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/sched/sysctl.h>
  30 #include <linux/seq_file.h>
  31 #include <linux/slab.h>
  32 #include <linux/spinlock.h>
  33 #include <linux/string.h>
  34
  35 #include <asm/kfence.h>
  36
  37 #include "kfence.h"
  38
  39 /* Disables KFENCE on the first warning assuming an irrecoverable error. */
  40 #define KFENCE_WARN_ON(cond)                                                   \
  41         ({                                                                     \
  42                 const bool __cond = WARN_ON(cond);                             \
  43                 if (unlikely(__cond)) {                                        \
  44                         WRITE_ONCE(kfence_enabled, false);                     \
  45                         disabled_by_warn = true;                               \
  46                 }                                                              \
  47                 __cond;                                                        \
  48         })
  49
  50 /* === Data ================================================================= */
  51
  52 static bool kfence_enabled __read_mostly;
  53 static bool disabled_by_warn __read_mostly;
  54
  55 unsigned long kfence_sample_interval __read_mostly = CONFIG_KFENCE_SAMPLE_INTERVAL;
  56 EXPORT_SYMBOL_GPL(kfence_sample_interval); /* Export for test modules. */
  57
  58 #ifdef MODULE_PARAM_PREFIX
  59 #undef MODULE_PARAM_PREFIX
  60 #endif
  61 #define MODULE_PARAM_PREFIX "kfence."
  62
  63 static int kfence_enable_late(void);
  64 static int param_set_sample_interval(const char *val, const struct kernel_param *kp)
  65 {
  66         unsigned long num;
  67         int ret = kstrtoul(val, 0, &num);
  68
  69         if (ret < 0)
  70                 return ret;
  71
  72         /* Using 0 to indicate KFENCE is disabled. */
  73         if (!num && READ_ONCE(kfence_enabled)) {
  74                 pr_info("disabled\n");
  75                 WRITE_ONCE(kfence_enabled, false);
  76         }
  77
  78         *((unsigned long *)kp->arg) = num;
  79
  80         if (num && !READ_ONCE(kfence_enabled) && system_state != SYSTEM_BOOTING)
  81                 return disabled_by_warn ? -EINVAL : kfence_enable_late();
  82         return 0;
  83 }
  84
  85 static int param_get_sample_interval(char *buffer, const struct kernel_param *kp)
  86 {
  87         if (!READ_ONCE(kfence_enabled))
  88                 return sprintf(buffer, "0\n");
  89
  90         return param_get_ulong(buffer, kp);
  91 }
  92
  93 static const struct kernel_param_ops sample_interval_param_ops = {
  94         .set = param_set_sample_interval,
  95         .get = param_get_sample_interval,
  96 };
  97 module_param_cb(sample_interval, &sample_interval_param_ops, &kfence_sample_interval, 0600);
  98
  99 /* Pool usage% threshold when currently covered allocations are skipped. */
 100 static unsigned long kfence_skip_covered_thresh __read_mostly = 75;
 101 module_param_named(skip_covered_thresh, kfence_skip_covered_thresh, ulong, 0644);
 102
 103 /* If true, use a deferrable timer. */
 104 static bool kfence_deferrable __read_mostly = IS_ENABLED(CONFIG_KFENCE_DEFERRABLE);
 105 module_param_named(deferrable, kfence_deferrable, bool, 0444);
 106
 107 /* If true, check all canary bytes on panic. */
 108 static bool kfence_check_on_panic __read_mostly;
 109 module_param_named(check_on_panic, kfence_check_on_panic, bool, 0444);
 110
 111 /* The pool of pages used for guard pages and objects. */
 112 char *__kfence_pool __read_mostly;
 113 EXPORT_SYMBOL(__kfence_pool); /* Export for test modules. */
 114
 115 /*
 116  * Per-object metadata, with one-to-one mapping of object metadata to
 117  * backing pages (in __kfence_pool).
 118  */
 119 static_assert(CONFIG_KFENCE_NUM_OBJECTS > 0);
 120 struct kfence_metadata kfence_metadata[CONFIG_KFENCE_NUM_OBJECTS];
 121
 122 /* Freelist with available objects. */
 123 static struct list_head kfence_freelist = LIST_HEAD_INIT(kfence_freelist);
 124 static DEFINE_RAW_SPINLOCK(kfence_freelist_lock); /* Lock protecting freelist. */
 125
 126 /*
 127  * The static key to set up a KFENCE allocation; or if static keys are not used
 128  * to gate allocations, to avoid a load and compare if KFENCE is disabled.
 129  */
 130 DEFINE_STATIC_KEY_FALSE(kfence_allocation_key);
 131
 132 /* Gates the allocation, ensuring only one succeeds in a given period. */
 133 atomic_t kfence_allocation_gate = ATOMIC_INIT(1);
 134
 135 /*
 136  * A Counting Bloom filter of allocation coverage: limits currently covered
 137  * allocations of the same source filling up the pool.
 138  *
 139  * Assuming a range of 15%-85% unique allocations in the pool at any point in
 140  * time, the below parameters provide a probablity of 0.02-0.33 for false
 141  * positive hits respectively:
 142  *
 143  *      P(alloc_traces) = (1 - e^(-HNUM * (alloc_traces / SIZE)) ^ HNUM
 144  */
 145 #define ALLOC_COVERED_HNUM      2
 146 #define ALLOC_COVERED_ORDER     (const_ilog2(CONFIG_KFENCE_NUM_OBJECTS) + 2)
 147 #define ALLOC_COVERED_SIZE      (1 << ALLOC_COVERED_ORDER)
 148 #define ALLOC_COVERED_HNEXT(h)  hash_32(h, ALLOC_COVERED_ORDER)
 149 #define ALLOC_COVERED_MASK      (ALLOC_COVERED_SIZE - 1)
 150 static atomic_t alloc_covered[ALLOC_COVERED_SIZE];
 151
 152 /* Stack depth used to determine uniqueness of an allocation. */
 153 #define UNIQUE_ALLOC_STACK_DEPTH ((size_t)8)
 154
 155 /*
 156  * Randomness for stack hashes, making the same collisions across reboots and
 157  * different machines less likely.
 158  */
 159 static u32 stack_hash_seed __ro_after_init;
 160
 161 /* Statistics counters for debugfs. */
 162 enum kfence_counter_id {
 163         KFENCE_COUNTER_ALLOCATED,
 164         KFENCE_COUNTER_ALLOCS,
 165         KFENCE_COUNTER_FREES,
 166         KFENCE_COUNTER_ZOMBIES,
 167         KFENCE_COUNTER_BUGS,
 168         KFENCE_COUNTER_SKIP_INCOMPAT,
 169         KFENCE_COUNTER_SKIP_CAPACITY,
 170         KFENCE_COUNTER_SKIP_COVERED,
 171         KFENCE_COUNTER_COUNT,
 172 };
 173 static atomic_long_t counters[KFENCE_COUNTER_COUNT];
 174 static const char *const counter_names[] = {
 175         [KFENCE_COUNTER_ALLOCATED]      = "currently allocated",
 176         [KFENCE_COUNTER_ALLOCS]         = "total allocations",
 177         [KFENCE_COUNTER_FREES]          = "total frees",
 178         [KFENCE_COUNTER_ZOMBIES]        = "zombie allocations",
 179         [KFENCE_COUNTER_BUGS]           = "total bugs",
 180         [KFENCE_COUNTER_SKIP_INCOMPAT]  = "skipped allocations (incompatible)",
 181         [KFENCE_COUNTER_SKIP_CAPACITY]  = "skipped allocations (capacity)",
 182         [KFENCE_COUNTER_SKIP_COVERED]   = "skipped allocations (covered)",
 183 };
 184 static_assert(ARRAY_SIZE(counter_names) == KFENCE_COUNTER_COUNT);
 185
 186 /* === Internals ============================================================ */
 187
 188 static inline bool should_skip_covered(void)
 189 {
 190         unsigned long thresh = (CONFIG_KFENCE_NUM_OBJECTS * kfence_skip_covered_thresh) / 100;
 191
 192         return atomic_long_read(&counters[KFENCE_COUNTER_ALLOCATED]) > thresh;
 193 }
 194
 195 static u32 get_alloc_stack_hash(unsigned long *stack_entries, size_t num_entries)
 196 {
 197         num_entries = min(num_entries, UNIQUE_ALLOC_STACK_DEPTH);
 198         num_entries = filter_irq_stacks(stack_entries, num_entries);
 199         return jhash(stack_entries, num_entries * sizeof(stack_entries[0]), stack_hash_seed);
 200 }
 201
 202 /*
 203  * Adds (or subtracts) count @val for allocation stack trace hash
 204  * @alloc_stack_hash from Counting Bloom filter.
 205  */
 206 static void alloc_covered_add(u32 alloc_stack_hash, int val)
 207 {
 208         int i;
 209
 210         for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
 211                 atomic_add(val, &alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]);
 212                 alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
 213         }
 214 }
 215
 216 /*
 217  * Returns true if the allocation stack trace hash @alloc_stack_hash is
 218  * currently contained (non-zero count) in Counting Bloom filter.
 219  */
 220 static bool alloc_covered_contains(u32 alloc_stack_hash)
 221 {
 222         int i;
 223
 224         for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
 225                 if (!atomic_read(&alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]))
 226                         return false;
 227                 alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
 228         }
 229
 230         return true;
 231 }
 232
 233 static bool kfence_protect(unsigned long addr)
 234 {
 235         return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), true));
 236 }
 237
 238 static bool kfence_unprotect(unsigned long addr)
 239 {
 240         return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), false));
 241 }
 242
 243 static inline unsigned long metadata_to_pageaddr(const struct kfence_metadata *meta)
 244 {
 245         unsigned long offset = (meta - kfence_metadata + 1) * PAGE_SIZE * 2;
 246         unsigned long pageaddr = (unsigned long)&__kfence_pool[offset];
 247
 248         /* The checks do not affect performance; only called from slow-paths. */
 249
 250         /* Only call with a pointer into kfence_metadata. */
 251         if (KFENCE_WARN_ON(meta < kfence_metadata ||
 252                            meta >= kfence_metadata + CONFIG_KFENCE_NUM_OBJECTS))
 253                 return 0;
 254
 255         /*
 256          * This metadata object only ever maps to 1 page; verify that the stored
 257          * address is in the expected range.
 258          */
 259         if (KFENCE_WARN_ON(ALIGN_DOWN(meta->addr, PAGE_SIZE) != pageaddr))
 260                 return 0;
 261
 262         return pageaddr;
 263 }
 264
 265 /*
 266  * Update the object's metadata state, including updating the alloc/free stacks
 267  * depending on the state transition.
 268  */
 269 static noinline void
 270 metadata_update_state(struct kfence_metadata *meta, enum kfence_object_state next,
 271                       unsigned long *stack_entries, size_t num_stack_entries)
 272 {
 273         struct kfence_track *track =
 274                 next == KFENCE_OBJECT_FREED ? &meta->free_track : &meta->alloc_track;
 275
 276         lockdep_assert_held(&meta->lock);
 277
 278         if (stack_entries) {
 279                 memcpy(track->stack_entries, stack_entries,
 280                        num_stack_entries * sizeof(stack_entries[0]));
 281         } else {
 282                 /*
 283                  * Skip over 1 (this) functions; noinline ensures we do not
 284                  * accidentally skip over the caller by never inlining.
 285                  */
 286                 num_stack_entries = stack_trace_save(track->stack_entries, KFENCE_STACK_DEPTH, 1);
 287         }
 288         track->num_stack_entries = num_stack_entries;
 289         track->pid = task_pid_nr(current);
 290         track->cpu = raw_smp_processor_id();
 291         track->ts_nsec = local_clock(); /* Same source as printk timestamps. */
 292
 293         /*
 294          * Pairs with READ_ONCE() in
 295          *      kfence_shutdown_cache(),
 296          *      kfence_handle_page_fault().
 297          */
 298         WRITE_ONCE(meta->state, next);
 299 }
 300
 301 /* Write canary byte to @addr. */
 302 static inline bool set_canary_byte(u8 *addr)
 303 {
 304         *addr = KFENCE_CANARY_PATTERN(addr);
 305         return true;
 306 }
 307
 308 /* Check canary byte at @addr. */
 309 static inline bool check_canary_byte(u8 *addr)
 310 {
 311         struct kfence_metadata *meta;
 312         unsigned long flags;
 313
 314         if (likely(*addr == KFENCE_CANARY_PATTERN(addr)))
 315                 return true;
 316
 317         atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
 318
 319         meta = addr_to_metadata((unsigned long)addr);
 320         raw_spin_lock_irqsave(&meta->lock, flags);
 321         kfence_report_error((unsigned long)addr, false, NULL, meta, KFENCE_ERROR_CORRUPTION);
 322         raw_spin_unlock_irqrestore(&meta->lock, flags);
 323
 324         return false;
 325 }
 326
 327 /* __always_inline this to ensure we won't do an indirect call to fn. */
 328 static __always_inline void for_each_canary(const struct kfence_metadata *meta, bool (*fn)(u8 *))
 329 {
 330         const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
 331         unsigned long addr;
 332
 333         /*
 334          * We'll iterate over each canary byte per-side until fn() returns
 335          * false. However, we'll still iterate over the canary bytes to the
 336          * right of the object even if there was an error in the canary bytes to
 337          * the left of the object. Specifically, if check_canary_byte()
 338          * generates an error, showing both sides might give more clues as to
 339          * what the error is about when displaying which bytes were corrupted.
 340          */
 341
 342         /* Apply to left of object. */
 343         for (addr = pageaddr; addr < meta->addr; addr++) {
 344                 if (!fn((u8 *)addr))
 345                         break;
 346         }
 347
 348         /* Apply to right of object. */
 349         for (addr = meta->addr + meta->size; addr < pageaddr + PAGE_SIZE; addr++) {
 350                 if (!fn((u8 *)addr))
 351                         break;
 352         }
 353 }
 354
 355 static void *kfence_guarded_alloc(struct kmem_cache *cache, size_t size, gfp_t gfp,
 356                                   unsigned long *stack_entries, size_t num_stack_entries,
 357                                   u32 alloc_stack_hash)
 358 {
 359         struct kfence_metadata *meta = NULL;
 360         unsigned long flags;
 361         struct slab *slab;
 362         void *addr;
 363         const bool random_right_allocate = prandom_u32_max(2);
 364         const bool random_fault = CONFIG_KFENCE_STRESS_TEST_FAULTS &&
 365                                   !prandom_u32_max(CONFIG_KFENCE_STRESS_TEST_FAULTS);
 366
 367         /* Try to obtain a free object. */
 368         raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
 369         if (!list_empty(&kfence_freelist)) {
 370                 meta = list_entry(kfence_freelist.next, struct kfence_metadata, list);
 371                 list_del_init(&meta->list);
 372         }
 373         raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
 374         if (!meta) {
 375                 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_CAPACITY]);
 376                 return NULL;
 377         }
 378
 379         if (unlikely(!raw_spin_trylock_irqsave(&meta->lock, flags))) {
 380                 /*
 381                  * This is extremely unlikely -- we are reporting on a
 382                  * use-after-free, which locked meta->lock, and the reporting
 383                  * code via printk calls kmalloc() which ends up in
 384                  * kfence_alloc() and tries to grab the same object that we're
 385                  * reporting on. While it has never been observed, lockdep does
 386                  * report that there is a possibility of deadlock. Fix it by
 387                  * using trylock and bailing out gracefully.
 388                  */
 389                 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
 390                 /* Put the object back on the freelist. */
 391                 list_add_tail(&meta->list, &kfence_freelist);
 392                 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
 393
 394                 return NULL;
 395         }
 396
 397         meta->addr = metadata_to_pageaddr(meta);
 398         /* Unprotect if we're reusing this page. */
 399         if (meta->state == KFENCE_OBJECT_FREED)
 400                 kfence_unprotect(meta->addr);
 401
 402         /*
 403          * Note: for allocations made before RNG initialization, will always
 404          * return zero. We still benefit from enabling KFENCE as early as
 405          * possible, even when the RNG is not yet available, as this will allow
 406          * KFENCE to detect bugs due to earlier allocations. The only downside
 407          * is that the out-of-bounds accesses detected are deterministic for
 408          * such allocations.
 409          */
 410         if (random_right_allocate) {
 411                 /* Allocate on the "right" side, re-calculate address. */
 412                 meta->addr += PAGE_SIZE - size;
 413                 meta->addr = ALIGN_DOWN(meta->addr, cache->align);
 414         }
 415
 416         addr = (void *)meta->addr;
 417
 418         /* Update remaining metadata. */
 419         metadata_update_state(meta, KFENCE_OBJECT_ALLOCATED, stack_entries, num_stack_entries);
 420         /* Pairs with READ_ONCE() in kfence_shutdown_cache(). */
 421         WRITE_ONCE(meta->cache, cache);
 422         meta->size = size;
 423         meta->alloc_stack_hash = alloc_stack_hash;
 424         raw_spin_unlock_irqrestore(&meta->lock, flags);
 425
 426         alloc_covered_add(alloc_stack_hash, 1);
 427
 428         /* Set required slab fields. */
 429         slab = virt_to_slab((void *)meta->addr);
 430         slab->slab_cache = cache;
 431 #if defined(CONFIG_SLUB)
 432         slab->objects = 1;
 433 #elif defined(CONFIG_SLAB)
 434         slab->s_mem = addr;
 435 #endif
 436
 437         /* Memory initialization. */
 438         for_each_canary(meta, set_canary_byte);
 439
 440         /*
 441          * We check slab_want_init_on_alloc() ourselves, rather than letting
 442          * SL*B do the initialization, as otherwise we might overwrite KFENCE's
 443          * redzone.
 444          */
 445         if (unlikely(slab_want_init_on_alloc(gfp, cache)))
 446                 memzero_explicit(addr, size);
 447         if (cache->ctor)
 448                 cache->ctor(addr);
 449
 450         if (random_fault)
 451                 kfence_protect(meta->addr); /* Random "faults" by protecting the object. */
 452
 453         atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCATED]);
 454         atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCS]);
 455
 456         return addr;
 457 }
 458
 459 static void kfence_guarded_free(void *addr, struct kfence_metadata *meta, bool zombie)
 460 {
 461         struct kcsan_scoped_access assert_page_exclusive;
 462         unsigned long flags;
 463         bool init;
 464
 465         raw_spin_lock_irqsave(&meta->lock, flags);
 466
 467         if (meta->state != KFENCE_OBJECT_ALLOCATED || meta->addr != (unsigned long)addr) {
 468                 /* Invalid or double-free, bail out. */
 469                 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
 470                 kfence_report_error((unsigned long)addr, false, NULL, meta,
 471                                     KFENCE_ERROR_INVALID_FREE);
 472                 raw_spin_unlock_irqrestore(&meta->lock, flags);
 473                 return;
 474         }
 475
 476         /* Detect racy use-after-free, or incorrect reallocation of this page by KFENCE. */
 477         kcsan_begin_scoped_access((void *)ALIGN_DOWN((unsigned long)addr, PAGE_SIZE), PAGE_SIZE,
 478                                   KCSAN_ACCESS_SCOPED | KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT,
 479                                   &assert_page_exclusive);
 480
 481         if (CONFIG_KFENCE_STRESS_TEST_FAULTS)
 482                 kfence_unprotect((unsigned long)addr); /* To check canary bytes. */
 483
 484         /* Restore page protection if there was an OOB access. */
 485         if (meta->unprotected_page) {
 486                 memzero_explicit((void *)ALIGN_DOWN(meta->unprotected_page, PAGE_SIZE), PAGE_SIZE);
 487                 kfence_protect(meta->unprotected_page);
 488                 meta->unprotected_page = 0;
 489         }
 490
 491         /* Mark the object as freed. */
 492         metadata_update_state(meta, KFENCE_OBJECT_FREED, NULL, 0);
 493         init = slab_want_init_on_free(meta->cache);
 494         raw_spin_unlock_irqrestore(&meta->lock, flags);
 495
 496         alloc_covered_add(meta->alloc_stack_hash, -1);
 497
 498         /* Check canary bytes for memory corruption. */
 499         for_each_canary(meta, check_canary_byte);
 500
 501         /*
 502          * Clear memory if init-on-free is set. While we protect the page, the
 503          * data is still there, and after a use-after-free is detected, we
 504          * unprotect the page, so the data is still accessible.
 505          */
 506         if (!zombie && unlikely(init))
 507                 memzero_explicit(addr, meta->size);
 508
 509         /* Protect to detect use-after-frees. */
 510         kfence_protect((unsigned long)addr);
 511
 512         kcsan_end_scoped_access(&assert_page_exclusive);
 513         if (!zombie) {
 514                 /* Add it to the tail of the freelist for reuse. */
 515                 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
 516                 KFENCE_WARN_ON(!list_empty(&meta->list));
 517                 list_add_tail(&meta->list, &kfence_freelist);
 518                 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
 519
 520                 atomic_long_dec(&counters[KFENCE_COUNTER_ALLOCATED]);
 521                 atomic_long_inc(&counters[KFENCE_COUNTER_FREES]);
 522         } else {
 523                 /* See kfence_shutdown_cache(). */
 524                 atomic_long_inc(&counters[KFENCE_COUNTER_ZOMBIES]);
 525         }
 526 }
 527
 528 static void rcu_guarded_free(struct rcu_head *h)
 529 {
 530         struct kfence_metadata *meta = container_of(h, struct kfence_metadata, rcu_head);
 531
 532         kfence_guarded_free((void *)meta->addr, meta, false);
 533 }
 534
 535 /*
 536  * Initialization of the KFENCE pool after its allocation.
 537  * Returns 0 on success; otherwise returns the address up to
 538  * which partial initialization succeeded.
 539  */
 540 static unsigned long kfence_init_pool(void)
 541 {
 542         unsigned long addr = (unsigned long)__kfence_pool;
 543         struct page *pages;
 544         int i;
 545
 546         if (!arch_kfence_init_pool())
 547                 return addr;
 548
 549         pages = virt_to_page(__kfence_pool);
 550
 551         /*
 552          * Set up object pages: they must have PG_slab set, to avoid freeing
 553          * these as real pages.
 554          *
 555          * We also want to avoid inserting kfence_free() in the kfree()
 556          * fast-path in SLUB, and therefore need to ensure kfree() correctly
 557          * enters __slab_free() slow-path.
 558          */
 559         for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
 560                 struct slab *slab = page_slab(&pages[i]);
 561
 562                 if (!i || (i % 2))
 563                         continue;
 564
 565                 /* Verify we do not have a compound head page. */
 566                 if (WARN_ON(compound_head(&pages[i]) != &pages[i]))
 567                         return addr;
 568
 569                 __folio_set_slab(slab_folio(slab));
 570 #ifdef CONFIG_MEMCG
 571                 slab->memcg_data = (unsigned long)&kfence_metadata[i / 2 - 1].objcg |
 572                                    MEMCG_DATA_OBJCGS;
 573 #endif
 574         }
 575
 576         /*
 577          * Protect the first 2 pages. The first page is mostly unnecessary, and
 578          * merely serves as an extended guard page. However, adding one
 579          * additional page in the beginning gives us an even number of pages,
 580          * which simplifies the mapping of address to metadata index.
 581          */
 582         for (i = 0; i < 2; i++) {
 583                 if (unlikely(!kfence_protect(addr)))
 584                         return addr;
 585
 586                 addr += PAGE_SIZE;
 587         }
 588
 589         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 590                 struct kfence_metadata *meta = &kfence_metadata[i];
 591
 592                 /* Initialize metadata. */
 593                 INIT_LIST_HEAD(&meta->list);
 594                 raw_spin_lock_init(&meta->lock);
 595                 meta->state = KFENCE_OBJECT_UNUSED;
 596                 meta->addr = addr; /* Initialize for validation in metadata_to_pageaddr(). */
 597                 list_add_tail(&meta->list, &kfence_freelist);
 598
 599                 /* Protect the right redzone. */
 600                 if (unlikely(!kfence_protect(addr + PAGE_SIZE)))
 601                         return addr;
 602
 603                 addr += 2 * PAGE_SIZE;
 604         }
 605
 606         return 0;
 607 }
 608
 609 static bool __init kfence_init_pool_early(void)
 610 {
 611         unsigned long addr;
 612
 613         if (!__kfence_pool)
 614                 return false;
 615
 616         addr = kfence_init_pool();
 617
 618         if (!addr) {
 619                 /*
 620                  * The pool is live and will never be deallocated from this point on.
 621                  * Ignore the pool object from the kmemleak phys object tree, as it would
 622                  * otherwise overlap with allocations returned by kfence_alloc(), which
 623                  * are registered with kmemleak through the slab post-alloc hook.
 624                  */
 625                 kmemleak_ignore_phys(__pa(__kfence_pool));
 626                 return true;
 627         }
 628
 629         /*
 630          * Only release unprotected pages, and do not try to go back and change
 631          * page attributes due to risk of failing to do so as well. If changing
 632          * page attributes for some pages fails, it is very likely that it also
 633          * fails for the first page, and therefore expect addr==__kfence_pool in
 634          * most failure cases.
 635          */
 636         for (char *p = (char *)addr; p < __kfence_pool + KFENCE_POOL_SIZE; p += PAGE_SIZE) {
 637                 struct slab *slab = virt_to_slab(p);
 638
 639                 if (!slab)
 640                         continue;
 641 #ifdef CONFIG_MEMCG
 642                 slab->memcg_data = 0;
 643 #endif
 644                 __folio_clear_slab(slab_folio(slab));
 645         }
 646         memblock_free_late(__pa(addr), KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool));
 647         __kfence_pool = NULL;
 648         return false;
 649 }
 650
 651 static bool kfence_init_pool_late(void)
 652 {
 653         unsigned long addr, free_size;
 654
 655         addr = kfence_init_pool();
 656
 657         if (!addr)
 658                 return true;
 659
 660         /* Same as above. */
 661         free_size = KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool);
 662 #ifdef CONFIG_CONTIG_ALLOC
 663         free_contig_range(page_to_pfn(virt_to_page((void *)addr)), free_size / PAGE_SIZE);
 664 #else
 665         free_pages_exact((void *)addr, free_size);
 666 #endif
 667         __kfence_pool = NULL;
 668         return false;
 669 }
 670
 671 /* === DebugFS Interface ==================================================== */
 672
 673 static int stats_show(struct seq_file *seq, void *v)
 674 {
 675         int i;
 676
 677         seq_printf(seq, "enabled: %i\n", READ_ONCE(kfence_enabled));
 678         for (i = 0; i < KFENCE_COUNTER_COUNT; i++)
 679                 seq_printf(seq, "%s: %ld\n", counter_names[i], atomic_long_read(&counters[i]));
 680
 681         return 0;
 682 }
 683 DEFINE_SHOW_ATTRIBUTE(stats);
 684
 685 /*
 686  * debugfs seq_file operations for /sys/kernel/debug/kfence/objects.
 687  * start_object() and next_object() return the object index + 1, because NULL is used
 688  * to stop iteration.
 689  */
 690 static void *start_object(struct seq_file *seq, loff_t *pos)
 691 {
 692         if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
 693                 return (void *)((long)*pos + 1);
 694         return NULL;
 695 }
 696
 697 static void stop_object(struct seq_file *seq, void *v)
 698 {
 699 }
 700
 701 static void *next_object(struct seq_file *seq, void *v, loff_t *pos)
 702 {
 703         ++*pos;
 704         if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
 705                 return (void *)((long)*pos + 1);
 706         return NULL;
 707 }
 708
 709 static int show_object(struct seq_file *seq, void *v)
 710 {
 711         struct kfence_metadata *meta = &kfence_metadata[(long)v - 1];
 712         unsigned long flags;
 713
 714         raw_spin_lock_irqsave(&meta->lock, flags);
 715         kfence_print_object(seq, meta);
 716         raw_spin_unlock_irqrestore(&meta->lock, flags);
 717         seq_puts(seq, "---------------------------------\n");
 718
 719         return 0;
 720 }
 721
 722 static const struct seq_operations objects_sops = {
 723         .start = start_object,
 724         .next = next_object,
 725         .stop = stop_object,
 726         .show = show_object,
 727 };
 728 DEFINE_SEQ_ATTRIBUTE(objects);
 729
 730 static int __init kfence_debugfs_init(void)
 731 {
 732         struct dentry *kfence_dir = debugfs_create_dir("kfence", NULL);
 733
 734         debugfs_create_file("stats", 0444, kfence_dir, NULL, &stats_fops);
 735         debugfs_create_file("objects", 0400, kfence_dir, NULL, &objects_fops);
 736         return 0;
 737 }
 738
 739 late_initcall(kfence_debugfs_init);
 740
 741 /* === Panic Notifier ====================================================== */
 742
 743 static void kfence_check_all_canary(void)
 744 {
 745         int i;
 746
 747         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 748                 struct kfence_metadata *meta = &kfence_metadata[i];
 749
 750                 if (meta->state == KFENCE_OBJECT_ALLOCATED)
 751                         for_each_canary(meta, check_canary_byte);
 752         }
 753 }
 754
 755 static int kfence_check_canary_callback(struct notifier_block *nb,
 756                                         unsigned long reason, void *arg)
 757 {
 758         kfence_check_all_canary();
 759         return NOTIFY_OK;
 760 }
 761
 762 static struct notifier_block kfence_check_canary_notifier = {
 763         .notifier_call = kfence_check_canary_callback,
 764 };
 765
 766 /* === Allocation Gate Timer ================================================ */
 767
 768 static struct delayed_work kfence_timer;
 769
 770 #ifdef CONFIG_KFENCE_STATIC_KEYS
 771 /* Wait queue to wake up allocation-gate timer task. */
 772 static DECLARE_WAIT_QUEUE_HEAD(allocation_wait);
 773
 774 static void wake_up_kfence_timer(struct irq_work *work)
 775 {
 776         wake_up(&allocation_wait);
 777 }
 778 static DEFINE_IRQ_WORK(wake_up_kfence_timer_work, wake_up_kfence_timer);
 779 #endif
 780
 781 /*
 782  * Set up delayed work, which will enable and disable the static key. We need to
 783  * use a work queue (rather than a simple timer), since enabling and disabling a
 784  * static key cannot be done from an interrupt.
 785  *
 786  * Note: Toggling a static branch currently causes IPIs, and here we'll end up
 787  * with a total of 2 IPIs to all CPUs. If this ends up a problem in future (with
 788  * more aggressive sampling intervals), we could get away with a variant that
 789  * avoids IPIs, at the cost of not immediately capturing allocations if the
 790  * instructions remain cached.
 791  */
 792 static void toggle_allocation_gate(struct work_struct *work)
 793 {
 794         if (!READ_ONCE(kfence_enabled))
 795                 return;
 796
 797         atomic_set(&kfence_allocation_gate, 0);
 798 #ifdef CONFIG_KFENCE_STATIC_KEYS
 799         /* Enable static key, and await allocation to happen. */
 800         static_branch_enable(&kfence_allocation_key);
 801
 802         if (sysctl_hung_task_timeout_secs) {
 803                 /*
 804                  * During low activity with no allocations we might wait a
 805                  * while; let's avoid the hung task warning.
 806                  */
 807                 wait_event_idle_timeout(allocation_wait, atomic_read(&kfence_allocation_gate),
 808                                         sysctl_hung_task_timeout_secs * HZ / 2);
 809         } else {
 810                 wait_event_idle(allocation_wait, atomic_read(&kfence_allocation_gate));
 811         }
 812
 813         /* Disable static key and reset timer. */
 814         static_branch_disable(&kfence_allocation_key);
 815 #endif
 816         queue_delayed_work(system_unbound_wq, &kfence_timer,
 817                            msecs_to_jiffies(kfence_sample_interval));
 818 }
 819
 820 /* === Public interface ===================================================== */
 821
 822 void __init kfence_alloc_pool(void)
 823 {
 824         if (!kfence_sample_interval)
 825                 return;
 826
 827         __kfence_pool = memblock_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
 828
 829         if (!__kfence_pool)
 830                 pr_err("failed to allocate pool\n");
 831 }
 832
 833 static void kfence_init_enable(void)
 834 {
 835         if (!IS_ENABLED(CONFIG_KFENCE_STATIC_KEYS))
 836                 static_branch_enable(&kfence_allocation_key);
 837
 838         if (kfence_deferrable)
 839                 INIT_DEFERRABLE_WORK(&kfence_timer, toggle_allocation_gate);
 840         else
 841                 INIT_DELAYED_WORK(&kfence_timer, toggle_allocation_gate);
 842
 843         if (kfence_check_on_panic)
 844                 atomic_notifier_chain_register(&panic_notifier_list, &kfence_check_canary_notifier);
 845
 846         WRITE_ONCE(kfence_enabled, true);
 847         queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
 848
 849         pr_info("initialized - using %lu bytes for %d objects at 0x%p-0x%p\n", KFENCE_POOL_SIZE,
 850                 CONFIG_KFENCE_NUM_OBJECTS, (void *)__kfence_pool,
 851                 (void *)(__kfence_pool + KFENCE_POOL_SIZE));
 852 }
 853
 854 void __init kfence_init(void)
 855 {
 856         stack_hash_seed = get_random_u32();
 857
 858         /* Setting kfence_sample_interval to 0 on boot disables KFENCE. */
 859         if (!kfence_sample_interval)
 860                 return;
 861
 862         if (!kfence_init_pool_early()) {
 863                 pr_err("%s failed\n", __func__);
 864                 return;
 865         }
 866
 867         kfence_init_enable();
 868 }
 869
 870 static int kfence_init_late(void)
 871 {
 872         const unsigned long nr_pages = KFENCE_POOL_SIZE / PAGE_SIZE;
 873 #ifdef CONFIG_CONTIG_ALLOC
 874         struct page *pages;
 875
 876         pages = alloc_contig_pages(nr_pages, GFP_KERNEL, first_online_node, NULL);
 877         if (!pages)
 878                 return -ENOMEM;
 879         __kfence_pool = page_to_virt(pages);
 880 #else
 881         if (nr_pages > MAX_ORDER_NR_PAGES) {
 882                 pr_warn("KFENCE_NUM_OBJECTS too large for buddy allocator\n");
 883                 return -EINVAL;
 884         }
 885         __kfence_pool = alloc_pages_exact(KFENCE_POOL_SIZE, GFP_KERNEL);
 886         if (!__kfence_pool)
 887                 return -ENOMEM;
 888 #endif
 889
 890         if (!kfence_init_pool_late()) {
 891                 pr_err("%s failed\n", __func__);
 892                 return -EBUSY;
 893         }
 894
 895         kfence_init_enable();
 896         return 0;
 897 }
 898
 899 static int kfence_enable_late(void)
 900 {
 901         if (!__kfence_pool)
 902                 return kfence_init_late();
 903
 904         WRITE_ONCE(kfence_enabled, true);
 905         queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
 906         pr_info("re-enabled\n");
 907         return 0;
 908 }
 909
 910 void kfence_shutdown_cache(struct kmem_cache *s)
 911 {
 912         unsigned long flags;
 913         struct kfence_metadata *meta;
 914         int i;
 915
 916         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 917                 bool in_use;
 918
 919                 meta = &kfence_metadata[i];
 920
 921                 /*
 922                  * If we observe some inconsistent cache and state pair where we
 923                  * should have returned false here, cache destruction is racing
 924                  * with either kmem_cache_alloc() or kmem_cache_free(). Taking
 925                  * the lock will not help, as different critical section
 926                  * serialization will have the same outcome.
 927                  */
 928                 if (READ_ONCE(meta->cache) != s ||
 929                     READ_ONCE(meta->state) != KFENCE_OBJECT_ALLOCATED)
 930                         continue;
 931
 932                 raw_spin_lock_irqsave(&meta->lock, flags);
 933                 in_use = meta->cache == s && meta->state == KFENCE_OBJECT_ALLOCATED;
 934                 raw_spin_unlock_irqrestore(&meta->lock, flags);
 935
 936                 if (in_use) {
 937                         /*
 938                          * This cache still has allocations, and we should not
 939                          * release them back into the freelist so they can still
 940                          * safely be used and retain the kernel's default
 941                          * behaviour of keeping the allocations alive (leak the
 942                          * cache); however, they effectively become "zombie
 943                          * allocations" as the KFENCE objects are the only ones
 944                          * still in use and the owning cache is being destroyed.
 945                          *
 946                          * We mark them freed, so that any subsequent use shows
 947                          * more useful error messages that will include stack
 948                          * traces of the user of the object, the original
 949                          * allocation, and caller to shutdown_cache().
 950                          */
 951                         kfence_guarded_free((void *)meta->addr, meta, /*zombie=*/true);
 952                 }
 953         }
 954
 955         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 956                 meta = &kfence_metadata[i];
 957
 958                 /* See above. */
 959                 if (READ_ONCE(meta->cache) != s || READ_ONCE(meta->state) != KFENCE_OBJECT_FREED)
 960                         continue;
 961
 962                 raw_spin_lock_irqsave(&meta->lock, flags);
 963                 if (meta->cache == s && meta->state == KFENCE_OBJECT_FREED)
 964                         meta->cache = NULL;
 965                 raw_spin_unlock_irqrestore(&meta->lock, flags);
 966         }
 967 }
 968
 969 void *__kfence_alloc(struct kmem_cache *s, size_t size, gfp_t flags)
 970 {
 971         unsigned long stack_entries[KFENCE_STACK_DEPTH];
 972         size_t num_stack_entries;
 973         u32 alloc_stack_hash;
 974
 975         /*
 976          * Perform size check before switching kfence_allocation_gate, so that
 977          * we don't disable KFENCE without making an allocation.
 978          */
 979         if (size > PAGE_SIZE) {
 980                 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
 981                 return NULL;
 982         }
 983
 984         /*
 985          * Skip allocations from non-default zones, including DMA. We cannot
 986          * guarantee that pages in the KFENCE pool will have the requested
 987          * properties (e.g. reside in DMAable memory).
 988          */
 989         if ((flags & GFP_ZONEMASK) ||
 990             (s->flags & (SLAB_CACHE_DMA | SLAB_CACHE_DMA32))) {
 991                 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
 992                 return NULL;
 993         }
 994
 995         /*
 996          * Skip allocations for this slab, if KFENCE has been disabled for
 997          * this slab.
 998          */
 999         if (s->flags & SLAB_SKIP_KFENCE)
1000                 return NULL;
1001
1002         if (atomic_inc_return(&kfence_allocation_gate) > 1)
1003                 return NULL;
1004 #ifdef CONFIG_KFENCE_STATIC_KEYS
1005         /*
1006          * waitqueue_active() is fully ordered after the update of
1007          * kfence_allocation_gate per atomic_inc_return().
1008          */
1009         if (waitqueue_active(&allocation_wait)) {
1010                 /*
1011                  * Calling wake_up() here may deadlock when allocations happen
1012                  * from within timer code. Use an irq_work to defer it.
1013                  */
1014                 irq_work_queue(&wake_up_kfence_timer_work);
1015         }
1016 #endif
1017
1018         if (!READ_ONCE(kfence_enabled))
1019                 return NULL;
1020
1021         num_stack_entries = stack_trace_save(stack_entries, KFENCE_STACK_DEPTH, 0);
1022
1023         /*
1024          * Do expensive check for coverage of allocation in slow-path after
1025          * allocation_gate has already become non-zero, even though it might
1026          * mean not making any allocation within a given sample interval.
1027          *
1028          * This ensures reasonable allocation coverage when the pool is almost
1029          * full, including avoiding long-lived allocations of the same source
1030          * filling up the pool (e.g. pagecache allocations).
1031          */
1032         alloc_stack_hash = get_alloc_stack_hash(stack_entries, num_stack_entries);
1033         if (should_skip_covered() && alloc_covered_contains(alloc_stack_hash)) {
1034                 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_COVERED]);
1035                 return NULL;
1036         }
1037
1038         return kfence_guarded_alloc(s, size, flags, stack_entries, num_stack_entries,
1039                                     alloc_stack_hash);
1040 }
1041
1042 size_t kfence_ksize(const void *addr)
1043 {
1044         const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1045
1046         /*
1047          * Read locklessly -- if there is a race with __kfence_alloc(), this is
1048          * either a use-after-free or invalid access.
1049          */
1050         return meta ? meta->size : 0;
1051 }
1052
1053 void *kfence_object_start(const void *addr)
1054 {
1055         const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1056
1057         /*
1058          * Read locklessly -- if there is a race with __kfence_alloc(), this is
1059          * either a use-after-free or invalid access.
1060          */
1061         return meta ? (void *)meta->addr : NULL;
1062 }
1063
1064 void __kfence_free(void *addr)
1065 {
1066         struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1067
1068 #ifdef CONFIG_MEMCG
1069         KFENCE_WARN_ON(meta->objcg);
1070 #endif
1071         /*
1072          * If the objects of the cache are SLAB_TYPESAFE_BY_RCU, defer freeing
1073          * the object, as the object page may be recycled for other-typed
1074          * objects once it has been freed. meta->cache may be NULL if the cache
1075          * was destroyed.
1076          */
1077         if (unlikely(meta->cache && (meta->cache->flags & SLAB_TYPESAFE_BY_RCU)))
1078                 call_rcu(&meta->rcu_head, rcu_guarded_free);
1079         else
1080                 kfence_guarded_free(addr, meta, false);
1081 }
1082
1083 bool kfence_handle_page_fault(unsigned long addr, bool is_write, struct pt_regs *regs)
1084 {
1085         const int page_index = (addr - (unsigned long)__kfence_pool) / PAGE_SIZE;
1086         struct kfence_metadata *to_report = NULL;
1087         enum kfence_error_type error_type;
1088         unsigned long flags;
1089
1090         if (!is_kfence_address((void *)addr))
1091                 return false;
1092
1093         if (!READ_ONCE(kfence_enabled)) /* If disabled at runtime ... */
1094                 return kfence_unprotect(addr); /* ... unprotect and proceed. */
1095
1096         atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
1097
1098         if (page_index % 2) {
1099                 /* This is a redzone, report a buffer overflow. */
1100                 struct kfence_metadata *meta;
1101                 int distance = 0;
1102
1103                 meta = addr_to_metadata(addr - PAGE_SIZE);
1104                 if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
1105                         to_report = meta;
1106                         /* Data race ok; distance calculation approximate. */
1107                         distance = addr - data_race(meta->addr + meta->size);
1108                 }
1109
1110                 meta = addr_to_metadata(addr + PAGE_SIZE);
1111                 if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
1112                         /* Data race ok; distance calculation approximate. */
1113                         if (!to_report || distance > data_race(meta->addr) - addr)
1114                                 to_report = meta;
1115                 }
1116
1117                 if (!to_report)
1118                         goto out;
1119
1120                 raw_spin_lock_irqsave(&to_report->lock, flags);
1121                 to_report->unprotected_page = addr;
1122                 error_type = KFENCE_ERROR_OOB;
1123
1124                 /*
1125                  * If the object was freed before we took the look we can still
1126                  * report this as an OOB -- the report will simply show the
1127                  * stacktrace of the free as well.
1128                  */
1129         } else {
1130                 to_report = addr_to_metadata(addr);
1131                 if (!to_report)
1132                         goto out;
1133
1134                 raw_spin_lock_irqsave(&to_report->lock, flags);
1135                 error_type = KFENCE_ERROR_UAF;
1136                 /*
1137                  * We may race with __kfence_alloc(), and it is possible that a
1138                  * freed object may be reallocated. We simply report this as a
1139                  * use-after-free, with the stack trace showing the place where
1140                  * the object was re-allocated.
1141                  */
1142         }
1143
1144 out:
1145         if (to_report) {
1146                 kfence_report_error(addr, is_write, regs, to_report, error_type);
1147                 raw_spin_unlock_irqrestore(&to_report->lock, flags);
1148         } else {
1149                 /* This may be a UAF or OOB access, but we can't be sure. */
1150                 kfence_report_error(addr, is_write, regs, NULL, KFENCE_ERROR_INVALID);
1151         }
1152
1153         return kfence_unprotect(addr); /* Unprotect and let access proceed. */
1154 }