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 29 bool __kasan_check_read(const volatile void *p, unsigned int size) 30 { 31 return kasan_check_range((unsigned long)p, size, false, _RET_IP_); 32 } 33 EXPORT_SYMBOL(__kasan_check_read); 34 35 bool __kasan_check_write(const volatile void *p, unsigned int size) 36 { 37 return kasan_check_range((unsigned long)p, size, true, _RET_IP_); 38 } 39 EXPORT_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 51 void *memset(void *addr, int c, size_t len) 52 { 53 if (!kasan_check_range((unsigned long)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 61 void *memmove(void *dest, const void *src, size_t len) 62 { 63 if (!kasan_check_range((unsigned long)src, len, false, _RET_IP_) || 64 !kasan_check_range((unsigned long)dest, len, true, _RET_IP_)) 65 return NULL; 66 67 return __memmove(dest, src, len); 68 } 69 #endif 70 71 #undef memcpy 72 void *memcpy(void *dest, const void *src, size_t len) 73 { 74 if (!kasan_check_range((unsigned long)src, len, false, _RET_IP_) || 75 !kasan_check_range((unsigned long)dest, len, true, _RET_IP_)) 76 return NULL; 77 78 return __memcpy(dest, src, len); 79 } 80 #endif 81 82 void *__asan_memset(void *addr, int c, size_t len) 83 { 84 if (!kasan_check_range((unsigned long)addr, len, true, _RET_IP_)) 85 return NULL; 86 87 return __memset(addr, c, len); 88 } 89 EXPORT_SYMBOL(__asan_memset); 90 91 #ifdef __HAVE_ARCH_MEMMOVE 92 void *__asan_memmove(void *dest, const void *src, size_t len) 93 { 94 if (!kasan_check_range((unsigned long)src, len, false, _RET_IP_) || 95 !kasan_check_range((unsigned long)dest, len, true, _RET_IP_)) 96 return NULL; 97 98 return __memmove(dest, src, len); 99 } 100 EXPORT_SYMBOL(__asan_memmove); 101 #endif 102 103 void *__asan_memcpy(void *dest, const void *src, size_t len) 104 { 105 if (!kasan_check_range((unsigned long)src, len, false, _RET_IP_) || 106 !kasan_check_range((unsigned long)dest, len, true, _RET_IP_)) 107 return NULL; 108 109 return __memcpy(dest, src, len); 110 } 111 EXPORT_SYMBOL(__asan_memcpy); 112 113 #ifdef CONFIG_KASAN_SW_TAGS 114 void *__hwasan_memset(void *addr, int c, size_t len) __alias(__asan_memset); 115 EXPORT_SYMBOL(__hwasan_memset); 116 #ifdef __HAVE_ARCH_MEMMOVE 117 void *__hwasan_memmove(void *dest, const void *src, size_t len) __alias(__asan_memmove); 118 EXPORT_SYMBOL(__hwasan_memmove); 119 #endif 120 void *__hwasan_memcpy(void *dest, const void *src, size_t len) __alias(__asan_memcpy); 121 EXPORT_SYMBOL(__hwasan_memcpy); 122 #endif 123 124 void 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_object_data) pass tagged 134 * addresses to this function. 135 */ 136 addr = kasan_reset_tag(addr); 137 138 /* Skip KFENCE memory if called explicitly outside of sl*b. */ 139 if (is_kfence_address(addr)) 140 return; 141 142 if (WARN_ON((unsigned long)addr & KASAN_GRANULE_MASK)) 143 return; 144 if (WARN_ON(size & KASAN_GRANULE_MASK)) 145 return; 146 147 shadow_start = kasan_mem_to_shadow(addr); 148 shadow_end = kasan_mem_to_shadow(addr + size); 149 150 __memset(shadow_start, value, shadow_end - shadow_start); 151 } 152 EXPORT_SYMBOL(kasan_poison); 153 154 #ifdef CONFIG_KASAN_GENERIC 155 void kasan_poison_last_granule(const void *addr, size_t size) 156 { 157 if (!kasan_arch_is_ready()) 158 return; 159 160 if (size & KASAN_GRANULE_MASK) { 161 u8 *shadow = (u8 *)kasan_mem_to_shadow(addr + size); 162 *shadow = size & KASAN_GRANULE_MASK; 163 } 164 } 165 #endif 166 167 void kasan_unpoison(const void *addr, size_t size, bool init) 168 { 169 u8 tag = get_tag(addr); 170 171 /* 172 * Perform shadow offset calculation based on untagged address, as 173 * some of the callers (e.g. kasan_unpoison_object_data) pass tagged 174 * addresses to this function. 175 */ 176 addr = kasan_reset_tag(addr); 177 178 /* 179 * Skip KFENCE memory if called explicitly outside of sl*b. Also note 180 * that calls to ksize(), where size is not a multiple of machine-word 181 * size, would otherwise poison the invalid portion of the word. 182 */ 183 if (is_kfence_address(addr)) 184 return; 185 186 if (WARN_ON((unsigned long)addr & KASAN_GRANULE_MASK)) 187 return; 188 189 /* Unpoison all granules that cover the object. */ 190 kasan_poison(addr, round_up(size, KASAN_GRANULE_SIZE), tag, false); 191 192 /* Partially poison the last granule for the generic mode. */ 193 if (IS_ENABLED(CONFIG_KASAN_GENERIC)) 194 kasan_poison_last_granule(addr, size); 195 } 196 197 #ifdef CONFIG_MEMORY_HOTPLUG 198 static bool shadow_mapped(unsigned long addr) 199 { 200 pgd_t *pgd = pgd_offset_k(addr); 201 p4d_t *p4d; 202 pud_t *pud; 203 pmd_t *pmd; 204 pte_t *pte; 205 206 if (pgd_none(*pgd)) 207 return false; 208 p4d = p4d_offset(pgd, addr); 209 if (p4d_none(*p4d)) 210 return false; 211 pud = pud_offset(p4d, addr); 212 if (pud_none(*pud)) 213 return false; 214 215 /* 216 * We can't use pud_large() or pud_huge(), the first one is 217 * arch-specific, the last one depends on HUGETLB_PAGE. So let's abuse 218 * pud_bad(), if pud is bad then it's bad because it's huge. 219 */ 220 if (pud_bad(*pud)) 221 return true; 222 pmd = pmd_offset(pud, addr); 223 if (pmd_none(*pmd)) 224 return false; 225 226 if (pmd_bad(*pmd)) 227 return true; 228 pte = pte_offset_kernel(pmd, addr); 229 return !pte_none(*pte); 230 } 231 232 static int __meminit kasan_mem_notifier(struct notifier_block *nb, 233 unsigned long action, void *data) 234 { 235 struct memory_notify *mem_data = data; 236 unsigned long nr_shadow_pages, start_kaddr, shadow_start; 237 unsigned long shadow_end, shadow_size; 238 239 nr_shadow_pages = mem_data->nr_pages >> KASAN_SHADOW_SCALE_SHIFT; 240 start_kaddr = (unsigned long)pfn_to_kaddr(mem_data->start_pfn); 241 shadow_start = (unsigned long)kasan_mem_to_shadow((void *)start_kaddr); 242 shadow_size = nr_shadow_pages << PAGE_SHIFT; 243 shadow_end = shadow_start + shadow_size; 244 245 if (WARN_ON(mem_data->nr_pages % KASAN_GRANULE_SIZE) || 246 WARN_ON(start_kaddr % KASAN_MEMORY_PER_SHADOW_PAGE)) 247 return NOTIFY_BAD; 248 249 switch (action) { 250 case MEM_GOING_ONLINE: { 251 void *ret; 252 253 /* 254 * If shadow is mapped already than it must have been mapped 255 * during the boot. This could happen if we onlining previously 256 * offlined memory. 257 */ 258 if (shadow_mapped(shadow_start)) 259 return NOTIFY_OK; 260 261 ret = __vmalloc_node_range(shadow_size, PAGE_SIZE, shadow_start, 262 shadow_end, GFP_KERNEL, 263 PAGE_KERNEL, VM_NO_GUARD, 264 pfn_to_nid(mem_data->start_pfn), 265 __builtin_return_address(0)); 266 if (!ret) 267 return NOTIFY_BAD; 268 269 kmemleak_ignore(ret); 270 return NOTIFY_OK; 271 } 272 case MEM_CANCEL_ONLINE: 273 case MEM_OFFLINE: { 274 struct vm_struct *vm; 275 276 /* 277 * shadow_start was either mapped during boot by kasan_init() 278 * or during memory online by __vmalloc_node_range(). 279 * In the latter case we can use vfree() to free shadow. 280 * Non-NULL result of the find_vm_area() will tell us if 281 * that was the second case. 282 * 283 * Currently it's not possible to free shadow mapped 284 * during boot by kasan_init(). It's because the code 285 * to do that hasn't been written yet. So we'll just 286 * leak the memory. 287 */ 288 vm = find_vm_area((void *)shadow_start); 289 if (vm) 290 vfree((void *)shadow_start); 291 } 292 } 293 294 return NOTIFY_OK; 295 } 296 297 static int __init kasan_memhotplug_init(void) 298 { 299 hotplug_memory_notifier(kasan_mem_notifier, DEFAULT_CALLBACK_PRI); 300 301 return 0; 302 } 303 304 core_initcall(kasan_memhotplug_init); 305 #endif 306 307 #ifdef CONFIG_KASAN_VMALLOC 308 309 void __init __weak kasan_populate_early_vm_area_shadow(void *start, 310 unsigned long size) 311 { 312 } 313 314 static int kasan_populate_vmalloc_pte(pte_t *ptep, unsigned long addr, 315 void *unused) 316 { 317 unsigned long page; 318 pte_t pte; 319 320 if (likely(!pte_none(*ptep))) 321 return 0; 322 323 page = __get_free_page(GFP_KERNEL); 324 if (!page) 325 return -ENOMEM; 326 327 memset((void *)page, KASAN_VMALLOC_INVALID, PAGE_SIZE); 328 pte = pfn_pte(PFN_DOWN(__pa(page)), PAGE_KERNEL); 329 330 spin_lock(&init_mm.page_table_lock); 331 if (likely(pte_none(*ptep))) { 332 set_pte_at(&init_mm, addr, ptep, pte); 333 page = 0; 334 } 335 spin_unlock(&init_mm.page_table_lock); 336 if (page) 337 free_page(page); 338 return 0; 339 } 340 341 int kasan_populate_vmalloc(unsigned long addr, unsigned long size) 342 { 343 unsigned long shadow_start, shadow_end; 344 int ret; 345 346 if (!kasan_arch_is_ready()) 347 return 0; 348 349 if (!is_vmalloc_or_module_addr((void *)addr)) 350 return 0; 351 352 shadow_start = (unsigned long)kasan_mem_to_shadow((void *)addr); 353 shadow_end = (unsigned long)kasan_mem_to_shadow((void *)addr + size); 354 355 /* 356 * User Mode Linux maps enough shadow memory for all of virtual memory 357 * at boot, so doesn't need to allocate more on vmalloc, just clear it. 358 * 359 * The remaining CONFIG_UML checks in this file exist for the same 360 * reason. 361 */ 362 if (IS_ENABLED(CONFIG_UML)) { 363 __memset((void *)shadow_start, KASAN_VMALLOC_INVALID, shadow_end - shadow_start); 364 return 0; 365 } 366 367 shadow_start = PAGE_ALIGN_DOWN(shadow_start); 368 shadow_end = PAGE_ALIGN(shadow_end); 369 370 ret = apply_to_page_range(&init_mm, shadow_start, 371 shadow_end - shadow_start, 372 kasan_populate_vmalloc_pte, NULL); 373 if (ret) 374 return ret; 375 376 flush_cache_vmap(shadow_start, shadow_end); 377 378 /* 379 * We need to be careful about inter-cpu effects here. Consider: 380 * 381 * CPU#0 CPU#1 382 * WRITE_ONCE(p, vmalloc(100)); while (x = READ_ONCE(p)) ; 383 * p[99] = 1; 384 * 385 * With compiler instrumentation, that ends up looking like this: 386 * 387 * CPU#0 CPU#1 388 * // vmalloc() allocates memory 389 * // let a = area->addr 390 * // we reach kasan_populate_vmalloc 391 * // and call kasan_unpoison: 392 * STORE shadow(a), unpoison_val 393 * ... 394 * STORE shadow(a+99), unpoison_val x = LOAD p 395 * // rest of vmalloc process <data dependency> 396 * STORE p, a LOAD shadow(x+99) 397 * 398 * If there is no barrier between the end of unpoisoning the shadow 399 * and the store of the result to p, the stores could be committed 400 * in a different order by CPU#0, and CPU#1 could erroneously observe 401 * poison in the shadow. 402 * 403 * We need some sort of barrier between the stores. 404 * 405 * In the vmalloc() case, this is provided by a smp_wmb() in 406 * clear_vm_uninitialized_flag(). In the per-cpu allocator and in 407 * get_vm_area() and friends, the caller gets shadow allocated but 408 * doesn't have any pages mapped into the virtual address space that 409 * has been reserved. Mapping those pages in will involve taking and 410 * releasing a page-table lock, which will provide the barrier. 411 */ 412 413 return 0; 414 } 415 416 static int kasan_depopulate_vmalloc_pte(pte_t *ptep, unsigned long addr, 417 void *unused) 418 { 419 unsigned long page; 420 421 page = (unsigned long)__va(pte_pfn(*ptep) << PAGE_SHIFT); 422 423 spin_lock(&init_mm.page_table_lock); 424 425 if (likely(!pte_none(*ptep))) { 426 pte_clear(&init_mm, addr, ptep); 427 free_page(page); 428 } 429 spin_unlock(&init_mm.page_table_lock); 430 431 return 0; 432 } 433 434 /* 435 * Release the backing for the vmalloc region [start, end), which 436 * lies within the free region [free_region_start, free_region_end). 437 * 438 * This can be run lazily, long after the region was freed. It runs 439 * under vmap_area_lock, so it's not safe to interact with the vmalloc/vmap 440 * infrastructure. 441 * 442 * How does this work? 443 * ------------------- 444 * 445 * We have a region that is page aligned, labeled as A. 446 * That might not map onto the shadow in a way that is page-aligned: 447 * 448 * start end 449 * v v 450 * |????????|????????|AAAAAAAA|AA....AA|AAAAAAAA|????????| < vmalloc 451 * -------- -------- -------- -------- -------- 452 * | | | | | 453 * | | | /-------/ | 454 * \-------\|/------/ |/---------------/ 455 * ||| || 456 * |??AAAAAA|AAAAAAAA|AA??????| < shadow 457 * (1) (2) (3) 458 * 459 * First we align the start upwards and the end downwards, so that the 460 * shadow of the region aligns with shadow page boundaries. In the 461 * example, this gives us the shadow page (2). This is the shadow entirely 462 * covered by this allocation. 463 * 464 * Then we have the tricky bits. We want to know if we can free the 465 * partially covered shadow pages - (1) and (3) in the example. For this, 466 * we are given the start and end of the free region that contains this 467 * allocation. Extending our previous example, we could have: 468 * 469 * free_region_start free_region_end 470 * | start end | 471 * v v v v 472 * |FFFFFFFF|FFFFFFFF|AAAAAAAA|AA....AA|AAAAAAAA|FFFFFFFF| < vmalloc 473 * -------- -------- -------- -------- -------- 474 * | | | | | 475 * | | | /-------/ | 476 * \-------\|/------/ |/---------------/ 477 * ||| || 478 * |FFAAAAAA|AAAAAAAA|AAF?????| < shadow 479 * (1) (2) (3) 480 * 481 * Once again, we align the start of the free region up, and the end of 482 * the free region down so that the shadow is page aligned. So we can free 483 * page (1) - we know no allocation currently uses anything in that page, 484 * because all of it is in the vmalloc free region. But we cannot free 485 * page (3), because we can't be sure that the rest of it is unused. 486 * 487 * We only consider pages that contain part of the original region for 488 * freeing: we don't try to free other pages from the free region or we'd 489 * end up trying to free huge chunks of virtual address space. 490 * 491 * Concurrency 492 * ----------- 493 * 494 * How do we know that we're not freeing a page that is simultaneously 495 * being used for a fresh allocation in kasan_populate_vmalloc(_pte)? 496 * 497 * We _can_ have kasan_release_vmalloc and kasan_populate_vmalloc running 498 * at the same time. While we run under free_vmap_area_lock, the population 499 * code does not. 500 * 501 * free_vmap_area_lock instead operates to ensure that the larger range 502 * [free_region_start, free_region_end) is safe: because __alloc_vmap_area and 503 * the per-cpu region-finding algorithm both run under free_vmap_area_lock, 504 * no space identified as free will become used while we are running. This 505 * means that so long as we are careful with alignment and only free shadow 506 * pages entirely covered by the free region, we will not run in to any 507 * trouble - any simultaneous allocations will be for disjoint regions. 508 */ 509 void kasan_release_vmalloc(unsigned long start, unsigned long end, 510 unsigned long free_region_start, 511 unsigned long free_region_end) 512 { 513 void *shadow_start, *shadow_end; 514 unsigned long region_start, region_end; 515 unsigned long size; 516 517 if (!kasan_arch_is_ready()) 518 return; 519 520 region_start = ALIGN(start, KASAN_MEMORY_PER_SHADOW_PAGE); 521 region_end = ALIGN_DOWN(end, KASAN_MEMORY_PER_SHADOW_PAGE); 522 523 free_region_start = ALIGN(free_region_start, KASAN_MEMORY_PER_SHADOW_PAGE); 524 525 if (start != region_start && 526 free_region_start < region_start) 527 region_start -= KASAN_MEMORY_PER_SHADOW_PAGE; 528 529 free_region_end = ALIGN_DOWN(free_region_end, KASAN_MEMORY_PER_SHADOW_PAGE); 530 531 if (end != region_end && 532 free_region_end > region_end) 533 region_end += KASAN_MEMORY_PER_SHADOW_PAGE; 534 535 shadow_start = kasan_mem_to_shadow((void *)region_start); 536 shadow_end = kasan_mem_to_shadow((void *)region_end); 537 538 if (shadow_end > shadow_start) { 539 size = shadow_end - shadow_start; 540 if (IS_ENABLED(CONFIG_UML)) { 541 __memset(shadow_start, KASAN_SHADOW_INIT, shadow_end - shadow_start); 542 return; 543 } 544 apply_to_existing_page_range(&init_mm, 545 (unsigned long)shadow_start, 546 size, kasan_depopulate_vmalloc_pte, 547 NULL); 548 flush_tlb_kernel_range((unsigned long)shadow_start, 549 (unsigned long)shadow_end); 550 } 551 } 552 553 void *__kasan_unpoison_vmalloc(const void *start, unsigned long size, 554 kasan_vmalloc_flags_t flags) 555 { 556 /* 557 * Software KASAN modes unpoison both VM_ALLOC and non-VM_ALLOC 558 * mappings, so the KASAN_VMALLOC_VM_ALLOC flag is ignored. 559 * Software KASAN modes can't optimize zeroing memory by combining it 560 * with setting memory tags, so the KASAN_VMALLOC_INIT flag is ignored. 561 */ 562 563 if (!kasan_arch_is_ready()) 564 return (void *)start; 565 566 if (!is_vmalloc_or_module_addr(start)) 567 return (void *)start; 568 569 /* 570 * Don't tag executable memory with the tag-based mode. 571 * The kernel doesn't tolerate having the PC register tagged. 572 */ 573 if (IS_ENABLED(CONFIG_KASAN_SW_TAGS) && 574 !(flags & KASAN_VMALLOC_PROT_NORMAL)) 575 return (void *)start; 576 577 start = set_tag(start, kasan_random_tag()); 578 kasan_unpoison(start, size, false); 579 return (void *)start; 580 } 581 582 /* 583 * Poison the shadow for a vmalloc region. Called as part of the 584 * freeing process at the time the region is freed. 585 */ 586 void __kasan_poison_vmalloc(const void *start, unsigned long size) 587 { 588 if (!kasan_arch_is_ready()) 589 return; 590 591 if (!is_vmalloc_or_module_addr(start)) 592 return; 593 594 size = round_up(size, KASAN_GRANULE_SIZE); 595 kasan_poison(start, size, KASAN_VMALLOC_INVALID, false); 596 } 597 598 #else /* CONFIG_KASAN_VMALLOC */ 599 600 int kasan_alloc_module_shadow(void *addr, size_t size, gfp_t gfp_mask) 601 { 602 void *ret; 603 size_t scaled_size; 604 size_t shadow_size; 605 unsigned long shadow_start; 606 607 shadow_start = (unsigned long)kasan_mem_to_shadow(addr); 608 scaled_size = (size + KASAN_GRANULE_SIZE - 1) >> 609 KASAN_SHADOW_SCALE_SHIFT; 610 shadow_size = round_up(scaled_size, PAGE_SIZE); 611 612 if (WARN_ON(!PAGE_ALIGNED(shadow_start))) 613 return -EINVAL; 614 615 if (IS_ENABLED(CONFIG_UML)) { 616 __memset((void *)shadow_start, KASAN_SHADOW_INIT, shadow_size); 617 return 0; 618 } 619 620 ret = __vmalloc_node_range(shadow_size, 1, shadow_start, 621 shadow_start + shadow_size, 622 GFP_KERNEL, 623 PAGE_KERNEL, VM_NO_GUARD, NUMA_NO_NODE, 624 __builtin_return_address(0)); 625 626 if (ret) { 627 struct vm_struct *vm = find_vm_area(addr); 628 __memset(ret, KASAN_SHADOW_INIT, shadow_size); 629 vm->flags |= VM_KASAN; 630 kmemleak_ignore(ret); 631 632 if (vm->flags & VM_DEFER_KMEMLEAK) 633 kmemleak_vmalloc(vm, size, gfp_mask); 634 635 return 0; 636 } 637 638 return -ENOMEM; 639 } 640 641 void kasan_free_module_shadow(const struct vm_struct *vm) 642 { 643 if (IS_ENABLED(CONFIG_UML)) 644 return; 645 646 if (vm->flags & VM_KASAN) 647 vfree(kasan_mem_to_shadow(vm->addr)); 648 } 649 650 #endif 651