 nr_shadow_pages = mem_data->nr_pages >> KASAN_SHADOW_SCALE_SHIFT;
221 start_kaddr = (unsigned long)pfn_to_kaddr(pfn: mem_data->start_pfn);
222 shadow_start = (unsigned long)kasan_mem_to_shadow((void *)start_kaddr);
223 shadow_size = nr_shadow_pages << PAGE_SHIFT;
224 shadow_end = shadow_start + shadow_size;
225
226 if (WARN_ON(mem_data->nr_pages % KASAN_GRANULE_SIZE) ||
227 WARN_ON(start_kaddr % KASAN_MEMORY_PER_SHADOW_PAGE))
228 return NOTIFY_BAD;
229
230 switch (action) {
231 case MEM_GOING_ONLINE: {
232 void *ret;
233
234 /*
235 * If shadow is mapped already than it must have been mapped
236 * during the boot. This could happen if we onlining previously
237 * offlined memory.
238 */
239 if (shadow_mapped(addr: shadow_start))
240 return NOTIFY_OK;
241
242 ret = __vmalloc_node_range(size: shadow_size, PAGE_SIZE, start: shadow_start,
243 end: shadow_end, GFP_KERNEL,
244 PAGE_KERNEL, VM_NO_GUARD,
245 pfn_to_nid(mem_data->start_pfn),
246 caller: __builtin_return_address(0));
247 if (!ret)
248 return NOTIFY_BAD;
249
250 kmemleak_ignore(ptr: ret);
251 return NOTIFY_OK;
252 }
253 case MEM_CANCEL_ONLINE:
254 case MEM_OFFLINE: {
255 struct vm_struct *vm;
256
257 /*
258 * shadow_start was either mapped during boot by kasan_init()
259 * or during memory online by __vmalloc_node_range().
260 * In the latter case we can use vfree() to free shadow.
261 * Non-NULL result of the find_vm_area() will tell us if
262 * that was the second case.
263 *
264 * Currently it's not possible to free shadow mapped
265 * during boot by kasan_init(). It's because the code
266 * to do that hasn't been written yet. So we'll just
267 * leak the memory.
268 */
269 vm = find_vm_area(addr: (void *)shadow_start);
270 if (vm)
271 vfree(addr: (void *)shadow_start);
272 }
273 }
274
275 return NOTIFY_OK;
276}
277
278static int __init kasan_memhotplug_init(void)
279{
280 hotplug_memory_notifier(kasan_mem_notifier, DEFAULT_CALLBACK_PRI);
281
282 return 0;
283}
284
285core_initcall(kasan_memhotplug_init);
286#endif
287
288#ifdef CONFIG_KASAN_VMALLOC
289
290void __init __weak kasan_populate_early_vm_area_shadow(void *start,
291 unsigned long size)
292{
293}
294
295static int kasan_populate_vmalloc_pte(pte_t *ptep, unsigned long addr,
296 void *unused)
297{
298 unsigned long page;
299 pte_t pte;
300
301 if (likely(!pte_none(ptep_get(ptep))))
302 return 0;
303
304 page = __get_free_page(GFP_KERNEL);
305 if (!page)
306 return -ENOMEM;
307
308 __memset((void *)page, KASAN_VMALLOC_INVALID, PAGE_SIZE);
309 pte = pfn_pte(PFN_DOWN(__pa(page)), PAGE_KERNEL);
310
311 spin_lock(&init_mm.page_table_lock);
312 if (likely(pte_none(ptep_get(ptep)))) {
313 set_pte_at(&init_mm, addr, ptep, pte);
314 page = 0;
315 }
316 spin_unlock(&init_mm.page_table_lock);
317 if (page)
318 free_page(page);
319 return 0;
320}
321
322int kasan_populate_vmalloc(unsigned long addr, unsigned long size)
323{
324 unsigned long shadow_start, shadow_end;
325 int ret;
326
327 if (!kasan_arch_is_ready())
328 return 0;
329
330 if (!is_vmalloc_or_module_addr((void *)addr))
331 return 0;
332
333 shadow_start = (unsigned long)kasan_mem_to_shadow((void *)addr);
334 shadow_end = (unsigned long)kasan_mem_to_shadow((void *)addr + size);
335
336 /*
337 * User Mode Linux maps enough shadow memory for all of virtual memory
338 * at boot, so doesn't need to allocate more on vmalloc, just clear it.
339 *
340 * The remaining CONFIG_UML checks in this file exist for the same
341 * reason.
342 */
343 if (IS_ENABLED(CONFIG_UML)) {
344 __memset((void *)shadow_start, KASAN_VMALLOC_INVALID, shadow_end - shadow_start);
345 return 0;
346 }
347
348 shadow_start = PAGE_ALIGN_DOWN(shadow_start);
349 shadow_end = PAGE_ALIGN(shadow_end);
350
351 ret = apply_to_page_range(&init_mm, shadow_start,
352 shadow_end - shadow_start,
353 kasan_populate_vmalloc_pte, NULL);
354 if (ret)
355 return ret;
356
357 flush_cache_vmap(shadow_start, shadow_end);
358
359 /*
360 * We need to be careful about inter-cpu effects here. Consider:
361 *
362 * CPU#0 CPU#1
363 * WRITE_ONCE(p, vmalloc(100)); while (x = READ_ONCE(p)) ;
364 * p[99] = 1;
365 *
366 * With compiler instrumentation, that ends up looking like this:
367 *
368 * CPU#0 CPU#1
369 * // vmalloc() allocates memory
370 * // let a = area->addr
371 * // we reach kasan_populate_vmalloc
372 * // and call kasan_unpoison:
373 * STORE shadow(a), unpoison_val
374 * ...
375 * STORE shadow(a+99), unpoison_val x = LOAD p
376 * // rest of vmalloc process <data dependency>
377 * STORE p, a LOAD shadow(x+99)
378 *
379 * If there is no barrier between the end of unpoisoning the shadow
380 * and the store of the result to p, the stores could be committed
381 * in a different order by CPU#0, and CPU#1 could erroneously observe
382 * poison in the shadow.
383 *
384 * We need some sort of barrier between the stores.
385 *
386 * In the vmalloc() case, this is provided by a smp_wmb() in
387 * clear_vm_uninitialized_flag(). In the per-cpu allocator and in
388 * get_vm_area() and friends, the caller gets shadow allocated but
389 * doesn't have any pages mapped into the virtual address space that
390 * has been reserved. Mapping those pages in will involve taking and
391 * releasing a page-table lock, which will provide the barrier.
392 */
393
394 return 0;
395}
396
397static int kasan_depopulate_vmalloc_pte(pte_t *ptep, unsigned long addr,
398 void *unused)
399{
400 unsigned long page;
401
402 page = (unsigned long)__va(pte_pfn(ptep_get(ptep)) << PAGE_SHIFT);
403
404 spin_lock(&init_mm.page_table_lock);
405
406 if (likely(!pte_none(ptep_get(ptep)))) {
407 pte_clear(&init_mm, addr, ptep);
408 free_page(page);
409 }
410 spin_unlock(&init_mm.page_table_lock);
411
412 return 0;
413}
414
415/*
416 * Release the backing for the vmalloc region [start, end), which
417 * lies within the free region [free_region_start, free_region_end).
418 *
419 * This can be run lazily, long after the region was freed. It runs
420 * under vmap_area_lock, so it's not safe to interact with the vmalloc/vmap
421 * infrastructure.
422 *
423 * How does this work?
424 * -------------------
425 *
426 * We have a region that is page aligned, labeled as A.
427 * That might not map onto the shadow in a way that is page-aligned:
428 *
429 * start end
430 * v v
431 * |????????|????????|AAAAAAAA|AA....AA|AAAAAAAA|????????| < vmalloc
432 * -------- -------- -------- -------- --------
433 * | | | | |
434 * | | | /-------/ |
435 * \-------\|/------/ |/---------------/
436 * ||| ||
437 * |??AAAAAA|AAAAAAAA|AA??????| < shadow
438 * (1) (2) (3)
439 *
440 * First we align the start upwards and the end downwards, so that the
441 * shadow of the region aligns with shadow page boundaries. In the
442 * example, this gives us the shadow page (2). This is the shadow entirely
443 * covered by this allocation.
444 *
445 * Then we have the tricky bits. We want to know if we can free the
446 * partially covered shadow pages - (1) and (3) in the example. For this,
447 * we are given the start and end of the free region that contains this
448 * allocation. Extending our previous example, we could have:
449 *
450 * free_region_start free_region_end
451 * | start end |
452 * v v v v
453 * |FFFFFFFF|FFFFFFFF|AAAAAAAA|AA....AA|AAAAAAAA|FFFFFFFF| < vmalloc
454 * -------- -------- -------- -------- --------
455 * | | | | |
456 * | | | /-------/ |
457 * \-------\|/------/ |/---------------/
458 * ||| ||
459 * |FFAAAAAA|AAAAAAAA|AAF?????| < shadow
460 * (1) (2) (3)
461 *
462 * Once again, we align the start of the free region up, and the end of
463 * the free region down so that the shadow is page aligned. So we can free
464 * page (1) - we know no allocation currently uses anything in that page,
465 * because all of it is in the vmalloc free region. But we cannot free
466 * page (3), because we can't be sure that the rest of it is unused.
467 *
468 * We only consider pages that contain part of the original region for
469 * freeing: we don't try to free other pages from the free region or we'd
470 * end up trying to free huge chunks of virtual address space.
471 *
472 * Concurrency
473 * -----------
474 *
475 * How do we know that we're not freeing a page that is simultaneously
476 * being used for a fresh allocation in kasan_populate_vmalloc(_pte)?
477 *
478 * We _can_ have kasan_release_vmalloc and kasan_populate_vmalloc running
479 * at the same time. While we run under free_vmap_area_lock, the population
480 * code does not.
481 *
482 * free_vmap_area_lock instead operates to ensure that the larger range
483 * [free_region_start, free_region_end) is safe: because __alloc_vmap_area and
484 * the per-cpu region-finding algorithm both run under free_vmap_area_lock,
485 * no space identified as free will become used while we are running. This
486 * means that so long as we are careful with alignment and only free shadow
487 * pages entirely covered by the free region, we will not run in to any
488 * trouble - any simultaneous allocations will be for disjoint regions.
489 */
490void kasan_release_vmalloc(unsigned long start, unsigned long end,
491 unsigned long free_region_start,
492 unsigned long free_region_end)
493{
494 void *shadow_start, *shadow_end;
495 unsigned long region_start, region_end;
496 unsigned long size;
497
498 if (!kasan_arch_is_ready())
499 return;
500
501 region_start = ALIGN(start, KASAN_MEMORY_PER_SHADOW_PAGE);
502 region_end = ALIGN_DOWN(end, KASAN_MEMORY_PER_SHADOW_PAGE);
503
504 free_region_start = ALIGN(free_region_start, KASAN_MEMORY_PER_SHADOW_PAGE);
505
506 if (start != region_start &&
507 free_region_start < region_start)
508 region_start -= KASAN_MEMORY_PER_SHADOW_PAGE;
509
510 free_region_end = ALIGN_DOWN(free_region_end, KASAN_MEMORY_PER_SHADOW_PAGE);
511
512 if (end != region_end &&
513 free_region_end > region_end)
514 region_end += KASAN_MEMORY_PER_SHADOW_PAGE;
515
516 shadow_start = kasan_mem_to_shadow((void *)region_start);
517 shadow_end = kasan_mem_to_shadow((void *)region_end);
518
519 if (shadow_end > shadow_start) {
520 size = shadow_end - shadow_start;
521 if (IS_ENABLED(CONFIG_UML)) {
522 __memset(shadow_start, KASAN_SHADOW_INIT, shadow_end - shadow_start);
523 return;
524 }
525 apply_to_existing_page_range(&init_mm,
526 (unsigned long)shadow_start,
527 size, kasan_depopulate_vmalloc_pte,
528 NULL);
529 flush_tlb_kernel_range((unsigned long)shadow_start,
530 (unsigned long)shadow_end);
531 }
532}
533
534void *__kasan_unpoison_vmalloc(const void *start, unsigned long size,
535 kasan_vmalloc_flags_t flags)
536{
537 /*
538 * Software KASAN modes unpoison both VM_ALLOC and non-VM_ALLOC
539 * mappings, so the KASAN_VMALLOC_VM_ALLOC flag is ignored.
540 * Software KASAN modes can't optimize zeroing memory by combining it
541 * with setting memory tags, so the KASAN_VMALLOC_INIT flag is ignored.
542 */
543
544 if (!kasan_arch_is_ready())
545 return (void *)start;
546
547 if (!is_vmalloc_or_module_addr(start))
548 return (void *)start;
549
550 /*
551 * Don't tag executable memory with the tag-based mode.
552 * The kernel doesn't tolerate having the PC register tagged.
553 */
554 if (IS_ENABLED(CONFIG_KASAN_SW_TAGS) &&
555 !(flags & KASAN_VMALLOC_PROT_NORMAL))
556 return (void *)start;
557
558 start = set_tag(start, kasan_random_tag());
559 kasan_unpoison(start, size, false);
560 return (void *)start;
561}
562
563/*
564 * Poison the shadow for a vmalloc region. Called as part of the
565 * freeing process at the time the region is freed.
566 */
567void __kasan_poison_vmalloc(const void *start, unsigned long size)
568{
569 if (!kasan_arch_is_ready())
570 return;
571
572 if (!is_vmalloc_or_module_addr(start))
573 return;
574
575 size = round_up(size, KASAN_GRANULE_SIZE);
576 kasan_poison(start, size, KASAN_VMALLOC_INVALID, false);
577}
578
579#else /* CONFIG_KASAN_VMALLOC */
580
581int kasan_alloc_module_shadow(void *addr, size_t size, gfp_t gfp_mask)
582{
583 void *ret;
584 size_t scaled_size;
585 size_t shadow_size;
586 unsigned long shadow_start;
587
588 shadow_start = (unsigned long)kasan_mem_to_shadow(addr);
589 scaled_size = (size + KASAN_GRANULE_SIZE - 1) >>
590 KASAN_SHADOW_SCALE_SHIFT;
591 shadow_size = round_up(scaled_size, PAGE_SIZE);
592
593 if (WARN_ON(!PAGE_ALIGNED(shadow_start)))
594 return -EINVAL;
595
596 if (IS_ENABLED(CONFIG_UML)) {
597 __memset(s: (void *)shadow_start, c: KASAN_SHADOW_INIT, n: shadow_size);
598 return 0;
599 }
600
601 ret = __vmalloc_node_range(size: shadow_size, align: 1, start: shadow_start,
602 end: shadow_start + shadow_size,
603 GFP_KERNEL,
604 PAGE_KERNEL, VM_NO_GUARD, NUMA_NO_NODE,
605 caller: __builtin_return_address(0));
606
607 if (ret) {
608 struct vm_struct *vm = find_vm_area(addr);
609 __memset(s: ret, c: KASAN_SHADOW_INIT, n: shadow_size);
610 vm->flags |= VM_KASAN;
611 kmemleak_ignore(ptr: ret);
612
613 if (vm->flags & VM_DEFER_KMEMLEAK)
614 kmemleak_vmalloc(area: vm, size, gfp: gfp_mask);
615
616 return 0;
617 }
618
619 return -ENOMEM;
620}
621
622void kasan_free_module_shadow(const struct vm_struct *vm)
623{
624 if (IS_ENABLED(CONFIG_UML))
625 return;
626
627 if (vm->flags & VM_KASAN)
628 vfree(addr: kasan_mem_to_shadow(vm->addr));
629}
630
631#endif
632

source code of linux/mm/kasan/shadow.c