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         if (!kasan_arch_is_ready())
  77                 return;
  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         addr = kasan_reset_tag(addr);
  85
  86         /* Skip KFENCE memory if called explicitly outside of sl*b. */
  87         if (is_kfence_address(addr))
  88                 return;
  89
  90         if (WARN_ON((unsigned long)addr & KASAN_GRANULE_MASK))
  91                 return;
  92         if (WARN_ON(size & KASAN_GRANULE_MASK))
  93                 return;
  94
  95         shadow_start = kasan_mem_to_shadow(addr);
  96         shadow_end = kasan_mem_to_shadow(addr + size);
  97
  98         __memset(shadow_start, value, shadow_end - shadow_start);
  99 }
 100 EXPORT_SYMBOL(kasan_poison);
 101
 102 #ifdef CONFIG_KASAN_GENERIC
 103 void kasan_poison_last_granule(const void *addr, size_t size)
 104 {
 105         if (!kasan_arch_is_ready())
 106                 return;
 107
 108         if (size & KASAN_GRANULE_MASK) {
 109                 u8 *shadow = (u8 *)kasan_mem_to_shadow(addr + size);
 110                 *shadow = size & KASAN_GRANULE_MASK;
 111         }
 112 }
 113 #endif
 114
 115 void kasan_unpoison(const void *addr, size_t size, bool init)
 116 {
 117         u8 tag = get_tag(addr);
 118
 119         /*
 120          * Perform shadow offset calculation based on untagged address, as
 121          * some of the callers (e.g. kasan_unpoison_object_data) pass tagged
 122          * addresses to this function.
 123          */
 124         addr = kasan_reset_tag(addr);
 125
 126         /*
 127          * Skip KFENCE memory if called explicitly outside of sl*b. Also note
 128          * that calls to ksize(), where size is not a multiple of machine-word
 129          * size, would otherwise poison the invalid portion of the word.
 130          */
 131         if (is_kfence_address(addr))
 132                 return;
 133
 134         if (WARN_ON((unsigned long)addr & KASAN_GRANULE_MASK))
 135                 return;
 136
 137         /* Unpoison all granules that cover the object. */
 138         kasan_poison(addr, round_up(size, KASAN_GRANULE_SIZE), tag, false);
 139
 140         /* Partially poison the last granule for the generic mode. */
 141         if (IS_ENABLED(CONFIG_KASAN_GENERIC))
 142                 kasan_poison_last_granule(addr, size);
 143 }
 144
 145 #ifdef CONFIG_MEMORY_HOTPLUG
 146 static bool shadow_mapped(unsigned long addr)
 147 {
 148         pgd_t *pgd = pgd_offset_k(addr);
 149         p4d_t *p4d;
 150         pud_t *pud;
 151         pmd_t *pmd;
 152         pte_t *pte;
 153
 154         if (pgd_none(*pgd))
 155                 return false;
 156         p4d = p4d_offset(pgd, addr);
 157         if (p4d_none(*p4d))
 158                 return false;
 159         pud = pud_offset(p4d, addr);
 160         if (pud_none(*pud))
 161                 return false;
 162
 163         /*
 164          * We can't use pud_large() or pud_huge(), the first one is
 165          * arch-specific, the last one depends on HUGETLB_PAGE.  So let's abuse
 166          * pud_bad(), if pud is bad then it's bad because it's huge.
 167          */
 168         if (pud_bad(*pud))
 169                 return true;
 170         pmd = pmd_offset(pud, addr);
 171         if (pmd_none(*pmd))
 172                 return false;
 173
 174         if (pmd_bad(*pmd))
 175                 return true;
 176         pte = pte_offset_kernel(pmd, addr);
 177         return !pte_none(*pte);
 178 }
 179
 180 static int __meminit kasan_mem_notifier(struct notifier_block *nb,
 181                         unsigned long action, void *data)
 182 {
 183         struct memory_notify *mem_data = data;
 184         unsigned long nr_shadow_pages, start_kaddr, shadow_start;
 185         unsigned long shadow_end, shadow_size;
 186
 187         nr_shadow_pages = mem_data->nr_pages >> KASAN_SHADOW_SCALE_SHIFT;
 188         start_kaddr = (unsigned long)pfn_to_kaddr(mem_data->start_pfn);
 189         shadow_start = (unsigned long)kasan_mem_to_shadow((void *)start_kaddr);
 190         shadow_size = nr_shadow_pages << PAGE_SHIFT;
 191         shadow_end = shadow_start + shadow_size;
 192
 193         if (WARN_ON(mem_data->nr_pages % KASAN_GRANULE_SIZE) ||
 194                 WARN_ON(start_kaddr % KASAN_MEMORY_PER_SHADOW_PAGE))
 195                 return NOTIFY_BAD;
 196
 197         switch (action) {
 198         case MEM_GOING_ONLINE: {
 199                 void *ret;
 200
 201                 /*
 202                  * If shadow is mapped already than it must have been mapped
 203                  * during the boot. This could happen if we onlining previously
 204                  * offlined memory.
 205                  */
 206                 if (shadow_mapped(shadow_start))
 207                         return NOTIFY_OK;
 208
 209                 ret = __vmalloc_node_range(shadow_size, PAGE_SIZE, shadow_start,
 210                                         shadow_end, GFP_KERNEL,
 211                                         PAGE_KERNEL, VM_NO_GUARD,
 212                                         pfn_to_nid(mem_data->start_pfn),
 213                                         __builtin_return_address(0));
 214                 if (!ret)
 215                         return NOTIFY_BAD;
 216
 217                 kmemleak_ignore(ret);
 218                 return NOTIFY_OK;
 219         }
 220         case MEM_CANCEL_ONLINE:
 221         case MEM_OFFLINE: {
 222                 struct vm_struct *vm;
 223
 224                 /*
 225                  * shadow_start was either mapped during boot by kasan_init()
 226                  * or during memory online by __vmalloc_node_range().
 227                  * In the latter case we can use vfree() to free shadow.
 228                  * Non-NULL result of the find_vm_area() will tell us if
 229                  * that was the second case.
 230                  *
 231                  * Currently it's not possible to free shadow mapped
 232                  * during boot by kasan_init(). It's because the code
 233                  * to do that hasn't been written yet. So we'll just
 234                  * leak the memory.
 235                  */
 236                 vm = find_vm_area((void *)shadow_start);
 237                 if (vm)
 238                         vfree((void *)shadow_start);
 239         }
 240         }
 241
 242         return NOTIFY_OK;
 243 }
 244
 245 static int __init kasan_memhotplug_init(void)
 246 {
 247         hotplug_memory_notifier(kasan_mem_notifier, 0);
 248
 249         return 0;
 250 }
 251
 252 core_initcall(kasan_memhotplug_init);
 253 #endif
 254
 255 #ifdef CONFIG_KASAN_VMALLOC
 256
 257 void __init __weak kasan_populate_early_vm_area_shadow(void *start,
 258                                                        unsigned long size)
 259 {
 260 }
 261
 262 static int kasan_populate_vmalloc_pte(pte_t *ptep, unsigned long addr,
 263                                       void *unused)
 264 {
 265         unsigned long page;
 266         pte_t pte;
 267
 268         if (likely(!pte_none(*ptep)))
 269                 return 0;
 270
 271         page = __get_free_page(GFP_KERNEL);
 272         if (!page)
 273                 return -ENOMEM;
 274
 275         memset((void *)page, KASAN_VMALLOC_INVALID, PAGE_SIZE);
 276         pte = pfn_pte(PFN_DOWN(__pa(page)), PAGE_KERNEL);
 277
 278         spin_lock(&init_mm.page_table_lock);
 279         if (likely(pte_none(*ptep))) {
 280                 set_pte_at(&init_mm, addr, ptep, pte);
 281                 page = 0;
 282         }
 283         spin_unlock(&init_mm.page_table_lock);
 284         if (page)
 285                 free_page(page);
 286         return 0;
 287 }
 288
 289 int kasan_populate_vmalloc(unsigned long addr, unsigned long size)
 290 {
 291         unsigned long shadow_start, shadow_end;
 292         int ret;
 293
 294         if (!is_vmalloc_or_module_addr((void *)addr))
 295                 return 0;
 296
 297         shadow_start = (unsigned long)kasan_mem_to_shadow((void *)addr);
 298         shadow_start = ALIGN_DOWN(shadow_start, PAGE_SIZE);
 299         shadow_end = (unsigned long)kasan_mem_to_shadow((void *)addr + size);
 300         shadow_end = ALIGN(shadow_end, PAGE_SIZE);
 301
 302         ret = apply_to_page_range(&init_mm, shadow_start,
 303                                   shadow_end - shadow_start,
 304                                   kasan_populate_vmalloc_pte, NULL);
 305         if (ret)
 306                 return ret;
 307
 308         flush_cache_vmap(shadow_start, shadow_end);
 309
 310         /*
 311          * We need to be careful about inter-cpu effects here. Consider:
 312          *
 313          *   CPU#0                                CPU#1
 314          * WRITE_ONCE(p, vmalloc(100));         while (x = READ_ONCE(p)) ;
 315          *                                      p[99] = 1;
 316          *
 317          * With compiler instrumentation, that ends up looking like this:
 318          *
 319          *   CPU#0                                CPU#1
 320          * // vmalloc() allocates memory
 321          * // let a = area->addr
 322          * // we reach kasan_populate_vmalloc
 323          * // and call kasan_unpoison:
 324          * STORE shadow(a), unpoison_val
 325          * ...
 326          * STORE shadow(a+99), unpoison_val     x = LOAD p
 327          * // rest of vmalloc process           <data dependency>
 328          * STORE p, a                           LOAD shadow(x+99)
 329          *
 330          * If there is no barrier between the end of unpoisoning the shadow
 331          * and the store of the result to p, the stores could be committed
 332          * in a different order by CPU#0, and CPU#1 could erroneously observe
 333          * poison in the shadow.
 334          *
 335          * We need some sort of barrier between the stores.
 336          *
 337          * In the vmalloc() case, this is provided by a smp_wmb() in
 338          * clear_vm_uninitialized_flag(). In the per-cpu allocator and in
 339          * get_vm_area() and friends, the caller gets shadow allocated but
 340          * doesn't have any pages mapped into the virtual address space that
 341          * has been reserved. Mapping those pages in will involve taking and
 342          * releasing a page-table lock, which will provide the barrier.
 343          */
 344
 345         return 0;
 346 }
 347
 348 static int kasan_depopulate_vmalloc_pte(pte_t *ptep, unsigned long addr,
 349                                         void *unused)
 350 {
 351         unsigned long page;
 352
 353         page = (unsigned long)__va(pte_pfn(*ptep) << PAGE_SHIFT);
 354
 355         spin_lock(&init_mm.page_table_lock);
 356
 357         if (likely(!pte_none(*ptep))) {
 358                 pte_clear(&init_mm, addr, ptep);
 359                 free_page(page);
 360         }
 361         spin_unlock(&init_mm.page_table_lock);
 362
 363         return 0;
 364 }
 365
 366 /*
 367  * Release the backing for the vmalloc region [start, end), which
 368  * lies within the free region [free_region_start, free_region_end).
 369  *
 370  * This can be run lazily, long after the region was freed. It runs
 371  * under vmap_area_lock, so it's not safe to interact with the vmalloc/vmap
 372  * infrastructure.
 373  *
 374  * How does this work?
 375  * -------------------
 376  *
 377  * We have a region that is page aligned, labeled as A.
 378  * That might not map onto the shadow in a way that is page-aligned:
 379  *
 380  *                    start                     end
 381  *                    v                         v
 382  * |????????|????????|AAAAAAAA|AA....AA|AAAAAAAA|????????| < vmalloc
 383  *  -------- -------- --------          -------- --------
 384  *      |        |       |                 |        |
 385  *      |        |       |         /-------/        |
 386  *      \-------\|/------/         |/---------------/
 387  *              |||                ||
 388  *             |??AAAAAA|AAAAAAAA|AA??????|                < shadow
 389  *                 (1)      (2)      (3)
 390  *
 391  * First we align the start upwards and the end downwards, so that the
 392  * shadow of the region aligns with shadow page boundaries. In the
 393  * example, this gives us the shadow page (2). This is the shadow entirely
 394  * covered by this allocation.
 395  *
 396  * Then we have the tricky bits. We want to know if we can free the
 397  * partially covered shadow pages - (1) and (3) in the example. For this,
 398  * we are given the start and end of the free region that contains this
 399  * allocation. Extending our previous example, we could have:
 400  *
 401  *  free_region_start                                    free_region_end
 402  *  |                 start                     end      |
 403  *  v                 v                         v        v
 404  * |FFFFFFFF|FFFFFFFF|AAAAAAAA|AA....AA|AAAAAAAA|FFFFFFFF| < vmalloc
 405  *  -------- -------- --------          -------- --------
 406  *      |        |       |                 |        |
 407  *      |        |       |         /-------/        |
 408  *      \-------\|/------/         |/---------------/
 409  *              |||                ||
 410  *             |FFAAAAAA|AAAAAAAA|AAF?????|                < shadow
 411  *                 (1)      (2)      (3)
 412  *
 413  * Once again, we align the start of the free region up, and the end of
 414  * the free region down so that the shadow is page aligned. So we can free
 415  * page (1) - we know no allocation currently uses anything in that page,
 416  * because all of it is in the vmalloc free region. But we cannot free
 417  * page (3), because we can't be sure that the rest of it is unused.
 418  *
 419  * We only consider pages that contain part of the original region for
 420  * freeing: we don't try to free other pages from the free region or we'd
 421  * end up trying to free huge chunks of virtual address space.
 422  *
 423  * Concurrency
 424  * -----------
 425  *
 426  * How do we know that we're not freeing a page that is simultaneously
 427  * being used for a fresh allocation in kasan_populate_vmalloc(_pte)?
 428  *
 429  * We _can_ have kasan_release_vmalloc and kasan_populate_vmalloc running
 430  * at the same time. While we run under free_vmap_area_lock, the population
 431  * code does not.
 432  *
 433  * free_vmap_area_lock instead operates to ensure that the larger range
 434  * [free_region_start, free_region_end) is safe: because __alloc_vmap_area and
 435  * the per-cpu region-finding algorithm both run under free_vmap_area_lock,
 436  * no space identified as free will become used while we are running. This
 437  * means that so long as we are careful with alignment and only free shadow
 438  * pages entirely covered by the free region, we will not run in to any
 439  * trouble - any simultaneous allocations will be for disjoint regions.
 440  */
 441 void kasan_release_vmalloc(unsigned long start, unsigned long end,
 442                            unsigned long free_region_start,
 443                            unsigned long free_region_end)
 444 {
 445         void *shadow_start, *shadow_end;
 446         unsigned long region_start, region_end;
 447         unsigned long size;
 448
 449         region_start = ALIGN(start, KASAN_MEMORY_PER_SHADOW_PAGE);
 450         region_end = ALIGN_DOWN(end, KASAN_MEMORY_PER_SHADOW_PAGE);
 451
 452         free_region_start = ALIGN(free_region_start, KASAN_MEMORY_PER_SHADOW_PAGE);
 453
 454         if (start != region_start &&
 455             free_region_start < region_start)
 456                 region_start -= KASAN_MEMORY_PER_SHADOW_PAGE;
 457
 458         free_region_end = ALIGN_DOWN(free_region_end, KASAN_MEMORY_PER_SHADOW_PAGE);
 459
 460         if (end != region_end &&
 461             free_region_end > region_end)
 462                 region_end += KASAN_MEMORY_PER_SHADOW_PAGE;
 463
 464         shadow_start = kasan_mem_to_shadow((void *)region_start);
 465         shadow_end = kasan_mem_to_shadow((void *)region_end);
 466
 467         if (shadow_end > shadow_start) {
 468                 size = shadow_end - shadow_start;
 469                 apply_to_existing_page_range(&init_mm,
 470                                              (unsigned long)shadow_start,
 471                                              size, kasan_depopulate_vmalloc_pte,
 472                                              NULL);
 473                 flush_tlb_kernel_range((unsigned long)shadow_start,
 474                                        (unsigned long)shadow_end);
 475         }
 476 }
 477
 478 void *__kasan_unpoison_vmalloc(const void *start, unsigned long size,
 479                                kasan_vmalloc_flags_t flags)
 480 {
 481         /*
 482          * Software KASAN modes unpoison both VM_ALLOC and non-VM_ALLOC
 483          * mappings, so the KASAN_VMALLOC_VM_ALLOC flag is ignored.
 484          * Software KASAN modes can't optimize zeroing memory by combining it
 485          * with setting memory tags, so the KASAN_VMALLOC_INIT flag is ignored.
 486          */
 487
 488         if (!is_vmalloc_or_module_addr(start))
 489                 return (void *)start;
 490
 491         /*
 492          * Don't tag executable memory with the tag-based mode.
 493          * The kernel doesn't tolerate having the PC register tagged.
 494          */
 495         if (IS_ENABLED(CONFIG_KASAN_SW_TAGS) &&
 496             !(flags & KASAN_VMALLOC_PROT_NORMAL))
 497                 return (void *)start;
 498
 499         start = set_tag(start, kasan_random_tag());
 500         kasan_unpoison(start, size, false);
 501         return (void *)start;
 502 }
 503
 504 /*
 505  * Poison the shadow for a vmalloc region. Called as part of the
 506  * freeing process at the time the region is freed.
 507  */
 508 void __kasan_poison_vmalloc(const void *start, unsigned long size)
 509 {
 510         if (!is_vmalloc_or_module_addr(start))
 511                 return;
 512
 513         size = round_up(size, KASAN_GRANULE_SIZE);
 514         kasan_poison(start, size, KASAN_VMALLOC_INVALID, false);
 515 }
 516
 517 #else /* CONFIG_KASAN_VMALLOC */
 518
 519 int kasan_alloc_module_shadow(void *addr, size_t size, gfp_t gfp_mask)
 520 {
 521         void *ret;
 522         size_t scaled_size;
 523         size_t shadow_size;
 524         unsigned long shadow_start;
 525
 526         shadow_start = (unsigned long)kasan_mem_to_shadow(addr);
 527         scaled_size = (size + KASAN_GRANULE_SIZE - 1) >>
 528                                 KASAN_SHADOW_SCALE_SHIFT;
 529         shadow_size = round_up(scaled_size, PAGE_SIZE);
 530
 531         if (WARN_ON(!PAGE_ALIGNED(shadow_start)))
 532                 return -EINVAL;
 533
 534         ret = __vmalloc_node_range(shadow_size, 1, shadow_start,
 535                         shadow_start + shadow_size,
 536                         GFP_KERNEL,
 537                         PAGE_KERNEL, VM_NO_GUARD, NUMA_NO_NODE,
 538                         __builtin_return_address(0));
 539
 540         if (ret) {
 541                 struct vm_struct *vm = find_vm_area(addr);
 542                 __memset(ret, KASAN_SHADOW_INIT, shadow_size);
 543                 vm->flags |= VM_KASAN;
 544                 kmemleak_ignore(ret);
 545
 546                 if (vm->flags & VM_DEFER_KMEMLEAK)
 547                         kmemleak_vmalloc(vm, size, gfp_mask);
 548
 549                 return 0;
 550         }
 551
 552         return -ENOMEM;
 553 }
 554
 555 void kasan_free_module_shadow(const struct vm_struct *vm)
 556 {
 557         if (vm->flags & VM_KASAN)
 558                 vfree(kasan_mem_to_shadow(vm->addr));
 559 }
 560
 561 #endif