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/kasan-enabled.h>
  17#include <linux/kcsan-checks.h>
  18#include <linux/kfence.h>
  19#include <linux/kmemleak.h>
  20#include <linux/list.h>
  21#include <linux/lockdep.h>
  22#include <linux/log2.h>
  23#include <linux/memblock.h>
  24#include <linux/moduleparam.h>
  25#include <linux/nodemask.h>
  26#include <linux/notifier.h>
  27#include <linux/panic_notifier.h>
  28#include <linux/random.h>
  29#include <linux/rcupdate.h>
  30#include <linux/reboot.h>
  31#include <linux/sched/clock.h>
  32#include <linux/seq_file.h>
  33#include <linux/slab.h>
  34#include <linux/spinlock.h>
  35#include <linux/string.h>
  36
  37#include <asm/kfence.h>
  38
  39#include "kfence.h"
  40
  41/* Disables KFENCE on the first warning assuming an irrecoverable error. */
  42#define KFENCE_WARN_ON(cond)                                                   \
  43	({                                                                     \
  44		const bool __cond = WARN_ON(cond);                             \
  45		if (unlikely(__cond)) {                                        \
  46			WRITE_ONCE(kfence_enabled, false);                     \
  47			disabled_by_warn = true;                               \
  48		}                                                              \
  49		__cond;                                                        \
  50	})
  51
  52/* === Data ================================================================= */
  53
  54static bool kfence_enabled __read_mostly;
  55static bool disabled_by_warn __read_mostly;
  56
  57unsigned long kfence_sample_interval __read_mostly = CONFIG_KFENCE_SAMPLE_INTERVAL;
  58EXPORT_SYMBOL_GPL(kfence_sample_interval); /* Export for test modules. */
  59
  60#ifdef MODULE_PARAM_PREFIX
  61#undef MODULE_PARAM_PREFIX
  62#endif
  63#define MODULE_PARAM_PREFIX "kfence."
  64
  65static int kfence_enable_late(void);
  66static int param_set_sample_interval(const char *val, const struct kernel_param *kp)
  67{
  68	unsigned long num;
  69	int ret = kstrtoul(val, 0, &num);
  70
  71	if (ret < 0)
  72		return ret;
  73
  74	/* Using 0 to indicate KFENCE is disabled. */
  75	if (!num && READ_ONCE(kfence_enabled)) {
  76		pr_info("disabled\n");
  77		WRITE_ONCE(kfence_enabled, false);
  78	}
  79
  80	*((unsigned long *)kp->arg) = num;
  81
  82	if (num && !READ_ONCE(kfence_enabled) && system_state != SYSTEM_BOOTING)
  83		return disabled_by_warn ? -EINVAL : kfence_enable_late();
  84	return 0;
  85}
  86
  87static int param_get_sample_interval(char *buffer, const struct kernel_param *kp)
  88{
  89	if (!READ_ONCE(kfence_enabled))
  90		return sprintf(buffer, "0\n");
  91
  92	return param_get_ulong(buffer, kp);
  93}
  94
  95static const struct kernel_param_ops sample_interval_param_ops = {
  96	.set = param_set_sample_interval,
  97	.get = param_get_sample_interval,
  98};
  99module_param_cb(sample_interval, &sample_interval_param_ops, &kfence_sample_interval, 0600);
 100
 101/* Pool usage% threshold when currently covered allocations are skipped. */
 102static unsigned long kfence_skip_covered_thresh __read_mostly = 75;
 103module_param_named(skip_covered_thresh, kfence_skip_covered_thresh, ulong, 0644);
 104
 105/* Allocation burst count: number of excess KFENCE allocations per sample. */
 106static unsigned int kfence_burst __read_mostly;
 107module_param_named(burst, kfence_burst, uint, 0644);
 108
 109/* If true, use a deferrable timer. */
 110static bool kfence_deferrable __read_mostly = IS_ENABLED(CONFIG_KFENCE_DEFERRABLE);
 111module_param_named(deferrable, kfence_deferrable, bool, 0444);
 112
 113/* If true, check all canary bytes on panic. */
 114static bool kfence_check_on_panic __read_mostly;
 115module_param_named(check_on_panic, kfence_check_on_panic, bool, 0444);
 116
 117/* The pool of pages used for guard pages and objects. */
 118char *__kfence_pool __read_mostly;
 119EXPORT_SYMBOL(__kfence_pool); /* Export for test modules. */
 120
 121/*
 122 * Per-object metadata, with one-to-one mapping of object metadata to
 123 * backing pages (in __kfence_pool).
 124 */
 125static_assert(CONFIG_KFENCE_NUM_OBJECTS > 0);
 126struct kfence_metadata *kfence_metadata __read_mostly;
 127
 128/*
 129 * If kfence_metadata is not NULL, it may be accessed by kfence_shutdown_cache().
 130 * So introduce kfence_metadata_init to initialize metadata, and then make
 131 * kfence_metadata visible after initialization is successful. This prevents
 132 * potential UAF or access to uninitialized metadata.
 133 */
 134static struct kfence_metadata *kfence_metadata_init __read_mostly;
 135
 136/* Freelist with available objects. */
 137static struct list_head kfence_freelist = LIST_HEAD_INIT(kfence_freelist);
 138static DEFINE_RAW_SPINLOCK(kfence_freelist_lock); /* Lock protecting freelist. */
 139
 140/*
 141 * The static key to set up a KFENCE allocation; or if static keys are not used
 142 * to gate allocations, to avoid a load and compare if KFENCE is disabled.
 143 */
 144DEFINE_STATIC_KEY_FALSE(kfence_allocation_key);
 145
 146/* Gates the allocation, ensuring only one succeeds in a given period. */
 147atomic_t kfence_allocation_gate = ATOMIC_INIT(1);
 148
 149/*
 150 * A Counting Bloom filter of allocation coverage: limits currently covered
 151 * allocations of the same source filling up the pool.
 152 *
 153 * Assuming a range of 15%-85% unique allocations in the pool at any point in
 154 * time, the below parameters provide a probablity of 0.02-0.33 for false
 155 * positive hits respectively:
 156 *
 157 *	P(alloc_traces) = (1 - e^(-HNUM * (alloc_traces / SIZE)) ^ HNUM
 158 */
 159#define ALLOC_COVERED_HNUM	2
 160#define ALLOC_COVERED_ORDER	(const_ilog2(CONFIG_KFENCE_NUM_OBJECTS) + 2)
 161#define ALLOC_COVERED_SIZE	(1 << ALLOC_COVERED_ORDER)
 162#define ALLOC_COVERED_HNEXT(h)	hash_32(h, ALLOC_COVERED_ORDER)
 163#define ALLOC_COVERED_MASK	(ALLOC_COVERED_SIZE - 1)
 164static atomic_t alloc_covered[ALLOC_COVERED_SIZE];
 165
 166/* Stack depth used to determine uniqueness of an allocation. */
 167#define UNIQUE_ALLOC_STACK_DEPTH ((size_t)8)
 168
 169/*
 170 * Randomness for stack hashes, making the same collisions across reboots and
 171 * different machines less likely.
 172 */
 173static u32 stack_hash_seed __ro_after_init;
 174
 175/* Statistics counters for debugfs. */
 176enum kfence_counter_id {
 177	KFENCE_COUNTER_ALLOCATED,
 178	KFENCE_COUNTER_ALLOCS,
 179	KFENCE_COUNTER_FREES,
 180	KFENCE_COUNTER_ZOMBIES,
 181	KFENCE_COUNTER_BUGS,
 182	KFENCE_COUNTER_SKIP_INCOMPAT,
 183	KFENCE_COUNTER_SKIP_CAPACITY,
 184	KFENCE_COUNTER_SKIP_COVERED,
 185	KFENCE_COUNTER_COUNT,
 186};
 187static atomic_long_t counters[KFENCE_COUNTER_COUNT];
 188static const char *const counter_names[] = {
 189	[KFENCE_COUNTER_ALLOCATED]	= "currently allocated",
 190	[KFENCE_COUNTER_ALLOCS]		= "total allocations",
 191	[KFENCE_COUNTER_FREES]		= "total frees",
 192	[KFENCE_COUNTER_ZOMBIES]	= "zombie allocations",
 193	[KFENCE_COUNTER_BUGS]		= "total bugs",
 194	[KFENCE_COUNTER_SKIP_INCOMPAT]	= "skipped allocations (incompatible)",
 195	[KFENCE_COUNTER_SKIP_CAPACITY]	= "skipped allocations (capacity)",
 196	[KFENCE_COUNTER_SKIP_COVERED]	= "skipped allocations (covered)",
 197};
 198static_assert(ARRAY_SIZE(counter_names) == KFENCE_COUNTER_COUNT);
 199
 200/* === Internals ============================================================ */
 201
 202static inline bool should_skip_covered(void)
 203{
 204	unsigned long thresh = (CONFIG_KFENCE_NUM_OBJECTS * kfence_skip_covered_thresh) / 100;
 205
 206	return atomic_long_read(&counters[KFENCE_COUNTER_ALLOCATED]) > thresh;
 207}
 208
 209static u32 get_alloc_stack_hash(unsigned long *stack_entries, size_t num_entries)
 210{
 211	num_entries = min(num_entries, UNIQUE_ALLOC_STACK_DEPTH);
 212	num_entries = filter_irq_stacks(stack_entries, num_entries);
 213	return jhash(stack_entries, num_entries * sizeof(stack_entries[0]), stack_hash_seed);
 214}
 215
 216/*
 217 * Adds (or subtracts) count @val for allocation stack trace hash
 218 * @alloc_stack_hash from Counting Bloom filter.
 219 */
 220static void alloc_covered_add(u32 alloc_stack_hash, int val)
 221{
 222	int i;
 223
 224	for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
 225		atomic_add(val, &alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]);
 226		alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
 227	}
 228}
 229
 230/*
 231 * Returns true if the allocation stack trace hash @alloc_stack_hash is
 232 * currently contained (non-zero count) in Counting Bloom filter.
 233 */
 234static bool alloc_covered_contains(u32 alloc_stack_hash)
 235{
 236	int i;
 237
 238	for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
 239		if (!atomic_read(&alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]))
 240			return false;
 241		alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
 242	}
 243
 244	return true;
 245}
 246
 247static bool kfence_protect(unsigned long addr)
 248{
 249	return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), true));
 250}
 251
 252static bool kfence_unprotect(unsigned long addr)
 253{
 254	return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), false));
 255}
 256
 257static inline unsigned long metadata_to_pageaddr(const struct kfence_metadata *meta)
 258{
 259	unsigned long offset = (meta - kfence_metadata + 1) * PAGE_SIZE * 2;
 260	unsigned long pageaddr = (unsigned long)&__kfence_pool[offset];
 261
 262	/* The checks do not affect performance; only called from slow-paths. */
 263
 264	/* Only call with a pointer into kfence_metadata. */
 265	if (KFENCE_WARN_ON(meta < kfence_metadata ||
 266			   meta >= kfence_metadata + CONFIG_KFENCE_NUM_OBJECTS))
 267		return 0;
 268
 269	/*
 270	 * This metadata object only ever maps to 1 page; verify that the stored
 271	 * address is in the expected range.
 272	 */
 273	if (KFENCE_WARN_ON(ALIGN_DOWN(meta->addr, PAGE_SIZE) != pageaddr))
 274		return 0;
 275
 276	return pageaddr;
 277}
 278
 279static inline bool kfence_obj_allocated(const struct kfence_metadata *meta)
 280{
 281	enum kfence_object_state state = READ_ONCE(meta->state);
 282
 283	return state == KFENCE_OBJECT_ALLOCATED || state == KFENCE_OBJECT_RCU_FREEING;
 284}
 285
 286/*
 287 * Update the object's metadata state, including updating the alloc/free stacks
 288 * depending on the state transition.
 289 */
 290static noinline void
 291metadata_update_state(struct kfence_metadata *meta, enum kfence_object_state next,
 292		      unsigned long *stack_entries, size_t num_stack_entries)
 293{
 294	struct kfence_track *track =
 295		next == KFENCE_OBJECT_ALLOCATED ? &meta->alloc_track : &meta->free_track;
 296
 297	lockdep_assert_held(&meta->lock);
 298
 299	/* Stack has been saved when calling rcu, skip. */
 300	if (READ_ONCE(meta->state) == KFENCE_OBJECT_RCU_FREEING)
 301		goto out;
 302
 303	if (stack_entries) {
 304		memcpy(track->stack_entries, stack_entries,
 305		       num_stack_entries * sizeof(stack_entries[0]));
 306	} else {
 307		/*
 308		 * Skip over 1 (this) functions; noinline ensures we do not
 309		 * accidentally skip over the caller by never inlining.
 310		 */
 311		num_stack_entries = stack_trace_save(track->stack_entries, KFENCE_STACK_DEPTH, 1);
 312	}
 313	track->num_stack_entries = num_stack_entries;
 314	track->pid = task_pid_nr(current);
 315	track->cpu = raw_smp_processor_id();
 316	track->ts_nsec = local_clock(); /* Same source as printk timestamps. */
 317
 318out:
 319	/*
 320	 * Pairs with READ_ONCE() in
 321	 *	kfence_shutdown_cache(),
 322	 *	kfence_handle_page_fault().
 323	 */
 324	WRITE_ONCE(meta->state, next);
 325}
 326
 327#ifdef CONFIG_KMSAN
 328#define check_canary_attributes noinline __no_kmsan_checks
 329#else
 330#define check_canary_attributes inline
 331#endif
 332
 333/* Check canary byte at @addr. */
 334static check_canary_attributes bool check_canary_byte(u8 *addr)
 335{
 336	struct kfence_metadata *meta;
 337	unsigned long flags;
 338
 339	if (likely(*addr == KFENCE_CANARY_PATTERN_U8(addr)))
 340		return true;
 341
 342	atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
 343
 344	meta = addr_to_metadata((unsigned long)addr);
 345	raw_spin_lock_irqsave(&meta->lock, flags);
 346	kfence_report_error((unsigned long)addr, false, NULL, meta, KFENCE_ERROR_CORRUPTION);
 347	raw_spin_unlock_irqrestore(&meta->lock, flags);
 348
 349	return false;
 350}
 351
 352static inline void set_canary(const struct kfence_metadata *meta)
 353{
 354	const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
 355	unsigned long addr = pageaddr;
 356
 357	/*
 358	 * The canary may be written to part of the object memory, but it does
 359	 * not affect it. The user should initialize the object before using it.
 360	 */
 361	for (; addr < meta->addr; addr += sizeof(u64))
 362		*((u64 *)addr) = KFENCE_CANARY_PATTERN_U64;
 363
 364	addr = ALIGN_DOWN(meta->addr + meta->size, sizeof(u64));
 365	for (; addr - pageaddr < PAGE_SIZE; addr += sizeof(u64))
 366		*((u64 *)addr) = KFENCE_CANARY_PATTERN_U64;
 367}
 368
 369static check_canary_attributes void
 370check_canary(const struct kfence_metadata *meta)
 371{
 372	const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
 373	unsigned long addr = pageaddr;
 374
 375	/*
 376	 * We'll iterate over each canary byte per-side until a corrupted byte
 377	 * is found. However, we'll still iterate over the canary bytes to the
 378	 * right of the object even if there was an error in the canary bytes to
 379	 * the left of the object. Specifically, if check_canary_byte()
 380	 * generates an error, showing both sides might give more clues as to
 381	 * what the error is about when displaying which bytes were corrupted.
 382	 */
 383
 384	/* Apply to left of object. */
 385	for (; meta->addr - addr >= sizeof(u64); addr += sizeof(u64)) {
 386		if (unlikely(*((u64 *)addr) != KFENCE_CANARY_PATTERN_U64))
 387			break;
 388	}
 389
 390	/*
 391	 * If the canary is corrupted in a certain 64 bytes, or the canary
 392	 * memory cannot be completely covered by multiple consecutive 64 bytes,
 393	 * it needs to be checked one by one.
 394	 */
 395	for (; addr < meta->addr; addr++) {
 396		if (unlikely(!check_canary_byte((u8 *)addr)))
 397			break;
 398	}
 399
 400	/* Apply to right of object. */
 401	for (addr = meta->addr + meta->size; addr % sizeof(u64) != 0; addr++) {
 402		if (unlikely(!check_canary_byte((u8 *)addr)))
 403			return;
 404	}
 405	for (; addr - pageaddr < PAGE_SIZE; addr += sizeof(u64)) {
 406		if (unlikely(*((u64 *)addr) != KFENCE_CANARY_PATTERN_U64)) {
 407
 408			for (; addr - pageaddr < PAGE_SIZE; addr++) {
 409				if (!check_canary_byte((u8 *)addr))
 410					return;
 411			}
 412		}
 413	}
 414}
 415
 416static void *kfence_guarded_alloc(struct kmem_cache *cache, size_t size, gfp_t gfp,
 417				  unsigned long *stack_entries, size_t num_stack_entries,
 418				  u32 alloc_stack_hash)
 419{
 420	struct kfence_metadata *meta = NULL;
 421	unsigned long flags;
 422	struct slab *slab;
 423	void *addr;
 424	const bool random_right_allocate = get_random_u32_below(2);
 425	const bool random_fault = CONFIG_KFENCE_STRESS_TEST_FAULTS &&
 426				  !get_random_u32_below(CONFIG_KFENCE_STRESS_TEST_FAULTS);
 427
 428	/* Try to obtain a free object. */
 429	raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
 430	if (!list_empty(&kfence_freelist)) {
 431		meta = list_entry(kfence_freelist.next, struct kfence_metadata, list);
 432		list_del_init(&meta->list);
 433	}
 434	raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
 435	if (!meta) {
 436		atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_CAPACITY]);
 437		return NULL;
 438	}
 439
 440	if (unlikely(!raw_spin_trylock_irqsave(&meta->lock, flags))) {
 441		/*
 442		 * This is extremely unlikely -- we are reporting on a
 443		 * use-after-free, which locked meta->lock, and the reporting
 444		 * code via printk calls kmalloc() which ends up in
 445		 * kfence_alloc() and tries to grab the same object that we're
 446		 * reporting on. While it has never been observed, lockdep does
 447		 * report that there is a possibility of deadlock. Fix it by
 448		 * using trylock and bailing out gracefully.
 449		 */
 450		raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
 451		/* Put the object back on the freelist. */
 452		list_add_tail(&meta->list, &kfence_freelist);
 453		raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
 454
 455		return NULL;
 456	}
 457
 458	meta->addr = metadata_to_pageaddr(meta);
 459	/* Unprotect if we're reusing this page. */
 460	if (meta->state == KFENCE_OBJECT_FREED)
 461		kfence_unprotect(meta->addr);
 462
 463	/*
 464	 * Note: for allocations made before RNG initialization, will always
 465	 * return zero. We still benefit from enabling KFENCE as early as
 466	 * possible, even when the RNG is not yet available, as this will allow
 467	 * KFENCE to detect bugs due to earlier allocations. The only downside
 468	 * is that the out-of-bounds accesses detected are deterministic for
 469	 * such allocations.
 470	 */
 471	if (random_right_allocate) {
 472		/* Allocate on the "right" side, re-calculate address. */
 473		meta->addr += PAGE_SIZE - size;
 474		meta->addr = ALIGN_DOWN(meta->addr, cache->align);
 475	}
 476
 477	addr = (void *)meta->addr;
 478
 479	/* Update remaining metadata. */
 480	metadata_update_state(meta, KFENCE_OBJECT_ALLOCATED, stack_entries, num_stack_entries);
 481	/* Pairs with READ_ONCE() in kfence_shutdown_cache(). */
 482	WRITE_ONCE(meta->cache, cache);
 483	meta->size = size;
 484	meta->alloc_stack_hash = alloc_stack_hash;
 485	raw_spin_unlock_irqrestore(&meta->lock, flags);
 486
 487	alloc_covered_add(alloc_stack_hash, 1);
 488
 489	/* Set required slab fields. */
 490	slab = virt_to_slab((void *)meta->addr);
 491	slab->slab_cache = cache;
 492	slab->objects = 1;
 493
 494	/* Memory initialization. */
 495	set_canary(meta);
 496
 497	/*
 498	 * We check slab_want_init_on_alloc() ourselves, rather than letting
 499	 * SL*B do the initialization, as otherwise we might overwrite KFENCE's
 500	 * redzone.
 501	 */
 502	if (unlikely(slab_want_init_on_alloc(gfp, cache)))
 503		memzero_explicit(addr, size);
 504	if (cache->ctor)
 505		cache->ctor(addr);
 506
 507	if (random_fault)
 508		kfence_protect(meta->addr); /* Random "faults" by protecting the object. */
 509
 510	atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCATED]);
 511	atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCS]);
 512
 513	return addr;
 514}
 515
 516static void kfence_guarded_free(void *addr, struct kfence_metadata *meta, bool zombie)
 517{
 518	struct kcsan_scoped_access assert_page_exclusive;
 519	unsigned long flags;
 520	bool init;
 521
 522	raw_spin_lock_irqsave(&meta->lock, flags);
 523
 524	if (!kfence_obj_allocated(meta) || meta->addr != (unsigned long)addr) {
 525		/* Invalid or double-free, bail out. */
 526		atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
 527		kfence_report_error((unsigned long)addr, false, NULL, meta,
 528				    KFENCE_ERROR_INVALID_FREE);
 529		raw_spin_unlock_irqrestore(&meta->lock, flags);
 530		return;
 531	}
 532
 533	/* Detect racy use-after-free, or incorrect reallocation of this page by KFENCE. */
 534	kcsan_begin_scoped_access((void *)ALIGN_DOWN((unsigned long)addr, PAGE_SIZE), PAGE_SIZE,
 535				  KCSAN_ACCESS_SCOPED | KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT,
 536				  &assert_page_exclusive);
 537
 538	if (CONFIG_KFENCE_STRESS_TEST_FAULTS)
 539		kfence_unprotect((unsigned long)addr); /* To check canary bytes. */
 540
 541	/* Restore page protection if there was an OOB access. */
 542	if (meta->unprotected_page) {
 543		memzero_explicit((void *)ALIGN_DOWN(meta->unprotected_page, PAGE_SIZE), PAGE_SIZE);
 544		kfence_protect(meta->unprotected_page);
 545		meta->unprotected_page = 0;
 546	}
 547
 548	/* Mark the object as freed. */
 549	metadata_update_state(meta, KFENCE_OBJECT_FREED, NULL, 0);
 550	init = slab_want_init_on_free(meta->cache);
 551	raw_spin_unlock_irqrestore(&meta->lock, flags);
 552
 553	alloc_covered_add(meta->alloc_stack_hash, -1);
 554
 555	/* Check canary bytes for memory corruption. */
 556	check_canary(meta);
 557
 558	/*
 559	 * Clear memory if init-on-free is set. While we protect the page, the
 560	 * data is still there, and after a use-after-free is detected, we
 561	 * unprotect the page, so the data is still accessible.
 562	 */
 563	if (!zombie && unlikely(init))
 564		memzero_explicit(addr, meta->size);
 565
 566	/* Protect to detect use-after-frees. */
 567	kfence_protect((unsigned long)addr);
 568
 569	kcsan_end_scoped_access(&assert_page_exclusive);
 570	if (!zombie) {
 571		/* Add it to the tail of the freelist for reuse. */
 572		raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
 573		KFENCE_WARN_ON(!list_empty(&meta->list));
 574		list_add_tail(&meta->list, &kfence_freelist);
 575		raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
 576
 577		atomic_long_dec(&counters[KFENCE_COUNTER_ALLOCATED]);
 578		atomic_long_inc(&counters[KFENCE_COUNTER_FREES]);
 579	} else {
 580		/* See kfence_shutdown_cache(). */
 581		atomic_long_inc(&counters[KFENCE_COUNTER_ZOMBIES]);
 582	}
 583}
 584
 585static void rcu_guarded_free(struct rcu_head *h)
 586{
 587	struct kfence_metadata *meta = container_of(h, struct kfence_metadata, rcu_head);
 588
 589	kfence_guarded_free((void *)meta->addr, meta, false);
 590}
 591
 592/*
 593 * Initialization of the KFENCE pool after its allocation.
 594 * Returns 0 on success; otherwise returns the address up to
 595 * which partial initialization succeeded.
 596 */
 597static unsigned long kfence_init_pool(void)
 598{
 599	unsigned long addr, start_pfn;
 600	int i, rand;
 601
 602	if (!arch_kfence_init_pool())
 603		return (unsigned long)__kfence_pool;
 604
 605	addr = (unsigned long)__kfence_pool;
 606	start_pfn = PHYS_PFN(virt_to_phys(__kfence_pool));
 607
 608	/*
 609	 * Set up object pages: they must have PGTY_slab set to avoid freeing
 610	 * them as real pages.
 611	 *
 612	 * We also want to avoid inserting kfence_free() in the kfree()
 613	 * fast-path in SLUB, and therefore need to ensure kfree() correctly
 614	 * enters __slab_free() slow-path.
 615	 */
 616	for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
 617		struct page *page;
 618
 619		if (!i || (i % 2))
 620			continue;
 621
 622		page = pfn_to_page(start_pfn + i);
 623		__SetPageSlab(page);
 624#ifdef CONFIG_MEMCG
 625		struct slab *slab = page_slab(page);
 626		slab->obj_exts = (unsigned long)&kfence_metadata_init[i / 2 - 1].obj_exts |
 627				 MEMCG_DATA_OBJEXTS;
 628#endif
 629	}
 630
 631	/*
 632	 * Protect the first 2 pages. The first page is mostly unnecessary, and
 633	 * merely serves as an extended guard page. However, adding one
 634	 * additional page in the beginning gives us an even number of pages,
 635	 * which simplifies the mapping of address to metadata index.
 636	 */
 637	for (i = 0; i < 2; i++) {
 638		if (unlikely(!kfence_protect(addr)))
 639			return addr;
 640
 641		addr += PAGE_SIZE;
 642	}
 643
 644	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 645		struct kfence_metadata *meta = &kfence_metadata_init[i];
 646
 647		/* Initialize metadata. */
 648		INIT_LIST_HEAD(&meta->list);
 649		raw_spin_lock_init(&meta->lock);
 650		meta->state = KFENCE_OBJECT_UNUSED;
 651		/* Use addr to randomize the freelist. */
 652		meta->addr = i;
 653
 654		/* Protect the right redzone. */
 655		if (unlikely(!kfence_protect(addr + 2 * i * PAGE_SIZE + PAGE_SIZE)))
 656			goto reset_slab;
 657	}
 658
 659	for (i = CONFIG_KFENCE_NUM_OBJECTS; i > 0; i--) {
 660		rand = get_random_u32_below(i);
 661		swap(kfence_metadata_init[i - 1].addr, kfence_metadata_init[rand].addr);
 662	}
 663
 664	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 665		struct kfence_metadata *meta_1 = &kfence_metadata_init[i];
 666		struct kfence_metadata *meta_2 = &kfence_metadata_init[meta_1->addr];
 667
 668		list_add_tail(&meta_2->list, &kfence_freelist);
 669	}
 670	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 671		kfence_metadata_init[i].addr = addr;
 672		addr += 2 * PAGE_SIZE;
 673	}
 674
 675	/*
 676	 * Make kfence_metadata visible only when initialization is successful.
 677	 * Otherwise, if the initialization fails and kfence_metadata is freed,
 678	 * it may cause UAF in kfence_shutdown_cache().
 679	 */
 680	smp_store_release(&kfence_metadata, kfence_metadata_init);
 681	return 0;
 682
 683reset_slab:
 684	addr += 2 * i * PAGE_SIZE;
 685	for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
 686		struct page *page;
 687
 688		if (!i || (i % 2))
 689			continue;
 690
 691		page = pfn_to_page(start_pfn + i);
 692#ifdef CONFIG_MEMCG
 693		struct slab *slab = page_slab(page);
 694		slab->obj_exts = 0;
 695#endif
 696		__ClearPageSlab(page);
 697	}
 698
 699	return addr;
 700}
 701
 702static bool __init kfence_init_pool_early(void)
 703{
 704	unsigned long addr;
 705
 706	if (!__kfence_pool)
 707		return false;
 708
 709	addr = kfence_init_pool();
 710
 711	if (!addr) {
 712		/*
 713		 * The pool is live and will never be deallocated from this point on.
 714		 * Ignore the pool object from the kmemleak phys object tree, as it would
 715		 * otherwise overlap with allocations returned by kfence_alloc(), which
 716		 * are registered with kmemleak through the slab post-alloc hook.
 717		 */
 718		kmemleak_ignore_phys(__pa(__kfence_pool));
 719		return true;
 720	}
 721
 722	/*
 723	 * Only release unprotected pages, and do not try to go back and change
 724	 * page attributes due to risk of failing to do so as well. If changing
 725	 * page attributes for some pages fails, it is very likely that it also
 726	 * fails for the first page, and therefore expect addr==__kfence_pool in
 727	 * most failure cases.
 728	 */
 729	memblock_free_late(__pa(addr), KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool));
 730	__kfence_pool = NULL;
 731
 732	memblock_free_late(__pa(kfence_metadata_init), KFENCE_METADATA_SIZE);
 733	kfence_metadata_init = NULL;
 734
 735	return false;
 736}
 737
 738/* === DebugFS Interface ==================================================== */
 739
 740static int stats_show(struct seq_file *seq, void *v)
 741{
 742	int i;
 743
 744	seq_printf(seq, "enabled: %i\n", READ_ONCE(kfence_enabled));
 745	for (i = 0; i < KFENCE_COUNTER_COUNT; i++)
 746		seq_printf(seq, "%s: %ld\n", counter_names[i], atomic_long_read(&counters[i]));
 747
 748	return 0;
 749}
 750DEFINE_SHOW_ATTRIBUTE(stats);
 751
 752/*
 753 * debugfs seq_file operations for /sys/kernel/debug/kfence/objects.
 754 * start_object() and next_object() return the object index + 1, because NULL is used
 755 * to stop iteration.
 756 */
 757static void *start_object(struct seq_file *seq, loff_t *pos)
 758{
 759	if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
 760		return (void *)((long)*pos + 1);
 761	return NULL;
 762}
 763
 764static void stop_object(struct seq_file *seq, void *v)
 765{
 766}
 767
 768static void *next_object(struct seq_file *seq, void *v, loff_t *pos)
 769{
 770	++*pos;
 771	if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
 772		return (void *)((long)*pos + 1);
 773	return NULL;
 774}
 775
 776static int show_object(struct seq_file *seq, void *v)
 777{
 778	struct kfence_metadata *meta = &kfence_metadata[(long)v - 1];
 779	unsigned long flags;
 780
 781	raw_spin_lock_irqsave(&meta->lock, flags);
 782	kfence_print_object(seq, meta);
 783	raw_spin_unlock_irqrestore(&meta->lock, flags);
 784	seq_puts(seq, "---------------------------------\n");
 785
 786	return 0;
 787}
 788
 789static const struct seq_operations objects_sops = {
 790	.start = start_object,
 791	.next = next_object,
 792	.stop = stop_object,
 793	.show = show_object,
 794};
 795DEFINE_SEQ_ATTRIBUTE(objects);
 796
 797static int kfence_debugfs_init(void)
 798{
 799	struct dentry *kfence_dir;
 800
 801	if (!READ_ONCE(kfence_enabled))
 802		return 0;
 803
 804	kfence_dir = debugfs_create_dir("kfence", NULL);
 805	debugfs_create_file("stats", 0444, kfence_dir, NULL, &stats_fops);
 806	debugfs_create_file("objects", 0400, kfence_dir, NULL, &objects_fops);
 807	return 0;
 808}
 809
 810late_initcall(kfence_debugfs_init);
 811
 812/* === Panic Notifier ====================================================== */
 813
 814static void kfence_check_all_canary(void)
 815{
 816	int i;
 817
 818	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 819		struct kfence_metadata *meta = &kfence_metadata[i];
 820
 821		if (kfence_obj_allocated(meta))
 822			check_canary(meta);
 823	}
 824}
 825
 826static int kfence_check_canary_callback(struct notifier_block *nb,
 827					unsigned long reason, void *arg)
 828{
 829	kfence_check_all_canary();
 830	return NOTIFY_OK;
 831}
 832
 833static struct notifier_block kfence_check_canary_notifier = {
 834	.notifier_call = kfence_check_canary_callback,
 835};
 836
 837/* === Allocation Gate Timer ================================================ */
 838
 839static struct delayed_work kfence_timer;
 840
 841#ifdef CONFIG_KFENCE_STATIC_KEYS
 842/* Wait queue to wake up allocation-gate timer task. */
 843static DECLARE_WAIT_QUEUE_HEAD(allocation_wait);
 844
 845static int kfence_reboot_callback(struct notifier_block *nb,
 846				  unsigned long action, void *data)
 847{
 848	/*
 849	 * Disable kfence to avoid static keys IPI synchronization during
 850	 * late shutdown/kexec
 851	 */
 852	WRITE_ONCE(kfence_enabled, false);
 853	/* Cancel any pending timer work */
 854	cancel_delayed_work(&kfence_timer);
 855	/*
 856	 * Wake up any blocked toggle_allocation_gate() so it can complete
 857	 * early while the system is still able to handle IPIs.
 858	 */
 859	wake_up(&allocation_wait);
 860
 861	return NOTIFY_OK;
 862}
 863
 864static struct notifier_block kfence_reboot_notifier = {
 865	.notifier_call = kfence_reboot_callback,
 866	.priority = INT_MAX, /* Run early to stop timers ASAP */
 867};
 868
 869static void wake_up_kfence_timer(struct irq_work *work)
 870{
 871	wake_up(&allocation_wait);
 872}
 873static DEFINE_IRQ_WORK(wake_up_kfence_timer_work, wake_up_kfence_timer);
 874#endif
 875
 876/*
 877 * Set up delayed work, which will enable and disable the static key. We need to
 878 * use a work queue (rather than a simple timer), since enabling and disabling a
 879 * static key cannot be done from an interrupt.
 880 *
 881 * Note: Toggling a static branch currently causes IPIs, and here we'll end up
 882 * with a total of 2 IPIs to all CPUs. If this ends up a problem in future (with
 883 * more aggressive sampling intervals), we could get away with a variant that
 884 * avoids IPIs, at the cost of not immediately capturing allocations if the
 885 * instructions remain cached.
 886 */
 887static void toggle_allocation_gate(struct work_struct *work)
 888{
 889	if (!READ_ONCE(kfence_enabled))
 890		return;
 891
 892	atomic_set(&kfence_allocation_gate, -kfence_burst);
 893#ifdef CONFIG_KFENCE_STATIC_KEYS
 894	/* Enable static key, and await allocation to happen. */
 895	static_branch_enable(&kfence_allocation_key);
 896
 897	wait_event_idle(allocation_wait,
 898			atomic_read(&kfence_allocation_gate) > 0 ||
 899			!READ_ONCE(kfence_enabled));
 900
 901	/* Disable static key and reset timer. */
 902	static_branch_disable(&kfence_allocation_key);
 903#endif
 904	queue_delayed_work(system_unbound_wq, &kfence_timer,
 905			   msecs_to_jiffies(kfence_sample_interval));
 906}
 907
 908/* === Public interface ===================================================== */
 909
 910void __init kfence_alloc_pool_and_metadata(void)
 911{
 912	if (!kfence_sample_interval)
 913		return;
 914
 915	/*
 916	 * If KASAN hardware tags are enabled, disable KFENCE, because it
 917	 * does not support MTE yet.
 918	 */
 919	if (kasan_hw_tags_enabled()) {
 920		pr_info("disabled as KASAN HW tags are enabled\n");
 921		if (__kfence_pool) {
 922			memblock_free(__kfence_pool, KFENCE_POOL_SIZE);
 923			__kfence_pool = NULL;
 924		}
 925		kfence_sample_interval = 0;
 926		return;
 927	}
 928
 929	/*
 930	 * If the pool has already been initialized by arch, there is no need to
 931	 * re-allocate the memory pool.
 932	 */
 933	if (!__kfence_pool)
 934		__kfence_pool = memblock_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
 935
 936	if (!__kfence_pool) {
 937		pr_err("failed to allocate pool\n");
 938		return;
 939	}
 940
 941	/* The memory allocated by memblock has been zeroed out. */
 942	kfence_metadata_init = memblock_alloc(KFENCE_METADATA_SIZE, PAGE_SIZE);
 943	if (!kfence_metadata_init) {
 944		pr_err("failed to allocate metadata\n");
 945		memblock_free(__kfence_pool, KFENCE_POOL_SIZE);
 946		__kfence_pool = NULL;
 947	}
 948}
 949
 950static void kfence_init_enable(void)
 951{
 952	if (!IS_ENABLED(CONFIG_KFENCE_STATIC_KEYS))
 953		static_branch_enable(&kfence_allocation_key);
 954
 955	if (kfence_deferrable)
 956		INIT_DEFERRABLE_WORK(&kfence_timer, toggle_allocation_gate);
 957	else
 958		INIT_DELAYED_WORK(&kfence_timer, toggle_allocation_gate);
 959
 960	if (kfence_check_on_panic)
 961		atomic_notifier_chain_register(&panic_notifier_list, &kfence_check_canary_notifier);
 962
 963#ifdef CONFIG_KFENCE_STATIC_KEYS
 964	register_reboot_notifier(&kfence_reboot_notifier);
 965#endif
 966
 967	WRITE_ONCE(kfence_enabled, true);
 968	queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
 969
 970	pr_info("initialized - using %lu bytes for %d objects at 0x%p-0x%p\n", KFENCE_POOL_SIZE,
 971		CONFIG_KFENCE_NUM_OBJECTS, (void *)__kfence_pool,
 972		(void *)(__kfence_pool + KFENCE_POOL_SIZE));
 973}
 974
 975void __init kfence_init(void)
 976{
 977	stack_hash_seed = get_random_u32();
 978
 979	/* Setting kfence_sample_interval to 0 on boot disables KFENCE. */
 980	if (!kfence_sample_interval)
 981		return;
 982
 983	if (!kfence_init_pool_early()) {
 984		pr_err("%s failed\n", __func__);
 985		return;
 986	}
 987
 988	kfence_init_enable();
 989}
 990
 991static int kfence_init_late(void)
 992{
 993	const unsigned long nr_pages_pool = KFENCE_POOL_SIZE / PAGE_SIZE;
 994	const unsigned long nr_pages_meta = KFENCE_METADATA_SIZE / PAGE_SIZE;
 995	unsigned long addr = (unsigned long)__kfence_pool;
 996	unsigned long free_size = KFENCE_POOL_SIZE;
 997	int err = -ENOMEM;
 998
 999#ifdef CONFIG_CONTIG_ALLOC
1000	struct page *pages;
1001
1002	pages = alloc_contig_pages(nr_pages_pool, GFP_KERNEL | __GFP_SKIP_KASAN,
1003				   first_online_node, NULL);
1004	if (!pages)
1005		return -ENOMEM;
1006
1007	__kfence_pool = page_to_virt(pages);
1008	pages = alloc_contig_pages(nr_pages_meta, GFP_KERNEL | __GFP_SKIP_KASAN,
1009				   first_online_node, NULL);
1010	if (pages)
1011		kfence_metadata_init = page_to_virt(pages);
1012#else
1013	if (nr_pages_pool > MAX_ORDER_NR_PAGES ||
1014	    nr_pages_meta > MAX_ORDER_NR_PAGES) {
1015		pr_warn("KFENCE_NUM_OBJECTS too large for buddy allocator\n");
1016		return -EINVAL;
1017	}
1018
1019	__kfence_pool = alloc_pages_exact(KFENCE_POOL_SIZE,
1020					  GFP_KERNEL | __GFP_SKIP_KASAN);
1021	if (!__kfence_pool)
1022		return -ENOMEM;
1023
1024	kfence_metadata_init = alloc_pages_exact(KFENCE_METADATA_SIZE,
1025						 GFP_KERNEL | __GFP_SKIP_KASAN);
1026#endif
1027
1028	if (!kfence_metadata_init)
1029		goto free_pool;
1030
1031	memzero_explicit(kfence_metadata_init, KFENCE_METADATA_SIZE);
1032	addr = kfence_init_pool();
1033	if (!addr) {
1034		kfence_init_enable();
1035		kfence_debugfs_init();
1036		return 0;
1037	}
1038
1039	pr_err("%s failed\n", __func__);
1040	free_size = KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool);
1041	err = -EBUSY;
1042
1043#ifdef CONFIG_CONTIG_ALLOC
1044	free_contig_range(page_to_pfn(virt_to_page((void *)kfence_metadata_init)),
1045			  nr_pages_meta);
1046free_pool:
1047	free_contig_range(page_to_pfn(virt_to_page((void *)addr)),
1048			  free_size / PAGE_SIZE);
1049#else
1050	free_pages_exact((void *)kfence_metadata_init, KFENCE_METADATA_SIZE);
1051free_pool:
1052	free_pages_exact((void *)addr, free_size);
1053#endif
1054
1055	kfence_metadata_init = NULL;
1056	__kfence_pool = NULL;
1057	return err;
1058}
1059
1060static int kfence_enable_late(void)
1061{
1062	if (!__kfence_pool)
1063		return kfence_init_late();
1064
1065	WRITE_ONCE(kfence_enabled, true);
1066	queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
1067	pr_info("re-enabled\n");
1068	return 0;
1069}
1070
1071void kfence_shutdown_cache(struct kmem_cache *s)
1072{
1073	unsigned long flags;
1074	struct kfence_metadata *meta;
1075	int i;
1076
1077	/* Pairs with release in kfence_init_pool(). */
1078	if (!smp_load_acquire(&kfence_metadata))
1079		return;
1080
1081	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
1082		bool in_use;
1083
1084		meta = &kfence_metadata[i];
1085
1086		/*
1087		 * If we observe some inconsistent cache and state pair where we
1088		 * should have returned false here, cache destruction is racing
1089		 * with either kmem_cache_alloc() or kmem_cache_free(). Taking
1090		 * the lock will not help, as different critical section
1091		 * serialization will have the same outcome.
1092		 */
1093		if (READ_ONCE(meta->cache) != s || !kfence_obj_allocated(meta))
1094			continue;
1095
1096		raw_spin_lock_irqsave(&meta->lock, flags);
1097		in_use = meta->cache == s && kfence_obj_allocated(meta);
1098		raw_spin_unlock_irqrestore(&meta->lock, flags);
1099
1100		if (in_use) {
1101			/*
1102			 * This cache still has allocations, and we should not
1103			 * release them back into the freelist so they can still
1104			 * safely be used and retain the kernel's default
1105			 * behaviour of keeping the allocations alive (leak the
1106			 * cache); however, they effectively become "zombie
1107			 * allocations" as the KFENCE objects are the only ones
1108			 * still in use and the owning cache is being destroyed.
1109			 *
1110			 * We mark them freed, so that any subsequent use shows
1111			 * more useful error messages that will include stack
1112			 * traces of the user of the object, the original
1113			 * allocation, and caller to shutdown_cache().
1114			 */
1115			kfence_guarded_free((void *)meta->addr, meta, /*zombie=*/true);
1116		}
1117	}
1118
1119	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
1120		meta = &kfence_metadata[i];
1121
1122		/* See above. */
1123		if (READ_ONCE(meta->cache) != s || READ_ONCE(meta->state) != KFENCE_OBJECT_FREED)
1124			continue;
1125
1126		raw_spin_lock_irqsave(&meta->lock, flags);
1127		if (meta->cache == s && meta->state == KFENCE_OBJECT_FREED)
1128			meta->cache = NULL;
1129		raw_spin_unlock_irqrestore(&meta->lock, flags);
1130	}
1131}
1132
1133void *__kfence_alloc(struct kmem_cache *s, size_t size, gfp_t flags)
1134{
1135	unsigned long stack_entries[KFENCE_STACK_DEPTH];
1136	size_t num_stack_entries;
1137	u32 alloc_stack_hash;
1138	int allocation_gate;
1139
1140	/*
1141	 * Perform size check before switching kfence_allocation_gate, so that
1142	 * we don't disable KFENCE without making an allocation.
1143	 */
1144	if (size > PAGE_SIZE) {
1145		atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
1146		return NULL;
1147	}
1148
1149	/*
1150	 * Skip allocations from non-default zones, including DMA. We cannot
1151	 * guarantee that pages in the KFENCE pool will have the requested
1152	 * properties (e.g. reside in DMAable memory).
1153	 */
1154	if ((flags & GFP_ZONEMASK) ||
1155	    ((flags & __GFP_THISNODE) && num_online_nodes() > 1) ||
1156	    (s->flags & (SLAB_CACHE_DMA | SLAB_CACHE_DMA32))) {
1157		atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
1158		return NULL;
1159	}
1160
1161	/*
1162	 * Skip allocations for this slab, if KFENCE has been disabled for
1163	 * this slab.
1164	 */
1165	if (s->flags & SLAB_SKIP_KFENCE)
1166		return NULL;
1167
1168	allocation_gate = atomic_inc_return(&kfence_allocation_gate);
1169	if (allocation_gate > 1)
1170		return NULL;
1171#ifdef CONFIG_KFENCE_STATIC_KEYS
1172	/*
1173	 * waitqueue_active() is fully ordered after the update of
1174	 * kfence_allocation_gate per atomic_inc_return().
1175	 */
1176	if (allocation_gate == 1 && waitqueue_active(&allocation_wait)) {
1177		/*
1178		 * Calling wake_up() here may deadlock when allocations happen
1179		 * from within timer code. Use an irq_work to defer it.
1180		 */
1181		irq_work_queue(&wake_up_kfence_timer_work);
1182	}
1183#endif
1184
1185	if (!READ_ONCE(kfence_enabled))
1186		return NULL;
1187
1188	num_stack_entries = stack_trace_save(stack_entries, KFENCE_STACK_DEPTH, 0);
1189
1190	/*
1191	 * Do expensive check for coverage of allocation in slow-path after
1192	 * allocation_gate has already become non-zero, even though it might
1193	 * mean not making any allocation within a given sample interval.
1194	 *
1195	 * This ensures reasonable allocation coverage when the pool is almost
1196	 * full, including avoiding long-lived allocations of the same source
1197	 * filling up the pool (e.g. pagecache allocations).
1198	 */
1199	alloc_stack_hash = get_alloc_stack_hash(stack_entries, num_stack_entries);
1200	if (should_skip_covered() && alloc_covered_contains(alloc_stack_hash)) {
1201		atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_COVERED]);
1202		return NULL;
1203	}
1204
1205	return kfence_guarded_alloc(s, size, flags, stack_entries, num_stack_entries,
1206				    alloc_stack_hash);
1207}
1208
1209size_t kfence_ksize(const void *addr)
1210{
1211	const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1212
1213	/*
1214	 * Read locklessly -- if there is a race with __kfence_alloc(), this is
1215	 * either a use-after-free or invalid access.
1216	 */
1217	return meta ? meta->size : 0;
1218}
1219
1220void *kfence_object_start(const void *addr)
1221{
1222	const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1223
1224	/*
1225	 * Read locklessly -- if there is a race with __kfence_alloc(), this is
1226	 * either a use-after-free or invalid access.
1227	 */
1228	return meta ? (void *)meta->addr : NULL;
1229}
1230
1231void __kfence_free(void *addr)
1232{
1233	struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1234
1235#ifdef CONFIG_MEMCG
1236	KFENCE_WARN_ON(meta->obj_exts.objcg);
1237#endif
1238	/*
1239	 * If the objects of the cache are SLAB_TYPESAFE_BY_RCU, defer freeing
1240	 * the object, as the object page may be recycled for other-typed
1241	 * objects once it has been freed. meta->cache may be NULL if the cache
1242	 * was destroyed.
1243	 * Save the stack trace here so that reports show where the user freed
1244	 * the object.
1245	 */
1246	if (unlikely(meta->cache && (meta->cache->flags & SLAB_TYPESAFE_BY_RCU))) {
1247		unsigned long flags;
1248
1249		raw_spin_lock_irqsave(&meta->lock, flags);
1250		metadata_update_state(meta, KFENCE_OBJECT_RCU_FREEING, NULL, 0);
1251		raw_spin_unlock_irqrestore(&meta->lock, flags);
1252		call_rcu(&meta->rcu_head, rcu_guarded_free);
1253	} else {
1254		kfence_guarded_free(addr, meta, false);
1255	}
1256}
1257
1258bool kfence_handle_page_fault(unsigned long addr, bool is_write, struct pt_regs *regs)
1259{
1260	const int page_index = (addr - (unsigned long)__kfence_pool) / PAGE_SIZE;
1261	struct kfence_metadata *to_report = NULL;
1262	enum kfence_error_type error_type;
1263	unsigned long flags;
1264
1265	if (!is_kfence_address((void *)addr))
1266		return false;
1267
1268	if (!READ_ONCE(kfence_enabled)) /* If disabled at runtime ... */
1269		return kfence_unprotect(addr); /* ... unprotect and proceed. */
1270
1271	atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
1272
1273	if (page_index % 2) {
1274		/* This is a redzone, report a buffer overflow. */
1275		struct kfence_metadata *meta;
1276		int distance = 0;
1277
1278		meta = addr_to_metadata(addr - PAGE_SIZE);
1279		if (meta && kfence_obj_allocated(meta)) {
1280			to_report = meta;
1281			/* Data race ok; distance calculation approximate. */
1282			distance = addr - data_race(meta->addr + meta->size);
1283		}
1284
1285		meta = addr_to_metadata(addr + PAGE_SIZE);
1286		if (meta && kfence_obj_allocated(meta)) {
1287			/* Data race ok; distance calculation approximate. */
1288			if (!to_report || distance > data_race(meta->addr) - addr)
1289				to_report = meta;
1290		}
1291
1292		if (!to_report)
1293			goto out;
1294
1295		raw_spin_lock_irqsave(&to_report->lock, flags);
1296		to_report->unprotected_page = addr;
1297		error_type = KFENCE_ERROR_OOB;
1298
1299		/*
1300		 * If the object was freed before we took the look we can still
1301		 * report this as an OOB -- the report will simply show the
1302		 * stacktrace of the free as well.
1303		 */
1304	} else {
1305		to_report = addr_to_metadata(addr);
1306		if (!to_report)
1307			goto out;
1308
1309		raw_spin_lock_irqsave(&to_report->lock, flags);
1310		error_type = KFENCE_ERROR_UAF;
1311		/*
1312		 * We may race with __kfence_alloc(), and it is possible that a
1313		 * freed object may be reallocated. We simply report this as a
1314		 * use-after-free, with the stack trace showing the place where
1315		 * the object was re-allocated.
1316		 */
1317	}
1318
1319out:
1320	if (to_report) {
1321		kfence_report_error(addr, is_write, regs, to_report, error_type);
1322		raw_spin_unlock_irqrestore(&to_report->lock, flags);
1323	} else {
1324		/* This may be a UAF or OOB access, but we can't be sure. */
1325		kfence_report_error(addr, is_write, regs, NULL, KFENCE_ERROR_INVALID);
1326	}
1327
1328	return kfence_unprotect(addr); /* Unprotect and let access proceed. */
1329}