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/random.h>
  25 #include <linux/rcupdate.h>
  26 #include <linux/sched/clock.h>
  27 #include <linux/sched/sysctl.h>
  28 #include <linux/seq_file.h>
  29 #include <linux/slab.h>
  30 #include <linux/spinlock.h>
  31 #include <linux/string.h>
  32
  33 #include <asm/kfence.h>
  34
  35 #include "kfence.h"
  36
  37 /* Disables KFENCE on the first warning assuming an irrecoverable error. */
  38 #define KFENCE_WARN_ON(cond)                                                   \
  39         ({                                                                     \
  40                 const bool __cond = WARN_ON(cond);                             \
  41                 if (unlikely(__cond))                                          \
  42                         WRITE_ONCE(kfence_enabled, false);                     \
  43                 __cond;                                                        \
  44         })
  45
  46 /* === Data ================================================================= */
  47
  48 static bool kfence_enabled __read_mostly;
  49
  50 unsigned long kfence_sample_interval __read_mostly = CONFIG_KFENCE_SAMPLE_INTERVAL;
  51 EXPORT_SYMBOL_GPL(kfence_sample_interval); /* Export for test modules. */
  52
  53 #ifdef MODULE_PARAM_PREFIX
  54 #undef MODULE_PARAM_PREFIX
  55 #endif
  56 #define MODULE_PARAM_PREFIX "kfence."
  57
  58 static int param_set_sample_interval(const char *val, const struct kernel_param *kp)
  59 {
  60         unsigned long num;
  61         int ret = kstrtoul(val, 0, &num);
  62
  63         if (ret < 0)
  64                 return ret;
  65
  66         if (!num) /* Using 0 to indicate KFENCE is disabled. */
  67                 WRITE_ONCE(kfence_enabled, false);
  68         else if (!READ_ONCE(kfence_enabled) && system_state != SYSTEM_BOOTING)
  69                 return -EINVAL; /* Cannot (re-)enable KFENCE on-the-fly. */
  70
  71         *((unsigned long *)kp->arg) = num;
  72         return 0;
  73 }
  74
  75 static int param_get_sample_interval(char *buffer, const struct kernel_param *kp)
  76 {
  77         if (!READ_ONCE(kfence_enabled))
  78                 return sprintf(buffer, "0\n");
  79
  80         return param_get_ulong(buffer, kp);
  81 }
  82
  83 static const struct kernel_param_ops sample_interval_param_ops = {
  84         .set = param_set_sample_interval,
  85         .get = param_get_sample_interval,
  86 };
  87 module_param_cb(sample_interval, &sample_interval_param_ops, &kfence_sample_interval, 0600);
  88
  89 /* Pool usage% threshold when currently covered allocations are skipped. */
  90 static unsigned long kfence_skip_covered_thresh __read_mostly = 75;
  91 module_param_named(skip_covered_thresh, kfence_skip_covered_thresh, ulong, 0644);
  92
  93 /* The pool of pages used for guard pages and objects. */
  94 char *__kfence_pool __ro_after_init;
  95 EXPORT_SYMBOL(__kfence_pool); /* Export for test modules. */
  96
  97 /*
  98  * Per-object metadata, with one-to-one mapping of object metadata to
  99  * backing pages (in __kfence_pool).
 100  */
 101 static_assert(CONFIG_KFENCE_NUM_OBJECTS > 0);
 102 struct kfence_metadata kfence_metadata[CONFIG_KFENCE_NUM_OBJECTS];
 103
 104 /* Freelist with available objects. */
 105 static struct list_head kfence_freelist = LIST_HEAD_INIT(kfence_freelist);
 106 static DEFINE_RAW_SPINLOCK(kfence_freelist_lock); /* Lock protecting freelist. */
 107
 108 /*
 109  * The static key to set up a KFENCE allocation; or if static keys are not used
 110  * to gate allocations, to avoid a load and compare if KFENCE is disabled.
 111  */
 112 DEFINE_STATIC_KEY_FALSE(kfence_allocation_key);
 113
 114 /* Gates the allocation, ensuring only one succeeds in a given period. */
 115 atomic_t kfence_allocation_gate = ATOMIC_INIT(1);
 116
 117 /*
 118  * A Counting Bloom filter of allocation coverage: limits currently covered
 119  * allocations of the same source filling up the pool.
 120  *
 121  * Assuming a range of 15%-85% unique allocations in the pool at any point in
 122  * time, the below parameters provide a probablity of 0.02-0.33 for false
 123  * positive hits respectively:
 124  *
 125  *      P(alloc_traces) = (1 - e^(-HNUM * (alloc_traces / SIZE)) ^ HNUM
 126  */
 127 #define ALLOC_COVERED_HNUM      2
 128 #define ALLOC_COVERED_ORDER     (const_ilog2(CONFIG_KFENCE_NUM_OBJECTS) + 2)
 129 #define ALLOC_COVERED_SIZE      (1 << ALLOC_COVERED_ORDER)
 130 #define ALLOC_COVERED_HNEXT(h)  hash_32(h, ALLOC_COVERED_ORDER)
 131 #define ALLOC_COVERED_MASK      (ALLOC_COVERED_SIZE - 1)
 132 static atomic_t alloc_covered[ALLOC_COVERED_SIZE];
 133
 134 /* Stack depth used to determine uniqueness of an allocation. */
 135 #define UNIQUE_ALLOC_STACK_DEPTH ((size_t)8)
 136
 137 /*
 138  * Randomness for stack hashes, making the same collisions across reboots and
 139  * different machines less likely.
 140  */
 141 static u32 stack_hash_seed __ro_after_init;
 142
 143 /* Statistics counters for debugfs. */
 144 enum kfence_counter_id {
 145         KFENCE_COUNTER_ALLOCATED,
 146         KFENCE_COUNTER_ALLOCS,
 147         KFENCE_COUNTER_FREES,
 148         KFENCE_COUNTER_ZOMBIES,
 149         KFENCE_COUNTER_BUGS,
 150         KFENCE_COUNTER_SKIP_INCOMPAT,
 151         KFENCE_COUNTER_SKIP_CAPACITY,
 152         KFENCE_COUNTER_SKIP_COVERED,
 153         KFENCE_COUNTER_COUNT,
 154 };
 155 static atomic_long_t counters[KFENCE_COUNTER_COUNT];
 156 static const char *const counter_names[] = {
 157         [KFENCE_COUNTER_ALLOCATED]      = "currently allocated",
 158         [KFENCE_COUNTER_ALLOCS]         = "total allocations",
 159         [KFENCE_COUNTER_FREES]          = "total frees",
 160         [KFENCE_COUNTER_ZOMBIES]        = "zombie allocations",
 161         [KFENCE_COUNTER_BUGS]           = "total bugs",
 162         [KFENCE_COUNTER_SKIP_INCOMPAT]  = "skipped allocations (incompatible)",
 163         [KFENCE_COUNTER_SKIP_CAPACITY]  = "skipped allocations (capacity)",
 164         [KFENCE_COUNTER_SKIP_COVERED]   = "skipped allocations (covered)",
 165 };
 166 static_assert(ARRAY_SIZE(counter_names) == KFENCE_COUNTER_COUNT);
 167
 168 /* === Internals ============================================================ */
 169
 170 static inline bool should_skip_covered(void)
 171 {
 172         unsigned long thresh = (CONFIG_KFENCE_NUM_OBJECTS * kfence_skip_covered_thresh) / 100;
 173
 174         return atomic_long_read(&counters[KFENCE_COUNTER_ALLOCATED]) > thresh;
 175 }
 176
 177 static u32 get_alloc_stack_hash(unsigned long *stack_entries, size_t num_entries)
 178 {
 179         num_entries = min(num_entries, UNIQUE_ALLOC_STACK_DEPTH);
 180         num_entries = filter_irq_stacks(stack_entries, num_entries);
 181         return jhash(stack_entries, num_entries * sizeof(stack_entries[0]), stack_hash_seed);
 182 }
 183
 184 /*
 185  * Adds (or subtracts) count @val for allocation stack trace hash
 186  * @alloc_stack_hash from Counting Bloom filter.
 187  */
 188 static void alloc_covered_add(u32 alloc_stack_hash, int val)
 189 {
 190         int i;
 191
 192         for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
 193                 atomic_add(val, &alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]);
 194                 alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
 195         }
 196 }
 197
 198 /*
 199  * Returns true if the allocation stack trace hash @alloc_stack_hash is
 200  * currently contained (non-zero count) in Counting Bloom filter.
 201  */
 202 static bool alloc_covered_contains(u32 alloc_stack_hash)
 203 {
 204         int i;
 205
 206         for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
 207                 if (!atomic_read(&alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]))
 208                         return false;
 209                 alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
 210         }
 211
 212         return true;
 213 }
 214
 215 static bool kfence_protect(unsigned long addr)
 216 {
 217         return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), true));
 218 }
 219
 220 static bool kfence_unprotect(unsigned long addr)
 221 {
 222         return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), false));
 223 }
 224
 225 static inline struct kfence_metadata *addr_to_metadata(unsigned long addr)
 226 {
 227         long index;
 228
 229         /* The checks do not affect performance; only called from slow-paths. */
 230
 231         if (!is_kfence_address((void *)addr))
 232                 return NULL;
 233
 234         /*
 235          * May be an invalid index if called with an address at the edge of
 236          * __kfence_pool, in which case we would report an "invalid access"
 237          * error.
 238          */
 239         index = (addr - (unsigned long)__kfence_pool) / (PAGE_SIZE * 2) - 1;
 240         if (index < 0 || index >= CONFIG_KFENCE_NUM_OBJECTS)
 241                 return NULL;
 242
 243         return &kfence_metadata[index];
 244 }
 245
 246 static inline unsigned long metadata_to_pageaddr(const struct kfence_metadata *meta)
 247 {
 248         unsigned long offset = (meta - kfence_metadata + 1) * PAGE_SIZE * 2;
 249         unsigned long pageaddr = (unsigned long)&__kfence_pool[offset];
 250
 251         /* The checks do not affect performance; only called from slow-paths. */
 252
 253         /* Only call with a pointer into kfence_metadata. */
 254         if (KFENCE_WARN_ON(meta < kfence_metadata ||
 255                            meta >= kfence_metadata + CONFIG_KFENCE_NUM_OBJECTS))
 256                 return 0;
 257
 258         /*
 259          * This metadata object only ever maps to 1 page; verify that the stored
 260          * address is in the expected range.
 261          */
 262         if (KFENCE_WARN_ON(ALIGN_DOWN(meta->addr, PAGE_SIZE) != pageaddr))
 263                 return 0;
 264
 265         return pageaddr;
 266 }
 267
 268 /*
 269  * Update the object's metadata state, including updating the alloc/free stacks
 270  * depending on the state transition.
 271  */
 272 static noinline void
 273 metadata_update_state(struct kfence_metadata *meta, enum kfence_object_state next,
 274                       unsigned long *stack_entries, size_t num_stack_entries)
 275 {
 276         struct kfence_track *track =
 277                 next == KFENCE_OBJECT_FREED ? &meta->free_track : &meta->alloc_track;
 278
 279         lockdep_assert_held(&meta->lock);
 280
 281         if (stack_entries) {
 282                 memcpy(track->stack_entries, stack_entries,
 283                        num_stack_entries * sizeof(stack_entries[0]));
 284         } else {
 285                 /*
 286                  * Skip over 1 (this) functions; noinline ensures we do not
 287                  * accidentally skip over the caller by never inlining.
 288                  */
 289                 num_stack_entries = stack_trace_save(track->stack_entries, KFENCE_STACK_DEPTH, 1);
 290         }
 291         track->num_stack_entries = num_stack_entries;
 292         track->pid = task_pid_nr(current);
 293         track->cpu = raw_smp_processor_id();
 294         track->ts_nsec = local_clock(); /* Same source as printk timestamps. */
 295
 296         /*
 297          * Pairs with READ_ONCE() in
 298          *      kfence_shutdown_cache(),
 299          *      kfence_handle_page_fault().
 300          */
 301         WRITE_ONCE(meta->state, next);
 302 }
 303
 304 /* Write canary byte to @addr. */
 305 static inline bool set_canary_byte(u8 *addr)
 306 {
 307         *addr = KFENCE_CANARY_PATTERN(addr);
 308         return true;
 309 }
 310
 311 /* Check canary byte at @addr. */
 312 static inline bool check_canary_byte(u8 *addr)
 313 {
 314         struct kfence_metadata *meta;
 315         unsigned long flags;
 316
 317         if (likely(*addr == KFENCE_CANARY_PATTERN(addr)))
 318                 return true;
 319
 320         atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
 321
 322         meta = addr_to_metadata((unsigned long)addr);
 323         raw_spin_lock_irqsave(&meta->lock, flags);
 324         kfence_report_error((unsigned long)addr, false, NULL, meta, KFENCE_ERROR_CORRUPTION);
 325         raw_spin_unlock_irqrestore(&meta->lock, flags);
 326
 327         return false;
 328 }
 329
 330 /* __always_inline this to ensure we won't do an indirect call to fn. */
 331 static __always_inline void for_each_canary(const struct kfence_metadata *meta, bool (*fn)(u8 *))
 332 {
 333         const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
 334         unsigned long addr;
 335
 336         /*
 337          * We'll iterate over each canary byte per-side until fn() returns
 338          * false. However, we'll still iterate over the canary bytes to the
 339          * right of the object even if there was an error in the canary bytes to
 340          * the left of the object. Specifically, if check_canary_byte()
 341          * generates an error, showing both sides might give more clues as to
 342          * what the error is about when displaying which bytes were corrupted.
 343          */
 344
 345         /* Apply to left of object. */
 346         for (addr = pageaddr; addr < meta->addr; addr++) {
 347                 if (!fn((u8 *)addr))
 348                         break;
 349         }
 350
 351         /* Apply to right of object. */
 352         for (addr = meta->addr + meta->size; addr < pageaddr + PAGE_SIZE; addr++) {
 353                 if (!fn((u8 *)addr))
 354                         break;
 355         }
 356 }
 357
 358 static void *kfence_guarded_alloc(struct kmem_cache *cache, size_t size, gfp_t gfp,
 359                                   unsigned long *stack_entries, size_t num_stack_entries,
 360                                   u32 alloc_stack_hash)
 361 {
 362         struct kfence_metadata *meta = NULL;
 363         unsigned long flags;
 364         struct slab *slab;
 365         void *addr;
 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 (prandom_u32_max(2)) {
 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 (CONFIG_KFENCE_STRESS_TEST_FAULTS && !prandom_u32_max(CONFIG_KFENCE_STRESS_TEST_FAULTS))
 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 static bool __init kfence_init_pool(void)
 536 {
 537         unsigned long addr = (unsigned long)__kfence_pool;
 538         struct page *pages;
 539         int i;
 540
 541         if (!__kfence_pool)
 542                 return false;
 543
 544         if (!arch_kfence_init_pool())
 545                 goto err;
 546
 547         pages = virt_to_page(addr);
 548
 549         /*
 550          * Set up object pages: they must have PG_slab set, to avoid freeing
 551          * these as real pages.
 552          *
 553          * We also want to avoid inserting kfence_free() in the kfree()
 554          * fast-path in SLUB, and therefore need to ensure kfree() correctly
 555          * enters __slab_free() slow-path.
 556          */
 557         for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
 558                 if (!i || (i % 2))
 559                         continue;
 560
 561                 /* Verify we do not have a compound head page. */
 562                 if (WARN_ON(compound_head(&pages[i]) != &pages[i]))
 563                         goto err;
 564
 565                 __SetPageSlab(&pages[i]);
 566         }
 567
 568         /*
 569          * Protect the first 2 pages. The first page is mostly unnecessary, and
 570          * merely serves as an extended guard page. However, adding one
 571          * additional page in the beginning gives us an even number of pages,
 572          * which simplifies the mapping of address to metadata index.
 573          */
 574         for (i = 0; i < 2; i++) {
 575                 if (unlikely(!kfence_protect(addr)))
 576                         goto err;
 577
 578                 addr += PAGE_SIZE;
 579         }
 580
 581         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 582                 struct kfence_metadata *meta = &kfence_metadata[i];
 583
 584                 /* Initialize metadata. */
 585                 INIT_LIST_HEAD(&meta->list);
 586                 raw_spin_lock_init(&meta->lock);
 587                 meta->state = KFENCE_OBJECT_UNUSED;
 588                 meta->addr = addr; /* Initialize for validation in metadata_to_pageaddr(). */
 589                 list_add_tail(&meta->list, &kfence_freelist);
 590
 591                 /* Protect the right redzone. */
 592                 if (unlikely(!kfence_protect(addr + PAGE_SIZE)))
 593                         goto err;
 594
 595                 addr += 2 * PAGE_SIZE;
 596         }
 597
 598         /*
 599          * The pool is live and will never be deallocated from this point on.
 600          * Remove the pool object from the kmemleak object tree, as it would
 601          * otherwise overlap with allocations returned by kfence_alloc(), which
 602          * are registered with kmemleak through the slab post-alloc hook.
 603          */
 604         kmemleak_free(__kfence_pool);
 605
 606         return true;
 607
 608 err:
 609         /*
 610          * Only release unprotected pages, and do not try to go back and change
 611          * page attributes due to risk of failing to do so as well. If changing
 612          * page attributes for some pages fails, it is very likely that it also
 613          * fails for the first page, and therefore expect addr==__kfence_pool in
 614          * most failure cases.
 615          */
 616         memblock_free_late(__pa(addr), KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool));
 617         __kfence_pool = NULL;
 618         return false;
 619 }
 620
 621 /* === DebugFS Interface ==================================================== */
 622
 623 static int stats_show(struct seq_file *seq, void *v)
 624 {
 625         int i;
 626
 627         seq_printf(seq, "enabled: %i\n", READ_ONCE(kfence_enabled));
 628         for (i = 0; i < KFENCE_COUNTER_COUNT; i++)
 629                 seq_printf(seq, "%s: %ld\n", counter_names[i], atomic_long_read(&counters[i]));
 630
 631         return 0;
 632 }
 633 DEFINE_SHOW_ATTRIBUTE(stats);
 634
 635 /*
 636  * debugfs seq_file operations for /sys/kernel/debug/kfence/objects.
 637  * start_object() and next_object() return the object index + 1, because NULL is used
 638  * to stop iteration.
 639  */
 640 static void *start_object(struct seq_file *seq, loff_t *pos)
 641 {
 642         if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
 643                 return (void *)((long)*pos + 1);
 644         return NULL;
 645 }
 646
 647 static void stop_object(struct seq_file *seq, void *v)
 648 {
 649 }
 650
 651 static void *next_object(struct seq_file *seq, void *v, loff_t *pos)
 652 {
 653         ++*pos;
 654         if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
 655                 return (void *)((long)*pos + 1);
 656         return NULL;
 657 }
 658
 659 static int show_object(struct seq_file *seq, void *v)
 660 {
 661         struct kfence_metadata *meta = &kfence_metadata[(long)v - 1];
 662         unsigned long flags;
 663
 664         raw_spin_lock_irqsave(&meta->lock, flags);
 665         kfence_print_object(seq, meta);
 666         raw_spin_unlock_irqrestore(&meta->lock, flags);
 667         seq_puts(seq, "---------------------------------\n");
 668
 669         return 0;
 670 }
 671
 672 static const struct seq_operations object_seqops = {
 673         .start = start_object,
 674         .next = next_object,
 675         .stop = stop_object,
 676         .show = show_object,
 677 };
 678
 679 static int open_objects(struct inode *inode, struct file *file)
 680 {
 681         return seq_open(file, &object_seqops);
 682 }
 683
 684 static const struct file_operations objects_fops = {
 685         .open = open_objects,
 686         .read = seq_read,
 687         .llseek = seq_lseek,
 688         .release = seq_release,
 689 };
 690
 691 static int __init kfence_debugfs_init(void)
 692 {
 693         struct dentry *kfence_dir = debugfs_create_dir("kfence", NULL);
 694
 695         debugfs_create_file("stats", 0444, kfence_dir, NULL, &stats_fops);
 696         debugfs_create_file("objects", 0400, kfence_dir, NULL, &objects_fops);
 697         return 0;
 698 }
 699
 700 late_initcall(kfence_debugfs_init);
 701
 702 /* === Allocation Gate Timer ================================================ */
 703
 704 #ifdef CONFIG_KFENCE_STATIC_KEYS
 705 /* Wait queue to wake up allocation-gate timer task. */
 706 static DECLARE_WAIT_QUEUE_HEAD(allocation_wait);
 707
 708 static void wake_up_kfence_timer(struct irq_work *work)
 709 {
 710         wake_up(&allocation_wait);
 711 }
 712 static DEFINE_IRQ_WORK(wake_up_kfence_timer_work, wake_up_kfence_timer);
 713 #endif
 714
 715 /*
 716  * Set up delayed work, which will enable and disable the static key. We need to
 717  * use a work queue (rather than a simple timer), since enabling and disabling a
 718  * static key cannot be done from an interrupt.
 719  *
 720  * Note: Toggling a static branch currently causes IPIs, and here we'll end up
 721  * with a total of 2 IPIs to all CPUs. If this ends up a problem in future (with
 722  * more aggressive sampling intervals), we could get away with a variant that
 723  * avoids IPIs, at the cost of not immediately capturing allocations if the
 724  * instructions remain cached.
 725  */
 726 static struct delayed_work kfence_timer;
 727 static void toggle_allocation_gate(struct work_struct *work)
 728 {
 729         if (!READ_ONCE(kfence_enabled))
 730                 return;
 731
 732         atomic_set(&kfence_allocation_gate, 0);
 733 #ifdef CONFIG_KFENCE_STATIC_KEYS
 734         /* Enable static key, and await allocation to happen. */
 735         static_branch_enable(&kfence_allocation_key);
 736
 737         if (sysctl_hung_task_timeout_secs) {
 738                 /*
 739                  * During low activity with no allocations we might wait a
 740                  * while; let's avoid the hung task warning.
 741                  */
 742                 wait_event_idle_timeout(allocation_wait, atomic_read(&kfence_allocation_gate),
 743                                         sysctl_hung_task_timeout_secs * HZ / 2);
 744         } else {
 745                 wait_event_idle(allocation_wait, atomic_read(&kfence_allocation_gate));
 746         }
 747
 748         /* Disable static key and reset timer. */
 749         static_branch_disable(&kfence_allocation_key);
 750 #endif
 751         queue_delayed_work(system_unbound_wq, &kfence_timer,
 752                            msecs_to_jiffies(kfence_sample_interval));
 753 }
 754 static DECLARE_DELAYED_WORK(kfence_timer, toggle_allocation_gate);
 755
 756 /* === Public interface ===================================================== */
 757
 758 void __init kfence_alloc_pool(void)
 759 {
 760         if (!kfence_sample_interval)
 761                 return;
 762
 763         __kfence_pool = memblock_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
 764
 765         if (!__kfence_pool)
 766                 pr_err("failed to allocate pool\n");
 767 }
 768
 769 void __init kfence_init(void)
 770 {
 771         /* Setting kfence_sample_interval to 0 on boot disables KFENCE. */
 772         if (!kfence_sample_interval)
 773                 return;
 774
 775         stack_hash_seed = (u32)random_get_entropy();
 776         if (!kfence_init_pool()) {
 777                 pr_err("%s failed\n", __func__);
 778                 return;
 779         }
 780
 781         if (!IS_ENABLED(CONFIG_KFENCE_STATIC_KEYS))
 782                 static_branch_enable(&kfence_allocation_key);
 783         WRITE_ONCE(kfence_enabled, true);
 784         queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
 785         pr_info("initialized - using %lu bytes for %d objects at 0x%p-0x%p\n", KFENCE_POOL_SIZE,
 786                 CONFIG_KFENCE_NUM_OBJECTS, (void *)__kfence_pool,
 787                 (void *)(__kfence_pool + KFENCE_POOL_SIZE));
 788 }
 789
 790 void kfence_shutdown_cache(struct kmem_cache *s)
 791 {
 792         unsigned long flags;
 793         struct kfence_metadata *meta;
 794         int i;
 795
 796         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 797                 bool in_use;
 798
 799                 meta = &kfence_metadata[i];
 800
 801                 /*
 802                  * If we observe some inconsistent cache and state pair where we
 803                  * should have returned false here, cache destruction is racing
 804                  * with either kmem_cache_alloc() or kmem_cache_free(). Taking
 805                  * the lock will not help, as different critical section
 806                  * serialization will have the same outcome.
 807                  */
 808                 if (READ_ONCE(meta->cache) != s ||
 809                     READ_ONCE(meta->state) != KFENCE_OBJECT_ALLOCATED)
 810                         continue;
 811
 812                 raw_spin_lock_irqsave(&meta->lock, flags);
 813                 in_use = meta->cache == s && meta->state == KFENCE_OBJECT_ALLOCATED;
 814                 raw_spin_unlock_irqrestore(&meta->lock, flags);
 815
 816                 if (in_use) {
 817                         /*
 818                          * This cache still has allocations, and we should not
 819                          * release them back into the freelist so they can still
 820                          * safely be used and retain the kernel's default
 821                          * behaviour of keeping the allocations alive (leak the
 822                          * cache); however, they effectively become "zombie
 823                          * allocations" as the KFENCE objects are the only ones
 824                          * still in use and the owning cache is being destroyed.
 825                          *
 826                          * We mark them freed, so that any subsequent use shows
 827                          * more useful error messages that will include stack
 828                          * traces of the user of the object, the original
 829                          * allocation, and caller to shutdown_cache().
 830                          */
 831                         kfence_guarded_free((void *)meta->addr, meta, /*zombie=*/true);
 832                 }
 833         }
 834
 835         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 836                 meta = &kfence_metadata[i];
 837
 838                 /* See above. */
 839                 if (READ_ONCE(meta->cache) != s || READ_ONCE(meta->state) != KFENCE_OBJECT_FREED)
 840                         continue;
 841
 842                 raw_spin_lock_irqsave(&meta->lock, flags);
 843                 if (meta->cache == s && meta->state == KFENCE_OBJECT_FREED)
 844                         meta->cache = NULL;
 845                 raw_spin_unlock_irqrestore(&meta->lock, flags);
 846         }
 847 }
 848
 849 void *__kfence_alloc(struct kmem_cache *s, size_t size, gfp_t flags)
 850 {
 851         unsigned long stack_entries[KFENCE_STACK_DEPTH];
 852         size_t num_stack_entries;
 853         u32 alloc_stack_hash;
 854
 855         /*
 856          * Perform size check before switching kfence_allocation_gate, so that
 857          * we don't disable KFENCE without making an allocation.
 858          */
 859         if (size > PAGE_SIZE) {
 860                 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
 861                 return NULL;
 862         }
 863
 864         /*
 865          * Skip allocations from non-default zones, including DMA. We cannot
 866          * guarantee that pages in the KFENCE pool will have the requested
 867          * properties (e.g. reside in DMAable memory).
 868          */
 869         if ((flags & GFP_ZONEMASK) ||
 870             (s->flags & (SLAB_CACHE_DMA | SLAB_CACHE_DMA32))) {
 871                 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
 872                 return NULL;
 873         }
 874
 875         if (atomic_inc_return(&kfence_allocation_gate) > 1)
 876                 return NULL;
 877 #ifdef CONFIG_KFENCE_STATIC_KEYS
 878         /*
 879          * waitqueue_active() is fully ordered after the update of
 880          * kfence_allocation_gate per atomic_inc_return().
 881          */
 882         if (waitqueue_active(&allocation_wait)) {
 883                 /*
 884                  * Calling wake_up() here may deadlock when allocations happen
 885                  * from within timer code. Use an irq_work to defer it.
 886                  */
 887                 irq_work_queue(&wake_up_kfence_timer_work);
 888         }
 889 #endif
 890
 891         if (!READ_ONCE(kfence_enabled))
 892                 return NULL;
 893
 894         num_stack_entries = stack_trace_save(stack_entries, KFENCE_STACK_DEPTH, 0);
 895
 896         /*
 897          * Do expensive check for coverage of allocation in slow-path after
 898          * allocation_gate has already become non-zero, even though it might
 899          * mean not making any allocation within a given sample interval.
 900          *
 901          * This ensures reasonable allocation coverage when the pool is almost
 902          * full, including avoiding long-lived allocations of the same source
 903          * filling up the pool (e.g. pagecache allocations).
 904          */
 905         alloc_stack_hash = get_alloc_stack_hash(stack_entries, num_stack_entries);
 906         if (should_skip_covered() && alloc_covered_contains(alloc_stack_hash)) {
 907                 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_COVERED]);
 908                 return NULL;
 909         }
 910
 911         return kfence_guarded_alloc(s, size, flags, stack_entries, num_stack_entries,
 912                                     alloc_stack_hash);
 913 }
 914
 915 size_t kfence_ksize(const void *addr)
 916 {
 917         const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
 918
 919         /*
 920          * Read locklessly -- if there is a race with __kfence_alloc(), this is
 921          * either a use-after-free or invalid access.
 922          */
 923         return meta ? meta->size : 0;
 924 }
 925
 926 void *kfence_object_start(const void *addr)
 927 {
 928         const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
 929
 930         /*
 931          * Read locklessly -- if there is a race with __kfence_alloc(), this is
 932          * either a use-after-free or invalid access.
 933          */
 934         return meta ? (void *)meta->addr : NULL;
 935 }
 936
 937 void __kfence_free(void *addr)
 938 {
 939         struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
 940
 941         /*
 942          * If the objects of the cache are SLAB_TYPESAFE_BY_RCU, defer freeing
 943          * the object, as the object page may be recycled for other-typed
 944          * objects once it has been freed. meta->cache may be NULL if the cache
 945          * was destroyed.
 946          */
 947         if (unlikely(meta->cache && (meta->cache->flags & SLAB_TYPESAFE_BY_RCU)))
 948                 call_rcu(&meta->rcu_head, rcu_guarded_free);
 949         else
 950                 kfence_guarded_free(addr, meta, false);
 951 }
 952
 953 bool kfence_handle_page_fault(unsigned long addr, bool is_write, struct pt_regs *regs)
 954 {
 955         const int page_index = (addr - (unsigned long)__kfence_pool) / PAGE_SIZE;
 956         struct kfence_metadata *to_report = NULL;
 957         enum kfence_error_type error_type;
 958         unsigned long flags;
 959
 960         if (!is_kfence_address((void *)addr))
 961                 return false;
 962
 963         if (!READ_ONCE(kfence_enabled)) /* If disabled at runtime ... */
 964                 return kfence_unprotect(addr); /* ... unprotect and proceed. */
 965
 966         atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
 967
 968         if (page_index % 2) {
 969                 /* This is a redzone, report a buffer overflow. */
 970                 struct kfence_metadata *meta;
 971                 int distance = 0;
 972
 973                 meta = addr_to_metadata(addr - PAGE_SIZE);
 974                 if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
 975                         to_report = meta;
 976                         /* Data race ok; distance calculation approximate. */
 977                         distance = addr - data_race(meta->addr + meta->size);
 978                 }
 979
 980                 meta = addr_to_metadata(addr + PAGE_SIZE);
 981                 if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
 982                         /* Data race ok; distance calculation approximate. */
 983                         if (!to_report || distance > data_race(meta->addr) - addr)
 984                                 to_report = meta;
 985                 }
 986
 987                 if (!to_report)
 988                         goto out;
 989
 990                 raw_spin_lock_irqsave(&to_report->lock, flags);
 991                 to_report->unprotected_page = addr;
 992                 error_type = KFENCE_ERROR_OOB;
 993
 994                 /*
 995                  * If the object was freed before we took the look we can still
 996                  * report this as an OOB -- the report will simply show the
 997                  * stacktrace of the free as well.
 998                  */
 999         } else {
1000                 to_report = addr_to_metadata(addr);
1001                 if (!to_report)
1002                         goto out;
1003
1004                 raw_spin_lock_irqsave(&to_report->lock, flags);
1005                 error_type = KFENCE_ERROR_UAF;
1006                 /*
1007                  * We may race with __kfence_alloc(), and it is possible that a
1008                  * freed object may be reallocated. We simply report this as a
1009                  * use-after-free, with the stack trace showing the place where
1010                  * the object was re-allocated.
1011                  */
1012         }
1013
1014 out:
1015         if (to_report) {
1016                 kfence_report_error(addr, is_write, regs, to_report, error_type);
1017                 raw_spin_unlock_irqrestore(&to_report->lock, flags);
1018         } else {
1019                 /* This may be a UAF or OOB access, but we can't be sure. */
1020                 kfence_report_error(addr, is_write, regs, NULL, KFENCE_ERROR_INVALID);
1021         }
1022
1023         return kfence_unprotect(addr); /* Unprotect and let access proceed. */
1024 }