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