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