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