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/interrupt.h>
  19 #include <linux/init.h>
  20 #include <linux/kasan.h>
  21 #include <linux/kernel.h>
  22 #include <linux/kmemleak.h>
  23 #include <linux/linkage.h>
  24 #include <linux/memblock.h>
  25 #include <linux/memory.h>
  26 #include <linux/mm.h>
  27 #include <linux/module.h>
  28 #include <linux/printk.h>
  29 #include <linux/sched.h>
  30 #include <linux/sched/task_stack.h>
  31 #include <linux/slab.h>
  32 #include <linux/stacktrace.h>
  33 #include <linux/string.h>
  34 #include <linux/types.h>
  35 #include <linux/vmalloc.h>
  36 #include <linux/bug.h>
  37 #include <linux/uaccess.h>
  38
  39 #include <asm/cacheflush.h>
  40 #include <asm/tlbflush.h>
  41
  42 #include "kasan.h"
  43 #include "../slab.h"
  44
  45 static inline int in_irqentry_text(unsigned long ptr)
  46 {
  47         return (ptr >= (unsigned long)&__irqentry_text_start &&
  48                 ptr < (unsigned long)&__irqentry_text_end) ||
  49                 (ptr >= (unsigned long)&__softirqentry_text_start &&
  50                  ptr < (unsigned long)&__softirqentry_text_end);
  51 }
  52
  53 static inline unsigned int filter_irq_stacks(unsigned long *entries,
  54                                              unsigned int nr_entries)
  55 {
  56         unsigned int i;
  57
  58         for (i = 0; i < nr_entries; i++) {
  59                 if (in_irqentry_text(entries[i])) {
  60                         /* Include the irqentry function into the stack. */
  61                         return i + 1;
  62                 }
  63         }
  64         return nr_entries;
  65 }
  66
  67 static inline depot_stack_handle_t save_stack(gfp_t flags)
  68 {
  69         unsigned long entries[KASAN_STACK_DEPTH];
  70         unsigned int nr_entries;
  71
  72         nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 0);
  73         nr_entries = filter_irq_stacks(entries, nr_entries);
  74         return stack_depot_save(entries, nr_entries, flags);
  75 }
  76
  77 static inline void set_track(struct kasan_track *track, gfp_t flags)
  78 {
  79         track->pid = current->pid;
  80         track->stack = save_stack(flags);
  81 }
  82
  83 void kasan_enable_current(void)
  84 {
  85         current->kasan_depth++;
  86 }
  87
  88 void kasan_disable_current(void)
  89 {
  90         current->kasan_depth--;
  91 }
  92
  93 bool __kasan_check_read(const volatile void *p, unsigned int size)
  94 {
  95         return check_memory_region((unsigned long)p, size, false, _RET_IP_);
  96 }
  97 EXPORT_SYMBOL(__kasan_check_read);
  98
  99 bool __kasan_check_write(const volatile void *p, unsigned int size)
 100 {
 101         return check_memory_region((unsigned long)p, size, true, _RET_IP_);
 102 }
 103 EXPORT_SYMBOL(__kasan_check_write);
 104
 105 #undef memset
 106 void *memset(void *addr, int c, size_t len)
 107 {
 108         check_memory_region((unsigned long)addr, len, true, _RET_IP_);
 109
 110         return __memset(addr, c, len);
 111 }
 112
 113 #ifdef __HAVE_ARCH_MEMMOVE
 114 #undef memmove
 115 void *memmove(void *dest, const void *src, size_t len)
 116 {
 117         check_memory_region((unsigned long)src, len, false, _RET_IP_);
 118         check_memory_region((unsigned long)dest, len, true, _RET_IP_);
 119
 120         return __memmove(dest, src, len);
 121 }
 122 #endif
 123
 124 #undef memcpy
 125 void *memcpy(void *dest, const void *src, size_t len)
 126 {
 127         check_memory_region((unsigned long)src, len, false, _RET_IP_);
 128         check_memory_region((unsigned long)dest, len, true, _RET_IP_);
 129
 130         return __memcpy(dest, src, len);
 131 }
 132
 133 /*
 134  * Poisons the shadow memory for 'size' bytes starting from 'addr'.
 135  * Memory addresses should be aligned to KASAN_SHADOW_SCALE_SIZE.
 136  */
 137 void kasan_poison_shadow(const void *address, size_t size, u8 value)
 138 {
 139         void *shadow_start, *shadow_end;
 140
 141         /*
 142          * Perform shadow offset calculation based on untagged address, as
 143          * some of the callers (e.g. kasan_poison_object_data) pass tagged
 144          * addresses to this function.
 145          */
 146         address = reset_tag(address);
 147
 148         shadow_start = kasan_mem_to_shadow(address);
 149         shadow_end = kasan_mem_to_shadow(address + size);
 150
 151         __memset(shadow_start, value, shadow_end - shadow_start);
 152 }
 153
 154 void kasan_unpoison_shadow(const void *address, size_t size)
 155 {
 156         u8 tag = get_tag(address);
 157
 158         /*
 159          * Perform shadow offset calculation based on untagged address, as
 160          * some of the callers (e.g. kasan_unpoison_object_data) pass tagged
 161          * addresses to this function.
 162          */
 163         address = reset_tag(address);
 164
 165         kasan_poison_shadow(address, size, tag);
 166
 167         if (size & KASAN_SHADOW_MASK) {
 168                 u8 *shadow = (u8 *)kasan_mem_to_shadow(address + size);
 169
 170                 if (IS_ENABLED(CONFIG_KASAN_SW_TAGS))
 171                         *shadow = tag;
 172                 else
 173                         *shadow = size & KASAN_SHADOW_MASK;
 174         }
 175 }
 176
 177 static void __kasan_unpoison_stack(struct task_struct *task, const void *sp)
 178 {
 179         void *base = task_stack_page(task);
 180         size_t size = sp - base;
 181
 182         kasan_unpoison_shadow(base, size);
 183 }
 184
 185 /* Unpoison the entire stack for a task. */
 186 void kasan_unpoison_task_stack(struct task_struct *task)
 187 {
 188         __kasan_unpoison_stack(task, task_stack_page(task) + THREAD_SIZE);
 189 }
 190
 191 /* Unpoison the stack for the current task beyond a watermark sp value. */
 192 asmlinkage void kasan_unpoison_task_stack_below(const void *watermark)
 193 {
 194         /*
 195          * Calculate the task stack base address.  Avoid using 'current'
 196          * because this function is called by early resume code which hasn't
 197          * yet set up the percpu register (%gs).
 198          */
 199         void *base = (void *)((unsigned long)watermark & ~(THREAD_SIZE - 1));
 200
 201         kasan_unpoison_shadow(base, watermark - base);
 202 }
 203
 204 /*
 205  * Clear all poison for the region between the current SP and a provided
 206  * watermark value, as is sometimes required prior to hand-crafted asm function
 207  * returns in the middle of functions.
 208  */
 209 void kasan_unpoison_stack_above_sp_to(const void *watermark)
 210 {
 211         const void *sp = __builtin_frame_address(0);
 212         size_t size = watermark - sp;
 213
 214         if (WARN_ON(sp > watermark))
 215                 return;
 216         kasan_unpoison_shadow(sp, size);
 217 }
 218
 219 void kasan_alloc_pages(struct page *page, unsigned int order)
 220 {
 221         u8 tag;
 222         unsigned long i;
 223
 224         if (unlikely(PageHighMem(page)))
 225                 return;
 226
 227         tag = random_tag();
 228         for (i = 0; i < (1 << order); i++)
 229                 page_kasan_tag_set(page + i, tag);
 230         kasan_unpoison_shadow(page_address(page), PAGE_SIZE << order);
 231 }
 232
 233 void kasan_free_pages(struct page *page, unsigned int order)
 234 {
 235         if (likely(!PageHighMem(page)))
 236                 kasan_poison_shadow(page_address(page),
 237                                 PAGE_SIZE << order,
 238                                 KASAN_FREE_PAGE);
 239 }
 240
 241 /*
 242  * Adaptive redzone policy taken from the userspace AddressSanitizer runtime.
 243  * For larger allocations larger redzones are used.
 244  */
 245 static inline unsigned int optimal_redzone(unsigned int object_size)
 246 {
 247         if (IS_ENABLED(CONFIG_KASAN_SW_TAGS))
 248                 return 0;
 249
 250         return
 251                 object_size <= 64        - 16   ? 16 :
 252                 object_size <= 128       - 32   ? 32 :
 253                 object_size <= 512       - 64   ? 64 :
 254                 object_size <= 4096      - 128  ? 128 :
 255                 object_size <= (1 << 14) - 256  ? 256 :
 256                 object_size <= (1 << 15) - 512  ? 512 :
 257                 object_size <= (1 << 16) - 1024 ? 1024 : 2048;
 258 }
 259
 260 void kasan_cache_create(struct kmem_cache *cache, unsigned int *size,
 261                         slab_flags_t *flags)
 262 {
 263         unsigned int orig_size = *size;
 264         unsigned int redzone_size;
 265         int redzone_adjust;
 266
 267         /* Add alloc meta. */
 268         cache->kasan_info.alloc_meta_offset = *size;
 269         *size += sizeof(struct kasan_alloc_meta);
 270
 271         /* Add free meta. */
 272         if (IS_ENABLED(CONFIG_KASAN_GENERIC) &&
 273             (cache->flags & SLAB_TYPESAFE_BY_RCU || cache->ctor ||
 274              cache->object_size < sizeof(struct kasan_free_meta))) {
 275                 cache->kasan_info.free_meta_offset = *size;
 276                 *size += sizeof(struct kasan_free_meta);
 277         }
 278
 279         redzone_size = optimal_redzone(cache->object_size);
 280         redzone_adjust = redzone_size - (*size - cache->object_size);
 281         if (redzone_adjust > 0)
 282                 *size += redzone_adjust;
 283
 284         *size = min_t(unsigned int, KMALLOC_MAX_SIZE,
 285                         max(*size, cache->object_size + redzone_size));
 286
 287         /*
 288          * If the metadata doesn't fit, don't enable KASAN at all.
 289          */
 290         if (*size <= cache->kasan_info.alloc_meta_offset ||
 291                         *size <= cache->kasan_info.free_meta_offset) {
 292                 cache->kasan_info.alloc_meta_offset = 0;
 293                 cache->kasan_info.free_meta_offset = 0;
 294                 *size = orig_size;
 295                 return;
 296         }
 297
 298         *flags |= SLAB_KASAN;
 299 }
 300
 301 size_t kasan_metadata_size(struct kmem_cache *cache)
 302 {
 303         return (cache->kasan_info.alloc_meta_offset ?
 304                 sizeof(struct kasan_alloc_meta) : 0) +
 305                 (cache->kasan_info.free_meta_offset ?
 306                 sizeof(struct kasan_free_meta) : 0);
 307 }
 308
 309 struct kasan_alloc_meta *get_alloc_info(struct kmem_cache *cache,
 310                                         const void *object)
 311 {
 312         return (void *)object + cache->kasan_info.alloc_meta_offset;
 313 }
 314
 315 struct kasan_free_meta *get_free_info(struct kmem_cache *cache,
 316                                       const void *object)
 317 {
 318         BUILD_BUG_ON(sizeof(struct kasan_free_meta) > 32);
 319         return (void *)object + cache->kasan_info.free_meta_offset;
 320 }
 321
 322
 323 static void kasan_set_free_info(struct kmem_cache *cache,
 324                 void *object, u8 tag)
 325 {
 326         struct kasan_alloc_meta *alloc_meta;
 327         u8 idx = 0;
 328
 329         alloc_meta = get_alloc_info(cache, object);
 330
 331 #ifdef CONFIG_KASAN_SW_TAGS_IDENTIFY
 332         idx = alloc_meta->free_track_idx;
 333         alloc_meta->free_pointer_tag[idx] = tag;
 334         alloc_meta->free_track_idx = (idx + 1) % KASAN_NR_FREE_STACKS;
 335 #endif
 336
 337         set_track(&alloc_meta->free_track[idx], GFP_NOWAIT);
 338 }
 339
 340 void kasan_poison_slab(struct page *page)
 341 {
 342         unsigned long i;
 343
 344         for (i = 0; i < compound_nr(page); i++)
 345                 page_kasan_tag_reset(page + i);
 346         kasan_poison_shadow(page_address(page), page_size(page),
 347                         KASAN_KMALLOC_REDZONE);
 348 }
 349
 350 void kasan_unpoison_object_data(struct kmem_cache *cache, void *object)
 351 {
 352         kasan_unpoison_shadow(object, cache->object_size);
 353 }
 354
 355 void kasan_poison_object_data(struct kmem_cache *cache, void *object)
 356 {
 357         kasan_poison_shadow(object,
 358                         round_up(cache->object_size, KASAN_SHADOW_SCALE_SIZE),
 359                         KASAN_KMALLOC_REDZONE);
 360 }
 361
 362 /*
 363  * This function assigns a tag to an object considering the following:
 364  * 1. A cache might have a constructor, which might save a pointer to a slab
 365  *    object somewhere (e.g. in the object itself). We preassign a tag for
 366  *    each object in caches with constructors during slab creation and reuse
 367  *    the same tag each time a particular object is allocated.
 368  * 2. A cache might be SLAB_TYPESAFE_BY_RCU, which means objects can be
 369  *    accessed after being freed. We preassign tags for objects in these
 370  *    caches as well.
 371  * 3. For SLAB allocator we can't preassign tags randomly since the freelist
 372  *    is stored as an array of indexes instead of a linked list. Assign tags
 373  *    based on objects indexes, so that objects that are next to each other
 374  *    get different tags.
 375  */
 376 static u8 assign_tag(struct kmem_cache *cache, const void *object,
 377                         bool init, bool keep_tag)
 378 {
 379         /*
 380          * 1. When an object is kmalloc()'ed, two hooks are called:
 381          *    kasan_slab_alloc() and kasan_kmalloc(). We assign the
 382          *    tag only in the first one.
 383          * 2. We reuse the same tag for krealloc'ed objects.
 384          */
 385         if (keep_tag)
 386                 return get_tag(object);
 387
 388         /*
 389          * If the cache neither has a constructor nor has SLAB_TYPESAFE_BY_RCU
 390          * set, assign a tag when the object is being allocated (init == false).
 391          */
 392         if (!cache->ctor && !(cache->flags & SLAB_TYPESAFE_BY_RCU))
 393                 return init ? KASAN_TAG_KERNEL : random_tag();
 394
 395         /* For caches that either have a constructor or SLAB_TYPESAFE_BY_RCU: */
 396 #ifdef CONFIG_SLAB
 397         /* For SLAB assign tags based on the object index in the freelist. */
 398         return (u8)obj_to_index(cache, virt_to_page(object), (void *)object);
 399 #else
 400         /*
 401          * For SLUB assign a random tag during slab creation, otherwise reuse
 402          * the already assigned tag.
 403          */
 404         return init ? random_tag() : get_tag(object);
 405 #endif
 406 }
 407
 408 void * __must_check kasan_init_slab_obj(struct kmem_cache *cache,
 409                                                 const void *object)
 410 {
 411         struct kasan_alloc_meta *alloc_info;
 412
 413         if (!(cache->flags & SLAB_KASAN))
 414                 return (void *)object;
 415
 416         alloc_info = get_alloc_info(cache, object);
 417         __memset(alloc_info, 0, sizeof(*alloc_info));
 418
 419         if (IS_ENABLED(CONFIG_KASAN_SW_TAGS))
 420                 object = set_tag(object,
 421                                 assign_tag(cache, object, true, false));
 422
 423         return (void *)object;
 424 }
 425
 426 static inline bool shadow_invalid(u8 tag, s8 shadow_byte)
 427 {
 428         if (IS_ENABLED(CONFIG_KASAN_GENERIC))
 429                 return shadow_byte < 0 ||
 430                         shadow_byte >= KASAN_SHADOW_SCALE_SIZE;
 431
 432         /* else CONFIG_KASAN_SW_TAGS: */
 433         if ((u8)shadow_byte == KASAN_TAG_INVALID)
 434                 return true;
 435         if ((tag != KASAN_TAG_KERNEL) && (tag != (u8)shadow_byte))
 436                 return true;
 437
 438         return false;
 439 }
 440
 441 static bool __kasan_slab_free(struct kmem_cache *cache, void *object,
 442                               unsigned long ip, bool quarantine)
 443 {
 444         s8 shadow_byte;
 445         u8 tag;
 446         void *tagged_object;
 447         unsigned long rounded_up_size;
 448
 449         tag = get_tag(object);
 450         tagged_object = object;
 451         object = reset_tag(object);
 452
 453         if (unlikely(nearest_obj(cache, virt_to_head_page(object), object) !=
 454             object)) {
 455                 kasan_report_invalid_free(tagged_object, ip);
 456                 return true;
 457         }
 458
 459         /* RCU slabs could be legally used after free within the RCU period */
 460         if (unlikely(cache->flags & SLAB_TYPESAFE_BY_RCU))
 461                 return false;
 462
 463         shadow_byte = READ_ONCE(*(s8 *)kasan_mem_to_shadow(object));
 464         if (shadow_invalid(tag, shadow_byte)) {
 465                 kasan_report_invalid_free(tagged_object, ip);
 466                 return true;
 467         }
 468
 469         rounded_up_size = round_up(cache->object_size, KASAN_SHADOW_SCALE_SIZE);
 470         kasan_poison_shadow(object, rounded_up_size, KASAN_KMALLOC_FREE);
 471
 472         if ((IS_ENABLED(CONFIG_KASAN_GENERIC) && !quarantine) ||
 473                         unlikely(!(cache->flags & SLAB_KASAN)))
 474                 return false;
 475
 476         kasan_set_free_info(cache, object, tag);
 477
 478         quarantine_put(get_free_info(cache, object), cache);
 479
 480         return IS_ENABLED(CONFIG_KASAN_GENERIC);
 481 }
 482
 483 bool kasan_slab_free(struct kmem_cache *cache, void *object, unsigned long ip)
 484 {
 485         return __kasan_slab_free(cache, object, ip, true);
 486 }
 487
 488 static void *__kasan_kmalloc(struct kmem_cache *cache, const void *object,
 489                                 size_t size, gfp_t flags, bool keep_tag)
 490 {
 491         unsigned long redzone_start;
 492         unsigned long redzone_end;
 493         u8 tag = 0xff;
 494
 495         if (gfpflags_allow_blocking(flags))
 496                 quarantine_reduce();
 497
 498         if (unlikely(object == NULL))
 499                 return NULL;
 500
 501         redzone_start = round_up((unsigned long)(object + size),
 502                                 KASAN_SHADOW_SCALE_SIZE);
 503         redzone_end = round_up((unsigned long)object + cache->object_size,
 504                                 KASAN_SHADOW_SCALE_SIZE);
 505
 506         if (IS_ENABLED(CONFIG_KASAN_SW_TAGS))
 507                 tag = assign_tag(cache, object, false, keep_tag);
 508
 509         /* Tag is ignored in set_tag without CONFIG_KASAN_SW_TAGS */
 510         kasan_unpoison_shadow(set_tag(object, tag), size);
 511         kasan_poison_shadow((void *)redzone_start, redzone_end - redzone_start,
 512                 KASAN_KMALLOC_REDZONE);
 513
 514         if (cache->flags & SLAB_KASAN)
 515                 set_track(&get_alloc_info(cache, object)->alloc_track, flags);
 516
 517         return set_tag(object, tag);
 518 }
 519
 520 void * __must_check kasan_slab_alloc(struct kmem_cache *cache, void *object,
 521                                         gfp_t flags)
 522 {
 523         return __kasan_kmalloc(cache, object, cache->object_size, flags, false);
 524 }
 525
 526 void * __must_check kasan_kmalloc(struct kmem_cache *cache, const void *object,
 527                                 size_t size, gfp_t flags)
 528 {
 529         return __kasan_kmalloc(cache, object, size, flags, true);
 530 }
 531 EXPORT_SYMBOL(kasan_kmalloc);
 532
 533 void * __must_check kasan_kmalloc_large(const void *ptr, size_t size,
 534                                                 gfp_t flags)
 535 {
 536         struct page *page;
 537         unsigned long redzone_start;
 538         unsigned long redzone_end;
 539
 540         if (gfpflags_allow_blocking(flags))
 541                 quarantine_reduce();
 542
 543         if (unlikely(ptr == NULL))
 544                 return NULL;
 545
 546         page = virt_to_page(ptr);
 547         redzone_start = round_up((unsigned long)(ptr + size),
 548                                 KASAN_SHADOW_SCALE_SIZE);
 549         redzone_end = (unsigned long)ptr + page_size(page);
 550
 551         kasan_unpoison_shadow(ptr, size);
 552         kasan_poison_shadow((void *)redzone_start, redzone_end - redzone_start,
 553                 KASAN_PAGE_REDZONE);
 554
 555         return (void *)ptr;
 556 }
 557
 558 void * __must_check kasan_krealloc(const void *object, size_t size, gfp_t flags)
 559 {
 560         struct page *page;
 561
 562         if (unlikely(object == ZERO_SIZE_PTR))
 563                 return (void *)object;
 564
 565         page = virt_to_head_page(object);
 566
 567         if (unlikely(!PageSlab(page)))
 568                 return kasan_kmalloc_large(object, size, flags);
 569         else
 570                 return __kasan_kmalloc(page->slab_cache, object, size,
 571                                                 flags, true);
 572 }
 573
 574 void kasan_poison_kfree(void *ptr, unsigned long ip)
 575 {
 576         struct page *page;
 577
 578         page = virt_to_head_page(ptr);
 579
 580         if (unlikely(!PageSlab(page))) {
 581                 if (ptr != page_address(page)) {
 582                         kasan_report_invalid_free(ptr, ip);
 583                         return;
 584                 }
 585                 kasan_poison_shadow(ptr, page_size(page), KASAN_FREE_PAGE);
 586         } else {
 587                 __kasan_slab_free(page->slab_cache, ptr, ip, false);
 588         }
 589 }
 590
 591 void kasan_kfree_large(void *ptr, unsigned long ip)
 592 {
 593         if (ptr != page_address(virt_to_head_page(ptr)))
 594                 kasan_report_invalid_free(ptr, ip);
 595         /* The object will be poisoned by page_alloc. */
 596 }
 597
 598 #ifndef CONFIG_KASAN_VMALLOC
 599 int kasan_module_alloc(void *addr, size_t size)
 600 {
 601         void *ret;
 602         size_t scaled_size;
 603         size_t shadow_size;
 604         unsigned long shadow_start;
 605
 606         shadow_start = (unsigned long)kasan_mem_to_shadow(addr);
 607         scaled_size = (size + KASAN_SHADOW_MASK) >> KASAN_SHADOW_SCALE_SHIFT;
 608         shadow_size = round_up(scaled_size, PAGE_SIZE);
 609
 610         if (WARN_ON(!PAGE_ALIGNED(shadow_start)))
 611                 return -EINVAL;
 612
 613         ret = __vmalloc_node_range(shadow_size, 1, shadow_start,
 614                         shadow_start + shadow_size,
 615                         GFP_KERNEL,
 616                         PAGE_KERNEL, VM_NO_GUARD, NUMA_NO_NODE,
 617                         __builtin_return_address(0));
 618
 619         if (ret) {
 620                 __memset(ret, KASAN_SHADOW_INIT, shadow_size);
 621                 find_vm_area(addr)->flags |= VM_KASAN;
 622                 kmemleak_ignore(ret);
 623                 return 0;
 624         }
 625
 626         return -ENOMEM;
 627 }
 628
 629 void kasan_free_shadow(const struct vm_struct *vm)
 630 {
 631         if (vm->flags & VM_KASAN)
 632                 vfree(kasan_mem_to_shadow(vm->addr));
 633 }
 634 #endif
 635
 636 extern void __kasan_report(unsigned long addr, size_t size, bool is_write, unsigned long ip);
 637
 638 void kasan_report(unsigned long addr, size_t size, bool is_write, unsigned long ip)
 639 {
 640         unsigned long flags = user_access_save();
 641         __kasan_report(addr, size, is_write, ip);
 642         user_access_restore(flags);
 643 }
 644
 645 #ifdef CONFIG_MEMORY_HOTPLUG
 646 static bool shadow_mapped(unsigned long addr)
 647 {
 648         pgd_t *pgd = pgd_offset_k(addr);
 649         p4d_t *p4d;
 650         pud_t *pud;
 651         pmd_t *pmd;
 652         pte_t *pte;
 653
 654         if (pgd_none(*pgd))
 655                 return false;
 656         p4d = p4d_offset(pgd, addr);
 657         if (p4d_none(*p4d))
 658                 return false;
 659         pud = pud_offset(p4d, addr);
 660         if (pud_none(*pud))
 661                 return false;
 662
 663         /*
 664          * We can't use pud_large() or pud_huge(), the first one is
 665          * arch-specific, the last one depends on HUGETLB_PAGE.  So let's abuse
 666          * pud_bad(), if pud is bad then it's bad because it's huge.
 667          */
 668         if (pud_bad(*pud))
 669                 return true;
 670         pmd = pmd_offset(pud, addr);
 671         if (pmd_none(*pmd))
 672                 return false;
 673
 674         if (pmd_bad(*pmd))
 675                 return true;
 676         pte = pte_offset_kernel(pmd, addr);
 677         return !pte_none(*pte);
 678 }
 679
 680 static int __meminit kasan_mem_notifier(struct notifier_block *nb,
 681                         unsigned long action, void *data)
 682 {
 683         struct memory_notify *mem_data = data;
 684         unsigned long nr_shadow_pages, start_kaddr, shadow_start;
 685         unsigned long shadow_end, shadow_size;
 686
 687         nr_shadow_pages = mem_data->nr_pages >> KASAN_SHADOW_SCALE_SHIFT;
 688         start_kaddr = (unsigned long)pfn_to_kaddr(mem_data->start_pfn);
 689         shadow_start = (unsigned long)kasan_mem_to_shadow((void *)start_kaddr);
 690         shadow_size = nr_shadow_pages << PAGE_SHIFT;
 691         shadow_end = shadow_start + shadow_size;
 692
 693         if (WARN_ON(mem_data->nr_pages % KASAN_SHADOW_SCALE_SIZE) ||
 694                 WARN_ON(start_kaddr % (KASAN_SHADOW_SCALE_SIZE << PAGE_SHIFT)))
 695                 return NOTIFY_BAD;
 696
 697         switch (action) {
 698         case MEM_GOING_ONLINE: {
 699                 void *ret;
 700
 701                 /*
 702                  * If shadow is mapped already than it must have been mapped
 703                  * during the boot. This could happen if we onlining previously
 704                  * offlined memory.
 705                  */
 706                 if (shadow_mapped(shadow_start))
 707                         return NOTIFY_OK;
 708
 709                 ret = __vmalloc_node_range(shadow_size, PAGE_SIZE, shadow_start,
 710                                         shadow_end, GFP_KERNEL,
 711                                         PAGE_KERNEL, VM_NO_GUARD,
 712                                         pfn_to_nid(mem_data->start_pfn),
 713                                         __builtin_return_address(0));
 714                 if (!ret)
 715                         return NOTIFY_BAD;
 716
 717                 kmemleak_ignore(ret);
 718                 return NOTIFY_OK;
 719         }
 720         case MEM_CANCEL_ONLINE:
 721         case MEM_OFFLINE: {
 722                 struct vm_struct *vm;
 723
 724                 /*
 725                  * shadow_start was either mapped during boot by kasan_init()
 726                  * or during memory online by __vmalloc_node_range().
 727                  * In the latter case we can use vfree() to free shadow.
 728                  * Non-NULL result of the find_vm_area() will tell us if
 729                  * that was the second case.
 730                  *
 731                  * Currently it's not possible to free shadow mapped
 732                  * during boot by kasan_init(). It's because the code
 733                  * to do that hasn't been written yet. So we'll just
 734                  * leak the memory.
 735                  */
 736                 vm = find_vm_area((void *)shadow_start);
 737                 if (vm)
 738                         vfree((void *)shadow_start);
 739         }
 740         }
 741
 742         return NOTIFY_OK;
 743 }
 744
 745 static int __init kasan_memhotplug_init(void)
 746 {
 747         hotplug_memory_notifier(kasan_mem_notifier, 0);
 748
 749         return 0;
 750 }
 751
 752 core_initcall(kasan_memhotplug_init);
 753 #endif
 754
 755 #ifdef CONFIG_KASAN_VMALLOC
 756 static int kasan_populate_vmalloc_pte(pte_t *ptep, unsigned long addr,
 757                                       void *unused)
 758 {
 759         unsigned long page;
 760         pte_t pte;
 761
 762         if (likely(!pte_none(*ptep)))
 763                 return 0;
 764
 765         page = __get_free_page(GFP_KERNEL);
 766         if (!page)
 767                 return -ENOMEM;
 768
 769         memset((void *)page, KASAN_VMALLOC_INVALID, PAGE_SIZE);
 770         pte = pfn_pte(PFN_DOWN(__pa(page)), PAGE_KERNEL);
 771
 772         spin_lock(&init_mm.page_table_lock);
 773         if (likely(pte_none(*ptep))) {
 774                 set_pte_at(&init_mm, addr, ptep, pte);
 775                 page = 0;
 776         }
 777         spin_unlock(&init_mm.page_table_lock);
 778         if (page)
 779                 free_page(page);
 780         return 0;
 781 }
 782
 783 int kasan_populate_vmalloc(unsigned long addr, unsigned long size)
 784 {
 785         unsigned long shadow_start, shadow_end;
 786         int ret;
 787
 788         if (!is_vmalloc_or_module_addr((void *)addr))
 789                 return 0;
 790
 791         shadow_start = (unsigned long)kasan_mem_to_shadow((void *)addr);
 792         shadow_start = ALIGN_DOWN(shadow_start, PAGE_SIZE);
 793         shadow_end = (unsigned long)kasan_mem_to_shadow((void *)addr + size);
 794         shadow_end = ALIGN(shadow_end, PAGE_SIZE);
 795
 796         ret = apply_to_page_range(&init_mm, shadow_start,
 797                                   shadow_end - shadow_start,
 798                                   kasan_populate_vmalloc_pte, NULL);
 799         if (ret)
 800                 return ret;
 801
 802         flush_cache_vmap(shadow_start, shadow_end);
 803
 804         /*
 805          * We need to be careful about inter-cpu effects here. Consider:
 806          *
 807          *   CPU#0                                CPU#1
 808          * WRITE_ONCE(p, vmalloc(100));         while (x = READ_ONCE(p)) ;
 809          *                                      p[99] = 1;
 810          *
 811          * With compiler instrumentation, that ends up looking like this:
 812          *
 813          *   CPU#0                                CPU#1
 814          * // vmalloc() allocates memory
 815          * // let a = area->addr
 816          * // we reach kasan_populate_vmalloc
 817          * // and call kasan_unpoison_shadow:
 818          * STORE shadow(a), unpoison_val
 819          * ...
 820          * STORE shadow(a+99), unpoison_val     x = LOAD p
 821          * // rest of vmalloc process           <data dependency>
 822          * STORE p, a                           LOAD shadow(x+99)
 823          *
 824          * If there is no barrier between the end of unpoisioning the shadow
 825          * and the store of the result to p, the stores could be committed
 826          * in a different order by CPU#0, and CPU#1 could erroneously observe
 827          * poison in the shadow.
 828          *
 829          * We need some sort of barrier between the stores.
 830          *
 831          * In the vmalloc() case, this is provided by a smp_wmb() in
 832          * clear_vm_uninitialized_flag(). In the per-cpu allocator and in
 833          * get_vm_area() and friends, the caller gets shadow allocated but
 834          * doesn't have any pages mapped into the virtual address space that
 835          * has been reserved. Mapping those pages in will involve taking and
 836          * releasing a page-table lock, which will provide the barrier.
 837          */
 838
 839         return 0;
 840 }
 841
 842 /*
 843  * Poison the shadow for a vmalloc region. Called as part of the
 844  * freeing process at the time the region is freed.
 845  */
 846 void kasan_poison_vmalloc(const void *start, unsigned long size)
 847 {
 848         if (!is_vmalloc_or_module_addr(start))
 849                 return;
 850
 851         size = round_up(size, KASAN_SHADOW_SCALE_SIZE);
 852         kasan_poison_shadow(start, size, KASAN_VMALLOC_INVALID);
 853 }
 854
 855 void kasan_unpoison_vmalloc(const void *start, unsigned long size)
 856 {
 857         if (!is_vmalloc_or_module_addr(start))
 858                 return;
 859
 860         kasan_unpoison_shadow(start, size);
 861 }
 862
 863 static int kasan_depopulate_vmalloc_pte(pte_t *ptep, unsigned long addr,
 864                                         void *unused)
 865 {
 866         unsigned long page;
 867
 868         page = (unsigned long)__va(pte_pfn(*ptep) << PAGE_SHIFT);
 869
 870         spin_lock(&init_mm.page_table_lock);
 871
 872         if (likely(!pte_none(*ptep))) {
 873                 pte_clear(&init_mm, addr, ptep);
 874                 free_page(page);
 875         }
 876         spin_unlock(&init_mm.page_table_lock);
 877
 878         return 0;
 879 }
 880
 881 /*
 882  * Release the backing for the vmalloc region [start, end), which
 883  * lies within the free region [free_region_start, free_region_end).
 884  *
 885  * This can be run lazily, long after the region was freed. It runs
 886  * under vmap_area_lock, so it's not safe to interact with the vmalloc/vmap
 887  * infrastructure.
 888  *
 889  * How does this work?
 890  * -------------------
 891  *
 892  * We have a region that is page aligned, labelled as A.
 893  * That might not map onto the shadow in a way that is page-aligned:
 894  *
 895  *                    start                     end
 896  *                    v                         v
 897  * |????????|????????|AAAAAAAA|AA....AA|AAAAAAAA|????????| < vmalloc
 898  *  -------- -------- --------          -------- --------
 899  *      |        |       |                 |        |
 900  *      |        |       |         /-------/        |
 901  *      \-------\|/------/         |/---------------/
 902  *              |||                ||
 903  *             |??AAAAAA|AAAAAAAA|AA??????|                < shadow
 904  *                 (1)      (2)      (3)
 905  *
 906  * First we align the start upwards and the end downwards, so that the
 907  * shadow of the region aligns with shadow page boundaries. In the
 908  * example, this gives us the shadow page (2). This is the shadow entirely
 909  * covered by this allocation.
 910  *
 911  * Then we have the tricky bits. We want to know if we can free the
 912  * partially covered shadow pages - (1) and (3) in the example. For this,
 913  * we are given the start and end of the free region that contains this
 914  * allocation. Extending our previous example, we could have:
 915  *
 916  *  free_region_start                                    free_region_end
 917  *  |                 start                     end      |
 918  *  v                 v                         v        v
 919  * |FFFFFFFF|FFFFFFFF|AAAAAAAA|AA....AA|AAAAAAAA|FFFFFFFF| < vmalloc
 920  *  -------- -------- --------          -------- --------
 921  *      |        |       |                 |        |
 922  *      |        |       |         /-------/        |
 923  *      \-------\|/------/         |/---------------/
 924  *              |||                ||
 925  *             |FFAAAAAA|AAAAAAAA|AAF?????|                < shadow
 926  *                 (1)      (2)      (3)
 927  *
 928  * Once again, we align the start of the free region up, and the end of
 929  * the free region down so that the shadow is page aligned. So we can free
 930  * page (1) - we know no allocation currently uses anything in that page,
 931  * because all of it is in the vmalloc free region. But we cannot free
 932  * page (3), because we can't be sure that the rest of it is unused.
 933  *
 934  * We only consider pages that contain part of the original region for
 935  * freeing: we don't try to free other pages from the free region or we'd
 936  * end up trying to free huge chunks of virtual address space.
 937  *
 938  * Concurrency
 939  * -----------
 940  *
 941  * How do we know that we're not freeing a page that is simultaneously
 942  * being used for a fresh allocation in kasan_populate_vmalloc(_pte)?
 943  *
 944  * We _can_ have kasan_release_vmalloc and kasan_populate_vmalloc running
 945  * at the same time. While we run under free_vmap_area_lock, the population
 946  * code does not.
 947  *
 948  * free_vmap_area_lock instead operates to ensure that the larger range
 949  * [free_region_start, free_region_end) is safe: because __alloc_vmap_area and
 950  * the per-cpu region-finding algorithm both run under free_vmap_area_lock,
 951  * no space identified as free will become used while we are running. This
 952  * means that so long as we are careful with alignment and only free shadow
 953  * pages entirely covered by the free region, we will not run in to any
 954  * trouble - any simultaneous allocations will be for disjoint regions.
 955  */
 956 void kasan_release_vmalloc(unsigned long start, unsigned long end,
 957                            unsigned long free_region_start,
 958                            unsigned long free_region_end)
 959 {
 960         void *shadow_start, *shadow_end;
 961         unsigned long region_start, region_end;
 962         unsigned long size;
 963
 964         region_start = ALIGN(start, PAGE_SIZE * KASAN_SHADOW_SCALE_SIZE);
 965         region_end = ALIGN_DOWN(end, PAGE_SIZE * KASAN_SHADOW_SCALE_SIZE);
 966
 967         free_region_start = ALIGN(free_region_start,
 968                                   PAGE_SIZE * KASAN_SHADOW_SCALE_SIZE);
 969
 970         if (start != region_start &&
 971             free_region_start < region_start)
 972                 region_start -= PAGE_SIZE * KASAN_SHADOW_SCALE_SIZE;
 973
 974         free_region_end = ALIGN_DOWN(free_region_end,
 975                                      PAGE_SIZE * KASAN_SHADOW_SCALE_SIZE);
 976
 977         if (end != region_end &&
 978             free_region_end > region_end)
 979                 region_end += PAGE_SIZE * KASAN_SHADOW_SCALE_SIZE;
 980
 981         shadow_start = kasan_mem_to_shadow((void *)region_start);
 982         shadow_end = kasan_mem_to_shadow((void *)region_end);
 983
 984         if (shadow_end > shadow_start) {
 985                 size = shadow_end - shadow_start;
 986                 apply_to_existing_page_range(&init_mm,
 987                                              (unsigned long)shadow_start,
 988                                              size, kasan_depopulate_vmalloc_pte,
 989                                              NULL);
 990                 flush_tlb_kernel_range((unsigned long)shadow_start,
 991                                        (unsigned long)shadow_end);
 992         }
 993 }
 994 #endif