xref: /linux/mm/kfence/core.c (revision 68550cbc6129159b7a6434796b721e8b66ee12f6)
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 page *page;
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 struct page fields. */
428 	page = virt_to_page(meta->addr);
429 	page->slab_cache = cache;
430 	if (IS_ENABLED(CONFIG_SLUB))
431 		page->objects = 1;
432 	if (IS_ENABLED(CONFIG_SLAB))
433 		page->s_mem = addr;
434 
435 	/* Memory initialization. */
436 	for_each_canary(meta, set_canary_byte);
437 
438 	/*
439 	 * We check slab_want_init_on_alloc() ourselves, rather than letting
440 	 * SL*B do the initialization, as otherwise we might overwrite KFENCE's
441 	 * redzone.
442 	 */
443 	if (unlikely(slab_want_init_on_alloc(gfp, cache)))
444 		memzero_explicit(addr, size);
445 	if (cache->ctor)
446 		cache->ctor(addr);
447 
448 	if (CONFIG_KFENCE_STRESS_TEST_FAULTS && !prandom_u32_max(CONFIG_KFENCE_STRESS_TEST_FAULTS))
449 		kfence_protect(meta->addr); /* Random "faults" by protecting the object. */
450 
451 	atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCATED]);
452 	atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCS]);
453 
454 	return addr;
455 }
456 
457 static void kfence_guarded_free(void *addr, struct kfence_metadata *meta, bool zombie)
458 {
459 	struct kcsan_scoped_access assert_page_exclusive;
460 	unsigned long flags;
461 	bool init;
462 
463 	raw_spin_lock_irqsave(&meta->lock, flags);
464 
465 	if (meta->state != KFENCE_OBJECT_ALLOCATED || meta->addr != (unsigned long)addr) {
466 		/* Invalid or double-free, bail out. */
467 		atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
468 		kfence_report_error((unsigned long)addr, false, NULL, meta,
469 				    KFENCE_ERROR_INVALID_FREE);
470 		raw_spin_unlock_irqrestore(&meta->lock, flags);
471 		return;
472 	}
473 
474 	/* Detect racy use-after-free, or incorrect reallocation of this page by KFENCE. */
475 	kcsan_begin_scoped_access((void *)ALIGN_DOWN((unsigned long)addr, PAGE_SIZE), PAGE_SIZE,
476 				  KCSAN_ACCESS_SCOPED | KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT,
477 				  &assert_page_exclusive);
478 
479 	if (CONFIG_KFENCE_STRESS_TEST_FAULTS)
480 		kfence_unprotect((unsigned long)addr); /* To check canary bytes. */
481 
482 	/* Restore page protection if there was an OOB access. */
483 	if (meta->unprotected_page) {
484 		memzero_explicit((void *)ALIGN_DOWN(meta->unprotected_page, PAGE_SIZE), PAGE_SIZE);
485 		kfence_protect(meta->unprotected_page);
486 		meta->unprotected_page = 0;
487 	}
488 
489 	/* Mark the object as freed. */
490 	metadata_update_state(meta, KFENCE_OBJECT_FREED, NULL, 0);
491 	init = slab_want_init_on_free(meta->cache);
492 	raw_spin_unlock_irqrestore(&meta->lock, flags);
493 
494 	alloc_covered_add(meta->alloc_stack_hash, -1);
495 
496 	/* Check canary bytes for memory corruption. */
497 	for_each_canary(meta, check_canary_byte);
498 
499 	/*
500 	 * Clear memory if init-on-free is set. While we protect the page, the
501 	 * data is still there, and after a use-after-free is detected, we
502 	 * unprotect the page, so the data is still accessible.
503 	 */
504 	if (!zombie && unlikely(init))
505 		memzero_explicit(addr, meta->size);
506 
507 	/* Protect to detect use-after-frees. */
508 	kfence_protect((unsigned long)addr);
509 
510 	kcsan_end_scoped_access(&assert_page_exclusive);
511 	if (!zombie) {
512 		/* Add it to the tail of the freelist for reuse. */
513 		raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
514 		KFENCE_WARN_ON(!list_empty(&meta->list));
515 		list_add_tail(&meta->list, &kfence_freelist);
516 		raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
517 
518 		atomic_long_dec(&counters[KFENCE_COUNTER_ALLOCATED]);
519 		atomic_long_inc(&counters[KFENCE_COUNTER_FREES]);
520 	} else {
521 		/* See kfence_shutdown_cache(). */
522 		atomic_long_inc(&counters[KFENCE_COUNTER_ZOMBIES]);
523 	}
524 }
525 
526 static void rcu_guarded_free(struct rcu_head *h)
527 {
528 	struct kfence_metadata *meta = container_of(h, struct kfence_metadata, rcu_head);
529 
530 	kfence_guarded_free((void *)meta->addr, meta, false);
531 }
532 
533 static bool __init kfence_init_pool(void)
534 {
535 	unsigned long addr = (unsigned long)__kfence_pool;
536 	struct page *pages;
537 	int i;
538 
539 	if (!__kfence_pool)
540 		return false;
541 
542 	if (!arch_kfence_init_pool())
543 		goto err;
544 
545 	pages = virt_to_page(addr);
546 
547 	/*
548 	 * Set up object pages: they must have PG_slab set, to avoid freeing
549 	 * these as real pages.
550 	 *
551 	 * We also want to avoid inserting kfence_free() in the kfree()
552 	 * fast-path in SLUB, and therefore need to ensure kfree() correctly
553 	 * enters __slab_free() slow-path.
554 	 */
555 	for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
556 		if (!i || (i % 2))
557 			continue;
558 
559 		/* Verify we do not have a compound head page. */
560 		if (WARN_ON(compound_head(&pages[i]) != &pages[i]))
561 			goto err;
562 
563 		__SetPageSlab(&pages[i]);
564 	}
565 
566 	/*
567 	 * Protect the first 2 pages. The first page is mostly unnecessary, and
568 	 * merely serves as an extended guard page. However, adding one
569 	 * additional page in the beginning gives us an even number of pages,
570 	 * which simplifies the mapping of address to metadata index.
571 	 */
572 	for (i = 0; i < 2; i++) {
573 		if (unlikely(!kfence_protect(addr)))
574 			goto err;
575 
576 		addr += PAGE_SIZE;
577 	}
578 
579 	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
580 		struct kfence_metadata *meta = &kfence_metadata[i];
581 
582 		/* Initialize metadata. */
583 		INIT_LIST_HEAD(&meta->list);
584 		raw_spin_lock_init(&meta->lock);
585 		meta->state = KFENCE_OBJECT_UNUSED;
586 		meta->addr = addr; /* Initialize for validation in metadata_to_pageaddr(). */
587 		list_add_tail(&meta->list, &kfence_freelist);
588 
589 		/* Protect the right redzone. */
590 		if (unlikely(!kfence_protect(addr + PAGE_SIZE)))
591 			goto err;
592 
593 		addr += 2 * PAGE_SIZE;
594 	}
595 
596 	/*
597 	 * The pool is live and will never be deallocated from this point on.
598 	 * Remove the pool object from the kmemleak object tree, as it would
599 	 * otherwise overlap with allocations returned by kfence_alloc(), which
600 	 * are registered with kmemleak through the slab post-alloc hook.
601 	 */
602 	kmemleak_free(__kfence_pool);
603 
604 	return true;
605 
606 err:
607 	/*
608 	 * Only release unprotected pages, and do not try to go back and change
609 	 * page attributes due to risk of failing to do so as well. If changing
610 	 * page attributes for some pages fails, it is very likely that it also
611 	 * fails for the first page, and therefore expect addr==__kfence_pool in
612 	 * most failure cases.
613 	 */
614 	memblock_free_late(__pa(addr), KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool));
615 	__kfence_pool = NULL;
616 	return false;
617 }
618 
619 /* === DebugFS Interface ==================================================== */
620 
621 static int stats_show(struct seq_file *seq, void *v)
622 {
623 	int i;
624 
625 	seq_printf(seq, "enabled: %i\n", READ_ONCE(kfence_enabled));
626 	for (i = 0; i < KFENCE_COUNTER_COUNT; i++)
627 		seq_printf(seq, "%s: %ld\n", counter_names[i], atomic_long_read(&counters[i]));
628 
629 	return 0;
630 }
631 DEFINE_SHOW_ATTRIBUTE(stats);
632 
633 /*
634  * debugfs seq_file operations for /sys/kernel/debug/kfence/objects.
635  * start_object() and next_object() return the object index + 1, because NULL is used
636  * to stop iteration.
637  */
638 static void *start_object(struct seq_file *seq, loff_t *pos)
639 {
640 	if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
641 		return (void *)((long)*pos + 1);
642 	return NULL;
643 }
644 
645 static void stop_object(struct seq_file *seq, void *v)
646 {
647 }
648 
649 static void *next_object(struct seq_file *seq, void *v, loff_t *pos)
650 {
651 	++*pos;
652 	if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
653 		return (void *)((long)*pos + 1);
654 	return NULL;
655 }
656 
657 static int show_object(struct seq_file *seq, void *v)
658 {
659 	struct kfence_metadata *meta = &kfence_metadata[(long)v - 1];
660 	unsigned long flags;
661 
662 	raw_spin_lock_irqsave(&meta->lock, flags);
663 	kfence_print_object(seq, meta);
664 	raw_spin_unlock_irqrestore(&meta->lock, flags);
665 	seq_puts(seq, "---------------------------------\n");
666 
667 	return 0;
668 }
669 
670 static const struct seq_operations object_seqops = {
671 	.start = start_object,
672 	.next = next_object,
673 	.stop = stop_object,
674 	.show = show_object,
675 };
676 
677 static int open_objects(struct inode *inode, struct file *file)
678 {
679 	return seq_open(file, &object_seqops);
680 }
681 
682 static const struct file_operations objects_fops = {
683 	.open = open_objects,
684 	.read = seq_read,
685 	.llseek = seq_lseek,
686 };
687 
688 static int __init kfence_debugfs_init(void)
689 {
690 	struct dentry *kfence_dir = debugfs_create_dir("kfence", NULL);
691 
692 	debugfs_create_file("stats", 0444, kfence_dir, NULL, &stats_fops);
693 	debugfs_create_file("objects", 0400, kfence_dir, NULL, &objects_fops);
694 	return 0;
695 }
696 
697 late_initcall(kfence_debugfs_init);
698 
699 /* === Allocation Gate Timer ================================================ */
700 
701 #ifdef CONFIG_KFENCE_STATIC_KEYS
702 /* Wait queue to wake up allocation-gate timer task. */
703 static DECLARE_WAIT_QUEUE_HEAD(allocation_wait);
704 
705 static void wake_up_kfence_timer(struct irq_work *work)
706 {
707 	wake_up(&allocation_wait);
708 }
709 static DEFINE_IRQ_WORK(wake_up_kfence_timer_work, wake_up_kfence_timer);
710 #endif
711 
712 /*
713  * Set up delayed work, which will enable and disable the static key. We need to
714  * use a work queue (rather than a simple timer), since enabling and disabling a
715  * static key cannot be done from an interrupt.
716  *
717  * Note: Toggling a static branch currently causes IPIs, and here we'll end up
718  * with a total of 2 IPIs to all CPUs. If this ends up a problem in future (with
719  * more aggressive sampling intervals), we could get away with a variant that
720  * avoids IPIs, at the cost of not immediately capturing allocations if the
721  * instructions remain cached.
722  */
723 static struct delayed_work kfence_timer;
724 static void toggle_allocation_gate(struct work_struct *work)
725 {
726 	if (!READ_ONCE(kfence_enabled))
727 		return;
728 
729 	atomic_set(&kfence_allocation_gate, 0);
730 #ifdef CONFIG_KFENCE_STATIC_KEYS
731 	/* Enable static key, and await allocation to happen. */
732 	static_branch_enable(&kfence_allocation_key);
733 
734 	if (sysctl_hung_task_timeout_secs) {
735 		/*
736 		 * During low activity with no allocations we might wait a
737 		 * while; let's avoid the hung task warning.
738 		 */
739 		wait_event_idle_timeout(allocation_wait, atomic_read(&kfence_allocation_gate),
740 					sysctl_hung_task_timeout_secs * HZ / 2);
741 	} else {
742 		wait_event_idle(allocation_wait, atomic_read(&kfence_allocation_gate));
743 	}
744 
745 	/* Disable static key and reset timer. */
746 	static_branch_disable(&kfence_allocation_key);
747 #endif
748 	queue_delayed_work(system_unbound_wq, &kfence_timer,
749 			   msecs_to_jiffies(kfence_sample_interval));
750 }
751 static DECLARE_DELAYED_WORK(kfence_timer, toggle_allocation_gate);
752 
753 /* === Public interface ===================================================== */
754 
755 void __init kfence_alloc_pool(void)
756 {
757 	if (!kfence_sample_interval)
758 		return;
759 
760 	__kfence_pool = memblock_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
761 
762 	if (!__kfence_pool)
763 		pr_err("failed to allocate pool\n");
764 }
765 
766 void __init kfence_init(void)
767 {
768 	/* Setting kfence_sample_interval to 0 on boot disables KFENCE. */
769 	if (!kfence_sample_interval)
770 		return;
771 
772 	stack_hash_seed = (u32)random_get_entropy();
773 	if (!kfence_init_pool()) {
774 		pr_err("%s failed\n", __func__);
775 		return;
776 	}
777 
778 	if (!IS_ENABLED(CONFIG_KFENCE_STATIC_KEYS))
779 		static_branch_enable(&kfence_allocation_key);
780 	WRITE_ONCE(kfence_enabled, true);
781 	queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
782 	pr_info("initialized - using %lu bytes for %d objects at 0x%p-0x%p\n", KFENCE_POOL_SIZE,
783 		CONFIG_KFENCE_NUM_OBJECTS, (void *)__kfence_pool,
784 		(void *)(__kfence_pool + KFENCE_POOL_SIZE));
785 }
786 
787 void kfence_shutdown_cache(struct kmem_cache *s)
788 {
789 	unsigned long flags;
790 	struct kfence_metadata *meta;
791 	int i;
792 
793 	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
794 		bool in_use;
795 
796 		meta = &kfence_metadata[i];
797 
798 		/*
799 		 * If we observe some inconsistent cache and state pair where we
800 		 * should have returned false here, cache destruction is racing
801 		 * with either kmem_cache_alloc() or kmem_cache_free(). Taking
802 		 * the lock will not help, as different critical section
803 		 * serialization will have the same outcome.
804 		 */
805 		if (READ_ONCE(meta->cache) != s ||
806 		    READ_ONCE(meta->state) != KFENCE_OBJECT_ALLOCATED)
807 			continue;
808 
809 		raw_spin_lock_irqsave(&meta->lock, flags);
810 		in_use = meta->cache == s && meta->state == KFENCE_OBJECT_ALLOCATED;
811 		raw_spin_unlock_irqrestore(&meta->lock, flags);
812 
813 		if (in_use) {
814 			/*
815 			 * This cache still has allocations, and we should not
816 			 * release them back into the freelist so they can still
817 			 * safely be used and retain the kernel's default
818 			 * behaviour of keeping the allocations alive (leak the
819 			 * cache); however, they effectively become "zombie
820 			 * allocations" as the KFENCE objects are the only ones
821 			 * still in use and the owning cache is being destroyed.
822 			 *
823 			 * We mark them freed, so that any subsequent use shows
824 			 * more useful error messages that will include stack
825 			 * traces of the user of the object, the original
826 			 * allocation, and caller to shutdown_cache().
827 			 */
828 			kfence_guarded_free((void *)meta->addr, meta, /*zombie=*/true);
829 		}
830 	}
831 
832 	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
833 		meta = &kfence_metadata[i];
834 
835 		/* See above. */
836 		if (READ_ONCE(meta->cache) != s || READ_ONCE(meta->state) != KFENCE_OBJECT_FREED)
837 			continue;
838 
839 		raw_spin_lock_irqsave(&meta->lock, flags);
840 		if (meta->cache == s && meta->state == KFENCE_OBJECT_FREED)
841 			meta->cache = NULL;
842 		raw_spin_unlock_irqrestore(&meta->lock, flags);
843 	}
844 }
845 
846 void *__kfence_alloc(struct kmem_cache *s, size_t size, gfp_t flags)
847 {
848 	unsigned long stack_entries[KFENCE_STACK_DEPTH];
849 	size_t num_stack_entries;
850 	u32 alloc_stack_hash;
851 
852 	/*
853 	 * Perform size check before switching kfence_allocation_gate, so that
854 	 * we don't disable KFENCE without making an allocation.
855 	 */
856 	if (size > PAGE_SIZE) {
857 		atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
858 		return NULL;
859 	}
860 
861 	/*
862 	 * Skip allocations from non-default zones, including DMA. We cannot
863 	 * guarantee that pages in the KFENCE pool will have the requested
864 	 * properties (e.g. reside in DMAable memory).
865 	 */
866 	if ((flags & GFP_ZONEMASK) ||
867 	    (s->flags & (SLAB_CACHE_DMA | SLAB_CACHE_DMA32))) {
868 		atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
869 		return NULL;
870 	}
871 
872 	if (atomic_inc_return(&kfence_allocation_gate) > 1)
873 		return NULL;
874 #ifdef CONFIG_KFENCE_STATIC_KEYS
875 	/*
876 	 * waitqueue_active() is fully ordered after the update of
877 	 * kfence_allocation_gate per atomic_inc_return().
878 	 */
879 	if (waitqueue_active(&allocation_wait)) {
880 		/*
881 		 * Calling wake_up() here may deadlock when allocations happen
882 		 * from within timer code. Use an irq_work to defer it.
883 		 */
884 		irq_work_queue(&wake_up_kfence_timer_work);
885 	}
886 #endif
887 
888 	if (!READ_ONCE(kfence_enabled))
889 		return NULL;
890 
891 	num_stack_entries = stack_trace_save(stack_entries, KFENCE_STACK_DEPTH, 0);
892 
893 	/*
894 	 * Do expensive check for coverage of allocation in slow-path after
895 	 * allocation_gate has already become non-zero, even though it might
896 	 * mean not making any allocation within a given sample interval.
897 	 *
898 	 * This ensures reasonable allocation coverage when the pool is almost
899 	 * full, including avoiding long-lived allocations of the same source
900 	 * filling up the pool (e.g. pagecache allocations).
901 	 */
902 	alloc_stack_hash = get_alloc_stack_hash(stack_entries, num_stack_entries);
903 	if (should_skip_covered() && alloc_covered_contains(alloc_stack_hash)) {
904 		atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_COVERED]);
905 		return NULL;
906 	}
907 
908 	return kfence_guarded_alloc(s, size, flags, stack_entries, num_stack_entries,
909 				    alloc_stack_hash);
910 }
911 
912 size_t kfence_ksize(const void *addr)
913 {
914 	const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
915 
916 	/*
917 	 * Read locklessly -- if there is a race with __kfence_alloc(), this is
918 	 * either a use-after-free or invalid access.
919 	 */
920 	return meta ? meta->size : 0;
921 }
922 
923 void *kfence_object_start(const void *addr)
924 {
925 	const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
926 
927 	/*
928 	 * Read locklessly -- if there is a race with __kfence_alloc(), this is
929 	 * either a use-after-free or invalid access.
930 	 */
931 	return meta ? (void *)meta->addr : NULL;
932 }
933 
934 void __kfence_free(void *addr)
935 {
936 	struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
937 
938 	/*
939 	 * If the objects of the cache are SLAB_TYPESAFE_BY_RCU, defer freeing
940 	 * the object, as the object page may be recycled for other-typed
941 	 * objects once it has been freed. meta->cache may be NULL if the cache
942 	 * was destroyed.
943 	 */
944 	if (unlikely(meta->cache && (meta->cache->flags & SLAB_TYPESAFE_BY_RCU)))
945 		call_rcu(&meta->rcu_head, rcu_guarded_free);
946 	else
947 		kfence_guarded_free(addr, meta, false);
948 }
949 
950 bool kfence_handle_page_fault(unsigned long addr, bool is_write, struct pt_regs *regs)
951 {
952 	const int page_index = (addr - (unsigned long)__kfence_pool) / PAGE_SIZE;
953 	struct kfence_metadata *to_report = NULL;
954 	enum kfence_error_type error_type;
955 	unsigned long flags;
956 
957 	if (!is_kfence_address((void *)addr))
958 		return false;
959 
960 	if (!READ_ONCE(kfence_enabled)) /* If disabled at runtime ... */
961 		return kfence_unprotect(addr); /* ... unprotect and proceed. */
962 
963 	atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
964 
965 	if (page_index % 2) {
966 		/* This is a redzone, report a buffer overflow. */
967 		struct kfence_metadata *meta;
968 		int distance = 0;
969 
970 		meta = addr_to_metadata(addr - PAGE_SIZE);
971 		if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
972 			to_report = meta;
973 			/* Data race ok; distance calculation approximate. */
974 			distance = addr - data_race(meta->addr + meta->size);
975 		}
976 
977 		meta = addr_to_metadata(addr + PAGE_SIZE);
978 		if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
979 			/* Data race ok; distance calculation approximate. */
980 			if (!to_report || distance > data_race(meta->addr) - addr)
981 				to_report = meta;
982 		}
983 
984 		if (!to_report)
985 			goto out;
986 
987 		raw_spin_lock_irqsave(&to_report->lock, flags);
988 		to_report->unprotected_page = addr;
989 		error_type = KFENCE_ERROR_OOB;
990 
991 		/*
992 		 * If the object was freed before we took the look we can still
993 		 * report this as an OOB -- the report will simply show the
994 		 * stacktrace of the free as well.
995 		 */
996 	} else {
997 		to_report = addr_to_metadata(addr);
998 		if (!to_report)
999 			goto out;
1000 
1001 		raw_spin_lock_irqsave(&to_report->lock, flags);
1002 		error_type = KFENCE_ERROR_UAF;
1003 		/*
1004 		 * We may race with __kfence_alloc(), and it is possible that a
1005 		 * freed object may be reallocated. We simply report this as a
1006 		 * use-after-free, with the stack trace showing the place where
1007 		 * the object was re-allocated.
1008 		 */
1009 	}
1010 
1011 out:
1012 	if (to_report) {
1013 		kfence_report_error(addr, is_write, regs, to_report, error_type);
1014 		raw_spin_unlock_irqrestore(&to_report->lock, flags);
1015 	} else {
1016 		/* This may be a UAF or OOB access, but we can't be sure. */
1017 		kfence_report_error(addr, is_write, regs, NULL, KFENCE_ERROR_INVALID);
1018 	}
1019 
1020 	return kfence_unprotect(addr); /* Unprotect and let access proceed. */
1021 }
1022