mm/kasan/shadow.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * This file contains KASAN runtime code that manages shadow memory for
   4  * generic and software tag-based KASAN modes.
   5  *
   6  * Copyright (c) 2014 Samsung Electronics Co., Ltd.
   7  * Author: Andrey Ryabinin <ryabinin.a.a@gmail.com>
   8  *
   9  * Some code borrowed from https://github.com/xairy/kasan-prototype by
  10  *        Andrey Konovalov <andreyknvl@gmail.com>
  11  */
  12
  13 #include <linux/init.h>
  14 #include <linux/kasan.h>
  15 #include <linux/kernel.h>
  16 #include <linux/kfence.h>
  17 #include <linux/kmemleak.h>
  18 #include <linux/memory.h>
  19 #include <linux/mm.h>
  20 #include <linux/string.h>
  21 #include <linux/types.h>
  22 #include <linux/vmalloc.h>
  23
  24 #include <asm/cacheflush.h>
  25 #include <asm/tlbflush.h>
  26
  27 #include "kasan.h"
  28
  29 bool __kasan_check_read(const volatile void *p, unsigned int size)
  30 {
  31         return kasan_check_range((unsigned long)p, size, false, _RET_IP_);
  32 }
  33 EXPORT_SYMBOL(__kasan_check_read);
  34
  35 bool __kasan_check_write(const volatile void *p, unsigned int size)
  36 {
  37         return kasan_check_range((unsigned long)p, size, true, _RET_IP_);
  38 }
  39 EXPORT_SYMBOL(__kasan_check_write);
  40
  41 #undef memset
  42 void *memset(void *addr, int c, size_t len)
  43 {
  44         if (!kasan_check_range((unsigned long)addr, len, true, _RET_IP_))
  45                 return NULL;
  46
  47         return __memset(addr, c, len);
  48 }
  49
  50 #ifdef __HAVE_ARCH_MEMMOVE
  51 #undef memmove
  52 void *memmove(void *dest, const void *src, size_t len)
  53 {
  54         if (!kasan_check_range((unsigned long)src, len, false, _RET_IP_) ||
  55             !kasan_check_range((unsigned long)dest, len, true, _RET_IP_))
  56                 return NULL;
  57
  58         return __memmove(dest, src, len);
  59 }
  60 #endif
  61
  62 #undef memcpy
  63 void *memcpy(void *dest, const void *src, size_t len)
  64 {
  65         if (!kasan_check_range((unsigned long)src, len, false, _RET_IP_) ||
  66             !kasan_check_range((unsigned long)dest, len, true, _RET_IP_))
  67                 return NULL;
  68
  69         return __memcpy(dest, src, len);
  70 }
  71
  72 void kasan_poison(const void *addr, size_t size, u8 value, bool init)
  73 {
  74         void *shadow_start, *shadow_end;
  75
  76         /*
  77          * Perform shadow offset calculation based on untagged address, as
  78          * some of the callers (e.g. kasan_poison_object_data) pass tagged
  79          * addresses to this function.
  80          */
  81         addr = kasan_reset_tag(addr);
  82
  83         /* Skip KFENCE memory if called explicitly outside of sl*b. */
  84         if (is_kfence_address(addr))
  85                 return;
  86
  87         if (WARN_ON((unsigned long)addr & KASAN_GRANULE_MASK))
  88                 return;
  89         if (WARN_ON(size & KASAN_GRANULE_MASK))
  90                 return;
  91
  92         shadow_start = kasan_mem_to_shadow(addr);
  93         shadow_end = kasan_mem_to_shadow(addr + size);
  94
  95         __memset(shadow_start, value, shadow_end - shadow_start);
  96 }
  97 EXPORT_SYMBOL(kasan_poison);
  98
  99 #ifdef CONFIG_KASAN_GENERIC
 100 void kasan_poison_last_granule(const void *addr, size_t size)
 101 {
 102         if (size & KASAN_GRANULE_MASK) {
 103                 u8 *shadow = (u8 *)kasan_mem_to_shadow(addr + size);
 104                 *shadow = size & KASAN_GRANULE_MASK;
 105         }
 106 }
 107 #endif
 108
 109 void kasan_unpoison(const void *addr, size_t size, bool init)
 110 {
 111         u8 tag = get_tag(addr);
 112
 113         /*
 114          * Perform shadow offset calculation based on untagged address, as
 115          * some of the callers (e.g. kasan_unpoison_object_data) pass tagged
 116          * addresses to this function.
 117          */
 118         addr = kasan_reset_tag(addr);
 119
 120         /*
 121          * Skip KFENCE memory if called explicitly outside of sl*b. Also note
 122          * that calls to ksize(), where size is not a multiple of machine-word
 123          * size, would otherwise poison the invalid portion of the word.
 124          */
 125         if (is_kfence_address(addr))
 126                 return;
 127
 128         if (WARN_ON((unsigned long)addr & KASAN_GRANULE_MASK))
 129                 return;
 130
 131         /* Unpoison all granules that cover the object. */
 132         kasan_poison(addr, round_up(size, KASAN_GRANULE_SIZE), tag, false);
 133
 134         /* Partially poison the last granule for the generic mode. */
 135         if (IS_ENABLED(CONFIG_KASAN_GENERIC))
 136                 kasan_poison_last_granule(addr, size);
 137 }
 138
 139 #ifdef CONFIG_MEMORY_HOTPLUG
 140 static bool shadow_mapped(unsigned long addr)
 141 {
 142         pgd_t *pgd = pgd_offset_k(addr);
 143         p4d_t *p4d;
 144         pud_t *pud;
 145         pmd_t *pmd;
 146         pte_t *pte;
 147
 148         if (pgd_none(*pgd))
 149                 return false;
 150         p4d = p4d_offset(pgd, addr);
 151         if (p4d_none(*p4d))
 152                 return false;
 153         pud = pud_offset(p4d, addr);
 154         if (pud_none(*pud))
 155                 return false;
 156
 157         /*
 158          * We can't use pud_large() or pud_huge(), the first one is
 159          * arch-specific, the last one depends on HUGETLB_PAGE.  So let's abuse
 160          * pud_bad(), if pud is bad then it's bad because it's huge.
 161          */
 162         if (pud_bad(*pud))
 163                 return true;
 164         pmd = pmd_offset(pud, addr);
 165         if (pmd_none(*pmd))
 166                 return false;
 167
 168         if (pmd_bad(*pmd))
 169                 return true;
 170         pte = pte_offset_kernel(pmd, addr);
 171         return !pte_none(*pte);
 172 }
 173
 174 static int __meminit kasan_mem_notifier(struct notifier_block *nb,
 175                         unsigned long action, void *data)
 176 {
 177         struct memory_notify *mem_data = data;
 178         unsigned long nr_shadow_pages, start_kaddr, shadow_start;
 179         unsigned long shadow_end, shadow_size;
 180
 181         nr_shadow_pages = mem_data->nr_pages >> KASAN_SHADOW_SCALE_SHIFT;
 182         start_kaddr = (unsigned long)pfn_to_kaddr(mem_data->start_pfn);
 183         shadow_start = (unsigned long)kasan_mem_to_shadow((void *)start_kaddr);
 184         shadow_size = nr_shadow_pages << PAGE_SHIFT;
 185         shadow_end = shadow_start + shadow_size;
 186
 187         if (WARN_ON(mem_data->nr_pages % KASAN_GRANULE_SIZE) ||
 188                 WARN_ON(start_kaddr % KASAN_MEMORY_PER_SHADOW_PAGE))
 189                 return NOTIFY_BAD;
 190
 191         switch (action) {
 192         case MEM_GOING_ONLINE: {
 193                 void *ret;
 194
 195                 /*
 196                  * If shadow is mapped already than it must have been mapped
 197                  * during the boot. This could happen if we onlining previously
 198                  * offlined memory.
 199                  */
 200                 if (shadow_mapped(shadow_start))
 201                         return NOTIFY_OK;
 202
 203                 ret = __vmalloc_node_range(shadow_size, PAGE_SIZE, shadow_start,
 204                                         shadow_end, GFP_KERNEL,
 205                                         PAGE_KERNEL, VM_NO_GUARD,
 206                                         pfn_to_nid(mem_data->start_pfn),
 207                                         __builtin_return_address(0));
 208                 if (!ret)
 209                         return NOTIFY_BAD;
 210
 211                 kmemleak_ignore(ret);
 212                 return NOTIFY_OK;
 213         }
 214         case MEM_CANCEL_ONLINE:
 215         case MEM_OFFLINE: {
 216                 struct vm_struct *vm;
 217
 218                 /*
 219                  * shadow_start was either mapped during boot by kasan_init()
 220                  * or during memory online by __vmalloc_node_range().
 221                  * In the latter case we can use vfree() to free shadow.
 222                  * Non-NULL result of the find_vm_area() will tell us if
 223                  * that was the second case.
 224                  *
 225                  * Currently it's not possible to free shadow mapped
 226                  * during boot by kasan_init(). It's because the code
 227                  * to do that hasn't been written yet. So we'll just
 228                  * leak the memory.
 229                  */
 230                 vm = find_vm_area((void *)shadow_start);
 231                 if (vm)
 232                         vfree((void *)shadow_start);
 233         }
 234         }
 235
 236         return NOTIFY_OK;
 237 }
 238
 239 static int __init kasan_memhotplug_init(void)
 240 {
 241         hotplug_memory_notifier(kasan_mem_notifier, 0);
 242
 243         return 0;
 244 }
 245
 246 core_initcall(kasan_memhotplug_init);
 247 #endif
 248
 249 #ifdef CONFIG_KASAN_VMALLOC
 250
 251 static int kasan_populate_vmalloc_pte(pte_t *ptep, unsigned long addr,
 252                                       void *unused)
 253 {
 254         unsigned long page;
 255         pte_t pte;
 256
 257         if (likely(!pte_none(*ptep)))
 258                 return 0;
 259
 260         page = __get_free_page(GFP_KERNEL);
 261         if (!page)
 262                 return -ENOMEM;
 263
 264         memset((void *)page, KASAN_VMALLOC_INVALID, PAGE_SIZE);
 265         pte = pfn_pte(PFN_DOWN(__pa(page)), PAGE_KERNEL);
 266
 267         spin_lock(&init_mm.page_table_lock);
 268         if (likely(pte_none(*ptep))) {
 269                 set_pte_at(&init_mm, addr, ptep, pte);
 270                 page = 0;
 271         }
 272         spin_unlock(&init_mm.page_table_lock);
 273         if (page)
 274                 free_page(page);
 275         return 0;
 276 }
 277
 278 int kasan_populate_vmalloc(unsigned long addr, unsigned long size)
 279 {
 280         unsigned long shadow_start, shadow_end;
 281         int ret;
 282
 283         if (!is_vmalloc_or_module_addr((void *)addr))
 284                 return 0;
 285
 286         shadow_start = (unsigned long)kasan_mem_to_shadow((void *)addr);
 287         shadow_start = ALIGN_DOWN(shadow_start, PAGE_SIZE);
 288         shadow_end = (unsigned long)kasan_mem_to_shadow((void *)addr + size);
 289         shadow_end = ALIGN(shadow_end, PAGE_SIZE);
 290
 291         ret = apply_to_page_range(&init_mm, shadow_start,
 292                                   shadow_end - shadow_start,
 293                                   kasan_populate_vmalloc_pte, NULL);
 294         if (ret)
 295                 return ret;
 296
 297         flush_cache_vmap(shadow_start, shadow_end);
 298
 299         /*
 300          * We need to be careful about inter-cpu effects here. Consider:
 301          *
 302          *   CPU#0                                CPU#1
 303          * WRITE_ONCE(p, vmalloc(100));         while (x = READ_ONCE(p)) ;
 304          *                                      p[99] = 1;
 305          *
 306          * With compiler instrumentation, that ends up looking like this:
 307          *
 308          *   CPU#0                                CPU#1
 309          * // vmalloc() allocates memory
 310          * // let a = area->addr
 311          * // we reach kasan_populate_vmalloc
 312          * // and call kasan_unpoison:
 313          * STORE shadow(a), unpoison_val
 314          * ...
 315          * STORE shadow(a+99), unpoison_val     x = LOAD p
 316          * // rest of vmalloc process           <data dependency>
 317          * STORE p, a                           LOAD shadow(x+99)
 318          *
 319          * If there is no barrier between the end of unpoisoning the shadow
 320          * and the store of the result to p, the stores could be committed
 321          * in a different order by CPU#0, and CPU#1 could erroneously observe
 322          * poison in the shadow.
 323          *
 324          * We need some sort of barrier between the stores.
 325          *
 326          * In the vmalloc() case, this is provided by a smp_wmb() in
 327          * clear_vm_uninitialized_flag(). In the per-cpu allocator and in
 328          * get_vm_area() and friends, the caller gets shadow allocated but
 329          * doesn't have any pages mapped into the virtual address space that
 330          * has been reserved. Mapping those pages in will involve taking and
 331          * releasing a page-table lock, which will provide the barrier.
 332          */
 333
 334         return 0;
 335 }
 336
 337 /*
 338  * Poison the shadow for a vmalloc region. Called as part of the
 339  * freeing process at the time the region is freed.
 340  */
 341 void kasan_poison_vmalloc(const void *start, unsigned long size)
 342 {
 343         if (!is_vmalloc_or_module_addr(start))
 344                 return;
 345
 346         size = round_up(size, KASAN_GRANULE_SIZE);
 347         kasan_poison(start, size, KASAN_VMALLOC_INVALID, false);
 348 }
 349
 350 void kasan_unpoison_vmalloc(const void *start, unsigned long size)
 351 {
 352         if (!is_vmalloc_or_module_addr(start))
 353                 return;
 354
 355         kasan_unpoison(start, size, false);
 356 }
 357
 358 static int kasan_depopulate_vmalloc_pte(pte_t *ptep, unsigned long addr,
 359                                         void *unused)
 360 {
 361         unsigned long page;
 362
 363         page = (unsigned long)__va(pte_pfn(*ptep) << PAGE_SHIFT);
 364
 365         spin_lock(&init_mm.page_table_lock);
 366
 367         if (likely(!pte_none(*ptep))) {
 368                 pte_clear(&init_mm, addr, ptep);
 369                 free_page(page);
 370         }
 371         spin_unlock(&init_mm.page_table_lock);
 372
 373         return 0;
 374 }
 375
 376 /*
 377  * Release the backing for the vmalloc region [start, end), which
 378  * lies within the free region [free_region_start, free_region_end).
 379  *
 380  * This can be run lazily, long after the region was freed. It runs
 381  * under vmap_area_lock, so it's not safe to interact with the vmalloc/vmap
 382  * infrastructure.
 383  *
 384  * How does this work?
 385  * -------------------
 386  *
 387  * We have a region that is page aligned, labeled as A.
 388  * That might not map onto the shadow in a way that is page-aligned:
 389  *
 390  *                    start                     end
 391  *                    v                         v
 392  * |????????|????????|AAAAAAAA|AA....AA|AAAAAAAA|????????| < vmalloc
 393  *  -------- -------- --------          -------- --------
 394  *      |        |       |                 |        |
 395  *      |        |       |         /-------/        |
 396  *      \-------\|/------/         |/---------------/
 397  *              |||                ||
 398  *             |??AAAAAA|AAAAAAAA|AA??????|                < shadow
 399  *                 (1)      (2)      (3)
 400  *
 401  * First we align the start upwards and the end downwards, so that the
 402  * shadow of the region aligns with shadow page boundaries. In the
 403  * example, this gives us the shadow page (2). This is the shadow entirely
 404  * covered by this allocation.
 405  *
 406  * Then we have the tricky bits. We want to know if we can free the
 407  * partially covered shadow pages - (1) and (3) in the example. For this,
 408  * we are given the start and end of the free region that contains this
 409  * allocation. Extending our previous example, we could have:
 410  *
 411  *  free_region_start                                    free_region_end
 412  *  |                 start                     end      |
 413  *  v                 v                         v        v
 414  * |FFFFFFFF|FFFFFFFF|AAAAAAAA|AA....AA|AAAAAAAA|FFFFFFFF| < vmalloc
 415  *  -------- -------- --------          -------- --------
 416  *      |        |       |                 |        |
 417  *      |        |       |         /-------/        |
 418  *      \-------\|/------/         |/---------------/
 419  *              |||                ||
 420  *             |FFAAAAAA|AAAAAAAA|AAF?????|                < shadow
 421  *                 (1)      (2)      (3)
 422  *
 423  * Once again, we align the start of the free region up, and the end of
 424  * the free region down so that the shadow is page aligned. So we can free
 425  * page (1) - we know no allocation currently uses anything in that page,
 426  * because all of it is in the vmalloc free region. But we cannot free
 427  * page (3), because we can't be sure that the rest of it is unused.
 428  *
 429  * We only consider pages that contain part of the original region for
 430  * freeing: we don't try to free other pages from the free region or we'd
 431  * end up trying to free huge chunks of virtual address space.
 432  *
 433  * Concurrency
 434  * -----------
 435  *
 436  * How do we know that we're not freeing a page that is simultaneously
 437  * being used for a fresh allocation in kasan_populate_vmalloc(_pte)?
 438  *
 439  * We _can_ have kasan_release_vmalloc and kasan_populate_vmalloc running
 440  * at the same time. While we run under free_vmap_area_lock, the population
 441  * code does not.
 442  *
 443  * free_vmap_area_lock instead operates to ensure that the larger range
 444  * [free_region_start, free_region_end) is safe: because __alloc_vmap_area and
 445  * the per-cpu region-finding algorithm both run under free_vmap_area_lock,
 446  * no space identified as free will become used while we are running. This
 447  * means that so long as we are careful with alignment and only free shadow
 448  * pages entirely covered by the free region, we will not run in to any
 449  * trouble - any simultaneous allocations will be for disjoint regions.
 450  */
 451 void kasan_release_vmalloc(unsigned long start, unsigned long end,
 452                            unsigned long free_region_start,
 453                            unsigned long free_region_end)
 454 {
 455         void *shadow_start, *shadow_end;
 456         unsigned long region_start, region_end;
 457         unsigned long size;
 458
 459         region_start = ALIGN(start, KASAN_MEMORY_PER_SHADOW_PAGE);
 460         region_end = ALIGN_DOWN(end, KASAN_MEMORY_PER_SHADOW_PAGE);
 461
 462         free_region_start = ALIGN(free_region_start, KASAN_MEMORY_PER_SHADOW_PAGE);
 463
 464         if (start != region_start &&
 465             free_region_start < region_start)
 466                 region_start -= KASAN_MEMORY_PER_SHADOW_PAGE;
 467
 468         free_region_end = ALIGN_DOWN(free_region_end, KASAN_MEMORY_PER_SHADOW_PAGE);
 469
 470         if (end != region_end &&
 471             free_region_end > region_end)
 472                 region_end += KASAN_MEMORY_PER_SHADOW_PAGE;
 473
 474         shadow_start = kasan_mem_to_shadow((void *)region_start);
 475         shadow_end = kasan_mem_to_shadow((void *)region_end);
 476
 477         if (shadow_end > shadow_start) {
 478                 size = shadow_end - shadow_start;
 479                 apply_to_existing_page_range(&init_mm,
 480                                              (unsigned long)shadow_start,
 481                                              size, kasan_depopulate_vmalloc_pte,
 482                                              NULL);
 483                 flush_tlb_kernel_range((unsigned long)shadow_start,
 484                                        (unsigned long)shadow_end);
 485         }
 486 }
 487
 488 #else /* CONFIG_KASAN_VMALLOC */
 489
 490 int kasan_module_alloc(void *addr, size_t size)
 491 {
 492         void *ret;
 493         size_t scaled_size;
 494         size_t shadow_size;
 495         unsigned long shadow_start;
 496
 497         shadow_start = (unsigned long)kasan_mem_to_shadow(addr);
 498         scaled_size = (size + KASAN_GRANULE_SIZE - 1) >>
 499                                 KASAN_SHADOW_SCALE_SHIFT;
 500         shadow_size = round_up(scaled_size, PAGE_SIZE);
 501
 502         if (WARN_ON(!PAGE_ALIGNED(shadow_start)))
 503                 return -EINVAL;
 504
 505         ret = __vmalloc_node_range(shadow_size, 1, shadow_start,
 506                         shadow_start + shadow_size,
 507                         GFP_KERNEL,
 508                         PAGE_KERNEL, VM_NO_GUARD, NUMA_NO_NODE,
 509                         __builtin_return_address(0));
 510
 511         if (ret) {
 512                 __memset(ret, KASAN_SHADOW_INIT, shadow_size);
 513                 find_vm_area(addr)->flags |= VM_KASAN;
 514                 kmemleak_ignore(ret);
 515                 return 0;
 516         }
 517
 518         return -ENOMEM;
 519 }
 520
 521 void kasan_free_shadow(const struct vm_struct *vm)
 522 {
 523         if (vm->flags & VM_KASAN)
 524                 vfree(kasan_mem_to_shadow(vm->addr));
 525 }
 526
 527 #endif