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