xref: /linux/mm/kfence/core.c (revision 0a94608f0f7de9b1135ffea3546afe68eafef57f)
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 			disabled_by_warn = true;                               \
44 		}                                                              \
45 		__cond;                                                        \
46 	})
47 
48 /* === Data ================================================================= */
49 
50 static bool kfence_enabled __read_mostly;
51 static bool disabled_by_warn __read_mostly;
52 
53 unsigned long kfence_sample_interval __read_mostly = CONFIG_KFENCE_SAMPLE_INTERVAL;
54 EXPORT_SYMBOL_GPL(kfence_sample_interval); /* Export for test modules. */
55 
56 #ifdef MODULE_PARAM_PREFIX
57 #undef MODULE_PARAM_PREFIX
58 #endif
59 #define MODULE_PARAM_PREFIX "kfence."
60 
61 static int kfence_enable_late(void);
62 static int param_set_sample_interval(const char *val, const struct kernel_param *kp)
63 {
64 	unsigned long num;
65 	int ret = kstrtoul(val, 0, &num);
66 
67 	if (ret < 0)
68 		return ret;
69 
70 	if (!num) /* Using 0 to indicate KFENCE is disabled. */
71 		WRITE_ONCE(kfence_enabled, false);
72 
73 	*((unsigned long *)kp->arg) = num;
74 
75 	if (num && !READ_ONCE(kfence_enabled) && system_state != SYSTEM_BOOTING)
76 		return disabled_by_warn ? -EINVAL : kfence_enable_late();
77 	return 0;
78 }
79 
80 static int param_get_sample_interval(char *buffer, const struct kernel_param *kp)
81 {
82 	if (!READ_ONCE(kfence_enabled))
83 		return sprintf(buffer, "0\n");
84 
85 	return param_get_ulong(buffer, kp);
86 }
87 
88 static const struct kernel_param_ops sample_interval_param_ops = {
89 	.set = param_set_sample_interval,
90 	.get = param_get_sample_interval,
91 };
92 module_param_cb(sample_interval, &sample_interval_param_ops, &kfence_sample_interval, 0600);
93 
94 /* Pool usage% threshold when currently covered allocations are skipped. */
95 static unsigned long kfence_skip_covered_thresh __read_mostly = 75;
96 module_param_named(skip_covered_thresh, kfence_skip_covered_thresh, ulong, 0644);
97 
98 /* If true, use a deferrable timer. */
99 static bool kfence_deferrable __read_mostly = IS_ENABLED(CONFIG_KFENCE_DEFERRABLE);
100 module_param_named(deferrable, kfence_deferrable, bool, 0444);
101 
102 /* The pool of pages used for guard pages and objects. */
103 char *__kfence_pool __read_mostly;
104 EXPORT_SYMBOL(__kfence_pool); /* Export for test modules. */
105 
106 /*
107  * Per-object metadata, with one-to-one mapping of object metadata to
108  * backing pages (in __kfence_pool).
109  */
110 static_assert(CONFIG_KFENCE_NUM_OBJECTS > 0);
111 struct kfence_metadata kfence_metadata[CONFIG_KFENCE_NUM_OBJECTS];
112 
113 /* Freelist with available objects. */
114 static struct list_head kfence_freelist = LIST_HEAD_INIT(kfence_freelist);
115 static DEFINE_RAW_SPINLOCK(kfence_freelist_lock); /* Lock protecting freelist. */
116 
117 /*
118  * The static key to set up a KFENCE allocation; or if static keys are not used
119  * to gate allocations, to avoid a load and compare if KFENCE is disabled.
120  */
121 DEFINE_STATIC_KEY_FALSE(kfence_allocation_key);
122 
123 /* Gates the allocation, ensuring only one succeeds in a given period. */
124 atomic_t kfence_allocation_gate = ATOMIC_INIT(1);
125 
126 /*
127  * A Counting Bloom filter of allocation coverage: limits currently covered
128  * allocations of the same source filling up the pool.
129  *
130  * Assuming a range of 15%-85% unique allocations in the pool at any point in
131  * time, the below parameters provide a probablity of 0.02-0.33 for false
132  * positive hits respectively:
133  *
134  *	P(alloc_traces) = (1 - e^(-HNUM * (alloc_traces / SIZE)) ^ HNUM
135  */
136 #define ALLOC_COVERED_HNUM	2
137 #define ALLOC_COVERED_ORDER	(const_ilog2(CONFIG_KFENCE_NUM_OBJECTS) + 2)
138 #define ALLOC_COVERED_SIZE	(1 << ALLOC_COVERED_ORDER)
139 #define ALLOC_COVERED_HNEXT(h)	hash_32(h, ALLOC_COVERED_ORDER)
140 #define ALLOC_COVERED_MASK	(ALLOC_COVERED_SIZE - 1)
141 static atomic_t alloc_covered[ALLOC_COVERED_SIZE];
142 
143 /* Stack depth used to determine uniqueness of an allocation. */
144 #define UNIQUE_ALLOC_STACK_DEPTH ((size_t)8)
145 
146 /*
147  * Randomness for stack hashes, making the same collisions across reboots and
148  * different machines less likely.
149  */
150 static u32 stack_hash_seed __ro_after_init;
151 
152 /* Statistics counters for debugfs. */
153 enum kfence_counter_id {
154 	KFENCE_COUNTER_ALLOCATED,
155 	KFENCE_COUNTER_ALLOCS,
156 	KFENCE_COUNTER_FREES,
157 	KFENCE_COUNTER_ZOMBIES,
158 	KFENCE_COUNTER_BUGS,
159 	KFENCE_COUNTER_SKIP_INCOMPAT,
160 	KFENCE_COUNTER_SKIP_CAPACITY,
161 	KFENCE_COUNTER_SKIP_COVERED,
162 	KFENCE_COUNTER_COUNT,
163 };
164 static atomic_long_t counters[KFENCE_COUNTER_COUNT];
165 static const char *const counter_names[] = {
166 	[KFENCE_COUNTER_ALLOCATED]	= "currently allocated",
167 	[KFENCE_COUNTER_ALLOCS]		= "total allocations",
168 	[KFENCE_COUNTER_FREES]		= "total frees",
169 	[KFENCE_COUNTER_ZOMBIES]	= "zombie allocations",
170 	[KFENCE_COUNTER_BUGS]		= "total bugs",
171 	[KFENCE_COUNTER_SKIP_INCOMPAT]	= "skipped allocations (incompatible)",
172 	[KFENCE_COUNTER_SKIP_CAPACITY]	= "skipped allocations (capacity)",
173 	[KFENCE_COUNTER_SKIP_COVERED]	= "skipped allocations (covered)",
174 };
175 static_assert(ARRAY_SIZE(counter_names) == KFENCE_COUNTER_COUNT);
176 
177 /* === Internals ============================================================ */
178 
179 static inline bool should_skip_covered(void)
180 {
181 	unsigned long thresh = (CONFIG_KFENCE_NUM_OBJECTS * kfence_skip_covered_thresh) / 100;
182 
183 	return atomic_long_read(&counters[KFENCE_COUNTER_ALLOCATED]) > thresh;
184 }
185 
186 static u32 get_alloc_stack_hash(unsigned long *stack_entries, size_t num_entries)
187 {
188 	num_entries = min(num_entries, UNIQUE_ALLOC_STACK_DEPTH);
189 	num_entries = filter_irq_stacks(stack_entries, num_entries);
190 	return jhash(stack_entries, num_entries * sizeof(stack_entries[0]), stack_hash_seed);
191 }
192 
193 /*
194  * Adds (or subtracts) count @val for allocation stack trace hash
195  * @alloc_stack_hash from Counting Bloom filter.
196  */
197 static void alloc_covered_add(u32 alloc_stack_hash, int val)
198 {
199 	int i;
200 
201 	for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
202 		atomic_add(val, &alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]);
203 		alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
204 	}
205 }
206 
207 /*
208  * Returns true if the allocation stack trace hash @alloc_stack_hash is
209  * currently contained (non-zero count) in Counting Bloom filter.
210  */
211 static bool alloc_covered_contains(u32 alloc_stack_hash)
212 {
213 	int i;
214 
215 	for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
216 		if (!atomic_read(&alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]))
217 			return false;
218 		alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
219 	}
220 
221 	return true;
222 }
223 
224 static bool kfence_protect(unsigned long addr)
225 {
226 	return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), true));
227 }
228 
229 static bool kfence_unprotect(unsigned long addr)
230 {
231 	return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), false));
232 }
233 
234 static inline unsigned long metadata_to_pageaddr(const struct kfence_metadata *meta)
235 {
236 	unsigned long offset = (meta - kfence_metadata + 1) * PAGE_SIZE * 2;
237 	unsigned long pageaddr = (unsigned long)&__kfence_pool[offset];
238 
239 	/* The checks do not affect performance; only called from slow-paths. */
240 
241 	/* Only call with a pointer into kfence_metadata. */
242 	if (KFENCE_WARN_ON(meta < kfence_metadata ||
243 			   meta >= kfence_metadata + CONFIG_KFENCE_NUM_OBJECTS))
244 		return 0;
245 
246 	/*
247 	 * This metadata object only ever maps to 1 page; verify that the stored
248 	 * address is in the expected range.
249 	 */
250 	if (KFENCE_WARN_ON(ALIGN_DOWN(meta->addr, PAGE_SIZE) != pageaddr))
251 		return 0;
252 
253 	return pageaddr;
254 }
255 
256 /*
257  * Update the object's metadata state, including updating the alloc/free stacks
258  * depending on the state transition.
259  */
260 static noinline void
261 metadata_update_state(struct kfence_metadata *meta, enum kfence_object_state next,
262 		      unsigned long *stack_entries, size_t num_stack_entries)
263 {
264 	struct kfence_track *track =
265 		next == KFENCE_OBJECT_FREED ? &meta->free_track : &meta->alloc_track;
266 
267 	lockdep_assert_held(&meta->lock);
268 
269 	if (stack_entries) {
270 		memcpy(track->stack_entries, stack_entries,
271 		       num_stack_entries * sizeof(stack_entries[0]));
272 	} else {
273 		/*
274 		 * Skip over 1 (this) functions; noinline ensures we do not
275 		 * accidentally skip over the caller by never inlining.
276 		 */
277 		num_stack_entries = stack_trace_save(track->stack_entries, KFENCE_STACK_DEPTH, 1);
278 	}
279 	track->num_stack_entries = num_stack_entries;
280 	track->pid = task_pid_nr(current);
281 	track->cpu = raw_smp_processor_id();
282 	track->ts_nsec = local_clock(); /* Same source as printk timestamps. */
283 
284 	/*
285 	 * Pairs with READ_ONCE() in
286 	 *	kfence_shutdown_cache(),
287 	 *	kfence_handle_page_fault().
288 	 */
289 	WRITE_ONCE(meta->state, next);
290 }
291 
292 /* Write canary byte to @addr. */
293 static inline bool set_canary_byte(u8 *addr)
294 {
295 	*addr = KFENCE_CANARY_PATTERN(addr);
296 	return true;
297 }
298 
299 /* Check canary byte at @addr. */
300 static inline bool check_canary_byte(u8 *addr)
301 {
302 	struct kfence_metadata *meta;
303 	unsigned long flags;
304 
305 	if (likely(*addr == KFENCE_CANARY_PATTERN(addr)))
306 		return true;
307 
308 	atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
309 
310 	meta = addr_to_metadata((unsigned long)addr);
311 	raw_spin_lock_irqsave(&meta->lock, flags);
312 	kfence_report_error((unsigned long)addr, false, NULL, meta, KFENCE_ERROR_CORRUPTION);
313 	raw_spin_unlock_irqrestore(&meta->lock, flags);
314 
315 	return false;
316 }
317 
318 /* __always_inline this to ensure we won't do an indirect call to fn. */
319 static __always_inline void for_each_canary(const struct kfence_metadata *meta, bool (*fn)(u8 *))
320 {
321 	const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
322 	unsigned long addr;
323 
324 	/*
325 	 * We'll iterate over each canary byte per-side until fn() returns
326 	 * false. However, we'll still iterate over the canary bytes to the
327 	 * right of the object even if there was an error in the canary bytes to
328 	 * the left of the object. Specifically, if check_canary_byte()
329 	 * generates an error, showing both sides might give more clues as to
330 	 * what the error is about when displaying which bytes were corrupted.
331 	 */
332 
333 	/* Apply to left of object. */
334 	for (addr = pageaddr; addr < meta->addr; addr++) {
335 		if (!fn((u8 *)addr))
336 			break;
337 	}
338 
339 	/* Apply to right of object. */
340 	for (addr = meta->addr + meta->size; addr < pageaddr + PAGE_SIZE; addr++) {
341 		if (!fn((u8 *)addr))
342 			break;
343 	}
344 }
345 
346 static void *kfence_guarded_alloc(struct kmem_cache *cache, size_t size, gfp_t gfp,
347 				  unsigned long *stack_entries, size_t num_stack_entries,
348 				  u32 alloc_stack_hash)
349 {
350 	struct kfence_metadata *meta = NULL;
351 	unsigned long flags;
352 	struct slab *slab;
353 	void *addr;
354 
355 	/* Try to obtain a free object. */
356 	raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
357 	if (!list_empty(&kfence_freelist)) {
358 		meta = list_entry(kfence_freelist.next, struct kfence_metadata, list);
359 		list_del_init(&meta->list);
360 	}
361 	raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
362 	if (!meta) {
363 		atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_CAPACITY]);
364 		return NULL;
365 	}
366 
367 	if (unlikely(!raw_spin_trylock_irqsave(&meta->lock, flags))) {
368 		/*
369 		 * This is extremely unlikely -- we are reporting on a
370 		 * use-after-free, which locked meta->lock, and the reporting
371 		 * code via printk calls kmalloc() which ends up in
372 		 * kfence_alloc() and tries to grab the same object that we're
373 		 * reporting on. While it has never been observed, lockdep does
374 		 * report that there is a possibility of deadlock. Fix it by
375 		 * using trylock and bailing out gracefully.
376 		 */
377 		raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
378 		/* Put the object back on the freelist. */
379 		list_add_tail(&meta->list, &kfence_freelist);
380 		raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
381 
382 		return NULL;
383 	}
384 
385 	meta->addr = metadata_to_pageaddr(meta);
386 	/* Unprotect if we're reusing this page. */
387 	if (meta->state == KFENCE_OBJECT_FREED)
388 		kfence_unprotect(meta->addr);
389 
390 	/*
391 	 * Note: for allocations made before RNG initialization, will always
392 	 * return zero. We still benefit from enabling KFENCE as early as
393 	 * possible, even when the RNG is not yet available, as this will allow
394 	 * KFENCE to detect bugs due to earlier allocations. The only downside
395 	 * is that the out-of-bounds accesses detected are deterministic for
396 	 * such allocations.
397 	 */
398 	if (prandom_u32_max(2)) {
399 		/* Allocate on the "right" side, re-calculate address. */
400 		meta->addr += PAGE_SIZE - size;
401 		meta->addr = ALIGN_DOWN(meta->addr, cache->align);
402 	}
403 
404 	addr = (void *)meta->addr;
405 
406 	/* Update remaining metadata. */
407 	metadata_update_state(meta, KFENCE_OBJECT_ALLOCATED, stack_entries, num_stack_entries);
408 	/* Pairs with READ_ONCE() in kfence_shutdown_cache(). */
409 	WRITE_ONCE(meta->cache, cache);
410 	meta->size = size;
411 	meta->alloc_stack_hash = alloc_stack_hash;
412 	raw_spin_unlock_irqrestore(&meta->lock, flags);
413 
414 	alloc_covered_add(alloc_stack_hash, 1);
415 
416 	/* Set required slab fields. */
417 	slab = virt_to_slab((void *)meta->addr);
418 	slab->slab_cache = cache;
419 #if defined(CONFIG_SLUB)
420 	slab->objects = 1;
421 #elif defined(CONFIG_SLAB)
422 	slab->s_mem = addr;
423 #endif
424 
425 	/* Memory initialization. */
426 	for_each_canary(meta, set_canary_byte);
427 
428 	/*
429 	 * We check slab_want_init_on_alloc() ourselves, rather than letting
430 	 * SL*B do the initialization, as otherwise we might overwrite KFENCE's
431 	 * redzone.
432 	 */
433 	if (unlikely(slab_want_init_on_alloc(gfp, cache)))
434 		memzero_explicit(addr, size);
435 	if (cache->ctor)
436 		cache->ctor(addr);
437 
438 	if (CONFIG_KFENCE_STRESS_TEST_FAULTS && !prandom_u32_max(CONFIG_KFENCE_STRESS_TEST_FAULTS))
439 		kfence_protect(meta->addr); /* Random "faults" by protecting the object. */
440 
441 	atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCATED]);
442 	atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCS]);
443 
444 	return addr;
445 }
446 
447 static void kfence_guarded_free(void *addr, struct kfence_metadata *meta, bool zombie)
448 {
449 	struct kcsan_scoped_access assert_page_exclusive;
450 	unsigned long flags;
451 	bool init;
452 
453 	raw_spin_lock_irqsave(&meta->lock, flags);
454 
455 	if (meta->state != KFENCE_OBJECT_ALLOCATED || meta->addr != (unsigned long)addr) {
456 		/* Invalid or double-free, bail out. */
457 		atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
458 		kfence_report_error((unsigned long)addr, false, NULL, meta,
459 				    KFENCE_ERROR_INVALID_FREE);
460 		raw_spin_unlock_irqrestore(&meta->lock, flags);
461 		return;
462 	}
463 
464 	/* Detect racy use-after-free, or incorrect reallocation of this page by KFENCE. */
465 	kcsan_begin_scoped_access((void *)ALIGN_DOWN((unsigned long)addr, PAGE_SIZE), PAGE_SIZE,
466 				  KCSAN_ACCESS_SCOPED | KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT,
467 				  &assert_page_exclusive);
468 
469 	if (CONFIG_KFENCE_STRESS_TEST_FAULTS)
470 		kfence_unprotect((unsigned long)addr); /* To check canary bytes. */
471 
472 	/* Restore page protection if there was an OOB access. */
473 	if (meta->unprotected_page) {
474 		memzero_explicit((void *)ALIGN_DOWN(meta->unprotected_page, PAGE_SIZE), PAGE_SIZE);
475 		kfence_protect(meta->unprotected_page);
476 		meta->unprotected_page = 0;
477 	}
478 
479 	/* Mark the object as freed. */
480 	metadata_update_state(meta, KFENCE_OBJECT_FREED, NULL, 0);
481 	init = slab_want_init_on_free(meta->cache);
482 	raw_spin_unlock_irqrestore(&meta->lock, flags);
483 
484 	alloc_covered_add(meta->alloc_stack_hash, -1);
485 
486 	/* Check canary bytes for memory corruption. */
487 	for_each_canary(meta, check_canary_byte);
488 
489 	/*
490 	 * Clear memory if init-on-free is set. While we protect the page, the
491 	 * data is still there, and after a use-after-free is detected, we
492 	 * unprotect the page, so the data is still accessible.
493 	 */
494 	if (!zombie && unlikely(init))
495 		memzero_explicit(addr, meta->size);
496 
497 	/* Protect to detect use-after-frees. */
498 	kfence_protect((unsigned long)addr);
499 
500 	kcsan_end_scoped_access(&assert_page_exclusive);
501 	if (!zombie) {
502 		/* Add it to the tail of the freelist for reuse. */
503 		raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
504 		KFENCE_WARN_ON(!list_empty(&meta->list));
505 		list_add_tail(&meta->list, &kfence_freelist);
506 		raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
507 
508 		atomic_long_dec(&counters[KFENCE_COUNTER_ALLOCATED]);
509 		atomic_long_inc(&counters[KFENCE_COUNTER_FREES]);
510 	} else {
511 		/* See kfence_shutdown_cache(). */
512 		atomic_long_inc(&counters[KFENCE_COUNTER_ZOMBIES]);
513 	}
514 }
515 
516 static void rcu_guarded_free(struct rcu_head *h)
517 {
518 	struct kfence_metadata *meta = container_of(h, struct kfence_metadata, rcu_head);
519 
520 	kfence_guarded_free((void *)meta->addr, meta, false);
521 }
522 
523 /*
524  * Initialization of the KFENCE pool after its allocation.
525  * Returns 0 on success; otherwise returns the address up to
526  * which partial initialization succeeded.
527  */
528 static unsigned long kfence_init_pool(void)
529 {
530 	unsigned long addr = (unsigned long)__kfence_pool;
531 	struct page *pages;
532 	int i;
533 
534 	if (!arch_kfence_init_pool())
535 		return addr;
536 
537 	pages = virt_to_page(addr);
538 
539 	/*
540 	 * Set up object pages: they must have PG_slab set, to avoid freeing
541 	 * these as real pages.
542 	 *
543 	 * We also want to avoid inserting kfence_free() in the kfree()
544 	 * fast-path in SLUB, and therefore need to ensure kfree() correctly
545 	 * enters __slab_free() slow-path.
546 	 */
547 	for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
548 		struct slab *slab = page_slab(&pages[i]);
549 
550 		if (!i || (i % 2))
551 			continue;
552 
553 		/* Verify we do not have a compound head page. */
554 		if (WARN_ON(compound_head(&pages[i]) != &pages[i]))
555 			return addr;
556 
557 		__folio_set_slab(slab_folio(slab));
558 #ifdef CONFIG_MEMCG
559 		slab->memcg_data = (unsigned long)&kfence_metadata[i / 2 - 1].objcg |
560 				   MEMCG_DATA_OBJCGS;
561 #endif
562 	}
563 
564 	/*
565 	 * Protect the first 2 pages. The first page is mostly unnecessary, and
566 	 * merely serves as an extended guard page. However, adding one
567 	 * additional page in the beginning gives us an even number of pages,
568 	 * which simplifies the mapping of address to metadata index.
569 	 */
570 	for (i = 0; i < 2; i++) {
571 		if (unlikely(!kfence_protect(addr)))
572 			return addr;
573 
574 		addr += PAGE_SIZE;
575 	}
576 
577 	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
578 		struct kfence_metadata *meta = &kfence_metadata[i];
579 
580 		/* Initialize metadata. */
581 		INIT_LIST_HEAD(&meta->list);
582 		raw_spin_lock_init(&meta->lock);
583 		meta->state = KFENCE_OBJECT_UNUSED;
584 		meta->addr = addr; /* Initialize for validation in metadata_to_pageaddr(). */
585 		list_add_tail(&meta->list, &kfence_freelist);
586 
587 		/* Protect the right redzone. */
588 		if (unlikely(!kfence_protect(addr + PAGE_SIZE)))
589 			return addr;
590 
591 		addr += 2 * PAGE_SIZE;
592 	}
593 
594 	/*
595 	 * The pool is live and will never be deallocated from this point on.
596 	 * Remove the pool object from the kmemleak object tree, as it would
597 	 * otherwise overlap with allocations returned by kfence_alloc(), which
598 	 * are registered with kmemleak through the slab post-alloc hook.
599 	 */
600 	kmemleak_free(__kfence_pool);
601 
602 	return 0;
603 }
604 
605 static bool __init kfence_init_pool_early(void)
606 {
607 	unsigned long addr;
608 
609 	if (!__kfence_pool)
610 		return false;
611 
612 	addr = kfence_init_pool();
613 
614 	if (!addr)
615 		return true;
616 
617 	/*
618 	 * Only release unprotected pages, and do not try to go back and change
619 	 * page attributes due to risk of failing to do so as well. If changing
620 	 * page attributes for some pages fails, it is very likely that it also
621 	 * fails for the first page, and therefore expect addr==__kfence_pool in
622 	 * most failure cases.
623 	 */
624 	memblock_free_late(__pa(addr), KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool));
625 	__kfence_pool = NULL;
626 	return false;
627 }
628 
629 static bool kfence_init_pool_late(void)
630 {
631 	unsigned long addr, free_size;
632 
633 	addr = kfence_init_pool();
634 
635 	if (!addr)
636 		return true;
637 
638 	/* Same as above. */
639 	free_size = KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool);
640 #ifdef CONFIG_CONTIG_ALLOC
641 	free_contig_range(page_to_pfn(virt_to_page(addr)), free_size / PAGE_SIZE);
642 #else
643 	free_pages_exact((void *)addr, free_size);
644 #endif
645 	__kfence_pool = NULL;
646 	return false;
647 }
648 
649 /* === DebugFS Interface ==================================================== */
650 
651 static int stats_show(struct seq_file *seq, void *v)
652 {
653 	int i;
654 
655 	seq_printf(seq, "enabled: %i\n", READ_ONCE(kfence_enabled));
656 	for (i = 0; i < KFENCE_COUNTER_COUNT; i++)
657 		seq_printf(seq, "%s: %ld\n", counter_names[i], atomic_long_read(&counters[i]));
658 
659 	return 0;
660 }
661 DEFINE_SHOW_ATTRIBUTE(stats);
662 
663 /*
664  * debugfs seq_file operations for /sys/kernel/debug/kfence/objects.
665  * start_object() and next_object() return the object index + 1, because NULL is used
666  * to stop iteration.
667  */
668 static void *start_object(struct seq_file *seq, loff_t *pos)
669 {
670 	if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
671 		return (void *)((long)*pos + 1);
672 	return NULL;
673 }
674 
675 static void stop_object(struct seq_file *seq, void *v)
676 {
677 }
678 
679 static void *next_object(struct seq_file *seq, void *v, loff_t *pos)
680 {
681 	++*pos;
682 	if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
683 		return (void *)((long)*pos + 1);
684 	return NULL;
685 }
686 
687 static int show_object(struct seq_file *seq, void *v)
688 {
689 	struct kfence_metadata *meta = &kfence_metadata[(long)v - 1];
690 	unsigned long flags;
691 
692 	raw_spin_lock_irqsave(&meta->lock, flags);
693 	kfence_print_object(seq, meta);
694 	raw_spin_unlock_irqrestore(&meta->lock, flags);
695 	seq_puts(seq, "---------------------------------\n");
696 
697 	return 0;
698 }
699 
700 static const struct seq_operations object_seqops = {
701 	.start = start_object,
702 	.next = next_object,
703 	.stop = stop_object,
704 	.show = show_object,
705 };
706 
707 static int open_objects(struct inode *inode, struct file *file)
708 {
709 	return seq_open(file, &object_seqops);
710 }
711 
712 static const struct file_operations objects_fops = {
713 	.open = open_objects,
714 	.read = seq_read,
715 	.llseek = seq_lseek,
716 	.release = seq_release,
717 };
718 
719 static int __init kfence_debugfs_init(void)
720 {
721 	struct dentry *kfence_dir = debugfs_create_dir("kfence", NULL);
722 
723 	debugfs_create_file("stats", 0444, kfence_dir, NULL, &stats_fops);
724 	debugfs_create_file("objects", 0400, kfence_dir, NULL, &objects_fops);
725 	return 0;
726 }
727 
728 late_initcall(kfence_debugfs_init);
729 
730 /* === Allocation Gate Timer ================================================ */
731 
732 static struct delayed_work kfence_timer;
733 
734 #ifdef CONFIG_KFENCE_STATIC_KEYS
735 /* Wait queue to wake up allocation-gate timer task. */
736 static DECLARE_WAIT_QUEUE_HEAD(allocation_wait);
737 
738 static void wake_up_kfence_timer(struct irq_work *work)
739 {
740 	wake_up(&allocation_wait);
741 }
742 static DEFINE_IRQ_WORK(wake_up_kfence_timer_work, wake_up_kfence_timer);
743 #endif
744 
745 /*
746  * Set up delayed work, which will enable and disable the static key. We need to
747  * use a work queue (rather than a simple timer), since enabling and disabling a
748  * static key cannot be done from an interrupt.
749  *
750  * Note: Toggling a static branch currently causes IPIs, and here we'll end up
751  * with a total of 2 IPIs to all CPUs. If this ends up a problem in future (with
752  * more aggressive sampling intervals), we could get away with a variant that
753  * avoids IPIs, at the cost of not immediately capturing allocations if the
754  * instructions remain cached.
755  */
756 static void toggle_allocation_gate(struct work_struct *work)
757 {
758 	if (!READ_ONCE(kfence_enabled))
759 		return;
760 
761 	atomic_set(&kfence_allocation_gate, 0);
762 #ifdef CONFIG_KFENCE_STATIC_KEYS
763 	/* Enable static key, and await allocation to happen. */
764 	static_branch_enable(&kfence_allocation_key);
765 
766 	if (sysctl_hung_task_timeout_secs) {
767 		/*
768 		 * During low activity with no allocations we might wait a
769 		 * while; let's avoid the hung task warning.
770 		 */
771 		wait_event_idle_timeout(allocation_wait, atomic_read(&kfence_allocation_gate),
772 					sysctl_hung_task_timeout_secs * HZ / 2);
773 	} else {
774 		wait_event_idle(allocation_wait, atomic_read(&kfence_allocation_gate));
775 	}
776 
777 	/* Disable static key and reset timer. */
778 	static_branch_disable(&kfence_allocation_key);
779 #endif
780 	queue_delayed_work(system_unbound_wq, &kfence_timer,
781 			   msecs_to_jiffies(kfence_sample_interval));
782 }
783 
784 /* === Public interface ===================================================== */
785 
786 void __init kfence_alloc_pool(void)
787 {
788 	if (!kfence_sample_interval)
789 		return;
790 
791 	__kfence_pool = memblock_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
792 
793 	if (!__kfence_pool)
794 		pr_err("failed to allocate pool\n");
795 }
796 
797 static void kfence_init_enable(void)
798 {
799 	if (!IS_ENABLED(CONFIG_KFENCE_STATIC_KEYS))
800 		static_branch_enable(&kfence_allocation_key);
801 
802 	if (kfence_deferrable)
803 		INIT_DEFERRABLE_WORK(&kfence_timer, toggle_allocation_gate);
804 	else
805 		INIT_DELAYED_WORK(&kfence_timer, toggle_allocation_gate);
806 
807 	WRITE_ONCE(kfence_enabled, true);
808 	queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
809 
810 	pr_info("initialized - using %lu bytes for %d objects at 0x%p-0x%p\n", KFENCE_POOL_SIZE,
811 		CONFIG_KFENCE_NUM_OBJECTS, (void *)__kfence_pool,
812 		(void *)(__kfence_pool + KFENCE_POOL_SIZE));
813 }
814 
815 void __init kfence_init(void)
816 {
817 	stack_hash_seed = (u32)random_get_entropy();
818 
819 	/* Setting kfence_sample_interval to 0 on boot disables KFENCE. */
820 	if (!kfence_sample_interval)
821 		return;
822 
823 	if (!kfence_init_pool_early()) {
824 		pr_err("%s failed\n", __func__);
825 		return;
826 	}
827 
828 	kfence_init_enable();
829 }
830 
831 static int kfence_init_late(void)
832 {
833 	const unsigned long nr_pages = KFENCE_POOL_SIZE / PAGE_SIZE;
834 #ifdef CONFIG_CONTIG_ALLOC
835 	struct page *pages;
836 
837 	pages = alloc_contig_pages(nr_pages, GFP_KERNEL, first_online_node, NULL);
838 	if (!pages)
839 		return -ENOMEM;
840 	__kfence_pool = page_to_virt(pages);
841 #else
842 	if (nr_pages > MAX_ORDER_NR_PAGES) {
843 		pr_warn("KFENCE_NUM_OBJECTS too large for buddy allocator\n");
844 		return -EINVAL;
845 	}
846 	__kfence_pool = alloc_pages_exact(KFENCE_POOL_SIZE, GFP_KERNEL);
847 	if (!__kfence_pool)
848 		return -ENOMEM;
849 #endif
850 
851 	if (!kfence_init_pool_late()) {
852 		pr_err("%s failed\n", __func__);
853 		return -EBUSY;
854 	}
855 
856 	kfence_init_enable();
857 	return 0;
858 }
859 
860 static int kfence_enable_late(void)
861 {
862 	if (!__kfence_pool)
863 		return kfence_init_late();
864 
865 	WRITE_ONCE(kfence_enabled, true);
866 	queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
867 	return 0;
868 }
869 
870 void kfence_shutdown_cache(struct kmem_cache *s)
871 {
872 	unsigned long flags;
873 	struct kfence_metadata *meta;
874 	int i;
875 
876 	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
877 		bool in_use;
878 
879 		meta = &kfence_metadata[i];
880 
881 		/*
882 		 * If we observe some inconsistent cache and state pair where we
883 		 * should have returned false here, cache destruction is racing
884 		 * with either kmem_cache_alloc() or kmem_cache_free(). Taking
885 		 * the lock will not help, as different critical section
886 		 * serialization will have the same outcome.
887 		 */
888 		if (READ_ONCE(meta->cache) != s ||
889 		    READ_ONCE(meta->state) != KFENCE_OBJECT_ALLOCATED)
890 			continue;
891 
892 		raw_spin_lock_irqsave(&meta->lock, flags);
893 		in_use = meta->cache == s && meta->state == KFENCE_OBJECT_ALLOCATED;
894 		raw_spin_unlock_irqrestore(&meta->lock, flags);
895 
896 		if (in_use) {
897 			/*
898 			 * This cache still has allocations, and we should not
899 			 * release them back into the freelist so they can still
900 			 * safely be used and retain the kernel's default
901 			 * behaviour of keeping the allocations alive (leak the
902 			 * cache); however, they effectively become "zombie
903 			 * allocations" as the KFENCE objects are the only ones
904 			 * still in use and the owning cache is being destroyed.
905 			 *
906 			 * We mark them freed, so that any subsequent use shows
907 			 * more useful error messages that will include stack
908 			 * traces of the user of the object, the original
909 			 * allocation, and caller to shutdown_cache().
910 			 */
911 			kfence_guarded_free((void *)meta->addr, meta, /*zombie=*/true);
912 		}
913 	}
914 
915 	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
916 		meta = &kfence_metadata[i];
917 
918 		/* See above. */
919 		if (READ_ONCE(meta->cache) != s || READ_ONCE(meta->state) != KFENCE_OBJECT_FREED)
920 			continue;
921 
922 		raw_spin_lock_irqsave(&meta->lock, flags);
923 		if (meta->cache == s && meta->state == KFENCE_OBJECT_FREED)
924 			meta->cache = NULL;
925 		raw_spin_unlock_irqrestore(&meta->lock, flags);
926 	}
927 }
928 
929 void *__kfence_alloc(struct kmem_cache *s, size_t size, gfp_t flags)
930 {
931 	unsigned long stack_entries[KFENCE_STACK_DEPTH];
932 	size_t num_stack_entries;
933 	u32 alloc_stack_hash;
934 
935 	/*
936 	 * Perform size check before switching kfence_allocation_gate, so that
937 	 * we don't disable KFENCE without making an allocation.
938 	 */
939 	if (size > PAGE_SIZE) {
940 		atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
941 		return NULL;
942 	}
943 
944 	/*
945 	 * Skip allocations from non-default zones, including DMA. We cannot
946 	 * guarantee that pages in the KFENCE pool will have the requested
947 	 * properties (e.g. reside in DMAable memory).
948 	 */
949 	if ((flags & GFP_ZONEMASK) ||
950 	    (s->flags & (SLAB_CACHE_DMA | SLAB_CACHE_DMA32))) {
951 		atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
952 		return NULL;
953 	}
954 
955 	if (atomic_inc_return(&kfence_allocation_gate) > 1)
956 		return NULL;
957 #ifdef CONFIG_KFENCE_STATIC_KEYS
958 	/*
959 	 * waitqueue_active() is fully ordered after the update of
960 	 * kfence_allocation_gate per atomic_inc_return().
961 	 */
962 	if (waitqueue_active(&allocation_wait)) {
963 		/*
964 		 * Calling wake_up() here may deadlock when allocations happen
965 		 * from within timer code. Use an irq_work to defer it.
966 		 */
967 		irq_work_queue(&wake_up_kfence_timer_work);
968 	}
969 #endif
970 
971 	if (!READ_ONCE(kfence_enabled))
972 		return NULL;
973 
974 	num_stack_entries = stack_trace_save(stack_entries, KFENCE_STACK_DEPTH, 0);
975 
976 	/*
977 	 * Do expensive check for coverage of allocation in slow-path after
978 	 * allocation_gate has already become non-zero, even though it might
979 	 * mean not making any allocation within a given sample interval.
980 	 *
981 	 * This ensures reasonable allocation coverage when the pool is almost
982 	 * full, including avoiding long-lived allocations of the same source
983 	 * filling up the pool (e.g. pagecache allocations).
984 	 */
985 	alloc_stack_hash = get_alloc_stack_hash(stack_entries, num_stack_entries);
986 	if (should_skip_covered() && alloc_covered_contains(alloc_stack_hash)) {
987 		atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_COVERED]);
988 		return NULL;
989 	}
990 
991 	return kfence_guarded_alloc(s, size, flags, stack_entries, num_stack_entries,
992 				    alloc_stack_hash);
993 }
994 
995 size_t kfence_ksize(const void *addr)
996 {
997 	const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
998 
999 	/*
1000 	 * Read locklessly -- if there is a race with __kfence_alloc(), this is
1001 	 * either a use-after-free or invalid access.
1002 	 */
1003 	return meta ? meta->size : 0;
1004 }
1005 
1006 void *kfence_object_start(const void *addr)
1007 {
1008 	const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1009 
1010 	/*
1011 	 * Read locklessly -- if there is a race with __kfence_alloc(), this is
1012 	 * either a use-after-free or invalid access.
1013 	 */
1014 	return meta ? (void *)meta->addr : NULL;
1015 }
1016 
1017 void __kfence_free(void *addr)
1018 {
1019 	struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1020 
1021 #ifdef CONFIG_MEMCG
1022 	KFENCE_WARN_ON(meta->objcg);
1023 #endif
1024 	/*
1025 	 * If the objects of the cache are SLAB_TYPESAFE_BY_RCU, defer freeing
1026 	 * the object, as the object page may be recycled for other-typed
1027 	 * objects once it has been freed. meta->cache may be NULL if the cache
1028 	 * was destroyed.
1029 	 */
1030 	if (unlikely(meta->cache && (meta->cache->flags & SLAB_TYPESAFE_BY_RCU)))
1031 		call_rcu(&meta->rcu_head, rcu_guarded_free);
1032 	else
1033 		kfence_guarded_free(addr, meta, false);
1034 }
1035 
1036 bool kfence_handle_page_fault(unsigned long addr, bool is_write, struct pt_regs *regs)
1037 {
1038 	const int page_index = (addr - (unsigned long)__kfence_pool) / PAGE_SIZE;
1039 	struct kfence_metadata *to_report = NULL;
1040 	enum kfence_error_type error_type;
1041 	unsigned long flags;
1042 
1043 	if (!is_kfence_address((void *)addr))
1044 		return false;
1045 
1046 	if (!READ_ONCE(kfence_enabled)) /* If disabled at runtime ... */
1047 		return kfence_unprotect(addr); /* ... unprotect and proceed. */
1048 
1049 	atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
1050 
1051 	if (page_index % 2) {
1052 		/* This is a redzone, report a buffer overflow. */
1053 		struct kfence_metadata *meta;
1054 		int distance = 0;
1055 
1056 		meta = addr_to_metadata(addr - PAGE_SIZE);
1057 		if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
1058 			to_report = meta;
1059 			/* Data race ok; distance calculation approximate. */
1060 			distance = addr - data_race(meta->addr + meta->size);
1061 		}
1062 
1063 		meta = addr_to_metadata(addr + PAGE_SIZE);
1064 		if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
1065 			/* Data race ok; distance calculation approximate. */
1066 			if (!to_report || distance > data_race(meta->addr) - addr)
1067 				to_report = meta;
1068 		}
1069 
1070 		if (!to_report)
1071 			goto out;
1072 
1073 		raw_spin_lock_irqsave(&to_report->lock, flags);
1074 		to_report->unprotected_page = addr;
1075 		error_type = KFENCE_ERROR_OOB;
1076 
1077 		/*
1078 		 * If the object was freed before we took the look we can still
1079 		 * report this as an OOB -- the report will simply show the
1080 		 * stacktrace of the free as well.
1081 		 */
1082 	} else {
1083 		to_report = addr_to_metadata(addr);
1084 		if (!to_report)
1085 			goto out;
1086 
1087 		raw_spin_lock_irqsave(&to_report->lock, flags);
1088 		error_type = KFENCE_ERROR_UAF;
1089 		/*
1090 		 * We may race with __kfence_alloc(), and it is possible that a
1091 		 * freed object may be reallocated. We simply report this as a
1092 		 * use-after-free, with the stack trace showing the place where
1093 		 * the object was re-allocated.
1094 		 */
1095 	}
1096 
1097 out:
1098 	if (to_report) {
1099 		kfence_report_error(addr, is_write, regs, to_report, error_type);
1100 		raw_spin_unlock_irqrestore(&to_report->lock, flags);
1101 	} else {
1102 		/* This may be a UAF or OOB access, but we can't be sure. */
1103 		kfence_report_error(addr, is_write, regs, NULL, KFENCE_ERROR_INVALID);
1104 	}
1105 
1106 	return kfence_unprotect(addr); /* Unprotect and let access proceed. */
1107 }
1108