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