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