mm/kasan/common.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * This file contains common generic and tag-based KASAN code.
   4  *
   5  * Copyright (c) 2014 Samsung Electronics Co., Ltd.
   6  * Author: Andrey Ryabinin <ryabinin.a.a@gmail.com>
   7  *
   8  * Some code borrowed from https://github.com/xairy/kasan-prototype by
   9  *        Andrey Konovalov <andreyknvl@gmail.com>
  10  *
  11  * This program is free software; you can redistribute it and/or modify
  12  * it under the terms of the GNU General Public License version 2 as
  13  * published by the Free Software Foundation.
  14  *
  15  */
  16
  17 #include <linux/export.h>
  18 #include <linux/init.h>
  19 #include <linux/kasan.h>
  20 #include <linux/kernel.h>
  21 #include <linux/kmemleak.h>
  22 #include <linux/linkage.h>
  23 #include <linux/memblock.h>
  24 #include <linux/memory.h>
  25 #include <linux/mm.h>
  26 #include <linux/module.h>
  27 #include <linux/printk.h>
  28 #include <linux/sched.h>
  29 #include <linux/sched/task_stack.h>
  30 #include <linux/slab.h>
  31 #include <linux/stacktrace.h>
  32 #include <linux/string.h>
  33 #include <linux/types.h>
  34 #include <linux/vmalloc.h>
  35 #include <linux/bug.h>
  36
  37 #include <asm/cacheflush.h>
  38 #include <asm/tlbflush.h>
  39
  40 #include "kasan.h"
  41 #include "../slab.h"
  42
  43 depot_stack_handle_t kasan_save_stack(gfp_t flags)
  44 {
  45         unsigned long entries[KASAN_STACK_DEPTH];
  46         unsigned int nr_entries;
  47
  48         nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 0);
  49         nr_entries = filter_irq_stacks(entries, nr_entries);
  50         return stack_depot_save(entries, nr_entries, flags);
  51 }
  52
  53 void kasan_set_track(struct kasan_track *track, gfp_t flags)
  54 {
  55         track->pid = current->pid;
  56         track->stack = kasan_save_stack(flags);
  57 }
  58
  59 void kasan_enable_current(void)
  60 {
  61         current->kasan_depth++;
  62 }
  63
  64 void kasan_disable_current(void)
  65 {
  66         current->kasan_depth--;
  67 }
  68
  69 bool __kasan_check_read(const volatile void *p, unsigned int size)
  70 {
  71         return check_memory_region((unsigned long)p, size, false, _RET_IP_);
  72 }
  73 EXPORT_SYMBOL(__kasan_check_read);
  74
  75 bool __kasan_check_write(const volatile void *p, unsigned int size)
  76 {
  77         return check_memory_region((unsigned long)p, size, true, _RET_IP_);
  78 }
  79 EXPORT_SYMBOL(__kasan_check_write);
  80
  81 #undef memset
  82 void *memset(void *addr, int c, size_t len)
  83 {
  84         if (!check_memory_region((unsigned long)addr, len, true, _RET_IP_))
  85                 return NULL;
  86
  87         return __memset(addr, c, len);
  88 }
  89
  90 #ifdef __HAVE_ARCH_MEMMOVE
  91 #undef memmove
  92 void *memmove(void *dest, const void *src, size_t len)
  93 {
  94         if (!check_memory_region((unsigned long)src, len, false, _RET_IP_) ||
  95             !check_memory_region((unsigned long)dest, len, true, _RET_IP_))
  96                 return NULL;
  97
  98         return __memmove(dest, src, len);
  99 }
 100 #endif
 101
 102 #undef memcpy
 103 void *memcpy(void *dest, const void *src, size_t len)
 104 {
 105         if (!check_memory_region((unsigned long)src, len, false, _RET_IP_) ||
 106             !check_memory_region((unsigned long)dest, len, true, _RET_IP_))
 107                 return NULL;
 108
 109         return __memcpy(dest, src, len);
 110 }
 111
 112 /*
 113  * Poisons the shadow memory for 'size' bytes starting from 'addr'.
 114  * Memory addresses should be aligned to KASAN_SHADOW_SCALE_SIZE.
 115  */
 116 void kasan_poison_shadow(const void *address, size_t size, u8 value)
 117 {
 118         void *shadow_start, *shadow_end;
 119
 120         /*
 121          * Perform shadow offset calculation based on untagged address, as
 122          * some of the callers (e.g. kasan_poison_object_data) pass tagged
 123          * addresses to this function.
 124          */
 125         address = reset_tag(address);
 126
 127         shadow_start = kasan_mem_to_shadow(address);
 128         shadow_end = kasan_mem_to_shadow(address + size);
 129
 130         __memset(shadow_start, value, shadow_end - shadow_start);
 131 }
 132
 133 void kasan_unpoison_shadow(const void *address, size_t size)
 134 {
 135         u8 tag = get_tag(address);
 136
 137         /*
 138          * Perform shadow offset calculation based on untagged address, as
 139          * some of the callers (e.g. kasan_unpoison_object_data) pass tagged
 140          * addresses to this function.
 141          */
 142         address = reset_tag(address);
 143
 144         kasan_poison_shadow(address, size, tag);
 145
 146         if (size & KASAN_SHADOW_MASK) {
 147                 u8 *shadow = (u8 *)kasan_mem_to_shadow(address + size);
 148
 149                 if (IS_ENABLED(CONFIG_KASAN_SW_TAGS))
 150                         *shadow = tag;
 151                 else
 152                         *shadow = size & KASAN_SHADOW_MASK;
 153         }
 154 }
 155
 156 static void __kasan_unpoison_stack(struct task_struct *task, const void *sp)
 157 {
 158         void *base = task_stack_page(task);
 159         size_t size = sp - base;
 160
 161         kasan_unpoison_shadow(base, size);
 162 }
 163
 164 /* Unpoison the entire stack for a task. */
 165 void kasan_unpoison_task_stack(struct task_struct *task)
 166 {
 167         __kasan_unpoison_stack(task, task_stack_page(task) + THREAD_SIZE);
 168 }
 169
 170 /* Unpoison the stack for the current task beyond a watermark sp value. */
 171 asmlinkage void kasan_unpoison_task_stack_below(const void *watermark)
 172 {
 173         /*
 174          * Calculate the task stack base address.  Avoid using 'current'
 175          * because this function is called by early resume code which hasn't
 176          * yet set up the percpu register (%gs).
 177          */
 178         void *base = (void *)((unsigned long)watermark & ~(THREAD_SIZE - 1));
 179
 180         kasan_unpoison_shadow(base, watermark - base);
 181 }
 182
 183 void kasan_alloc_pages(struct page *page, unsigned int order)
 184 {
 185         u8 tag;
 186         unsigned long i;
 187
 188         if (unlikely(PageHighMem(page)))
 189                 return;
 190
 191         tag = random_tag();
 192         for (i = 0; i < (1 << order); i++)
 193                 page_kasan_tag_set(page + i, tag);
 194         kasan_unpoison_shadow(page_address(page), PAGE_SIZE << order);
 195 }
 196
 197 void kasan_free_pages(struct page *page, unsigned int order)
 198 {
 199         if (likely(!PageHighMem(page)))
 200                 kasan_poison_shadow(page_address(page),
 201                                 PAGE_SIZE << order,
 202                                 KASAN_FREE_PAGE);
 203 }
 204
 205 /*
 206  * Adaptive redzone policy taken from the userspace AddressSanitizer runtime.
 207  * For larger allocations larger redzones are used.
 208  */
 209 static inline unsigned int optimal_redzone(unsigned int object_size)
 210 {
 211         if (IS_ENABLED(CONFIG_KASAN_SW_TAGS))
 212                 return 0;
 213
 214         return
 215                 object_size <= 64        - 16   ? 16 :
 216                 object_size <= 128       - 32   ? 32 :
 217                 object_size <= 512       - 64   ? 64 :
 218                 object_size <= 4096      - 128  ? 128 :
 219                 object_size <= (1 << 14) - 256  ? 256 :
 220                 object_size <= (1 << 15) - 512  ? 512 :
 221                 object_size <= (1 << 16) - 1024 ? 1024 : 2048;
 222 }
 223
 224 void kasan_cache_create(struct kmem_cache *cache, unsigned int *size,
 225                         slab_flags_t *flags)
 226 {
 227         unsigned int orig_size = *size;
 228         unsigned int redzone_size;
 229         int redzone_adjust;
 230
 231         /* Add alloc meta. */
 232         cache->kasan_info.alloc_meta_offset = *size;
 233         *size += sizeof(struct kasan_alloc_meta);
 234
 235         /* Add free meta. */
 236         if (IS_ENABLED(CONFIG_KASAN_GENERIC) &&
 237             (cache->flags & SLAB_TYPESAFE_BY_RCU || cache->ctor ||
 238              cache->object_size < sizeof(struct kasan_free_meta))) {
 239                 cache->kasan_info.free_meta_offset = *size;
 240                 *size += sizeof(struct kasan_free_meta);
 241         }
 242
 243         redzone_size = optimal_redzone(cache->object_size);
 244         redzone_adjust = redzone_size - (*size - cache->object_size);
 245         if (redzone_adjust > 0)
 246                 *size += redzone_adjust;
 247
 248         *size = min_t(unsigned int, KMALLOC_MAX_SIZE,
 249                         max(*size, cache->object_size + redzone_size));
 250
 251         /*
 252          * If the metadata doesn't fit, don't enable KASAN at all.
 253          */
 254         if (*size <= cache->kasan_info.alloc_meta_offset ||
 255                         *size <= cache->kasan_info.free_meta_offset) {
 256                 cache->kasan_info.alloc_meta_offset = 0;
 257                 cache->kasan_info.free_meta_offset = 0;
 258                 *size = orig_size;
 259                 return;
 260         }
 261
 262         *flags |= SLAB_KASAN;
 263 }
 264
 265 size_t kasan_metadata_size(struct kmem_cache *cache)
 266 {
 267         return (cache->kasan_info.alloc_meta_offset ?
 268                 sizeof(struct kasan_alloc_meta) : 0) +
 269                 (cache->kasan_info.free_meta_offset ?
 270                 sizeof(struct kasan_free_meta) : 0);
 271 }
 272
 273 struct kasan_alloc_meta *get_alloc_info(struct kmem_cache *cache,
 274                                         const void *object)
 275 {
 276         return (void *)object + cache->kasan_info.alloc_meta_offset;
 277 }
 278
 279 struct kasan_free_meta *get_free_info(struct kmem_cache *cache,
 280                                       const void *object)
 281 {
 282         BUILD_BUG_ON(sizeof(struct kasan_free_meta) > 32);
 283         return (void *)object + cache->kasan_info.free_meta_offset;
 284 }
 285
 286 void kasan_poison_slab(struct page *page)
 287 {
 288         unsigned long i;
 289
 290         for (i = 0; i < compound_nr(page); i++)
 291                 page_kasan_tag_reset(page + i);
 292         kasan_poison_shadow(page_address(page), page_size(page),
 293                         KASAN_KMALLOC_REDZONE);
 294 }
 295
 296 void kasan_unpoison_object_data(struct kmem_cache *cache, void *object)
 297 {
 298         kasan_unpoison_shadow(object, cache->object_size);
 299 }
 300
 301 void kasan_poison_object_data(struct kmem_cache *cache, void *object)
 302 {
 303         kasan_poison_shadow(object,
 304                         round_up(cache->object_size, KASAN_SHADOW_SCALE_SIZE),
 305                         KASAN_KMALLOC_REDZONE);
 306 }
 307
 308 /*
 309  * This function assigns a tag to an object considering the following:
 310  * 1. A cache might have a constructor, which might save a pointer to a slab
 311  *    object somewhere (e.g. in the object itself). We preassign a tag for
 312  *    each object in caches with constructors during slab creation and reuse
 313  *    the same tag each time a particular object is allocated.
 314  * 2. A cache might be SLAB_TYPESAFE_BY_RCU, which means objects can be
 315  *    accessed after being freed. We preassign tags for objects in these
 316  *    caches as well.
 317  * 3. For SLAB allocator we can't preassign tags randomly since the freelist
 318  *    is stored as an array of indexes instead of a linked list. Assign tags
 319  *    based on objects indexes, so that objects that are next to each other
 320  *    get different tags.
 321  */
 322 static u8 assign_tag(struct kmem_cache *cache, const void *object,
 323                         bool init, bool keep_tag)
 324 {
 325         /*
 326          * 1. When an object is kmalloc()'ed, two hooks are called:
 327          *    kasan_slab_alloc() and kasan_kmalloc(). We assign the
 328          *    tag only in the first one.
 329          * 2. We reuse the same tag for krealloc'ed objects.
 330          */
 331         if (keep_tag)
 332                 return get_tag(object);
 333
 334         /*
 335          * If the cache neither has a constructor nor has SLAB_TYPESAFE_BY_RCU
 336          * set, assign a tag when the object is being allocated (init == false).
 337          */
 338         if (!cache->ctor && !(cache->flags & SLAB_TYPESAFE_BY_RCU))
 339                 return init ? KASAN_TAG_KERNEL : random_tag();
 340
 341         /* For caches that either have a constructor or SLAB_TYPESAFE_BY_RCU: */
 342 #ifdef CONFIG_SLAB
 343         /* For SLAB assign tags based on the object index in the freelist. */
 344         return (u8)obj_to_index(cache, virt_to_page(object), (void *)object);
 345 #else
 346         /*
 347          * For SLUB assign a random tag during slab creation, otherwise reuse
 348          * the already assigned tag.
 349          */
 350         return init ? random_tag() : get_tag(object);
 351 #endif
 352 }
 353
 354 void * __must_check kasan_init_slab_obj(struct kmem_cache *cache,
 355                                                 const void *object)
 356 {
 357         struct kasan_alloc_meta *alloc_info;
 358
 359         if (!(cache->flags & SLAB_KASAN))
 360                 return (void *)object;
 361
 362         alloc_info = get_alloc_info(cache, object);
 363         __memset(alloc_info, 0, sizeof(*alloc_info));
 364
 365         if (IS_ENABLED(CONFIG_KASAN_SW_TAGS))
 366                 object = set_tag(object,
 367                                 assign_tag(cache, object, true, false));
 368
 369         return (void *)object;
 370 }
 371
 372 static inline bool shadow_invalid(u8 tag, s8 shadow_byte)
 373 {
 374         if (IS_ENABLED(CONFIG_KASAN_GENERIC))
 375                 return shadow_byte < 0 ||
 376                         shadow_byte >= KASAN_SHADOW_SCALE_SIZE;
 377
 378         /* else CONFIG_KASAN_SW_TAGS: */
 379         if ((u8)shadow_byte == KASAN_TAG_INVALID)
 380                 return true;
 381         if ((tag != KASAN_TAG_KERNEL) && (tag != (u8)shadow_byte))
 382                 return true;
 383
 384         return false;
 385 }
 386
 387 static bool __kasan_slab_free(struct kmem_cache *cache, void *object,
 388                               unsigned long ip, bool quarantine)
 389 {
 390         s8 shadow_byte;
 391         u8 tag;
 392         void *tagged_object;
 393         unsigned long rounded_up_size;
 394
 395         tag = get_tag(object);
 396         tagged_object = object;
 397         object = reset_tag(object);
 398
 399         if (unlikely(nearest_obj(cache, virt_to_head_page(object), object) !=
 400             object)) {
 401                 kasan_report_invalid_free(tagged_object, ip);
 402                 return true;
 403         }
 404
 405         /* RCU slabs could be legally used after free within the RCU period */
 406         if (unlikely(cache->flags & SLAB_TYPESAFE_BY_RCU))
 407                 return false;
 408
 409         shadow_byte = READ_ONCE(*(s8 *)kasan_mem_to_shadow(object));
 410         if (shadow_invalid(tag, shadow_byte)) {
 411                 kasan_report_invalid_free(tagged_object, ip);
 412                 return true;
 413         }
 414
 415         rounded_up_size = round_up(cache->object_size, KASAN_SHADOW_SCALE_SIZE);
 416         kasan_poison_shadow(object, rounded_up_size, KASAN_KMALLOC_FREE);
 417
 418         if ((IS_ENABLED(CONFIG_KASAN_GENERIC) && !quarantine) ||
 419                         unlikely(!(cache->flags & SLAB_KASAN)))
 420                 return false;
 421
 422         kasan_set_free_info(cache, object, tag);
 423
 424         quarantine_put(get_free_info(cache, object), cache);
 425
 426         return IS_ENABLED(CONFIG_KASAN_GENERIC);
 427 }
 428
 429 bool kasan_slab_free(struct kmem_cache *cache, void *object, unsigned long ip)
 430 {
 431         return __kasan_slab_free(cache, object, ip, true);
 432 }
 433
 434 static void *__kasan_kmalloc(struct kmem_cache *cache, const void *object,
 435                                 size_t size, gfp_t flags, bool keep_tag)
 436 {
 437         unsigned long redzone_start;
 438         unsigned long redzone_end;
 439         u8 tag = 0xff;
 440
 441         if (gfpflags_allow_blocking(flags))
 442                 quarantine_reduce();
 443
 444         if (unlikely(object == NULL))
 445                 return NULL;
 446
 447         redzone_start = round_up((unsigned long)(object + size),
 448                                 KASAN_SHADOW_SCALE_SIZE);
 449         redzone_end = round_up((unsigned long)object + cache->object_size,
 450                                 KASAN_SHADOW_SCALE_SIZE);
 451
 452         if (IS_ENABLED(CONFIG_KASAN_SW_TAGS))
 453                 tag = assign_tag(cache, object, false, keep_tag);
 454
 455         /* Tag is ignored in set_tag without CONFIG_KASAN_SW_TAGS */
 456         kasan_unpoison_shadow(set_tag(object, tag), size);
 457         kasan_poison_shadow((void *)redzone_start, redzone_end - redzone_start,
 458                 KASAN_KMALLOC_REDZONE);
 459
 460         if (cache->flags & SLAB_KASAN)
 461                 kasan_set_track(&get_alloc_info(cache, object)->alloc_track, flags);
 462
 463         return set_tag(object, tag);
 464 }
 465
 466 void * __must_check kasan_slab_alloc(struct kmem_cache *cache, void *object,
 467                                         gfp_t flags)
 468 {
 469         return __kasan_kmalloc(cache, object, cache->object_size, flags, false);
 470 }
 471
 472 void * __must_check kasan_kmalloc(struct kmem_cache *cache, const void *object,
 473                                 size_t size, gfp_t flags)
 474 {
 475         return __kasan_kmalloc(cache, object, size, flags, true);
 476 }
 477 EXPORT_SYMBOL(kasan_kmalloc);
 478
 479 void * __must_check kasan_kmalloc_large(const void *ptr, size_t size,
 480                                                 gfp_t flags)
 481 {
 482         struct page *page;
 483         unsigned long redzone_start;
 484         unsigned long redzone_end;
 485
 486         if (gfpflags_allow_blocking(flags))
 487                 quarantine_reduce();
 488
 489         if (unlikely(ptr == NULL))
 490                 return NULL;
 491
 492         page = virt_to_page(ptr);
 493         redzone_start = round_up((unsigned long)(ptr + size),
 494                                 KASAN_SHADOW_SCALE_SIZE);
 495         redzone_end = (unsigned long)ptr + page_size(page);
 496
 497         kasan_unpoison_shadow(ptr, size);
 498         kasan_poison_shadow((void *)redzone_start, redzone_end - redzone_start,
 499                 KASAN_PAGE_REDZONE);
 500
 501         return (void *)ptr;
 502 }
 503
 504 void * __must_check kasan_krealloc(const void *object, size_t size, gfp_t flags)
 505 {
 506         struct page *page;
 507
 508         if (unlikely(object == ZERO_SIZE_PTR))
 509                 return (void *)object;
 510
 511         page = virt_to_head_page(object);
 512
 513         if (unlikely(!PageSlab(page)))
 514                 return kasan_kmalloc_large(object, size, flags);
 515         else
 516                 return __kasan_kmalloc(page->slab_cache, object, size,
 517                                                 flags, true);
 518 }
 519
 520 void kasan_poison_kfree(void *ptr, unsigned long ip)
 521 {
 522         struct page *page;
 523
 524         page = virt_to_head_page(ptr);
 525
 526         if (unlikely(!PageSlab(page))) {
 527                 if (ptr != page_address(page)) {
 528                         kasan_report_invalid_free(ptr, ip);
 529                         return;
 530                 }
 531                 kasan_poison_shadow(ptr, page_size(page), KASAN_FREE_PAGE);
 532         } else {
 533                 __kasan_slab_free(page->slab_cache, ptr, ip, false);
 534         }
 535 }
 536
 537 void kasan_kfree_large(void *ptr, unsigned long ip)
 538 {
 539         if (ptr != page_address(virt_to_head_page(ptr)))
 540                 kasan_report_invalid_free(ptr, ip);
 541         /* The object will be poisoned by page_alloc. */
 542 }
 543
 544 #ifndef CONFIG_KASAN_VMALLOC
 545 int kasan_module_alloc(void *addr, size_t size)
 546 {
 547         void *ret;
 548         size_t scaled_size;
 549         size_t shadow_size;
 550         unsigned long shadow_start;
 551
 552         shadow_start = (unsigned long)kasan_mem_to_shadow(addr);
 553         scaled_size = (size + KASAN_SHADOW_MASK) >> KASAN_SHADOW_SCALE_SHIFT;
 554         shadow_size = round_up(scaled_size, PAGE_SIZE);
 555
 556         if (WARN_ON(!PAGE_ALIGNED(shadow_start)))
 557                 return -EINVAL;
 558
 559         ret = __vmalloc_node_range(shadow_size, 1, shadow_start,
 560                         shadow_start + shadow_size,
 561                         GFP_KERNEL,
 562                         PAGE_KERNEL, VM_NO_GUARD, NUMA_NO_NODE,
 563                         __builtin_return_address(0));
 564
 565         if (ret) {
 566                 __memset(ret, KASAN_SHADOW_INIT, shadow_size);
 567                 find_vm_area(addr)->flags |= VM_KASAN;
 568                 kmemleak_ignore(ret);
 569                 return 0;
 570         }
 571
 572         return -ENOMEM;
 573 }
 574
 575 void kasan_free_shadow(const struct vm_struct *vm)
 576 {
 577         if (vm->flags & VM_KASAN)
 578                 vfree(kasan_mem_to_shadow(vm->addr));
 579 }
 580 #endif
 581
 582 #ifdef CONFIG_MEMORY_HOTPLUG
 583 static bool shadow_mapped(unsigned long addr)
 584 {
 585         pgd_t *pgd = pgd_offset_k(addr);
 586         p4d_t *p4d;
 587         pud_t *pud;
 588         pmd_t *pmd;
 589         pte_t *pte;
 590
 591         if (pgd_none(*pgd))
 592                 return false;
 593         p4d = p4d_offset(pgd, addr);
 594         if (p4d_none(*p4d))
 595                 return false;
 596         pud = pud_offset(p4d, addr);
 597         if (pud_none(*pud))
 598                 return false;
 599
 600         /*
 601          * We can't use pud_large() or pud_huge(), the first one is
 602          * arch-specific, the last one depends on HUGETLB_PAGE.  So let's abuse
 603          * pud_bad(), if pud is bad then it's bad because it's huge.
 604          */
 605         if (pud_bad(*pud))
 606                 return true;
 607         pmd = pmd_offset(pud, addr);
 608         if (pmd_none(*pmd))
 609                 return false;
 610
 611         if (pmd_bad(*pmd))
 612                 return true;
 613         pte = pte_offset_kernel(pmd, addr);
 614         return !pte_none(*pte);
 615 }
 616
 617 static int __meminit kasan_mem_notifier(struct notifier_block *nb,
 618                         unsigned long action, void *data)
 619 {
 620         struct memory_notify *mem_data = data;
 621         unsigned long nr_shadow_pages, start_kaddr, shadow_start;
 622         unsigned long shadow_end, shadow_size;
 623
 624         nr_shadow_pages = mem_data->nr_pages >> KASAN_SHADOW_SCALE_SHIFT;
 625         start_kaddr = (unsigned long)pfn_to_kaddr(mem_data->start_pfn);
 626         shadow_start = (unsigned long)kasan_mem_to_shadow((void *)start_kaddr);
 627         shadow_size = nr_shadow_pages << PAGE_SHIFT;
 628         shadow_end = shadow_start + shadow_size;
 629
 630         if (WARN_ON(mem_data->nr_pages % KASAN_SHADOW_SCALE_SIZE) ||
 631                 WARN_ON(start_kaddr % (KASAN_SHADOW_SCALE_SIZE << PAGE_SHIFT)))
 632                 return NOTIFY_BAD;
 633
 634         switch (action) {
 635         case MEM_GOING_ONLINE: {
 636                 void *ret;
 637
 638                 /*
 639                  * If shadow is mapped already than it must have been mapped
 640                  * during the boot. This could happen if we onlining previously
 641                  * offlined memory.
 642                  */
 643                 if (shadow_mapped(shadow_start))
 644                         return NOTIFY_OK;
 645
 646                 ret = __vmalloc_node_range(shadow_size, PAGE_SIZE, shadow_start,
 647                                         shadow_end, GFP_KERNEL,
 648                                         PAGE_KERNEL, VM_NO_GUARD,
 649                                         pfn_to_nid(mem_data->start_pfn),
 650                                         __builtin_return_address(0));
 651                 if (!ret)
 652                         return NOTIFY_BAD;
 653
 654                 kmemleak_ignore(ret);
 655                 return NOTIFY_OK;
 656         }
 657         case MEM_CANCEL_ONLINE:
 658         case MEM_OFFLINE: {
 659                 struct vm_struct *vm;
 660
 661                 /*
 662                  * shadow_start was either mapped during boot by kasan_init()
 663                  * or during memory online by __vmalloc_node_range().
 664                  * In the latter case we can use vfree() to free shadow.
 665                  * Non-NULL result of the find_vm_area() will tell us if
 666                  * that was the second case.
 667                  *
 668                  * Currently it's not possible to free shadow mapped
 669                  * during boot by kasan_init(). It's because the code
 670                  * to do that hasn't been written yet. So we'll just
 671                  * leak the memory.
 672                  */
 673                 vm = find_vm_area((void *)shadow_start);
 674                 if (vm)
 675                         vfree((void *)shadow_start);
 676         }
 677         }
 678
 679         return NOTIFY_OK;
 680 }
 681
 682 static int __init kasan_memhotplug_init(void)
 683 {
 684         hotplug_memory_notifier(kasan_mem_notifier, 0);
 685
 686         return 0;
 687 }
 688
 689 core_initcall(kasan_memhotplug_init);
 690 #endif
 691
 692 #ifdef CONFIG_KASAN_VMALLOC
 693 static int kasan_populate_vmalloc_pte(pte_t *ptep, unsigned long addr,
 694                                       void *unused)
 695 {
 696         unsigned long page;
 697         pte_t pte;
 698
 699         if (likely(!pte_none(*ptep)))
 700                 return 0;
 701
 702         page = __get_free_page(GFP_KERNEL);
 703         if (!page)
 704                 return -ENOMEM;
 705
 706         memset((void *)page, KASAN_VMALLOC_INVALID, PAGE_SIZE);
 707         pte = pfn_pte(PFN_DOWN(__pa(page)), PAGE_KERNEL);
 708
 709         spin_lock(&init_mm.page_table_lock);
 710         if (likely(pte_none(*ptep))) {
 711                 set_pte_at(&init_mm, addr, ptep, pte);
 712                 page = 0;
 713         }
 714         spin_unlock(&init_mm.page_table_lock);
 715         if (page)
 716                 free_page(page);
 717         return 0;
 718 }
 719
 720 int kasan_populate_vmalloc(unsigned long addr, unsigned long size)
 721 {
 722         unsigned long shadow_start, shadow_end;
 723         int ret;
 724
 725         if (!is_vmalloc_or_module_addr((void *)addr))
 726                 return 0;
 727
 728         shadow_start = (unsigned long)kasan_mem_to_shadow((void *)addr);
 729         shadow_start = ALIGN_DOWN(shadow_start, PAGE_SIZE);
 730         shadow_end = (unsigned long)kasan_mem_to_shadow((void *)addr + size);
 731         shadow_end = ALIGN(shadow_end, PAGE_SIZE);
 732
 733         ret = apply_to_page_range(&init_mm, shadow_start,
 734                                   shadow_end - shadow_start,
 735                                   kasan_populate_vmalloc_pte, NULL);
 736         if (ret)
 737                 return ret;
 738
 739         flush_cache_vmap(shadow_start, shadow_end);
 740
 741         /*
 742          * We need to be careful about inter-cpu effects here. Consider:
 743          *
 744          *   CPU#0                                CPU#1
 745          * WRITE_ONCE(p, vmalloc(100));         while (x = READ_ONCE(p)) ;
 746          *                                      p[99] = 1;
 747          *
 748          * With compiler instrumentation, that ends up looking like this:
 749          *
 750          *   CPU#0                                CPU#1
 751          * // vmalloc() allocates memory
 752          * // let a = area->addr
 753          * // we reach kasan_populate_vmalloc
 754          * // and call kasan_unpoison_shadow:
 755          * STORE shadow(a), unpoison_val
 756          * ...
 757          * STORE shadow(a+99), unpoison_val     x = LOAD p
 758          * // rest of vmalloc process           <data dependency>
 759          * STORE p, a                           LOAD shadow(x+99)
 760          *
 761          * If there is no barrier between the end of unpoisioning the shadow
 762          * and the store of the result to p, the stores could be committed
 763          * in a different order by CPU#0, and CPU#1 could erroneously observe
 764          * poison in the shadow.
 765          *
 766          * We need some sort of barrier between the stores.
 767          *
 768          * In the vmalloc() case, this is provided by a smp_wmb() in
 769          * clear_vm_uninitialized_flag(). In the per-cpu allocator and in
 770          * get_vm_area() and friends, the caller gets shadow allocated but
 771          * doesn't have any pages mapped into the virtual address space that
 772          * has been reserved. Mapping those pages in will involve taking and
 773          * releasing a page-table lock, which will provide the barrier.
 774          */
 775
 776         return 0;
 777 }
 778
 779 /*
 780  * Poison the shadow for a vmalloc region. Called as part of the
 781  * freeing process at the time the region is freed.
 782  */
 783 void kasan_poison_vmalloc(const void *start, unsigned long size)
 784 {
 785         if (!is_vmalloc_or_module_addr(start))
 786                 return;
 787
 788         size = round_up(size, KASAN_SHADOW_SCALE_SIZE);
 789         kasan_poison_shadow(start, size, KASAN_VMALLOC_INVALID);
 790 }
 791
 792 void kasan_unpoison_vmalloc(const void *start, unsigned long size)
 793 {
 794         if (!is_vmalloc_or_module_addr(start))
 795                 return;
 796
 797         kasan_unpoison_shadow(start, size);
 798 }
 799
 800 static int kasan_depopulate_vmalloc_pte(pte_t *ptep, unsigned long addr,
 801                                         void *unused)
 802 {
 803         unsigned long page;
 804
 805         page = (unsigned long)__va(pte_pfn(*ptep) << PAGE_SHIFT);
 806
 807         spin_lock(&init_mm.page_table_lock);
 808
 809         if (likely(!pte_none(*ptep))) {
 810                 pte_clear(&init_mm, addr, ptep);
 811                 free_page(page);
 812         }
 813         spin_unlock(&init_mm.page_table_lock);
 814
 815         return 0;
 816 }
 817
 818 /*
 819  * Release the backing for the vmalloc region [start, end), which
 820  * lies within the free region [free_region_start, free_region_end).
 821  *
 822  * This can be run lazily, long after the region was freed. It runs
 823  * under vmap_area_lock, so it's not safe to interact with the vmalloc/vmap
 824  * infrastructure.
 825  *
 826  * How does this work?
 827  * -------------------
 828  *
 829  * We have a region that is page aligned, labelled as A.
 830  * That might not map onto the shadow in a way that is page-aligned:
 831  *
 832  *                    start                     end
 833  *                    v                         v
 834  * |????????|????????|AAAAAAAA|AA....AA|AAAAAAAA|????????| < vmalloc
 835  *  -------- -------- --------          -------- --------
 836  *      |        |       |                 |        |
 837  *      |        |       |         /-------/        |
 838  *      \-------\|/------/         |/---------------/
 839  *              |||                ||
 840  *             |??AAAAAA|AAAAAAAA|AA??????|                < shadow
 841  *                 (1)      (2)      (3)
 842  *
 843  * First we align the start upwards and the end downwards, so that the
 844  * shadow of the region aligns with shadow page boundaries. In the
 845  * example, this gives us the shadow page (2). This is the shadow entirely
 846  * covered by this allocation.
 847  *
 848  * Then we have the tricky bits. We want to know if we can free the
 849  * partially covered shadow pages - (1) and (3) in the example. For this,
 850  * we are given the start and end of the free region that contains this
 851  * allocation. Extending our previous example, we could have:
 852  *
 853  *  free_region_start                                    free_region_end
 854  *  |                 start                     end      |
 855  *  v                 v                         v        v
 856  * |FFFFFFFF|FFFFFFFF|AAAAAAAA|AA....AA|AAAAAAAA|FFFFFFFF| < vmalloc
 857  *  -------- -------- --------          -------- --------
 858  *      |        |       |                 |        |
 859  *      |        |       |         /-------/        |
 860  *      \-------\|/------/         |/---------------/
 861  *              |||                ||
 862  *             |FFAAAAAA|AAAAAAAA|AAF?????|                < shadow
 863  *                 (1)      (2)      (3)
 864  *
 865  * Once again, we align the start of the free region up, and the end of
 866  * the free region down so that the shadow is page aligned. So we can free
 867  * page (1) - we know no allocation currently uses anything in that page,
 868  * because all of it is in the vmalloc free region. But we cannot free
 869  * page (3), because we can't be sure that the rest of it is unused.
 870  *
 871  * We only consider pages that contain part of the original region for
 872  * freeing: we don't try to free other pages from the free region or we'd
 873  * end up trying to free huge chunks of virtual address space.
 874  *
 875  * Concurrency
 876  * -----------
 877  *
 878  * How do we know that we're not freeing a page that is simultaneously
 879  * being used for a fresh allocation in kasan_populate_vmalloc(_pte)?
 880  *
 881  * We _can_ have kasan_release_vmalloc and kasan_populate_vmalloc running
 882  * at the same time. While we run under free_vmap_area_lock, the population
 883  * code does not.
 884  *
 885  * free_vmap_area_lock instead operates to ensure that the larger range
 886  * [free_region_start, free_region_end) is safe: because __alloc_vmap_area and
 887  * the per-cpu region-finding algorithm both run under free_vmap_area_lock,
 888  * no space identified as free will become used while we are running. This
 889  * means that so long as we are careful with alignment and only free shadow
 890  * pages entirely covered by the free region, we will not run in to any
 891  * trouble - any simultaneous allocations will be for disjoint regions.
 892  */
 893 void kasan_release_vmalloc(unsigned long start, unsigned long end,
 894                            unsigned long free_region_start,
 895                            unsigned long free_region_end)
 896 {
 897         void *shadow_start, *shadow_end;
 898         unsigned long region_start, region_end;
 899         unsigned long size;
 900
 901         region_start = ALIGN(start, PAGE_SIZE * KASAN_SHADOW_SCALE_SIZE);
 902         region_end = ALIGN_DOWN(end, PAGE_SIZE * KASAN_SHADOW_SCALE_SIZE);
 903
 904         free_region_start = ALIGN(free_region_start,
 905                                   PAGE_SIZE * KASAN_SHADOW_SCALE_SIZE);
 906
 907         if (start != region_start &&
 908             free_region_start < region_start)
 909                 region_start -= PAGE_SIZE * KASAN_SHADOW_SCALE_SIZE;
 910
 911         free_region_end = ALIGN_DOWN(free_region_end,
 912                                      PAGE_SIZE * KASAN_SHADOW_SCALE_SIZE);
 913
 914         if (end != region_end &&
 915             free_region_end > region_end)
 916                 region_end += PAGE_SIZE * KASAN_SHADOW_SCALE_SIZE;
 917
 918         shadow_start = kasan_mem_to_shadow((void *)region_start);
 919         shadow_end = kasan_mem_to_shadow((void *)region_end);
 920
 921         if (shadow_end > shadow_start) {
 922                 size = shadow_end - shadow_start;
 923                 apply_to_existing_page_range(&init_mm,
 924                                              (unsigned long)shadow_start,
 925                                              size, kasan_depopulate_vmalloc_pte,
 926                                              NULL);
 927                 flush_tlb_kernel_range((unsigned long)shadow_start,
 928                                        (unsigned long)shadow_end);
 929         }
 930 }
 931 #endif