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