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 #ifndef 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 #undef memset 48 void *memset(void *addr, int c, size_t len) 49 { 50 if (!kasan_check_range((unsigned long)addr, len, true, _RET_IP_)) 51 return NULL; 52 53 return __memset(addr, c, len); 54 } 55 56 #ifdef __HAVE_ARCH_MEMMOVE 57 #undef memmove 58 void *memmove(void *dest, const void *src, size_t len) 59 { 60 if (!kasan_check_range((unsigned long)src, len, false, _RET_IP_) || 61 !kasan_check_range((unsigned long)dest, len, true, _RET_IP_)) 62 return NULL; 63 64 return __memmove(dest, src, len); 65 } 66 #endif 67 68 #undef memcpy 69 void *memcpy(void *dest, const void *src, size_t len) 70 { 71 if (!kasan_check_range((unsigned long)src, len, false, _RET_IP_) || 72 !kasan_check_range((unsigned long)dest, len, true, _RET_IP_)) 73 return NULL; 74 75 return __memcpy(dest, src, len); 76 } 77 #endif 78 79 void *__asan_memset(void *addr, int c, size_t len) 80 { 81 if (!kasan_check_range((unsigned long)addr, len, true, _RET_IP_)) 82 return NULL; 83 84 return __memset(addr, c, len); 85 } 86 EXPORT_SYMBOL(__asan_memset); 87 88 #ifdef __HAVE_ARCH_MEMMOVE 89 void *__asan_memmove(void *dest, const void *src, size_t len) 90 { 91 if (!kasan_check_range((unsigned long)src, len, false, _RET_IP_) || 92 !kasan_check_range((unsigned long)dest, len, true, _RET_IP_)) 93 return NULL; 94 95 return __memmove(dest, src, len); 96 } 97 EXPORT_SYMBOL(__asan_memmove); 98 #endif 99 100 void *__asan_memcpy(void *dest, const void *src, size_t len) 101 { 102 if (!kasan_check_range((unsigned long)src, len, false, _RET_IP_) || 103 !kasan_check_range((unsigned long)dest, len, true, _RET_IP_)) 104 return NULL; 105 106 return __memcpy(dest, src, len); 107 } 108 EXPORT_SYMBOL(__asan_memcpy); 109 110 void kasan_poison(const void *addr, size_t size, u8 value, bool init) 111 { 112 void *shadow_start, *shadow_end; 113 114 if (!kasan_arch_is_ready()) 115 return; 116 117 /* 118 * Perform shadow offset calculation based on untagged address, as 119 * some of the callers (e.g. kasan_poison_object_data) pass tagged 120 * addresses to this function. 121 */ 122 addr = kasan_reset_tag(addr); 123 124 /* Skip KFENCE memory if called explicitly outside of sl*b. */ 125 if (is_kfence_address(addr)) 126 return; 127 128 if (WARN_ON((unsigned long)addr & KASAN_GRANULE_MASK)) 129 return; 130 if (WARN_ON(size & KASAN_GRANULE_MASK)) 131 return; 132 133 shadow_start = kasan_mem_to_shadow(addr); 134 shadow_end = kasan_mem_to_shadow(addr + size); 135 136 __memset(shadow_start, value, shadow_end - shadow_start); 137 } 138 EXPORT_SYMBOL(kasan_poison); 139 140 #ifdef CONFIG_KASAN_GENERIC 141 void kasan_poison_last_granule(const void *addr, size_t size) 142 { 143 if (!kasan_arch_is_ready()) 144 return; 145 146 if (size & KASAN_GRANULE_MASK) { 147 u8 *shadow = (u8 *)kasan_mem_to_shadow(addr + size); 148 *shadow = size & KASAN_GRANULE_MASK; 149 } 150 } 151 #endif 152 153 void kasan_unpoison(const void *addr, size_t size, bool init) 154 { 155 u8 tag = get_tag(addr); 156 157 /* 158 * Perform shadow offset calculation based on untagged address, as 159 * some of the callers (e.g. kasan_unpoison_object_data) pass tagged 160 * addresses to this function. 161 */ 162 addr = kasan_reset_tag(addr); 163 164 /* 165 * Skip KFENCE memory if called explicitly outside of sl*b. Also note 166 * that calls to ksize(), where size is not a multiple of machine-word 167 * size, would otherwise poison the invalid portion of the word. 168 */ 169 if (is_kfence_address(addr)) 170 return; 171 172 if (WARN_ON((unsigned long)addr & KASAN_GRANULE_MASK)) 173 return; 174 175 /* Unpoison all granules that cover the object. */ 176 kasan_poison(addr, round_up(size, KASAN_GRANULE_SIZE), tag, false); 177 178 /* Partially poison the last granule for the generic mode. */ 179 if (IS_ENABLED(CONFIG_KASAN_GENERIC)) 180 kasan_poison_last_granule(addr, size); 181 } 182 183 #ifdef CONFIG_MEMORY_HOTPLUG 184 static bool shadow_mapped(unsigned long addr) 185 { 186 pgd_t *pgd = pgd_offset_k(addr); 187 p4d_t *p4d; 188 pud_t *pud; 189 pmd_t *pmd; 190 pte_t *pte; 191 192 if (pgd_none(*pgd)) 193 return false; 194 p4d = p4d_offset(pgd, addr); 195 if (p4d_none(*p4d)) 196 return false; 197 pud = pud_offset(p4d, addr); 198 if (pud_none(*pud)) 199 return false; 200 201 /* 202 * We can't use pud_large() or pud_huge(), the first one is 203 * arch-specific, the last one depends on HUGETLB_PAGE. So let's abuse 204 * pud_bad(), if pud is bad then it's bad because it's huge. 205 */ 206 if (pud_bad(*pud)) 207 return true; 208 pmd = pmd_offset(pud, addr); 209 if (pmd_none(*pmd)) 210 return false; 211 212 if (pmd_bad(*pmd)) 213 return true; 214 pte = pte_offset_kernel(pmd, addr); 215 return !pte_none(*pte); 216 } 217 218 static int __meminit kasan_mem_notifier(struct notifier_block *nb, 219 unsigned long action, void *data) 220 { 221 struct memory_notify *mem_data = data; 222 unsigned long nr_shadow_pages, start_kaddr, shadow_start; 223 unsigned long shadow_end, shadow_size; 224 225 nr_shadow_pages = mem_data->nr_pages >> KASAN_SHADOW_SCALE_SHIFT; 226 start_kaddr = (unsigned long)pfn_to_kaddr(mem_data->start_pfn); 227 shadow_start = (unsigned long)kasan_mem_to_shadow((void *)start_kaddr); 228 shadow_size = nr_shadow_pages << PAGE_SHIFT; 229 shadow_end = shadow_start + shadow_size; 230 231 if (WARN_ON(mem_data->nr_pages % KASAN_GRANULE_SIZE) || 232 WARN_ON(start_kaddr % KASAN_MEMORY_PER_SHADOW_PAGE)) 233 return NOTIFY_BAD; 234 235 switch (action) { 236 case MEM_GOING_ONLINE: { 237 void *ret; 238 239 /* 240 * If shadow is mapped already than it must have been mapped 241 * during the boot. This could happen if we onlining previously 242 * offlined memory. 243 */ 244 if (shadow_mapped(shadow_start)) 245 return NOTIFY_OK; 246 247 ret = __vmalloc_node_range(shadow_size, PAGE_SIZE, shadow_start, 248 shadow_end, GFP_KERNEL, 249 PAGE_KERNEL, VM_NO_GUARD, 250 pfn_to_nid(mem_data->start_pfn), 251 __builtin_return_address(0)); 252 if (!ret) 253 return NOTIFY_BAD; 254 255 kmemleak_ignore(ret); 256 return NOTIFY_OK; 257 } 258 case MEM_CANCEL_ONLINE: 259 case MEM_OFFLINE: { 260 struct vm_struct *vm; 261 262 /* 263 * shadow_start was either mapped during boot by kasan_init() 264 * or during memory online by __vmalloc_node_range(). 265 * In the latter case we can use vfree() to free shadow. 266 * Non-NULL result of the find_vm_area() will tell us if 267 * that was the second case. 268 * 269 * Currently it's not possible to free shadow mapped 270 * during boot by kasan_init(). It's because the code 271 * to do that hasn't been written yet. So we'll just 272 * leak the memory. 273 */ 274 vm = find_vm_area((void *)shadow_start); 275 if (vm) 276 vfree((void *)shadow_start); 277 } 278 } 279 280 return NOTIFY_OK; 281 } 282 283 static int __init kasan_memhotplug_init(void) 284 { 285 hotplug_memory_notifier(kasan_mem_notifier, DEFAULT_CALLBACK_PRI); 286 287 return 0; 288 } 289 290 core_initcall(kasan_memhotplug_init); 291 #endif 292 293 #ifdef CONFIG_KASAN_VMALLOC 294 295 void __init __weak kasan_populate_early_vm_area_shadow(void *start, 296 unsigned long size) 297 { 298 } 299 300 static int kasan_populate_vmalloc_pte(pte_t *ptep, unsigned long addr, 301 void *unused) 302 { 303 unsigned long page; 304 pte_t pte; 305 306 if (likely(!pte_none(*ptep))) 307 return 0; 308 309 page = __get_free_page(GFP_KERNEL); 310 if (!page) 311 return -ENOMEM; 312 313 memset((void *)page, KASAN_VMALLOC_INVALID, PAGE_SIZE); 314 pte = pfn_pte(PFN_DOWN(__pa(page)), PAGE_KERNEL); 315 316 spin_lock(&init_mm.page_table_lock); 317 if (likely(pte_none(*ptep))) { 318 set_pte_at(&init_mm, addr, ptep, pte); 319 page = 0; 320 } 321 spin_unlock(&init_mm.page_table_lock); 322 if (page) 323 free_page(page); 324 return 0; 325 } 326 327 int kasan_populate_vmalloc(unsigned long addr, unsigned long size) 328 { 329 unsigned long shadow_start, shadow_end; 330 int ret; 331 332 if (!kasan_arch_is_ready()) 333 return 0; 334 335 if (!is_vmalloc_or_module_addr((void *)addr)) 336 return 0; 337 338 shadow_start = (unsigned long)kasan_mem_to_shadow((void *)addr); 339 shadow_end = (unsigned long)kasan_mem_to_shadow((void *)addr + size); 340 341 /* 342 * User Mode Linux maps enough shadow memory for all of virtual memory 343 * at boot, so doesn't need to allocate more on vmalloc, just clear it. 344 * 345 * The remaining CONFIG_UML checks in this file exist for the same 346 * reason. 347 */ 348 if (IS_ENABLED(CONFIG_UML)) { 349 __memset((void *)shadow_start, KASAN_VMALLOC_INVALID, shadow_end - shadow_start); 350 return 0; 351 } 352 353 shadow_start = PAGE_ALIGN_DOWN(shadow_start); 354 shadow_end = PAGE_ALIGN(shadow_end); 355 356 ret = apply_to_page_range(&init_mm, shadow_start, 357 shadow_end - shadow_start, 358 kasan_populate_vmalloc_pte, NULL); 359 if (ret) 360 return ret; 361 362 flush_cache_vmap(shadow_start, shadow_end); 363 364 /* 365 * We need to be careful about inter-cpu effects here. Consider: 366 * 367 * CPU#0 CPU#1 368 * WRITE_ONCE(p, vmalloc(100)); while (x = READ_ONCE(p)) ; 369 * p[99] = 1; 370 * 371 * With compiler instrumentation, that ends up looking like this: 372 * 373 * CPU#0 CPU#1 374 * // vmalloc() allocates memory 375 * // let a = area->addr 376 * // we reach kasan_populate_vmalloc 377 * // and call kasan_unpoison: 378 * STORE shadow(a), unpoison_val 379 * ... 380 * STORE shadow(a+99), unpoison_val x = LOAD p 381 * // rest of vmalloc process <data dependency> 382 * STORE p, a LOAD shadow(x+99) 383 * 384 * If there is no barrier between the end of unpoisoning the shadow 385 * and the store of the result to p, the stores could be committed 386 * in a different order by CPU#0, and CPU#1 could erroneously observe 387 * poison in the shadow. 388 * 389 * We need some sort of barrier between the stores. 390 * 391 * In the vmalloc() case, this is provided by a smp_wmb() in 392 * clear_vm_uninitialized_flag(). In the per-cpu allocator and in 393 * get_vm_area() and friends, the caller gets shadow allocated but 394 * doesn't have any pages mapped into the virtual address space that 395 * has been reserved. Mapping those pages in will involve taking and 396 * releasing a page-table lock, which will provide the barrier. 397 */ 398 399 return 0; 400 } 401 402 static int kasan_depopulate_vmalloc_pte(pte_t *ptep, unsigned long addr, 403 void *unused) 404 { 405 unsigned long page; 406 407 page = (unsigned long)__va(pte_pfn(*ptep) << PAGE_SHIFT); 408 409 spin_lock(&init_mm.page_table_lock); 410 411 if (likely(!pte_none(*ptep))) { 412 pte_clear(&init_mm, addr, ptep); 413 free_page(page); 414 } 415 spin_unlock(&init_mm.page_table_lock); 416 417 return 0; 418 } 419 420 /* 421 * Release the backing for the vmalloc region [start, end), which 422 * lies within the free region [free_region_start, free_region_end). 423 * 424 * This can be run lazily, long after the region was freed. It runs 425 * under vmap_area_lock, so it's not safe to interact with the vmalloc/vmap 426 * infrastructure. 427 * 428 * How does this work? 429 * ------------------- 430 * 431 * We have a region that is page aligned, labeled as A. 432 * That might not map onto the shadow in a way that is page-aligned: 433 * 434 * start end 435 * v v 436 * |????????|????????|AAAAAAAA|AA....AA|AAAAAAAA|????????| < vmalloc 437 * -------- -------- -------- -------- -------- 438 * | | | | | 439 * | | | /-------/ | 440 * \-------\|/------/ |/---------------/ 441 * ||| || 442 * |??AAAAAA|AAAAAAAA|AA??????| < shadow 443 * (1) (2) (3) 444 * 445 * First we align the start upwards and the end downwards, so that the 446 * shadow of the region aligns with shadow page boundaries. In the 447 * example, this gives us the shadow page (2). This is the shadow entirely 448 * covered by this allocation. 449 * 450 * Then we have the tricky bits. We want to know if we can free the 451 * partially covered shadow pages - (1) and (3) in the example. For this, 452 * we are given the start and end of the free region that contains this 453 * allocation. Extending our previous example, we could have: 454 * 455 * free_region_start free_region_end 456 * | start end | 457 * v v v v 458 * |FFFFFFFF|FFFFFFFF|AAAAAAAA|AA....AA|AAAAAAAA|FFFFFFFF| < vmalloc 459 * -------- -------- -------- -------- -------- 460 * | | | | | 461 * | | | /-------/ | 462 * \-------\|/------/ |/---------------/ 463 * ||| || 464 * |FFAAAAAA|AAAAAAAA|AAF?????| < shadow 465 * (1) (2) (3) 466 * 467 * Once again, we align the start of the free region up, and the end of 468 * the free region down so that the shadow is page aligned. So we can free 469 * page (1) - we know no allocation currently uses anything in that page, 470 * because all of it is in the vmalloc free region. But we cannot free 471 * page (3), because we can't be sure that the rest of it is unused. 472 * 473 * We only consider pages that contain part of the original region for 474 * freeing: we don't try to free other pages from the free region or we'd 475 * end up trying to free huge chunks of virtual address space. 476 * 477 * Concurrency 478 * ----------- 479 * 480 * How do we know that we're not freeing a page that is simultaneously 481 * being used for a fresh allocation in kasan_populate_vmalloc(_pte)? 482 * 483 * We _can_ have kasan_release_vmalloc and kasan_populate_vmalloc running 484 * at the same time. While we run under free_vmap_area_lock, the population 485 * code does not. 486 * 487 * free_vmap_area_lock instead operates to ensure that the larger range 488 * [free_region_start, free_region_end) is safe: because __alloc_vmap_area and 489 * the per-cpu region-finding algorithm both run under free_vmap_area_lock, 490 * no space identified as free will become used while we are running. This 491 * means that so long as we are careful with alignment and only free shadow 492 * pages entirely covered by the free region, we will not run in to any 493 * trouble - any simultaneous allocations will be for disjoint regions. 494 */ 495 void kasan_release_vmalloc(unsigned long start, unsigned long end, 496 unsigned long free_region_start, 497 unsigned long free_region_end) 498 { 499 void *shadow_start, *shadow_end; 500 unsigned long region_start, region_end; 501 unsigned long size; 502 503 if (!kasan_arch_is_ready()) 504 return; 505 506 region_start = ALIGN(start, KASAN_MEMORY_PER_SHADOW_PAGE); 507 region_end = ALIGN_DOWN(end, KASAN_MEMORY_PER_SHADOW_PAGE); 508 509 free_region_start = ALIGN(free_region_start, KASAN_MEMORY_PER_SHADOW_PAGE); 510 511 if (start != region_start && 512 free_region_start < region_start) 513 region_start -= KASAN_MEMORY_PER_SHADOW_PAGE; 514 515 free_region_end = ALIGN_DOWN(free_region_end, KASAN_MEMORY_PER_SHADOW_PAGE); 516 517 if (end != region_end && 518 free_region_end > region_end) 519 region_end += KASAN_MEMORY_PER_SHADOW_PAGE; 520 521 shadow_start = kasan_mem_to_shadow((void *)region_start); 522 shadow_end = kasan_mem_to_shadow((void *)region_end); 523 524 if (shadow_end > shadow_start) { 525 size = shadow_end - shadow_start; 526 if (IS_ENABLED(CONFIG_UML)) { 527 __memset(shadow_start, KASAN_SHADOW_INIT, shadow_end - shadow_start); 528 return; 529 } 530 apply_to_existing_page_range(&init_mm, 531 (unsigned long)shadow_start, 532 size, kasan_depopulate_vmalloc_pte, 533 NULL); 534 flush_tlb_kernel_range((unsigned long)shadow_start, 535 (unsigned long)shadow_end); 536 } 537 } 538 539 void *__kasan_unpoison_vmalloc(const void *start, unsigned long size, 540 kasan_vmalloc_flags_t flags) 541 { 542 /* 543 * Software KASAN modes unpoison both VM_ALLOC and non-VM_ALLOC 544 * mappings, so the KASAN_VMALLOC_VM_ALLOC flag is ignored. 545 * Software KASAN modes can't optimize zeroing memory by combining it 546 * with setting memory tags, so the KASAN_VMALLOC_INIT flag is ignored. 547 */ 548 549 if (!kasan_arch_is_ready()) 550 return (void *)start; 551 552 if (!is_vmalloc_or_module_addr(start)) 553 return (void *)start; 554 555 /* 556 * Don't tag executable memory with the tag-based mode. 557 * The kernel doesn't tolerate having the PC register tagged. 558 */ 559 if (IS_ENABLED(CONFIG_KASAN_SW_TAGS) && 560 !(flags & KASAN_VMALLOC_PROT_NORMAL)) 561 return (void *)start; 562 563 start = set_tag(start, kasan_random_tag()); 564 kasan_unpoison(start, size, false); 565 return (void *)start; 566 } 567 568 /* 569 * Poison the shadow for a vmalloc region. Called as part of the 570 * freeing process at the time the region is freed. 571 */ 572 void __kasan_poison_vmalloc(const void *start, unsigned long size) 573 { 574 if (!kasan_arch_is_ready()) 575 return; 576 577 if (!is_vmalloc_or_module_addr(start)) 578 return; 579 580 size = round_up(size, KASAN_GRANULE_SIZE); 581 kasan_poison(start, size, KASAN_VMALLOC_INVALID, false); 582 } 583 584 #else /* CONFIG_KASAN_VMALLOC */ 585 586 int kasan_alloc_module_shadow(void *addr, size_t size, gfp_t gfp_mask) 587 { 588 void *ret; 589 size_t scaled_size; 590 size_t shadow_size; 591 unsigned long shadow_start; 592 593 shadow_start = (unsigned long)kasan_mem_to_shadow(addr); 594 scaled_size = (size + KASAN_GRANULE_SIZE - 1) >> 595 KASAN_SHADOW_SCALE_SHIFT; 596 shadow_size = round_up(scaled_size, PAGE_SIZE); 597 598 if (WARN_ON(!PAGE_ALIGNED(shadow_start))) 599 return -EINVAL; 600 601 if (IS_ENABLED(CONFIG_UML)) { 602 __memset((void *)shadow_start, KASAN_SHADOW_INIT, shadow_size); 603 return 0; 604 } 605 606 ret = __vmalloc_node_range(shadow_size, 1, shadow_start, 607 shadow_start + shadow_size, 608 GFP_KERNEL, 609 PAGE_KERNEL, VM_NO_GUARD, NUMA_NO_NODE, 610 __builtin_return_address(0)); 611 612 if (ret) { 613 struct vm_struct *vm = find_vm_area(addr); 614 __memset(ret, KASAN_SHADOW_INIT, shadow_size); 615 vm->flags |= VM_KASAN; 616 kmemleak_ignore(ret); 617 618 if (vm->flags & VM_DEFER_KMEMLEAK) 619 kmemleak_vmalloc(vm, size, gfp_mask); 620 621 return 0; 622 } 623 624 return -ENOMEM; 625 } 626 627 void kasan_free_module_shadow(const struct vm_struct *vm) 628 { 629 if (IS_ENABLED(CONFIG_UML)) 630 return; 631 632 if (vm->flags & VM_KASAN) 633 vfree(kasan_mem_to_shadow(vm->addr)); 634 } 635 636 #endif 637