mm/kfence/core.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * KFENCE guarded object allocator and fault handling.
   4  *
   5  * Copyright (C) 2020, Google LLC.
   6  */
   7
   8 #define pr_fmt(fmt) "kfence: " fmt
   9
  10 #include <linux/atomic.h>
  11 #include <linux/bug.h>
  12 #include <linux/debugfs.h>
  13 #include <linux/hash.h>
  14 #include <linux/irq_work.h>
  15 #include <linux/jhash.h>
  16 #include <linux/kcsan-checks.h>
  17 #include <linux/kfence.h>
  18 #include <linux/kmemleak.h>
  19 #include <linux/list.h>
  20 #include <linux/lockdep.h>
  21 #include <linux/log2.h>
  22 #include <linux/memblock.h>
  23 #include <linux/moduleparam.h>
  24 #include <linux/notifier.h>
  25 #include <linux/panic_notifier.h>
  26 #include <linux/random.h>
  27 #include <linux/rcupdate.h>
  28 #include <linux/sched/clock.h>
  29 #include <linux/seq_file.h>
  30 #include <linux/slab.h>
  31 #include <linux/spinlock.h>
  32 #include <linux/string.h>
  33
  34 #include <asm/kfence.h>
  35
  36 #include "kfence.h"
  37
  38 /* Disables KFENCE on the first warning assuming an irrecoverable error. */
  39 #define KFENCE_WARN_ON(cond)                                                   \
  40         ({                                                                     \
  41                 const bool __cond = WARN_ON(cond);                             \
  42                 if (unlikely(__cond)) {                                        \
  43                         WRITE_ONCE(kfence_enabled, false);                     \
  44                         disabled_by_warn = true;                               \
  45                 }                                                              \
  46                 __cond;                                                        \
  47         })
  48
  49 /* === Data ================================================================= */
  50
  51 static bool kfence_enabled __read_mostly;
  52 static bool disabled_by_warn __read_mostly;
  53
  54 unsigned long kfence_sample_interval __read_mostly = CONFIG_KFENCE_SAMPLE_INTERVAL;
  55 EXPORT_SYMBOL_GPL(kfence_sample_interval); /* Export for test modules. */
  56
  57 #ifdef MODULE_PARAM_PREFIX
  58 #undef MODULE_PARAM_PREFIX
  59 #endif
  60 #define MODULE_PARAM_PREFIX "kfence."
  61
  62 static int kfence_enable_late(void);
  63 static int param_set_sample_interval(const char *val, const struct kernel_param *kp)
  64 {
  65         unsigned long num;
  66         int ret = kstrtoul(val, 0, &num);
  67
  68         if (ret < 0)
  69                 return ret;
  70
  71         /* Using 0 to indicate KFENCE is disabled. */
  72         if (!num && READ_ONCE(kfence_enabled)) {
  73                 pr_info("disabled\n");
  74                 WRITE_ONCE(kfence_enabled, false);
  75         }
  76
  77         *((unsigned long *)kp->arg) = num;
  78
  79         if (num && !READ_ONCE(kfence_enabled) && system_state != SYSTEM_BOOTING)
  80                 return disabled_by_warn ? -EINVAL : kfence_enable_late();
  81         return 0;
  82 }
  83
  84 static int param_get_sample_interval(char *buffer, const struct kernel_param *kp)
  85 {
  86         if (!READ_ONCE(kfence_enabled))
  87                 return sprintf(buffer, "0\n");
  88
  89         return param_get_ulong(buffer, kp);
  90 }
  91
  92 static const struct kernel_param_ops sample_interval_param_ops = {
  93         .set = param_set_sample_interval,
  94         .get = param_get_sample_interval,
  95 };
  96 module_param_cb(sample_interval, &sample_interval_param_ops, &kfence_sample_interval, 0600);
  97
  98 /* Pool usage% threshold when currently covered allocations are skipped. */
  99 static unsigned long kfence_skip_covered_thresh __read_mostly = 75;
 100 module_param_named(skip_covered_thresh, kfence_skip_covered_thresh, ulong, 0644);
 101
 102 /* If true, use a deferrable timer. */
 103 static bool kfence_deferrable __read_mostly = IS_ENABLED(CONFIG_KFENCE_DEFERRABLE);
 104 module_param_named(deferrable, kfence_deferrable, bool, 0444);
 105
 106 /* If true, check all canary bytes on panic. */
 107 static bool kfence_check_on_panic __read_mostly;
 108 module_param_named(check_on_panic, kfence_check_on_panic, bool, 0444);
 109
 110 /* The pool of pages used for guard pages and objects. */
 111 char *__kfence_pool __read_mostly;
 112 EXPORT_SYMBOL(__kfence_pool); /* Export for test modules. */
 113
 114 /*
 115  * Per-object metadata, with one-to-one mapping of object metadata to
 116  * backing pages (in __kfence_pool).
 117  */
 118 static_assert(CONFIG_KFENCE_NUM_OBJECTS > 0);
 119 struct kfence_metadata *kfence_metadata __read_mostly;
 120
 121 /*
 122  * If kfence_metadata is not NULL, it may be accessed by kfence_shutdown_cache().
 123  * So introduce kfence_metadata_init to initialize metadata, and then make
 124  * kfence_metadata visible after initialization is successful. This prevents
 125  * potential UAF or access to uninitialized metadata.
 126  */
 127 static struct kfence_metadata *kfence_metadata_init __read_mostly;
 128
 129 /* Freelist with available objects. */
 130 static struct list_head kfence_freelist = LIST_HEAD_INIT(kfence_freelist);
 131 static DEFINE_RAW_SPINLOCK(kfence_freelist_lock); /* Lock protecting freelist. */
 132
 133 /*
 134  * The static key to set up a KFENCE allocation; or if static keys are not used
 135  * to gate allocations, to avoid a load and compare if KFENCE is disabled.
 136  */
 137 DEFINE_STATIC_KEY_FALSE(kfence_allocation_key);
 138
 139 /* Gates the allocation, ensuring only one succeeds in a given period. */
 140 atomic_t kfence_allocation_gate = ATOMIC_INIT(1);
 141
 142 /*
 143  * A Counting Bloom filter of allocation coverage: limits currently covered
 144  * allocations of the same source filling up the pool.
 145  *
 146  * Assuming a range of 15%-85% unique allocations in the pool at any point in
 147  * time, the below parameters provide a probablity of 0.02-0.33 for false
 148  * positive hits respectively:
 149  *
 150  *      P(alloc_traces) = (1 - e^(-HNUM * (alloc_traces / SIZE)) ^ HNUM
 151  */
 152 #define ALLOC_COVERED_HNUM      2
 153 #define ALLOC_COVERED_ORDER     (const_ilog2(CONFIG_KFENCE_NUM_OBJECTS) + 2)
 154 #define ALLOC_COVERED_SIZE      (1 << ALLOC_COVERED_ORDER)
 155 #define ALLOC_COVERED_HNEXT(h)  hash_32(h, ALLOC_COVERED_ORDER)
 156 #define ALLOC_COVERED_MASK      (ALLOC_COVERED_SIZE - 1)
 157 static atomic_t alloc_covered[ALLOC_COVERED_SIZE];
 158
 159 /* Stack depth used to determine uniqueness of an allocation. */
 160 #define UNIQUE_ALLOC_STACK_DEPTH ((size_t)8)
 161
 162 /*
 163  * Randomness for stack hashes, making the same collisions across reboots and
 164  * different machines less likely.
 165  */
 166 static u32 stack_hash_seed __ro_after_init;
 167
 168 /* Statistics counters for debugfs. */
 169 enum kfence_counter_id {
 170         KFENCE_COUNTER_ALLOCATED,
 171         KFENCE_COUNTER_ALLOCS,
 172         KFENCE_COUNTER_FREES,
 173         KFENCE_COUNTER_ZOMBIES,
 174         KFENCE_COUNTER_BUGS,
 175         KFENCE_COUNTER_SKIP_INCOMPAT,
 176         KFENCE_COUNTER_SKIP_CAPACITY,
 177         KFENCE_COUNTER_SKIP_COVERED,
 178         KFENCE_COUNTER_COUNT,
 179 };
 180 static atomic_long_t counters[KFENCE_COUNTER_COUNT];
 181 static const char *const counter_names[] = {
 182         [KFENCE_COUNTER_ALLOCATED]      = "currently allocated",
 183         [KFENCE_COUNTER_ALLOCS]         = "total allocations",
 184         [KFENCE_COUNTER_FREES]          = "total frees",
 185         [KFENCE_COUNTER_ZOMBIES]        = "zombie allocations",
 186         [KFENCE_COUNTER_BUGS]           = "total bugs",
 187         [KFENCE_COUNTER_SKIP_INCOMPAT]  = "skipped allocations (incompatible)",
 188         [KFENCE_COUNTER_SKIP_CAPACITY]  = "skipped allocations (capacity)",
 189         [KFENCE_COUNTER_SKIP_COVERED]   = "skipped allocations (covered)",
 190 };
 191 static_assert(ARRAY_SIZE(counter_names) == KFENCE_COUNTER_COUNT);
 192
 193 /* === Internals ============================================================ */
 194
 195 static inline bool should_skip_covered(void)
 196 {
 197         unsigned long thresh = (CONFIG_KFENCE_NUM_OBJECTS * kfence_skip_covered_thresh) / 100;
 198
 199         return atomic_long_read(&counters[KFENCE_COUNTER_ALLOCATED]) > thresh;
 200 }
 201
 202 static u32 get_alloc_stack_hash(unsigned long *stack_entries, size_t num_entries)
 203 {
 204         num_entries = min(num_entries, UNIQUE_ALLOC_STACK_DEPTH);
 205         num_entries = filter_irq_stacks(stack_entries, num_entries);
 206         return jhash(stack_entries, num_entries * sizeof(stack_entries[0]), stack_hash_seed);
 207 }
 208
 209 /*
 210  * Adds (or subtracts) count @val for allocation stack trace hash
 211  * @alloc_stack_hash from Counting Bloom filter.
 212  */
 213 static void alloc_covered_add(u32 alloc_stack_hash, int val)
 214 {
 215         int i;
 216
 217         for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
 218                 atomic_add(val, &alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]);
 219                 alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
 220         }
 221 }
 222
 223 /*
 224  * Returns true if the allocation stack trace hash @alloc_stack_hash is
 225  * currently contained (non-zero count) in Counting Bloom filter.
 226  */
 227 static bool alloc_covered_contains(u32 alloc_stack_hash)
 228 {
 229         int i;
 230
 231         for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
 232                 if (!atomic_read(&alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]))
 233                         return false;
 234                 alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
 235         }
 236
 237         return true;
 238 }
 239
 240 static bool kfence_protect(unsigned long addr)
 241 {
 242         return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), true));
 243 }
 244
 245 static bool kfence_unprotect(unsigned long addr)
 246 {
 247         return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), false));
 248 }
 249
 250 static inline unsigned long metadata_to_pageaddr(const struct kfence_metadata *meta)
 251 {
 252         unsigned long offset = (meta - kfence_metadata + 1) * PAGE_SIZE * 2;
 253         unsigned long pageaddr = (unsigned long)&__kfence_pool[offset];
 254
 255         /* The checks do not affect performance; only called from slow-paths. */
 256
 257         /* Only call with a pointer into kfence_metadata. */
 258         if (KFENCE_WARN_ON(meta < kfence_metadata ||
 259                            meta >= kfence_metadata + CONFIG_KFENCE_NUM_OBJECTS))
 260                 return 0;
 261
 262         /*
 263          * This metadata object only ever maps to 1 page; verify that the stored
 264          * address is in the expected range.
 265          */
 266         if (KFENCE_WARN_ON(ALIGN_DOWN(meta->addr, PAGE_SIZE) != pageaddr))
 267                 return 0;
 268
 269         return pageaddr;
 270 }
 271
 272 /*
 273  * Update the object's metadata state, including updating the alloc/free stacks
 274  * depending on the state transition.
 275  */
 276 static noinline void
 277 metadata_update_state(struct kfence_metadata *meta, enum kfence_object_state next,
 278                       unsigned long *stack_entries, size_t num_stack_entries)
 279 {
 280         struct kfence_track *track =
 281                 next == KFENCE_OBJECT_FREED ? &meta->free_track : &meta->alloc_track;
 282
 283         lockdep_assert_held(&meta->lock);
 284
 285         if (stack_entries) {
 286                 memcpy(track->stack_entries, stack_entries,
 287                        num_stack_entries * sizeof(stack_entries[0]));
 288         } else {
 289                 /*
 290                  * Skip over 1 (this) functions; noinline ensures we do not
 291                  * accidentally skip over the caller by never inlining.
 292                  */
 293                 num_stack_entries = stack_trace_save(track->stack_entries, KFENCE_STACK_DEPTH, 1);
 294         }
 295         track->num_stack_entries = num_stack_entries;
 296         track->pid = task_pid_nr(current);
 297         track->cpu = raw_smp_processor_id();
 298         track->ts_nsec = local_clock(); /* Same source as printk timestamps. */
 299
 300         /*
 301          * Pairs with READ_ONCE() in
 302          *      kfence_shutdown_cache(),
 303          *      kfence_handle_page_fault().
 304          */
 305         WRITE_ONCE(meta->state, next);
 306 }
 307
 308 /* Check canary byte at @addr. */
 309 static inline bool check_canary_byte(u8 *addr)
 310 {
 311         struct kfence_metadata *meta;
 312         unsigned long flags;
 313
 314         if (likely(*addr == KFENCE_CANARY_PATTERN_U8(addr)))
 315                 return true;
 316
 317         atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
 318
 319         meta = addr_to_metadata((unsigned long)addr);
 320         raw_spin_lock_irqsave(&meta->lock, flags);
 321         kfence_report_error((unsigned long)addr, false, NULL, meta, KFENCE_ERROR_CORRUPTION);
 322         raw_spin_unlock_irqrestore(&meta->lock, flags);
 323
 324         return false;
 325 }
 326
 327 static inline void set_canary(const struct kfence_metadata *meta)
 328 {
 329         const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
 330         unsigned long addr = pageaddr;
 331
 332         /*
 333          * The canary may be written to part of the object memory, but it does
 334          * not affect it. The user should initialize the object before using it.
 335          */
 336         for (; addr < meta->addr; addr += sizeof(u64))
 337                 *((u64 *)addr) = KFENCE_CANARY_PATTERN_U64;
 338
 339         addr = ALIGN_DOWN(meta->addr + meta->size, sizeof(u64));
 340         for (; addr - pageaddr < PAGE_SIZE; addr += sizeof(u64))
 341                 *((u64 *)addr) = KFENCE_CANARY_PATTERN_U64;
 342 }
 343
 344 static inline void check_canary(const struct kfence_metadata *meta)
 345 {
 346         const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
 347         unsigned long addr = pageaddr;
 348
 349         /*
 350          * We'll iterate over each canary byte per-side until a corrupted byte
 351          * is found. However, we'll still iterate over the canary bytes to the
 352          * right of the object even if there was an error in the canary bytes to
 353          * the left of the object. Specifically, if check_canary_byte()
 354          * generates an error, showing both sides might give more clues as to
 355          * what the error is about when displaying which bytes were corrupted.
 356          */
 357
 358         /* Apply to left of object. */
 359         for (; meta->addr - addr >= sizeof(u64); addr += sizeof(u64)) {
 360                 if (unlikely(*((u64 *)addr) != KFENCE_CANARY_PATTERN_U64))
 361                         break;
 362         }
 363
 364         /*
 365          * If the canary is corrupted in a certain 64 bytes, or the canary
 366          * memory cannot be completely covered by multiple consecutive 64 bytes,
 367          * it needs to be checked one by one.
 368          */
 369         for (; addr < meta->addr; addr++) {
 370                 if (unlikely(!check_canary_byte((u8 *)addr)))
 371                         break;
 372         }
 373
 374         /* Apply to right of object. */
 375         for (addr = meta->addr + meta->size; addr % sizeof(u64) != 0; addr++) {
 376                 if (unlikely(!check_canary_byte((u8 *)addr)))
 377                         return;
 378         }
 379         for (; addr - pageaddr < PAGE_SIZE; addr += sizeof(u64)) {
 380                 if (unlikely(*((u64 *)addr) != KFENCE_CANARY_PATTERN_U64)) {
 381
 382                         for (; addr - pageaddr < PAGE_SIZE; addr++) {
 383                                 if (!check_canary_byte((u8 *)addr))
 384                                         return;
 385                         }
 386                 }
 387         }
 388 }
 389
 390 static void *kfence_guarded_alloc(struct kmem_cache *cache, size_t size, gfp_t gfp,
 391                                   unsigned long *stack_entries, size_t num_stack_entries,
 392                                   u32 alloc_stack_hash)
 393 {
 394         struct kfence_metadata *meta = NULL;
 395         unsigned long flags;
 396         struct slab *slab;
 397         void *addr;
 398         const bool random_right_allocate = get_random_u32_below(2);
 399         const bool random_fault = CONFIG_KFENCE_STRESS_TEST_FAULTS &&
 400                                   !get_random_u32_below(CONFIG_KFENCE_STRESS_TEST_FAULTS);
 401
 402         /* Try to obtain a free object. */
 403         raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
 404         if (!list_empty(&kfence_freelist)) {
 405                 meta = list_entry(kfence_freelist.next, struct kfence_metadata, list);
 406                 list_del_init(&meta->list);
 407         }
 408         raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
 409         if (!meta) {
 410                 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_CAPACITY]);
 411                 return NULL;
 412         }
 413
 414         if (unlikely(!raw_spin_trylock_irqsave(&meta->lock, flags))) {
 415                 /*
 416                  * This is extremely unlikely -- we are reporting on a
 417                  * use-after-free, which locked meta->lock, and the reporting
 418                  * code via printk calls kmalloc() which ends up in
 419                  * kfence_alloc() and tries to grab the same object that we're
 420                  * reporting on. While it has never been observed, lockdep does
 421                  * report that there is a possibility of deadlock. Fix it by
 422                  * using trylock and bailing out gracefully.
 423                  */
 424                 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
 425                 /* Put the object back on the freelist. */
 426                 list_add_tail(&meta->list, &kfence_freelist);
 427                 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
 428
 429                 return NULL;
 430         }
 431
 432         meta->addr = metadata_to_pageaddr(meta);
 433         /* Unprotect if we're reusing this page. */
 434         if (meta->state == KFENCE_OBJECT_FREED)
 435                 kfence_unprotect(meta->addr);
 436
 437         /*
 438          * Note: for allocations made before RNG initialization, will always
 439          * return zero. We still benefit from enabling KFENCE as early as
 440          * possible, even when the RNG is not yet available, as this will allow
 441          * KFENCE to detect bugs due to earlier allocations. The only downside
 442          * is that the out-of-bounds accesses detected are deterministic for
 443          * such allocations.
 444          */
 445         if (random_right_allocate) {
 446                 /* Allocate on the "right" side, re-calculate address. */
 447                 meta->addr += PAGE_SIZE - size;
 448                 meta->addr = ALIGN_DOWN(meta->addr, cache->align);
 449         }
 450
 451         addr = (void *)meta->addr;
 452
 453         /* Update remaining metadata. */
 454         metadata_update_state(meta, KFENCE_OBJECT_ALLOCATED, stack_entries, num_stack_entries);
 455         /* Pairs with READ_ONCE() in kfence_shutdown_cache(). */
 456         WRITE_ONCE(meta->cache, cache);
 457         meta->size = size;
 458         meta->alloc_stack_hash = alloc_stack_hash;
 459         raw_spin_unlock_irqrestore(&meta->lock, flags);
 460
 461         alloc_covered_add(alloc_stack_hash, 1);
 462
 463         /* Set required slab fields. */
 464         slab = virt_to_slab((void *)meta->addr);
 465         slab->slab_cache = cache;
 466 #if defined(CONFIG_SLUB)
 467         slab->objects = 1;
 468 #elif defined(CONFIG_SLAB)
 469         slab->s_mem = addr;
 470 #endif
 471
 472         /* Memory initialization. */
 473         set_canary(meta);
 474
 475         /*
 476          * We check slab_want_init_on_alloc() ourselves, rather than letting
 477          * SL*B do the initialization, as otherwise we might overwrite KFENCE's
 478          * redzone.
 479          */
 480         if (unlikely(slab_want_init_on_alloc(gfp, cache)))
 481                 memzero_explicit(addr, size);
 482         if (cache->ctor)
 483                 cache->ctor(addr);
 484
 485         if (random_fault)
 486                 kfence_protect(meta->addr); /* Random "faults" by protecting the object. */
 487
 488         atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCATED]);
 489         atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCS]);
 490
 491         return addr;
 492 }
 493
 494 static void kfence_guarded_free(void *addr, struct kfence_metadata *meta, bool zombie)
 495 {
 496         struct kcsan_scoped_access assert_page_exclusive;
 497         unsigned long flags;
 498         bool init;
 499
 500         raw_spin_lock_irqsave(&meta->lock, flags);
 501
 502         if (meta->state != KFENCE_OBJECT_ALLOCATED || meta->addr != (unsigned long)addr) {
 503                 /* Invalid or double-free, bail out. */
 504                 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
 505                 kfence_report_error((unsigned long)addr, false, NULL, meta,
 506                                     KFENCE_ERROR_INVALID_FREE);
 507                 raw_spin_unlock_irqrestore(&meta->lock, flags);
 508                 return;
 509         }
 510
 511         /* Detect racy use-after-free, or incorrect reallocation of this page by KFENCE. */
 512         kcsan_begin_scoped_access((void *)ALIGN_DOWN((unsigned long)addr, PAGE_SIZE), PAGE_SIZE,
 513                                   KCSAN_ACCESS_SCOPED | KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT,
 514                                   &assert_page_exclusive);
 515
 516         if (CONFIG_KFENCE_STRESS_TEST_FAULTS)
 517                 kfence_unprotect((unsigned long)addr); /* To check canary bytes. */
 518
 519         /* Restore page protection if there was an OOB access. */
 520         if (meta->unprotected_page) {
 521                 memzero_explicit((void *)ALIGN_DOWN(meta->unprotected_page, PAGE_SIZE), PAGE_SIZE);
 522                 kfence_protect(meta->unprotected_page);
 523                 meta->unprotected_page = 0;
 524         }
 525
 526         /* Mark the object as freed. */
 527         metadata_update_state(meta, KFENCE_OBJECT_FREED, NULL, 0);
 528         init = slab_want_init_on_free(meta->cache);
 529         raw_spin_unlock_irqrestore(&meta->lock, flags);
 530
 531         alloc_covered_add(meta->alloc_stack_hash, -1);
 532
 533         /* Check canary bytes for memory corruption. */
 534         check_canary(meta);
 535
 536         /*
 537          * Clear memory if init-on-free is set. While we protect the page, the
 538          * data is still there, and after a use-after-free is detected, we
 539          * unprotect the page, so the data is still accessible.
 540          */
 541         if (!zombie && unlikely(init))
 542                 memzero_explicit(addr, meta->size);
 543
 544         /* Protect to detect use-after-frees. */
 545         kfence_protect((unsigned long)addr);
 546
 547         kcsan_end_scoped_access(&assert_page_exclusive);
 548         if (!zombie) {
 549                 /* Add it to the tail of the freelist for reuse. */
 550                 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
 551                 KFENCE_WARN_ON(!list_empty(&meta->list));
 552                 list_add_tail(&meta->list, &kfence_freelist);
 553                 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
 554
 555                 atomic_long_dec(&counters[KFENCE_COUNTER_ALLOCATED]);
 556                 atomic_long_inc(&counters[KFENCE_COUNTER_FREES]);
 557         } else {
 558                 /* See kfence_shutdown_cache(). */
 559                 atomic_long_inc(&counters[KFENCE_COUNTER_ZOMBIES]);
 560         }
 561 }
 562
 563 static void rcu_guarded_free(struct rcu_head *h)
 564 {
 565         struct kfence_metadata *meta = container_of(h, struct kfence_metadata, rcu_head);
 566
 567         kfence_guarded_free((void *)meta->addr, meta, false);
 568 }
 569
 570 /*
 571  * Initialization of the KFENCE pool after its allocation.
 572  * Returns 0 on success; otherwise returns the address up to
 573  * which partial initialization succeeded.
 574  */
 575 static unsigned long kfence_init_pool(void)
 576 {
 577         unsigned long addr = (unsigned long)__kfence_pool;
 578         struct page *pages;
 579         int i;
 580
 581         if (!arch_kfence_init_pool())
 582                 return addr;
 583
 584         pages = virt_to_page(__kfence_pool);
 585
 586         /*
 587          * Set up object pages: they must have PG_slab set, to avoid freeing
 588          * these as real pages.
 589          *
 590          * We also want to avoid inserting kfence_free() in the kfree()
 591          * fast-path in SLUB, and therefore need to ensure kfree() correctly
 592          * enters __slab_free() slow-path.
 593          */
 594         for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
 595                 struct slab *slab = page_slab(nth_page(pages, i));
 596
 597                 if (!i || (i % 2))
 598                         continue;
 599
 600                 __folio_set_slab(slab_folio(slab));
 601 #ifdef CONFIG_MEMCG
 602                 slab->memcg_data = (unsigned long)&kfence_metadata_init[i / 2 - 1].objcg |
 603                                    MEMCG_DATA_OBJCGS;
 604 #endif
 605         }
 606
 607         /*
 608          * Protect the first 2 pages. The first page is mostly unnecessary, and
 609          * merely serves as an extended guard page. However, adding one
 610          * additional page in the beginning gives us an even number of pages,
 611          * which simplifies the mapping of address to metadata index.
 612          */
 613         for (i = 0; i < 2; i++) {
 614                 if (unlikely(!kfence_protect(addr)))
 615                         return addr;
 616
 617                 addr += PAGE_SIZE;
 618         }
 619
 620         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 621                 struct kfence_metadata *meta = &kfence_metadata_init[i];
 622
 623                 /* Initialize metadata. */
 624                 INIT_LIST_HEAD(&meta->list);
 625                 raw_spin_lock_init(&meta->lock);
 626                 meta->state = KFENCE_OBJECT_UNUSED;
 627                 meta->addr = addr; /* Initialize for validation in metadata_to_pageaddr(). */
 628                 list_add_tail(&meta->list, &kfence_freelist);
 629
 630                 /* Protect the right redzone. */
 631                 if (unlikely(!kfence_protect(addr + PAGE_SIZE)))
 632                         goto reset_slab;
 633
 634                 addr += 2 * PAGE_SIZE;
 635         }
 636
 637         /*
 638          * Make kfence_metadata visible only when initialization is successful.
 639          * Otherwise, if the initialization fails and kfence_metadata is freed,
 640          * it may cause UAF in kfence_shutdown_cache().
 641          */
 642         smp_store_release(&kfence_metadata, kfence_metadata_init);
 643         return 0;
 644
 645 reset_slab:
 646         for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
 647                 struct slab *slab = page_slab(nth_page(pages, i));
 648
 649                 if (!i || (i % 2))
 650                         continue;
 651 #ifdef CONFIG_MEMCG
 652                 slab->memcg_data = 0;
 653 #endif
 654                 __folio_clear_slab(slab_folio(slab));
 655         }
 656
 657         return addr;
 658 }
 659
 660 static bool __init kfence_init_pool_early(void)
 661 {
 662         unsigned long addr;
 663
 664         if (!__kfence_pool)
 665                 return false;
 666
 667         addr = kfence_init_pool();
 668
 669         if (!addr) {
 670                 /*
 671                  * The pool is live and will never be deallocated from this point on.
 672                  * Ignore the pool object from the kmemleak phys object tree, as it would
 673                  * otherwise overlap with allocations returned by kfence_alloc(), which
 674                  * are registered with kmemleak through the slab post-alloc hook.
 675                  */
 676                 kmemleak_ignore_phys(__pa(__kfence_pool));
 677                 return true;
 678         }
 679
 680         /*
 681          * Only release unprotected pages, and do not try to go back and change
 682          * page attributes due to risk of failing to do so as well. If changing
 683          * page attributes for some pages fails, it is very likely that it also
 684          * fails for the first page, and therefore expect addr==__kfence_pool in
 685          * most failure cases.
 686          */
 687         memblock_free_late(__pa(addr), KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool));
 688         __kfence_pool = NULL;
 689
 690         memblock_free_late(__pa(kfence_metadata_init), KFENCE_METADATA_SIZE);
 691         kfence_metadata_init = NULL;
 692
 693         return false;
 694 }
 695
 696 /* === DebugFS Interface ==================================================== */
 697
 698 static int stats_show(struct seq_file *seq, void *v)
 699 {
 700         int i;
 701
 702         seq_printf(seq, "enabled: %i\n", READ_ONCE(kfence_enabled));
 703         for (i = 0; i < KFENCE_COUNTER_COUNT; i++)
 704                 seq_printf(seq, "%s: %ld\n", counter_names[i], atomic_long_read(&counters[i]));
 705
 706         return 0;
 707 }
 708 DEFINE_SHOW_ATTRIBUTE(stats);
 709
 710 /*
 711  * debugfs seq_file operations for /sys/kernel/debug/kfence/objects.
 712  * start_object() and next_object() return the object index + 1, because NULL is used
 713  * to stop iteration.
 714  */
 715 static void *start_object(struct seq_file *seq, loff_t *pos)
 716 {
 717         if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
 718                 return (void *)((long)*pos + 1);
 719         return NULL;
 720 }
 721
 722 static void stop_object(struct seq_file *seq, void *v)
 723 {
 724 }
 725
 726 static void *next_object(struct seq_file *seq, void *v, loff_t *pos)
 727 {
 728         ++*pos;
 729         if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
 730                 return (void *)((long)*pos + 1);
 731         return NULL;
 732 }
 733
 734 static int show_object(struct seq_file *seq, void *v)
 735 {
 736         struct kfence_metadata *meta = &kfence_metadata[(long)v - 1];
 737         unsigned long flags;
 738
 739         raw_spin_lock_irqsave(&meta->lock, flags);
 740         kfence_print_object(seq, meta);
 741         raw_spin_unlock_irqrestore(&meta->lock, flags);
 742         seq_puts(seq, "---------------------------------\n");
 743
 744         return 0;
 745 }
 746
 747 static const struct seq_operations objects_sops = {
 748         .start = start_object,
 749         .next = next_object,
 750         .stop = stop_object,
 751         .show = show_object,
 752 };
 753 DEFINE_SEQ_ATTRIBUTE(objects);
 754
 755 static int kfence_debugfs_init(void)
 756 {
 757         struct dentry *kfence_dir;
 758
 759         if (!READ_ONCE(kfence_enabled))
 760                 return 0;
 761
 762         kfence_dir = debugfs_create_dir("kfence", NULL);
 763         debugfs_create_file("stats", 0444, kfence_dir, NULL, &stats_fops);
 764         debugfs_create_file("objects", 0400, kfence_dir, NULL, &objects_fops);
 765         return 0;
 766 }
 767
 768 late_initcall(kfence_debugfs_init);
 769
 770 /* === Panic Notifier ====================================================== */
 771
 772 static void kfence_check_all_canary(void)
 773 {
 774         int i;
 775
 776         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 777                 struct kfence_metadata *meta = &kfence_metadata[i];
 778
 779                 if (meta->state == KFENCE_OBJECT_ALLOCATED)
 780                         check_canary(meta);
 781         }
 782 }
 783
 784 static int kfence_check_canary_callback(struct notifier_block *nb,
 785                                         unsigned long reason, void *arg)
 786 {
 787         kfence_check_all_canary();
 788         return NOTIFY_OK;
 789 }
 790
 791 static struct notifier_block kfence_check_canary_notifier = {
 792         .notifier_call = kfence_check_canary_callback,
 793 };
 794
 795 /* === Allocation Gate Timer ================================================ */
 796
 797 static struct delayed_work kfence_timer;
 798
 799 #ifdef CONFIG_KFENCE_STATIC_KEYS
 800 /* Wait queue to wake up allocation-gate timer task. */
 801 static DECLARE_WAIT_QUEUE_HEAD(allocation_wait);
 802
 803 static void wake_up_kfence_timer(struct irq_work *work)
 804 {
 805         wake_up(&allocation_wait);
 806 }
 807 static DEFINE_IRQ_WORK(wake_up_kfence_timer_work, wake_up_kfence_timer);
 808 #endif
 809
 810 /*
 811  * Set up delayed work, which will enable and disable the static key. We need to
 812  * use a work queue (rather than a simple timer), since enabling and disabling a
 813  * static key cannot be done from an interrupt.
 814  *
 815  * Note: Toggling a static branch currently causes IPIs, and here we'll end up
 816  * with a total of 2 IPIs to all CPUs. If this ends up a problem in future (with
 817  * more aggressive sampling intervals), we could get away with a variant that
 818  * avoids IPIs, at the cost of not immediately capturing allocations if the
 819  * instructions remain cached.
 820  */
 821 static void toggle_allocation_gate(struct work_struct *work)
 822 {
 823         if (!READ_ONCE(kfence_enabled))
 824                 return;
 825
 826         atomic_set(&kfence_allocation_gate, 0);
 827 #ifdef CONFIG_KFENCE_STATIC_KEYS
 828         /* Enable static key, and await allocation to happen. */
 829         static_branch_enable(&kfence_allocation_key);
 830
 831         wait_event_idle(allocation_wait, atomic_read(&kfence_allocation_gate));
 832
 833         /* Disable static key and reset timer. */
 834         static_branch_disable(&kfence_allocation_key);
 835 #endif
 836         queue_delayed_work(system_unbound_wq, &kfence_timer,
 837                            msecs_to_jiffies(kfence_sample_interval));
 838 }
 839
 840 /* === Public interface ===================================================== */
 841
 842 void __init kfence_alloc_pool_and_metadata(void)
 843 {
 844         if (!kfence_sample_interval)
 845                 return;
 846
 847         /*
 848          * If the pool has already been initialized by arch, there is no need to
 849          * re-allocate the memory pool.
 850          */
 851         if (!__kfence_pool)
 852                 __kfence_pool = memblock_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
 853
 854         if (!__kfence_pool) {
 855                 pr_err("failed to allocate pool\n");
 856                 return;
 857         }
 858
 859         /* The memory allocated by memblock has been zeroed out. */
 860         kfence_metadata_init = memblock_alloc(KFENCE_METADATA_SIZE, PAGE_SIZE);
 861         if (!kfence_metadata_init) {
 862                 pr_err("failed to allocate metadata\n");
 863                 memblock_free(__kfence_pool, KFENCE_POOL_SIZE);
 864                 __kfence_pool = NULL;
 865         }
 866 }
 867
 868 static void kfence_init_enable(void)
 869 {
 870         if (!IS_ENABLED(CONFIG_KFENCE_STATIC_KEYS))
 871                 static_branch_enable(&kfence_allocation_key);
 872
 873         if (kfence_deferrable)
 874                 INIT_DEFERRABLE_WORK(&kfence_timer, toggle_allocation_gate);
 875         else
 876                 INIT_DELAYED_WORK(&kfence_timer, toggle_allocation_gate);
 877
 878         if (kfence_check_on_panic)
 879                 atomic_notifier_chain_register(&panic_notifier_list, &kfence_check_canary_notifier);
 880
 881         WRITE_ONCE(kfence_enabled, true);
 882         queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
 883
 884         pr_info("initialized - using %lu bytes for %d objects at 0x%p-0x%p\n", KFENCE_POOL_SIZE,
 885                 CONFIG_KFENCE_NUM_OBJECTS, (void *)__kfence_pool,
 886                 (void *)(__kfence_pool + KFENCE_POOL_SIZE));
 887 }
 888
 889 void __init kfence_init(void)
 890 {
 891         stack_hash_seed = get_random_u32();
 892
 893         /* Setting kfence_sample_interval to 0 on boot disables KFENCE. */
 894         if (!kfence_sample_interval)
 895                 return;
 896
 897         if (!kfence_init_pool_early()) {
 898                 pr_err("%s failed\n", __func__);
 899                 return;
 900         }
 901
 902         kfence_init_enable();
 903 }
 904
 905 static int kfence_init_late(void)
 906 {
 907         const unsigned long nr_pages_pool = KFENCE_POOL_SIZE / PAGE_SIZE;
 908         const unsigned long nr_pages_meta = KFENCE_METADATA_SIZE / PAGE_SIZE;
 909         unsigned long addr = (unsigned long)__kfence_pool;
 910         unsigned long free_size = KFENCE_POOL_SIZE;
 911         int err = -ENOMEM;
 912
 913 #ifdef CONFIG_CONTIG_ALLOC
 914         struct page *pages;
 915
 916         pages = alloc_contig_pages(nr_pages_pool, GFP_KERNEL, first_online_node,
 917                                    NULL);
 918         if (!pages)
 919                 return -ENOMEM;
 920
 921         __kfence_pool = page_to_virt(pages);
 922         pages = alloc_contig_pages(nr_pages_meta, GFP_KERNEL, first_online_node,
 923                                    NULL);
 924         if (pages)
 925                 kfence_metadata_init = page_to_virt(pages);
 926 #else
 927         if (nr_pages_pool > MAX_ORDER_NR_PAGES ||
 928             nr_pages_meta > MAX_ORDER_NR_PAGES) {
 929                 pr_warn("KFENCE_NUM_OBJECTS too large for buddy allocator\n");
 930                 return -EINVAL;
 931         }
 932
 933         __kfence_pool = alloc_pages_exact(KFENCE_POOL_SIZE, GFP_KERNEL);
 934         if (!__kfence_pool)
 935                 return -ENOMEM;
 936
 937         kfence_metadata_init = alloc_pages_exact(KFENCE_METADATA_SIZE, GFP_KERNEL);
 938 #endif
 939
 940         if (!kfence_metadata_init)
 941                 goto free_pool;
 942
 943         memzero_explicit(kfence_metadata_init, KFENCE_METADATA_SIZE);
 944         addr = kfence_init_pool();
 945         if (!addr) {
 946                 kfence_init_enable();
 947                 kfence_debugfs_init();
 948                 return 0;
 949         }
 950
 951         pr_err("%s failed\n", __func__);
 952         free_size = KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool);
 953         err = -EBUSY;
 954
 955 #ifdef CONFIG_CONTIG_ALLOC
 956         free_contig_range(page_to_pfn(virt_to_page((void *)kfence_metadata_init)),
 957                           nr_pages_meta);
 958 free_pool:
 959         free_contig_range(page_to_pfn(virt_to_page((void *)addr)),
 960                           free_size / PAGE_SIZE);
 961 #else
 962         free_pages_exact((void *)kfence_metadata_init, KFENCE_METADATA_SIZE);
 963 free_pool:
 964         free_pages_exact((void *)addr, free_size);
 965 #endif
 966
 967         kfence_metadata_init = NULL;
 968         __kfence_pool = NULL;
 969         return err;
 970 }
 971
 972 static int kfence_enable_late(void)
 973 {
 974         if (!__kfence_pool)
 975                 return kfence_init_late();
 976
 977         WRITE_ONCE(kfence_enabled, true);
 978         queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
 979         pr_info("re-enabled\n");
 980         return 0;
 981 }
 982
 983 void kfence_shutdown_cache(struct kmem_cache *s)
 984 {
 985         unsigned long flags;
 986         struct kfence_metadata *meta;
 987         int i;
 988
 989         /* Pairs with release in kfence_init_pool(). */
 990         if (!smp_load_acquire(&kfence_metadata))
 991                 return;
 992
 993         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 994                 bool in_use;
 995
 996                 meta = &kfence_metadata[i];
 997
 998                 /*
 999                  * If we observe some inconsistent cache and state pair where we
1000                  * should have returned false here, cache destruction is racing
1001                  * with either kmem_cache_alloc() or kmem_cache_free(). Taking
1002                  * the lock will not help, as different critical section
1003                  * serialization will have the same outcome.
1004                  */
1005                 if (READ_ONCE(meta->cache) != s ||
1006                     READ_ONCE(meta->state) != KFENCE_OBJECT_ALLOCATED)
1007                         continue;
1008
1009                 raw_spin_lock_irqsave(&meta->lock, flags);
1010                 in_use = meta->cache == s && meta->state == KFENCE_OBJECT_ALLOCATED;
1011                 raw_spin_unlock_irqrestore(&meta->lock, flags);
1012
1013                 if (in_use) {
1014                         /*
1015                          * This cache still has allocations, and we should not
1016                          * release them back into the freelist so they can still
1017                          * safely be used and retain the kernel's default
1018                          * behaviour of keeping the allocations alive (leak the
1019                          * cache); however, they effectively become "zombie
1020                          * allocations" as the KFENCE objects are the only ones
1021                          * still in use and the owning cache is being destroyed.
1022                          *
1023                          * We mark them freed, so that any subsequent use shows
1024                          * more useful error messages that will include stack
1025                          * traces of the user of the object, the original
1026                          * allocation, and caller to shutdown_cache().
1027                          */
1028                         kfence_guarded_free((void *)meta->addr, meta, /*zombie=*/true);
1029                 }
1030         }
1031
1032         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
1033                 meta = &kfence_metadata[i];
1034
1035                 /* See above. */
1036                 if (READ_ONCE(meta->cache) != s || READ_ONCE(meta->state) != KFENCE_OBJECT_FREED)
1037                         continue;
1038
1039                 raw_spin_lock_irqsave(&meta->lock, flags);
1040                 if (meta->cache == s && meta->state == KFENCE_OBJECT_FREED)
1041                         meta->cache = NULL;
1042                 raw_spin_unlock_irqrestore(&meta->lock, flags);
1043         }
1044 }
1045
1046 void *__kfence_alloc(struct kmem_cache *s, size_t size, gfp_t flags)
1047 {
1048         unsigned long stack_entries[KFENCE_STACK_DEPTH];
1049         size_t num_stack_entries;
1050         u32 alloc_stack_hash;
1051
1052         /*
1053          * Perform size check before switching kfence_allocation_gate, so that
1054          * we don't disable KFENCE without making an allocation.
1055          */
1056         if (size > PAGE_SIZE) {
1057                 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
1058                 return NULL;
1059         }
1060
1061         /*
1062          * Skip allocations from non-default zones, including DMA. We cannot
1063          * guarantee that pages in the KFENCE pool will have the requested
1064          * properties (e.g. reside in DMAable memory).
1065          */
1066         if ((flags & GFP_ZONEMASK) ||
1067             (s->flags & (SLAB_CACHE_DMA | SLAB_CACHE_DMA32))) {
1068                 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
1069                 return NULL;
1070         }
1071
1072         /*
1073          * Skip allocations for this slab, if KFENCE has been disabled for
1074          * this slab.
1075          */
1076         if (s->flags & SLAB_SKIP_KFENCE)
1077                 return NULL;
1078
1079         if (atomic_inc_return(&kfence_allocation_gate) > 1)
1080                 return NULL;
1081 #ifdef CONFIG_KFENCE_STATIC_KEYS
1082         /*
1083          * waitqueue_active() is fully ordered after the update of
1084          * kfence_allocation_gate per atomic_inc_return().
1085          */
1086         if (waitqueue_active(&allocation_wait)) {
1087                 /*
1088                  * Calling wake_up() here may deadlock when allocations happen
1089                  * from within timer code. Use an irq_work to defer it.
1090                  */
1091                 irq_work_queue(&wake_up_kfence_timer_work);
1092         }
1093 #endif
1094
1095         if (!READ_ONCE(kfence_enabled))
1096                 return NULL;
1097
1098         num_stack_entries = stack_trace_save(stack_entries, KFENCE_STACK_DEPTH, 0);
1099
1100         /*
1101          * Do expensive check for coverage of allocation in slow-path after
1102          * allocation_gate has already become non-zero, even though it might
1103          * mean not making any allocation within a given sample interval.
1104          *
1105          * This ensures reasonable allocation coverage when the pool is almost
1106          * full, including avoiding long-lived allocations of the same source
1107          * filling up the pool (e.g. pagecache allocations).
1108          */
1109         alloc_stack_hash = get_alloc_stack_hash(stack_entries, num_stack_entries);
1110         if (should_skip_covered() && alloc_covered_contains(alloc_stack_hash)) {
1111                 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_COVERED]);
1112                 return NULL;
1113         }
1114
1115         return kfence_guarded_alloc(s, size, flags, stack_entries, num_stack_entries,
1116                                     alloc_stack_hash);
1117 }
1118
1119 size_t kfence_ksize(const void *addr)
1120 {
1121         const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1122
1123         /*
1124          * Read locklessly -- if there is a race with __kfence_alloc(), this is
1125          * either a use-after-free or invalid access.
1126          */
1127         return meta ? meta->size : 0;
1128 }
1129
1130 void *kfence_object_start(const void *addr)
1131 {
1132         const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1133
1134         /*
1135          * Read locklessly -- if there is a race with __kfence_alloc(), this is
1136          * either a use-after-free or invalid access.
1137          */
1138         return meta ? (void *)meta->addr : NULL;
1139 }
1140
1141 void __kfence_free(void *addr)
1142 {
1143         struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1144
1145 #ifdef CONFIG_MEMCG
1146         KFENCE_WARN_ON(meta->objcg);
1147 #endif
1148         /*
1149          * If the objects of the cache are SLAB_TYPESAFE_BY_RCU, defer freeing
1150          * the object, as the object page may be recycled for other-typed
1151          * objects once it has been freed. meta->cache may be NULL if the cache
1152          * was destroyed.
1153          */
1154         if (unlikely(meta->cache && (meta->cache->flags & SLAB_TYPESAFE_BY_RCU)))
1155                 call_rcu(&meta->rcu_head, rcu_guarded_free);
1156         else
1157                 kfence_guarded_free(addr, meta, false);
1158 }
1159
1160 bool kfence_handle_page_fault(unsigned long addr, bool is_write, struct pt_regs *regs)
1161 {
1162         const int page_index = (addr - (unsigned long)__kfence_pool) / PAGE_SIZE;
1163         struct kfence_metadata *to_report = NULL;
1164         enum kfence_error_type error_type;
1165         unsigned long flags;
1166
1167         if (!is_kfence_address((void *)addr))
1168                 return false;
1169
1170         if (!READ_ONCE(kfence_enabled)) /* If disabled at runtime ... */
1171                 return kfence_unprotect(addr); /* ... unprotect and proceed. */
1172
1173         atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
1174
1175         if (page_index % 2) {
1176                 /* This is a redzone, report a buffer overflow. */
1177                 struct kfence_metadata *meta;
1178                 int distance = 0;
1179
1180                 meta = addr_to_metadata(addr - PAGE_SIZE);
1181                 if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
1182                         to_report = meta;
1183                         /* Data race ok; distance calculation approximate. */
1184                         distance = addr - data_race(meta->addr + meta->size);
1185                 }
1186
1187                 meta = addr_to_metadata(addr + PAGE_SIZE);
1188                 if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
1189                         /* Data race ok; distance calculation approximate. */
1190                         if (!to_report || distance > data_race(meta->addr) - addr)
1191                                 to_report = meta;
1192                 }
1193
1194                 if (!to_report)
1195                         goto out;
1196
1197                 raw_spin_lock_irqsave(&to_report->lock, flags);
1198                 to_report->unprotected_page = addr;
1199                 error_type = KFENCE_ERROR_OOB;
1200
1201                 /*
1202                  * If the object was freed before we took the look we can still
1203                  * report this as an OOB -- the report will simply show the
1204                  * stacktrace of the free as well.
1205                  */
1206         } else {
1207                 to_report = addr_to_metadata(addr);
1208                 if (!to_report)
1209                         goto out;
1210
1211                 raw_spin_lock_irqsave(&to_report->lock, flags);
1212                 error_type = KFENCE_ERROR_UAF;
1213                 /*
1214                  * We may race with __kfence_alloc(), and it is possible that a
1215                  * freed object may be reallocated. We simply report this as a
1216                  * use-after-free, with the stack trace showing the place where
1217                  * the object was re-allocated.
1218                  */
1219         }
1220
1221 out:
1222         if (to_report) {
1223                 kfence_report_error(addr, is_write, regs, to_report, error_type);
1224                 raw_spin_unlock_irqrestore(&to_report->lock, flags);
1225         } else {
1226                 /* This may be a UAF or OOB access, but we can't be sure. */
1227                 kfence_report_error(addr, is_write, regs, NULL, KFENCE_ERROR_INVALID);
1228         }
1229
1230         return kfence_unprotect(addr); /* Unprotect and let access proceed. */
1231 }