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 /* Check canary byte at @addr. */
 301 static inline bool check_canary_byte(u8 *addr)
 302 {
 303         struct kfence_metadata *meta;
 304         unsigned long flags;
 305
 306         if (likely(*addr == KFENCE_CANARY_PATTERN_U8(addr)))
 307                 return true;
 308
 309         atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
 310
 311         meta = addr_to_metadata((unsigned long)addr);
 312         raw_spin_lock_irqsave(&meta->lock, flags);
 313         kfence_report_error((unsigned long)addr, false, NULL, meta, KFENCE_ERROR_CORRUPTION);
 314         raw_spin_unlock_irqrestore(&meta->lock, flags);
 315
 316         return false;
 317 }
 318
 319 static inline void set_canary(const struct kfence_metadata *meta)
 320 {
 321         const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
 322         unsigned long addr = pageaddr;
 323
 324         /*
 325          * The canary may be written to part of the object memory, but it does
 326          * not affect it. The user should initialize the object before using it.
 327          */
 328         for (; addr < meta->addr; addr += sizeof(u64))
 329                 *((u64 *)addr) = KFENCE_CANARY_PATTERN_U64;
 330
 331         addr = ALIGN_DOWN(meta->addr + meta->size, sizeof(u64));
 332         for (; addr - pageaddr < PAGE_SIZE; addr += sizeof(u64))
 333                 *((u64 *)addr) = KFENCE_CANARY_PATTERN_U64;
 334 }
 335
 336 static inline void check_canary(const struct kfence_metadata *meta)
 337 {
 338         const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
 339         unsigned long addr = pageaddr;
 340
 341         /*
 342          * We'll iterate over each canary byte per-side until a corrupted byte
 343          * is found. However, we'll still iterate over the canary bytes to the
 344          * right of the object even if there was an error in the canary bytes to
 345          * the left of the object. Specifically, if check_canary_byte()
 346          * generates an error, showing both sides might give more clues as to
 347          * what the error is about when displaying which bytes were corrupted.
 348          */
 349
 350         /* Apply to left of object. */
 351         for (; meta->addr - addr >= sizeof(u64); addr += sizeof(u64)) {
 352                 if (unlikely(*((u64 *)addr) != KFENCE_CANARY_PATTERN_U64))
 353                         break;
 354         }
 355
 356         /*
 357          * If the canary is corrupted in a certain 64 bytes, or the canary
 358          * memory cannot be completely covered by multiple consecutive 64 bytes,
 359          * it needs to be checked one by one.
 360          */
 361         for (; addr < meta->addr; addr++) {
 362                 if (unlikely(!check_canary_byte((u8 *)addr)))
 363                         break;
 364         }
 365
 366         /* Apply to right of object. */
 367         for (addr = meta->addr + meta->size; addr % sizeof(u64) != 0; addr++) {
 368                 if (unlikely(!check_canary_byte((u8 *)addr)))
 369                         return;
 370         }
 371         for (; addr - pageaddr < PAGE_SIZE; addr += sizeof(u64)) {
 372                 if (unlikely(*((u64 *)addr) != KFENCE_CANARY_PATTERN_U64)) {
 373
 374                         for (; addr - pageaddr < PAGE_SIZE; addr++) {
 375                                 if (!check_canary_byte((u8 *)addr))
 376                                         return;
 377                         }
 378                 }
 379         }
 380 }
 381
 382 static void *kfence_guarded_alloc(struct kmem_cache *cache, size_t size, gfp_t gfp,
 383                                   unsigned long *stack_entries, size_t num_stack_entries,
 384                                   u32 alloc_stack_hash)
 385 {
 386         struct kfence_metadata *meta = NULL;
 387         unsigned long flags;
 388         struct slab *slab;
 389         void *addr;
 390         const bool random_right_allocate = get_random_u32_below(2);
 391         const bool random_fault = CONFIG_KFENCE_STRESS_TEST_FAULTS &&
 392                                   !get_random_u32_below(CONFIG_KFENCE_STRESS_TEST_FAULTS);
 393
 394         /* Try to obtain a free object. */
 395         raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
 396         if (!list_empty(&kfence_freelist)) {
 397                 meta = list_entry(kfence_freelist.next, struct kfence_metadata, list);
 398                 list_del_init(&meta->list);
 399         }
 400         raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
 401         if (!meta) {
 402                 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_CAPACITY]);
 403                 return NULL;
 404         }
 405
 406         if (unlikely(!raw_spin_trylock_irqsave(&meta->lock, flags))) {
 407                 /*
 408                  * This is extremely unlikely -- we are reporting on a
 409                  * use-after-free, which locked meta->lock, and the reporting
 410                  * code via printk calls kmalloc() which ends up in
 411                  * kfence_alloc() and tries to grab the same object that we're
 412                  * reporting on. While it has never been observed, lockdep does
 413                  * report that there is a possibility of deadlock. Fix it by
 414                  * using trylock and bailing out gracefully.
 415                  */
 416                 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
 417                 /* Put the object back on the freelist. */
 418                 list_add_tail(&meta->list, &kfence_freelist);
 419                 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
 420
 421                 return NULL;
 422         }
 423
 424         meta->addr = metadata_to_pageaddr(meta);
 425         /* Unprotect if we're reusing this page. */
 426         if (meta->state == KFENCE_OBJECT_FREED)
 427                 kfence_unprotect(meta->addr);
 428
 429         /*
 430          * Note: for allocations made before RNG initialization, will always
 431          * return zero. We still benefit from enabling KFENCE as early as
 432          * possible, even when the RNG is not yet available, as this will allow
 433          * KFENCE to detect bugs due to earlier allocations. The only downside
 434          * is that the out-of-bounds accesses detected are deterministic for
 435          * such allocations.
 436          */
 437         if (random_right_allocate) {
 438                 /* Allocate on the "right" side, re-calculate address. */
 439                 meta->addr += PAGE_SIZE - size;
 440                 meta->addr = ALIGN_DOWN(meta->addr, cache->align);
 441         }
 442
 443         addr = (void *)meta->addr;
 444
 445         /* Update remaining metadata. */
 446         metadata_update_state(meta, KFENCE_OBJECT_ALLOCATED, stack_entries, num_stack_entries);
 447         /* Pairs with READ_ONCE() in kfence_shutdown_cache(). */
 448         WRITE_ONCE(meta->cache, cache);
 449         meta->size = size;
 450         meta->alloc_stack_hash = alloc_stack_hash;
 451         raw_spin_unlock_irqrestore(&meta->lock, flags);
 452
 453         alloc_covered_add(alloc_stack_hash, 1);
 454
 455         /* Set required slab fields. */
 456         slab = virt_to_slab((void *)meta->addr);
 457         slab->slab_cache = cache;
 458 #if defined(CONFIG_SLUB)
 459         slab->objects = 1;
 460 #elif defined(CONFIG_SLAB)
 461         slab->s_mem = addr;
 462 #endif
 463
 464         /* Memory initialization. */
 465         set_canary(meta);
 466
 467         /*
 468          * We check slab_want_init_on_alloc() ourselves, rather than letting
 469          * SL*B do the initialization, as otherwise we might overwrite KFENCE's
 470          * redzone.
 471          */
 472         if (unlikely(slab_want_init_on_alloc(gfp, cache)))
 473                 memzero_explicit(addr, size);
 474         if (cache->ctor)
 475                 cache->ctor(addr);
 476
 477         if (random_fault)
 478                 kfence_protect(meta->addr); /* Random "faults" by protecting the object. */
 479
 480         atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCATED]);
 481         atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCS]);
 482
 483         return addr;
 484 }
 485
 486 static void kfence_guarded_free(void *addr, struct kfence_metadata *meta, bool zombie)
 487 {
 488         struct kcsan_scoped_access assert_page_exclusive;
 489         unsigned long flags;
 490         bool init;
 491
 492         raw_spin_lock_irqsave(&meta->lock, flags);
 493
 494         if (meta->state != KFENCE_OBJECT_ALLOCATED || meta->addr != (unsigned long)addr) {
 495                 /* Invalid or double-free, bail out. */
 496                 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
 497                 kfence_report_error((unsigned long)addr, false, NULL, meta,
 498                                     KFENCE_ERROR_INVALID_FREE);
 499                 raw_spin_unlock_irqrestore(&meta->lock, flags);
 500                 return;
 501         }
 502
 503         /* Detect racy use-after-free, or incorrect reallocation of this page by KFENCE. */
 504         kcsan_begin_scoped_access((void *)ALIGN_DOWN((unsigned long)addr, PAGE_SIZE), PAGE_SIZE,
 505                                   KCSAN_ACCESS_SCOPED | KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT,
 506                                   &assert_page_exclusive);
 507
 508         if (CONFIG_KFENCE_STRESS_TEST_FAULTS)
 509                 kfence_unprotect((unsigned long)addr); /* To check canary bytes. */
 510
 511         /* Restore page protection if there was an OOB access. */
 512         if (meta->unprotected_page) {
 513                 memzero_explicit((void *)ALIGN_DOWN(meta->unprotected_page, PAGE_SIZE), PAGE_SIZE);
 514                 kfence_protect(meta->unprotected_page);
 515                 meta->unprotected_page = 0;
 516         }
 517
 518         /* Mark the object as freed. */
 519         metadata_update_state(meta, KFENCE_OBJECT_FREED, NULL, 0);
 520         init = slab_want_init_on_free(meta->cache);
 521         raw_spin_unlock_irqrestore(&meta->lock, flags);
 522
 523         alloc_covered_add(meta->alloc_stack_hash, -1);
 524
 525         /* Check canary bytes for memory corruption. */
 526         check_canary(meta);
 527
 528         /*
 529          * Clear memory if init-on-free is set. While we protect the page, the
 530          * data is still there, and after a use-after-free is detected, we
 531          * unprotect the page, so the data is still accessible.
 532          */
 533         if (!zombie && unlikely(init))
 534                 memzero_explicit(addr, meta->size);
 535
 536         /* Protect to detect use-after-frees. */
 537         kfence_protect((unsigned long)addr);
 538
 539         kcsan_end_scoped_access(&assert_page_exclusive);
 540         if (!zombie) {
 541                 /* Add it to the tail of the freelist for reuse. */
 542                 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
 543                 KFENCE_WARN_ON(!list_empty(&meta->list));
 544                 list_add_tail(&meta->list, &kfence_freelist);
 545                 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
 546
 547                 atomic_long_dec(&counters[KFENCE_COUNTER_ALLOCATED]);
 548                 atomic_long_inc(&counters[KFENCE_COUNTER_FREES]);
 549         } else {
 550                 /* See kfence_shutdown_cache(). */
 551                 atomic_long_inc(&counters[KFENCE_COUNTER_ZOMBIES]);
 552         }
 553 }
 554
 555 static void rcu_guarded_free(struct rcu_head *h)
 556 {
 557         struct kfence_metadata *meta = container_of(h, struct kfence_metadata, rcu_head);
 558
 559         kfence_guarded_free((void *)meta->addr, meta, false);
 560 }
 561
 562 /*
 563  * Initialization of the KFENCE pool after its allocation.
 564  * Returns 0 on success; otherwise returns the address up to
 565  * which partial initialization succeeded.
 566  */
 567 static unsigned long kfence_init_pool(void)
 568 {
 569         unsigned long addr = (unsigned long)__kfence_pool;
 570         struct page *pages;
 571         int i;
 572
 573         if (!arch_kfence_init_pool())
 574                 return addr;
 575
 576         pages = virt_to_page(__kfence_pool);
 577
 578         /*
 579          * Set up object pages: they must have PG_slab set, to avoid freeing
 580          * these as real pages.
 581          *
 582          * We also want to avoid inserting kfence_free() in the kfree()
 583          * fast-path in SLUB, and therefore need to ensure kfree() correctly
 584          * enters __slab_free() slow-path.
 585          */
 586         for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
 587                 struct slab *slab = page_slab(nth_page(pages, i));
 588
 589                 if (!i || (i % 2))
 590                         continue;
 591
 592                 __folio_set_slab(slab_folio(slab));
 593 #ifdef CONFIG_MEMCG
 594                 slab->memcg_data = (unsigned long)&kfence_metadata[i / 2 - 1].objcg |
 595                                    MEMCG_DATA_OBJCGS;
 596 #endif
 597         }
 598
 599         /*
 600          * Protect the first 2 pages. The first page is mostly unnecessary, and
 601          * merely serves as an extended guard page. However, adding one
 602          * additional page in the beginning gives us an even number of pages,
 603          * which simplifies the mapping of address to metadata index.
 604          */
 605         for (i = 0; i < 2; i++) {
 606                 if (unlikely(!kfence_protect(addr)))
 607                         return addr;
 608
 609                 addr += PAGE_SIZE;
 610         }
 611
 612         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 613                 struct kfence_metadata *meta = &kfence_metadata[i];
 614
 615                 /* Initialize metadata. */
 616                 INIT_LIST_HEAD(&meta->list);
 617                 raw_spin_lock_init(&meta->lock);
 618                 meta->state = KFENCE_OBJECT_UNUSED;
 619                 meta->addr = addr; /* Initialize for validation in metadata_to_pageaddr(). */
 620                 list_add_tail(&meta->list, &kfence_freelist);
 621
 622                 /* Protect the right redzone. */
 623                 if (unlikely(!kfence_protect(addr + PAGE_SIZE)))
 624                         goto reset_slab;
 625
 626                 addr += 2 * PAGE_SIZE;
 627         }
 628
 629         return 0;
 630
 631 reset_slab:
 632         for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
 633                 struct slab *slab = page_slab(nth_page(pages, i));
 634
 635                 if (!i || (i % 2))
 636                         continue;
 637 #ifdef CONFIG_MEMCG
 638                 slab->memcg_data = 0;
 639 #endif
 640                 __folio_clear_slab(slab_folio(slab));
 641         }
 642
 643         return addr;
 644 }
 645
 646 static bool __init kfence_init_pool_early(void)
 647 {
 648         unsigned long addr;
 649
 650         if (!__kfence_pool)
 651                 return false;
 652
 653         addr = kfence_init_pool();
 654
 655         if (!addr) {
 656                 /*
 657                  * The pool is live and will never be deallocated from this point on.
 658                  * Ignore the pool object from the kmemleak phys object tree, as it would
 659                  * otherwise overlap with allocations returned by kfence_alloc(), which
 660                  * are registered with kmemleak through the slab post-alloc hook.
 661                  */
 662                 kmemleak_ignore_phys(__pa(__kfence_pool));
 663                 return true;
 664         }
 665
 666         /*
 667          * Only release unprotected pages, and do not try to go back and change
 668          * page attributes due to risk of failing to do so as well. If changing
 669          * page attributes for some pages fails, it is very likely that it also
 670          * fails for the first page, and therefore expect addr==__kfence_pool in
 671          * most failure cases.
 672          */
 673         memblock_free_late(__pa(addr), KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool));
 674         __kfence_pool = NULL;
 675         return false;
 676 }
 677
 678 static bool kfence_init_pool_late(void)
 679 {
 680         unsigned long addr, free_size;
 681
 682         addr = kfence_init_pool();
 683
 684         if (!addr)
 685                 return true;
 686
 687         /* Same as above. */
 688         free_size = KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool);
 689 #ifdef CONFIG_CONTIG_ALLOC
 690         free_contig_range(page_to_pfn(virt_to_page((void *)addr)), free_size / PAGE_SIZE);
 691 #else
 692         free_pages_exact((void *)addr, free_size);
 693 #endif
 694         __kfence_pool = NULL;
 695         return false;
 696 }
 697
 698 /* === DebugFS Interface ==================================================== */
 699
 700 static int stats_show(struct seq_file *seq, void *v)
 701 {
 702         int i;
 703
 704         seq_printf(seq, "enabled: %i\n", READ_ONCE(kfence_enabled));
 705         for (i = 0; i < KFENCE_COUNTER_COUNT; i++)
 706                 seq_printf(seq, "%s: %ld\n", counter_names[i], atomic_long_read(&counters[i]));
 707
 708         return 0;
 709 }
 710 DEFINE_SHOW_ATTRIBUTE(stats);
 711
 712 /*
 713  * debugfs seq_file operations for /sys/kernel/debug/kfence/objects.
 714  * start_object() and next_object() return the object index + 1, because NULL is used
 715  * to stop iteration.
 716  */
 717 static void *start_object(struct seq_file *seq, loff_t *pos)
 718 {
 719         if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
 720                 return (void *)((long)*pos + 1);
 721         return NULL;
 722 }
 723
 724 static void stop_object(struct seq_file *seq, void *v)
 725 {
 726 }
 727
 728 static void *next_object(struct seq_file *seq, void *v, loff_t *pos)
 729 {
 730         ++*pos;
 731         if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
 732                 return (void *)((long)*pos + 1);
 733         return NULL;
 734 }
 735
 736 static int show_object(struct seq_file *seq, void *v)
 737 {
 738         struct kfence_metadata *meta = &kfence_metadata[(long)v - 1];
 739         unsigned long flags;
 740
 741         raw_spin_lock_irqsave(&meta->lock, flags);
 742         kfence_print_object(seq, meta);
 743         raw_spin_unlock_irqrestore(&meta->lock, flags);
 744         seq_puts(seq, "---------------------------------\n");
 745
 746         return 0;
 747 }
 748
 749 static const struct seq_operations objects_sops = {
 750         .start = start_object,
 751         .next = next_object,
 752         .stop = stop_object,
 753         .show = show_object,
 754 };
 755 DEFINE_SEQ_ATTRIBUTE(objects);
 756
 757 static int kfence_debugfs_init(void)
 758 {
 759         struct dentry *kfence_dir;
 760
 761         if (!READ_ONCE(kfence_enabled))
 762                 return 0;
 763
 764         kfence_dir = debugfs_create_dir("kfence", NULL);
 765         debugfs_create_file("stats", 0444, kfence_dir, NULL, &stats_fops);
 766         debugfs_create_file("objects", 0400, kfence_dir, NULL, &objects_fops);
 767         return 0;
 768 }
 769
 770 late_initcall(kfence_debugfs_init);
 771
 772 /* === Panic Notifier ====================================================== */
 773
 774 static void kfence_check_all_canary(void)
 775 {
 776         int i;
 777
 778         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 779                 struct kfence_metadata *meta = &kfence_metadata[i];
 780
 781                 if (meta->state == KFENCE_OBJECT_ALLOCATED)
 782                         check_canary(meta);
 783         }
 784 }
 785
 786 static int kfence_check_canary_callback(struct notifier_block *nb,
 787                                         unsigned long reason, void *arg)
 788 {
 789         kfence_check_all_canary();
 790         return NOTIFY_OK;
 791 }
 792
 793 static struct notifier_block kfence_check_canary_notifier = {
 794         .notifier_call = kfence_check_canary_callback,
 795 };
 796
 797 /* === Allocation Gate Timer ================================================ */
 798
 799 static struct delayed_work kfence_timer;
 800
 801 #ifdef CONFIG_KFENCE_STATIC_KEYS
 802 /* Wait queue to wake up allocation-gate timer task. */
 803 static DECLARE_WAIT_QUEUE_HEAD(allocation_wait);
 804
 805 static void wake_up_kfence_timer(struct irq_work *work)
 806 {
 807         wake_up(&allocation_wait);
 808 }
 809 static DEFINE_IRQ_WORK(wake_up_kfence_timer_work, wake_up_kfence_timer);
 810 #endif
 811
 812 /*
 813  * Set up delayed work, which will enable and disable the static key. We need to
 814  * use a work queue (rather than a simple timer), since enabling and disabling a
 815  * static key cannot be done from an interrupt.
 816  *
 817  * Note: Toggling a static branch currently causes IPIs, and here we'll end up
 818  * with a total of 2 IPIs to all CPUs. If this ends up a problem in future (with
 819  * more aggressive sampling intervals), we could get away with a variant that
 820  * avoids IPIs, at the cost of not immediately capturing allocations if the
 821  * instructions remain cached.
 822  */
 823 static void toggle_allocation_gate(struct work_struct *work)
 824 {
 825         if (!READ_ONCE(kfence_enabled))
 826                 return;
 827
 828         atomic_set(&kfence_allocation_gate, 0);
 829 #ifdef CONFIG_KFENCE_STATIC_KEYS
 830         /* Enable static key, and await allocation to happen. */
 831         static_branch_enable(&kfence_allocation_key);
 832
 833         wait_event_idle(allocation_wait, atomic_read(&kfence_allocation_gate));
 834
 835         /* Disable static key and reset timer. */
 836         static_branch_disable(&kfence_allocation_key);
 837 #endif
 838         queue_delayed_work(system_unbound_wq, &kfence_timer,
 839                            msecs_to_jiffies(kfence_sample_interval));
 840 }
 841
 842 /* === Public interface ===================================================== */
 843
 844 void __init kfence_alloc_pool(void)
 845 {
 846         if (!kfence_sample_interval)
 847                 return;
 848
 849         /* if the pool has already been initialized by arch, skip the below. */
 850         if (__kfence_pool)
 851                 return;
 852
 853         __kfence_pool = memblock_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
 854
 855         if (!__kfence_pool)
 856                 pr_err("failed to allocate pool\n");
 857 }
 858
 859 static void kfence_init_enable(void)
 860 {
 861         if (!IS_ENABLED(CONFIG_KFENCE_STATIC_KEYS))
 862                 static_branch_enable(&kfence_allocation_key);
 863
 864         if (kfence_deferrable)
 865                 INIT_DEFERRABLE_WORK(&kfence_timer, toggle_allocation_gate);
 866         else
 867                 INIT_DELAYED_WORK(&kfence_timer, toggle_allocation_gate);
 868
 869         if (kfence_check_on_panic)
 870                 atomic_notifier_chain_register(&panic_notifier_list, &kfence_check_canary_notifier);
 871
 872         WRITE_ONCE(kfence_enabled, true);
 873         queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
 874
 875         pr_info("initialized - using %lu bytes for %d objects at 0x%p-0x%p\n", KFENCE_POOL_SIZE,
 876                 CONFIG_KFENCE_NUM_OBJECTS, (void *)__kfence_pool,
 877                 (void *)(__kfence_pool + KFENCE_POOL_SIZE));
 878 }
 879
 880 void __init kfence_init(void)
 881 {
 882         stack_hash_seed = get_random_u32();
 883
 884         /* Setting kfence_sample_interval to 0 on boot disables KFENCE. */
 885         if (!kfence_sample_interval)
 886                 return;
 887
 888         if (!kfence_init_pool_early()) {
 889                 pr_err("%s failed\n", __func__);
 890                 return;
 891         }
 892
 893         kfence_init_enable();
 894 }
 895
 896 static int kfence_init_late(void)
 897 {
 898         const unsigned long nr_pages = KFENCE_POOL_SIZE / PAGE_SIZE;
 899 #ifdef CONFIG_CONTIG_ALLOC
 900         struct page *pages;
 901
 902         pages = alloc_contig_pages(nr_pages, GFP_KERNEL, first_online_node, NULL);
 903         if (!pages)
 904                 return -ENOMEM;
 905         __kfence_pool = page_to_virt(pages);
 906 #else
 907         if (nr_pages > MAX_ORDER_NR_PAGES) {
 908                 pr_warn("KFENCE_NUM_OBJECTS too large for buddy allocator\n");
 909                 return -EINVAL;
 910         }
 911         __kfence_pool = alloc_pages_exact(KFENCE_POOL_SIZE, GFP_KERNEL);
 912         if (!__kfence_pool)
 913                 return -ENOMEM;
 914 #endif
 915
 916         if (!kfence_init_pool_late()) {
 917                 pr_err("%s failed\n", __func__);
 918                 return -EBUSY;
 919         }
 920
 921         kfence_init_enable();
 922         kfence_debugfs_init();
 923
 924         return 0;
 925 }
 926
 927 static int kfence_enable_late(void)
 928 {
 929         if (!__kfence_pool)
 930                 return kfence_init_late();
 931
 932         WRITE_ONCE(kfence_enabled, true);
 933         queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
 934         pr_info("re-enabled\n");
 935         return 0;
 936 }
 937
 938 void kfence_shutdown_cache(struct kmem_cache *s)
 939 {
 940         unsigned long flags;
 941         struct kfence_metadata *meta;
 942         int i;
 943
 944         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 945                 bool in_use;
 946
 947                 meta = &kfence_metadata[i];
 948
 949                 /*
 950                  * If we observe some inconsistent cache and state pair where we
 951                  * should have returned false here, cache destruction is racing
 952                  * with either kmem_cache_alloc() or kmem_cache_free(). Taking
 953                  * the lock will not help, as different critical section
 954                  * serialization will have the same outcome.
 955                  */
 956                 if (READ_ONCE(meta->cache) != s ||
 957                     READ_ONCE(meta->state) != KFENCE_OBJECT_ALLOCATED)
 958                         continue;
 959
 960                 raw_spin_lock_irqsave(&meta->lock, flags);
 961                 in_use = meta->cache == s && meta->state == KFENCE_OBJECT_ALLOCATED;
 962                 raw_spin_unlock_irqrestore(&meta->lock, flags);
 963
 964                 if (in_use) {
 965                         /*
 966                          * This cache still has allocations, and we should not
 967                          * release them back into the freelist so they can still
 968                          * safely be used and retain the kernel's default
 969                          * behaviour of keeping the allocations alive (leak the
 970                          * cache); however, they effectively become "zombie
 971                          * allocations" as the KFENCE objects are the only ones
 972                          * still in use and the owning cache is being destroyed.
 973                          *
 974                          * We mark them freed, so that any subsequent use shows
 975                          * more useful error messages that will include stack
 976                          * traces of the user of the object, the original
 977                          * allocation, and caller to shutdown_cache().
 978                          */
 979                         kfence_guarded_free((void *)meta->addr, meta, /*zombie=*/true);
 980                 }
 981         }
 982
 983         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 984                 meta = &kfence_metadata[i];
 985
 986                 /* See above. */
 987                 if (READ_ONCE(meta->cache) != s || READ_ONCE(meta->state) != KFENCE_OBJECT_FREED)
 988                         continue;
 989
 990                 raw_spin_lock_irqsave(&meta->lock, flags);
 991                 if (meta->cache == s && meta->state == KFENCE_OBJECT_FREED)
 992                         meta->cache = NULL;
 993                 raw_spin_unlock_irqrestore(&meta->lock, flags);
 994         }
 995 }
 996
 997 void *__kfence_alloc(struct kmem_cache *s, size_t size, gfp_t flags)
 998 {
 999         unsigned long stack_entries[KFENCE_STACK_DEPTH];
1000         size_t num_stack_entries;
1001         u32 alloc_stack_hash;
1002
1003         /*
1004          * Perform size check before switching kfence_allocation_gate, so that
1005          * we don't disable KFENCE without making an allocation.
1006          */
1007         if (size > PAGE_SIZE) {
1008                 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
1009                 return NULL;
1010         }
1011
1012         /*
1013          * Skip allocations from non-default zones, including DMA. We cannot
1014          * guarantee that pages in the KFENCE pool will have the requested
1015          * properties (e.g. reside in DMAable memory).
1016          */
1017         if ((flags & GFP_ZONEMASK) ||
1018             (s->flags & (SLAB_CACHE_DMA | SLAB_CACHE_DMA32))) {
1019                 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
1020                 return NULL;
1021         }
1022
1023         /*
1024          * Skip allocations for this slab, if KFENCE has been disabled for
1025          * this slab.
1026          */
1027         if (s->flags & SLAB_SKIP_KFENCE)
1028                 return NULL;
1029
1030         if (atomic_inc_return(&kfence_allocation_gate) > 1)
1031                 return NULL;
1032 #ifdef CONFIG_KFENCE_STATIC_KEYS
1033         /*
1034          * waitqueue_active() is fully ordered after the update of
1035          * kfence_allocation_gate per atomic_inc_return().
1036          */
1037         if (waitqueue_active(&allocation_wait)) {
1038                 /*
1039                  * Calling wake_up() here may deadlock when allocations happen
1040                  * from within timer code. Use an irq_work to defer it.
1041                  */
1042                 irq_work_queue(&wake_up_kfence_timer_work);
1043         }
1044 #endif
1045
1046         if (!READ_ONCE(kfence_enabled))
1047                 return NULL;
1048
1049         num_stack_entries = stack_trace_save(stack_entries, KFENCE_STACK_DEPTH, 0);
1050
1051         /*
1052          * Do expensive check for coverage of allocation in slow-path after
1053          * allocation_gate has already become non-zero, even though it might
1054          * mean not making any allocation within a given sample interval.
1055          *
1056          * This ensures reasonable allocation coverage when the pool is almost
1057          * full, including avoiding long-lived allocations of the same source
1058          * filling up the pool (e.g. pagecache allocations).
1059          */
1060         alloc_stack_hash = get_alloc_stack_hash(stack_entries, num_stack_entries);
1061         if (should_skip_covered() && alloc_covered_contains(alloc_stack_hash)) {
1062                 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_COVERED]);
1063                 return NULL;
1064         }
1065
1066         return kfence_guarded_alloc(s, size, flags, stack_entries, num_stack_entries,
1067                                     alloc_stack_hash);
1068 }
1069
1070 size_t kfence_ksize(const void *addr)
1071 {
1072         const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1073
1074         /*
1075          * Read locklessly -- if there is a race with __kfence_alloc(), this is
1076          * either a use-after-free or invalid access.
1077          */
1078         return meta ? meta->size : 0;
1079 }
1080
1081 void *kfence_object_start(const void *addr)
1082 {
1083         const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1084
1085         /*
1086          * Read locklessly -- if there is a race with __kfence_alloc(), this is
1087          * either a use-after-free or invalid access.
1088          */
1089         return meta ? (void *)meta->addr : NULL;
1090 }
1091
1092 void __kfence_free(void *addr)
1093 {
1094         struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1095
1096 #ifdef CONFIG_MEMCG
1097         KFENCE_WARN_ON(meta->objcg);
1098 #endif
1099         /*
1100          * If the objects of the cache are SLAB_TYPESAFE_BY_RCU, defer freeing
1101          * the object, as the object page may be recycled for other-typed
1102          * objects once it has been freed. meta->cache may be NULL if the cache
1103          * was destroyed.
1104          */
1105         if (unlikely(meta->cache && (meta->cache->flags & SLAB_TYPESAFE_BY_RCU)))
1106                 call_rcu(&meta->rcu_head, rcu_guarded_free);
1107         else
1108                 kfence_guarded_free(addr, meta, false);
1109 }
1110
1111 bool kfence_handle_page_fault(unsigned long addr, bool is_write, struct pt_regs *regs)
1112 {
1113         const int page_index = (addr - (unsigned long)__kfence_pool) / PAGE_SIZE;
1114         struct kfence_metadata *to_report = NULL;
1115         enum kfence_error_type error_type;
1116         unsigned long flags;
1117
1118         if (!is_kfence_address((void *)addr))
1119                 return false;
1120
1121         if (!READ_ONCE(kfence_enabled)) /* If disabled at runtime ... */
1122                 return kfence_unprotect(addr); /* ... unprotect and proceed. */
1123
1124         atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
1125
1126         if (page_index % 2) {
1127                 /* This is a redzone, report a buffer overflow. */
1128                 struct kfence_metadata *meta;
1129                 int distance = 0;
1130
1131                 meta = addr_to_metadata(addr - PAGE_SIZE);
1132                 if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
1133                         to_report = meta;
1134                         /* Data race ok; distance calculation approximate. */
1135                         distance = addr - data_race(meta->addr + meta->size);
1136                 }
1137
1138                 meta = addr_to_metadata(addr + PAGE_SIZE);
1139                 if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
1140                         /* Data race ok; distance calculation approximate. */
1141                         if (!to_report || distance > data_race(meta->addr) - addr)
1142                                 to_report = meta;
1143                 }
1144
1145                 if (!to_report)
1146                         goto out;
1147
1148                 raw_spin_lock_irqsave(&to_report->lock, flags);
1149                 to_report->unprotected_page = addr;
1150                 error_type = KFENCE_ERROR_OOB;
1151
1152                 /*
1153                  * If the object was freed before we took the look we can still
1154                  * report this as an OOB -- the report will simply show the
1155                  * stacktrace of the free as well.
1156                  */
1157         } else {
1158                 to_report = addr_to_metadata(addr);
1159                 if (!to_report)
1160                         goto out;
1161
1162                 raw_spin_lock_irqsave(&to_report->lock, flags);
1163                 error_type = KFENCE_ERROR_UAF;
1164                 /*
1165                  * We may race with __kfence_alloc(), and it is possible that a
1166                  * freed object may be reallocated. We simply report this as a
1167                  * use-after-free, with the stack trace showing the place where
1168                  * the object was re-allocated.
1169                  */
1170         }
1171
1172 out:
1173         if (to_report) {
1174                 kfence_report_error(addr, is_write, regs, to_report, error_type);
1175                 raw_spin_unlock_irqrestore(&to_report->lock, flags);
1176         } else {
1177                 /* This may be a UAF or OOB access, but we can't be sure. */
1178                 kfence_report_error(addr, is_write, regs, NULL, KFENCE_ERROR_INVALID);
1179         }
1180
1181         return kfence_unprotect(addr); /* Unprotect and let access proceed. */
1182 }