xref: /linux/mm/kfence/core.c (revision 8804d970fab45726b3c7cd7f240b31122aa94219)
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/nodemask.h>
25 #include <linux/notifier.h>
26 #include <linux/panic_notifier.h>
27 #include <linux/random.h>
28 #include <linux/rcupdate.h>
29 #include <linux/sched/clock.h>
30 #include <linux/seq_file.h>
31 #include <linux/slab.h>
32 #include <linux/spinlock.h>
33 #include <linux/string.h>
34 
35 #include <asm/kfence.h>
36 
37 #include "kfence.h"
38 
39 /* Disables KFENCE on the first warning assuming an irrecoverable error. */
40 #define KFENCE_WARN_ON(cond)                                                   \
41 	({                                                                     \
42 		const bool __cond = WARN_ON(cond);                             \
43 		if (unlikely(__cond)) {                                        \
44 			WRITE_ONCE(kfence_enabled, false);                     \
45 			disabled_by_warn = true;                               \
46 		}                                                              \
47 		__cond;                                                        \
48 	})
49 
50 /* === Data ================================================================= */
51 
52 static bool kfence_enabled __read_mostly;
53 static bool disabled_by_warn __read_mostly;
54 
55 unsigned long kfence_sample_interval __read_mostly = CONFIG_KFENCE_SAMPLE_INTERVAL;
56 EXPORT_SYMBOL_GPL(kfence_sample_interval); /* Export for test modules. */
57 
58 #ifdef MODULE_PARAM_PREFIX
59 #undef MODULE_PARAM_PREFIX
60 #endif
61 #define MODULE_PARAM_PREFIX "kfence."
62 
63 static int kfence_enable_late(void);
param_set_sample_interval(const char * val,const struct kernel_param * kp)64 static int param_set_sample_interval(const char *val, const struct kernel_param *kp)
65 {
66 	unsigned long num;
67 	int ret = kstrtoul(val, 0, &num);
68 
69 	if (ret < 0)
70 		return ret;
71 
72 	/* Using 0 to indicate KFENCE is disabled. */
73 	if (!num && READ_ONCE(kfence_enabled)) {
74 		pr_info("disabled\n");
75 		WRITE_ONCE(kfence_enabled, false);
76 	}
77 
78 	*((unsigned long *)kp->arg) = num;
79 
80 	if (num && !READ_ONCE(kfence_enabled) && system_state != SYSTEM_BOOTING)
81 		return disabled_by_warn ? -EINVAL : kfence_enable_late();
82 	return 0;
83 }
84 
param_get_sample_interval(char * buffer,const struct kernel_param * kp)85 static int param_get_sample_interval(char *buffer, const struct kernel_param *kp)
86 {
87 	if (!READ_ONCE(kfence_enabled))
88 		return sprintf(buffer, "0\n");
89 
90 	return param_get_ulong(buffer, kp);
91 }
92 
93 static const struct kernel_param_ops sample_interval_param_ops = {
94 	.set = param_set_sample_interval,
95 	.get = param_get_sample_interval,
96 };
97 module_param_cb(sample_interval, &sample_interval_param_ops, &kfence_sample_interval, 0600);
98 
99 /* Pool usage% threshold when currently covered allocations are skipped. */
100 static unsigned long kfence_skip_covered_thresh __read_mostly = 75;
101 module_param_named(skip_covered_thresh, kfence_skip_covered_thresh, ulong, 0644);
102 
103 /* Allocation burst count: number of excess KFENCE allocations per sample. */
104 static unsigned int kfence_burst __read_mostly;
105 module_param_named(burst, kfence_burst, uint, 0644);
106 
107 /* If true, use a deferrable timer. */
108 static bool kfence_deferrable __read_mostly = IS_ENABLED(CONFIG_KFENCE_DEFERRABLE);
109 module_param_named(deferrable, kfence_deferrable, bool, 0444);
110 
111 /* If true, check all canary bytes on panic. */
112 static bool kfence_check_on_panic __read_mostly;
113 module_param_named(check_on_panic, kfence_check_on_panic, bool, 0444);
114 
115 /* The pool of pages used for guard pages and objects. */
116 char *__kfence_pool __read_mostly;
117 EXPORT_SYMBOL(__kfence_pool); /* Export for test modules. */
118 
119 /*
120  * Per-object metadata, with one-to-one mapping of object metadata to
121  * backing pages (in __kfence_pool).
122  */
123 static_assert(CONFIG_KFENCE_NUM_OBJECTS > 0);
124 struct kfence_metadata *kfence_metadata __read_mostly;
125 
126 /*
127  * If kfence_metadata is not NULL, it may be accessed by kfence_shutdown_cache().
128  * So introduce kfence_metadata_init to initialize metadata, and then make
129  * kfence_metadata visible after initialization is successful. This prevents
130  * potential UAF or access to uninitialized metadata.
131  */
132 static struct kfence_metadata *kfence_metadata_init __read_mostly;
133 
134 /* Freelist with available objects. */
135 static struct list_head kfence_freelist = LIST_HEAD_INIT(kfence_freelist);
136 static DEFINE_RAW_SPINLOCK(kfence_freelist_lock); /* Lock protecting freelist. */
137 
138 /*
139  * The static key to set up a KFENCE allocation; or if static keys are not used
140  * to gate allocations, to avoid a load and compare if KFENCE is disabled.
141  */
142 DEFINE_STATIC_KEY_FALSE(kfence_allocation_key);
143 
144 /* Gates the allocation, ensuring only one succeeds in a given period. */
145 atomic_t kfence_allocation_gate = ATOMIC_INIT(1);
146 
147 /*
148  * A Counting Bloom filter of allocation coverage: limits currently covered
149  * allocations of the same source filling up the pool.
150  *
151  * Assuming a range of 15%-85% unique allocations in the pool at any point in
152  * time, the below parameters provide a probablity of 0.02-0.33 for false
153  * positive hits respectively:
154  *
155  *	P(alloc_traces) = (1 - e^(-HNUM * (alloc_traces / SIZE)) ^ HNUM
156  */
157 #define ALLOC_COVERED_HNUM	2
158 #define ALLOC_COVERED_ORDER	(const_ilog2(CONFIG_KFENCE_NUM_OBJECTS) + 2)
159 #define ALLOC_COVERED_SIZE	(1 << ALLOC_COVERED_ORDER)
160 #define ALLOC_COVERED_HNEXT(h)	hash_32(h, ALLOC_COVERED_ORDER)
161 #define ALLOC_COVERED_MASK	(ALLOC_COVERED_SIZE - 1)
162 static atomic_t alloc_covered[ALLOC_COVERED_SIZE];
163 
164 /* Stack depth used to determine uniqueness of an allocation. */
165 #define UNIQUE_ALLOC_STACK_DEPTH ((size_t)8)
166 
167 /*
168  * Randomness for stack hashes, making the same collisions across reboots and
169  * different machines less likely.
170  */
171 static u32 stack_hash_seed __ro_after_init;
172 
173 /* Statistics counters for debugfs. */
174 enum kfence_counter_id {
175 	KFENCE_COUNTER_ALLOCATED,
176 	KFENCE_COUNTER_ALLOCS,
177 	KFENCE_COUNTER_FREES,
178 	KFENCE_COUNTER_ZOMBIES,
179 	KFENCE_COUNTER_BUGS,
180 	KFENCE_COUNTER_SKIP_INCOMPAT,
181 	KFENCE_COUNTER_SKIP_CAPACITY,
182 	KFENCE_COUNTER_SKIP_COVERED,
183 	KFENCE_COUNTER_COUNT,
184 };
185 static atomic_long_t counters[KFENCE_COUNTER_COUNT];
186 static const char *const counter_names[] = {
187 	[KFENCE_COUNTER_ALLOCATED]	= "currently allocated",
188 	[KFENCE_COUNTER_ALLOCS]		= "total allocations",
189 	[KFENCE_COUNTER_FREES]		= "total frees",
190 	[KFENCE_COUNTER_ZOMBIES]	= "zombie allocations",
191 	[KFENCE_COUNTER_BUGS]		= "total bugs",
192 	[KFENCE_COUNTER_SKIP_INCOMPAT]	= "skipped allocations (incompatible)",
193 	[KFENCE_COUNTER_SKIP_CAPACITY]	= "skipped allocations (capacity)",
194 	[KFENCE_COUNTER_SKIP_COVERED]	= "skipped allocations (covered)",
195 };
196 static_assert(ARRAY_SIZE(counter_names) == KFENCE_COUNTER_COUNT);
197 
198 /* === Internals ============================================================ */
199 
should_skip_covered(void)200 static inline bool should_skip_covered(void)
201 {
202 	unsigned long thresh = (CONFIG_KFENCE_NUM_OBJECTS * kfence_skip_covered_thresh) / 100;
203 
204 	return atomic_long_read(&counters[KFENCE_COUNTER_ALLOCATED]) > thresh;
205 }
206 
get_alloc_stack_hash(unsigned long * stack_entries,size_t num_entries)207 static u32 get_alloc_stack_hash(unsigned long *stack_entries, size_t num_entries)
208 {
209 	num_entries = min(num_entries, UNIQUE_ALLOC_STACK_DEPTH);
210 	num_entries = filter_irq_stacks(stack_entries, num_entries);
211 	return jhash(stack_entries, num_entries * sizeof(stack_entries[0]), stack_hash_seed);
212 }
213 
214 /*
215  * Adds (or subtracts) count @val for allocation stack trace hash
216  * @alloc_stack_hash from Counting Bloom filter.
217  */
alloc_covered_add(u32 alloc_stack_hash,int val)218 static void alloc_covered_add(u32 alloc_stack_hash, int val)
219 {
220 	int i;
221 
222 	for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
223 		atomic_add(val, &alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]);
224 		alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
225 	}
226 }
227 
228 /*
229  * Returns true if the allocation stack trace hash @alloc_stack_hash is
230  * currently contained (non-zero count) in Counting Bloom filter.
231  */
alloc_covered_contains(u32 alloc_stack_hash)232 static bool alloc_covered_contains(u32 alloc_stack_hash)
233 {
234 	int i;
235 
236 	for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
237 		if (!atomic_read(&alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]))
238 			return false;
239 		alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
240 	}
241 
242 	return true;
243 }
244 
kfence_protect(unsigned long addr)245 static bool kfence_protect(unsigned long addr)
246 {
247 	return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), true));
248 }
249 
kfence_unprotect(unsigned long addr)250 static bool kfence_unprotect(unsigned long addr)
251 {
252 	return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), false));
253 }
254 
metadata_to_pageaddr(const struct kfence_metadata * meta)255 static inline unsigned long metadata_to_pageaddr(const struct kfence_metadata *meta)
256 {
257 	unsigned long offset = (meta - kfence_metadata + 1) * PAGE_SIZE * 2;
258 	unsigned long pageaddr = (unsigned long)&__kfence_pool[offset];
259 
260 	/* The checks do not affect performance; only called from slow-paths. */
261 
262 	/* Only call with a pointer into kfence_metadata. */
263 	if (KFENCE_WARN_ON(meta < kfence_metadata ||
264 			   meta >= kfence_metadata + CONFIG_KFENCE_NUM_OBJECTS))
265 		return 0;
266 
267 	/*
268 	 * This metadata object only ever maps to 1 page; verify that the stored
269 	 * address is in the expected range.
270 	 */
271 	if (KFENCE_WARN_ON(ALIGN_DOWN(meta->addr, PAGE_SIZE) != pageaddr))
272 		return 0;
273 
274 	return pageaddr;
275 }
276 
kfence_obj_allocated(const struct kfence_metadata * meta)277 static inline bool kfence_obj_allocated(const struct kfence_metadata *meta)
278 {
279 	enum kfence_object_state state = READ_ONCE(meta->state);
280 
281 	return state == KFENCE_OBJECT_ALLOCATED || state == KFENCE_OBJECT_RCU_FREEING;
282 }
283 
284 /*
285  * Update the object's metadata state, including updating the alloc/free stacks
286  * depending on the state transition.
287  */
288 static noinline void
metadata_update_state(struct kfence_metadata * meta,enum kfence_object_state next,unsigned long * stack_entries,size_t num_stack_entries)289 metadata_update_state(struct kfence_metadata *meta, enum kfence_object_state next,
290 		      unsigned long *stack_entries, size_t num_stack_entries)
291 {
292 	struct kfence_track *track =
293 		next == KFENCE_OBJECT_ALLOCATED ? &meta->alloc_track : &meta->free_track;
294 
295 	lockdep_assert_held(&meta->lock);
296 
297 	/* Stack has been saved when calling rcu, skip. */
298 	if (READ_ONCE(meta->state) == KFENCE_OBJECT_RCU_FREEING)
299 		goto out;
300 
301 	if (stack_entries) {
302 		memcpy(track->stack_entries, stack_entries,
303 		       num_stack_entries * sizeof(stack_entries[0]));
304 	} else {
305 		/*
306 		 * Skip over 1 (this) functions; noinline ensures we do not
307 		 * accidentally skip over the caller by never inlining.
308 		 */
309 		num_stack_entries = stack_trace_save(track->stack_entries, KFENCE_STACK_DEPTH, 1);
310 	}
311 	track->num_stack_entries = num_stack_entries;
312 	track->pid = task_pid_nr(current);
313 	track->cpu = raw_smp_processor_id();
314 	track->ts_nsec = local_clock(); /* Same source as printk timestamps. */
315 
316 out:
317 	/*
318 	 * Pairs with READ_ONCE() in
319 	 *	kfence_shutdown_cache(),
320 	 *	kfence_handle_page_fault().
321 	 */
322 	WRITE_ONCE(meta->state, next);
323 }
324 
325 #ifdef CONFIG_KMSAN
326 #define check_canary_attributes noinline __no_kmsan_checks
327 #else
328 #define check_canary_attributes inline
329 #endif
330 
331 /* Check canary byte at @addr. */
check_canary_byte(u8 * addr)332 static check_canary_attributes bool check_canary_byte(u8 *addr)
333 {
334 	struct kfence_metadata *meta;
335 	unsigned long flags;
336 
337 	if (likely(*addr == KFENCE_CANARY_PATTERN_U8(addr)))
338 		return true;
339 
340 	atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
341 
342 	meta = addr_to_metadata((unsigned long)addr);
343 	raw_spin_lock_irqsave(&meta->lock, flags);
344 	kfence_report_error((unsigned long)addr, false, NULL, meta, KFENCE_ERROR_CORRUPTION);
345 	raw_spin_unlock_irqrestore(&meta->lock, flags);
346 
347 	return false;
348 }
349 
set_canary(const struct kfence_metadata * meta)350 static inline void set_canary(const struct kfence_metadata *meta)
351 {
352 	const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
353 	unsigned long addr = pageaddr;
354 
355 	/*
356 	 * The canary may be written to part of the object memory, but it does
357 	 * not affect it. The user should initialize the object before using it.
358 	 */
359 	for (; addr < meta->addr; addr += sizeof(u64))
360 		*((u64 *)addr) = KFENCE_CANARY_PATTERN_U64;
361 
362 	addr = ALIGN_DOWN(meta->addr + meta->size, sizeof(u64));
363 	for (; addr - pageaddr < PAGE_SIZE; addr += sizeof(u64))
364 		*((u64 *)addr) = KFENCE_CANARY_PATTERN_U64;
365 }
366 
367 static check_canary_attributes void
check_canary(const struct kfence_metadata * meta)368 check_canary(const struct kfence_metadata *meta)
369 {
370 	const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
371 	unsigned long addr = pageaddr;
372 
373 	/*
374 	 * We'll iterate over each canary byte per-side until a corrupted byte
375 	 * is found. However, we'll still iterate over the canary bytes to the
376 	 * right of the object even if there was an error in the canary bytes to
377 	 * the left of the object. Specifically, if check_canary_byte()
378 	 * generates an error, showing both sides might give more clues as to
379 	 * what the error is about when displaying which bytes were corrupted.
380 	 */
381 
382 	/* Apply to left of object. */
383 	for (; meta->addr - addr >= sizeof(u64); addr += sizeof(u64)) {
384 		if (unlikely(*((u64 *)addr) != KFENCE_CANARY_PATTERN_U64))
385 			break;
386 	}
387 
388 	/*
389 	 * If the canary is corrupted in a certain 64 bytes, or the canary
390 	 * memory cannot be completely covered by multiple consecutive 64 bytes,
391 	 * it needs to be checked one by one.
392 	 */
393 	for (; addr < meta->addr; addr++) {
394 		if (unlikely(!check_canary_byte((u8 *)addr)))
395 			break;
396 	}
397 
398 	/* Apply to right of object. */
399 	for (addr = meta->addr + meta->size; addr % sizeof(u64) != 0; addr++) {
400 		if (unlikely(!check_canary_byte((u8 *)addr)))
401 			return;
402 	}
403 	for (; addr - pageaddr < PAGE_SIZE; addr += sizeof(u64)) {
404 		if (unlikely(*((u64 *)addr) != KFENCE_CANARY_PATTERN_U64)) {
405 
406 			for (; addr - pageaddr < PAGE_SIZE; addr++) {
407 				if (!check_canary_byte((u8 *)addr))
408 					return;
409 			}
410 		}
411 	}
412 }
413 
kfence_guarded_alloc(struct kmem_cache * cache,size_t size,gfp_t gfp,unsigned long * stack_entries,size_t num_stack_entries,u32 alloc_stack_hash)414 static void *kfence_guarded_alloc(struct kmem_cache *cache, size_t size, gfp_t gfp,
415 				  unsigned long *stack_entries, size_t num_stack_entries,
416 				  u32 alloc_stack_hash)
417 {
418 	struct kfence_metadata *meta = NULL;
419 	unsigned long flags;
420 	struct slab *slab;
421 	void *addr;
422 	const bool random_right_allocate = get_random_u32_below(2);
423 	const bool random_fault = CONFIG_KFENCE_STRESS_TEST_FAULTS &&
424 				  !get_random_u32_below(CONFIG_KFENCE_STRESS_TEST_FAULTS);
425 
426 	/* Try to obtain a free object. */
427 	raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
428 	if (!list_empty(&kfence_freelist)) {
429 		meta = list_entry(kfence_freelist.next, struct kfence_metadata, list);
430 		list_del_init(&meta->list);
431 	}
432 	raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
433 	if (!meta) {
434 		atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_CAPACITY]);
435 		return NULL;
436 	}
437 
438 	if (unlikely(!raw_spin_trylock_irqsave(&meta->lock, flags))) {
439 		/*
440 		 * This is extremely unlikely -- we are reporting on a
441 		 * use-after-free, which locked meta->lock, and the reporting
442 		 * code via printk calls kmalloc() which ends up in
443 		 * kfence_alloc() and tries to grab the same object that we're
444 		 * reporting on. While it has never been observed, lockdep does
445 		 * report that there is a possibility of deadlock. Fix it by
446 		 * using trylock and bailing out gracefully.
447 		 */
448 		raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
449 		/* Put the object back on the freelist. */
450 		list_add_tail(&meta->list, &kfence_freelist);
451 		raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
452 
453 		return NULL;
454 	}
455 
456 	meta->addr = metadata_to_pageaddr(meta);
457 	/* Unprotect if we're reusing this page. */
458 	if (meta->state == KFENCE_OBJECT_FREED)
459 		kfence_unprotect(meta->addr);
460 
461 	/*
462 	 * Note: for allocations made before RNG initialization, will always
463 	 * return zero. We still benefit from enabling KFENCE as early as
464 	 * possible, even when the RNG is not yet available, as this will allow
465 	 * KFENCE to detect bugs due to earlier allocations. The only downside
466 	 * is that the out-of-bounds accesses detected are deterministic for
467 	 * such allocations.
468 	 */
469 	if (random_right_allocate) {
470 		/* Allocate on the "right" side, re-calculate address. */
471 		meta->addr += PAGE_SIZE - size;
472 		meta->addr = ALIGN_DOWN(meta->addr, cache->align);
473 	}
474 
475 	addr = (void *)meta->addr;
476 
477 	/* Update remaining metadata. */
478 	metadata_update_state(meta, KFENCE_OBJECT_ALLOCATED, stack_entries, num_stack_entries);
479 	/* Pairs with READ_ONCE() in kfence_shutdown_cache(). */
480 	WRITE_ONCE(meta->cache, cache);
481 	meta->size = size;
482 	meta->alloc_stack_hash = alloc_stack_hash;
483 	raw_spin_unlock_irqrestore(&meta->lock, flags);
484 
485 	alloc_covered_add(alloc_stack_hash, 1);
486 
487 	/* Set required slab fields. */
488 	slab = virt_to_slab((void *)meta->addr);
489 	slab->slab_cache = cache;
490 	slab->objects = 1;
491 
492 	/* Memory initialization. */
493 	set_canary(meta);
494 
495 	/*
496 	 * We check slab_want_init_on_alloc() ourselves, rather than letting
497 	 * SL*B do the initialization, as otherwise we might overwrite KFENCE's
498 	 * redzone.
499 	 */
500 	if (unlikely(slab_want_init_on_alloc(gfp, cache)))
501 		memzero_explicit(addr, size);
502 	if (cache->ctor)
503 		cache->ctor(addr);
504 
505 	if (random_fault)
506 		kfence_protect(meta->addr); /* Random "faults" by protecting the object. */
507 
508 	atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCATED]);
509 	atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCS]);
510 
511 	return addr;
512 }
513 
kfence_guarded_free(void * addr,struct kfence_metadata * meta,bool zombie)514 static void kfence_guarded_free(void *addr, struct kfence_metadata *meta, bool zombie)
515 {
516 	struct kcsan_scoped_access assert_page_exclusive;
517 	unsigned long flags;
518 	bool init;
519 
520 	raw_spin_lock_irqsave(&meta->lock, flags);
521 
522 	if (!kfence_obj_allocated(meta) || meta->addr != (unsigned long)addr) {
523 		/* Invalid or double-free, bail out. */
524 		atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
525 		kfence_report_error((unsigned long)addr, false, NULL, meta,
526 				    KFENCE_ERROR_INVALID_FREE);
527 		raw_spin_unlock_irqrestore(&meta->lock, flags);
528 		return;
529 	}
530 
531 	/* Detect racy use-after-free, or incorrect reallocation of this page by KFENCE. */
532 	kcsan_begin_scoped_access((void *)ALIGN_DOWN((unsigned long)addr, PAGE_SIZE), PAGE_SIZE,
533 				  KCSAN_ACCESS_SCOPED | KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT,
534 				  &assert_page_exclusive);
535 
536 	if (CONFIG_KFENCE_STRESS_TEST_FAULTS)
537 		kfence_unprotect((unsigned long)addr); /* To check canary bytes. */
538 
539 	/* Restore page protection if there was an OOB access. */
540 	if (meta->unprotected_page) {
541 		memzero_explicit((void *)ALIGN_DOWN(meta->unprotected_page, PAGE_SIZE), PAGE_SIZE);
542 		kfence_protect(meta->unprotected_page);
543 		meta->unprotected_page = 0;
544 	}
545 
546 	/* Mark the object as freed. */
547 	metadata_update_state(meta, KFENCE_OBJECT_FREED, NULL, 0);
548 	init = slab_want_init_on_free(meta->cache);
549 	raw_spin_unlock_irqrestore(&meta->lock, flags);
550 
551 	alloc_covered_add(meta->alloc_stack_hash, -1);
552 
553 	/* Check canary bytes for memory corruption. */
554 	check_canary(meta);
555 
556 	/*
557 	 * Clear memory if init-on-free is set. While we protect the page, the
558 	 * data is still there, and after a use-after-free is detected, we
559 	 * unprotect the page, so the data is still accessible.
560 	 */
561 	if (!zombie && unlikely(init))
562 		memzero_explicit(addr, meta->size);
563 
564 	/* Protect to detect use-after-frees. */
565 	kfence_protect((unsigned long)addr);
566 
567 	kcsan_end_scoped_access(&assert_page_exclusive);
568 	if (!zombie) {
569 		/* Add it to the tail of the freelist for reuse. */
570 		raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
571 		KFENCE_WARN_ON(!list_empty(&meta->list));
572 		list_add_tail(&meta->list, &kfence_freelist);
573 		raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
574 
575 		atomic_long_dec(&counters[KFENCE_COUNTER_ALLOCATED]);
576 		atomic_long_inc(&counters[KFENCE_COUNTER_FREES]);
577 	} else {
578 		/* See kfence_shutdown_cache(). */
579 		atomic_long_inc(&counters[KFENCE_COUNTER_ZOMBIES]);
580 	}
581 }
582 
rcu_guarded_free(struct rcu_head * h)583 static void rcu_guarded_free(struct rcu_head *h)
584 {
585 	struct kfence_metadata *meta = container_of(h, struct kfence_metadata, rcu_head);
586 
587 	kfence_guarded_free((void *)meta->addr, meta, false);
588 }
589 
590 /*
591  * Initialization of the KFENCE pool after its allocation.
592  * Returns 0 on success; otherwise returns the address up to
593  * which partial initialization succeeded.
594  */
kfence_init_pool(void)595 static unsigned long kfence_init_pool(void)
596 {
597 	unsigned long addr, start_pfn;
598 	int i;
599 
600 	if (!arch_kfence_init_pool())
601 		return (unsigned long)__kfence_pool;
602 
603 	addr = (unsigned long)__kfence_pool;
604 	start_pfn = PHYS_PFN(virt_to_phys(__kfence_pool));
605 
606 	/*
607 	 * Set up object pages: they must have PGTY_slab set to avoid freeing
608 	 * them as real pages.
609 	 *
610 	 * We also want to avoid inserting kfence_free() in the kfree()
611 	 * fast-path in SLUB, and therefore need to ensure kfree() correctly
612 	 * enters __slab_free() slow-path.
613 	 */
614 	for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
615 		struct slab *slab;
616 
617 		if (!i || (i % 2))
618 			continue;
619 
620 		slab = page_slab(pfn_to_page(start_pfn + i));
621 		__folio_set_slab(slab_folio(slab));
622 #ifdef CONFIG_MEMCG
623 		slab->obj_exts = (unsigned long)&kfence_metadata_init[i / 2 - 1].obj_exts |
624 				 MEMCG_DATA_OBJEXTS;
625 #endif
626 	}
627 
628 	/*
629 	 * Protect the first 2 pages. The first page is mostly unnecessary, and
630 	 * merely serves as an extended guard page. However, adding one
631 	 * additional page in the beginning gives us an even number of pages,
632 	 * which simplifies the mapping of address to metadata index.
633 	 */
634 	for (i = 0; i < 2; i++) {
635 		if (unlikely(!kfence_protect(addr)))
636 			return addr;
637 
638 		addr += PAGE_SIZE;
639 	}
640 
641 	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
642 		struct kfence_metadata *meta = &kfence_metadata_init[i];
643 
644 		/* Initialize metadata. */
645 		INIT_LIST_HEAD(&meta->list);
646 		raw_spin_lock_init(&meta->lock);
647 		meta->state = KFENCE_OBJECT_UNUSED;
648 		meta->addr = addr; /* Initialize for validation in metadata_to_pageaddr(). */
649 		list_add_tail(&meta->list, &kfence_freelist);
650 
651 		/* Protect the right redzone. */
652 		if (unlikely(!kfence_protect(addr + PAGE_SIZE)))
653 			goto reset_slab;
654 
655 		addr += 2 * PAGE_SIZE;
656 	}
657 
658 	/*
659 	 * Make kfence_metadata visible only when initialization is successful.
660 	 * Otherwise, if the initialization fails and kfence_metadata is freed,
661 	 * it may cause UAF in kfence_shutdown_cache().
662 	 */
663 	smp_store_release(&kfence_metadata, kfence_metadata_init);
664 	return 0;
665 
666 reset_slab:
667 	for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
668 		struct slab *slab;
669 
670 		if (!i || (i % 2))
671 			continue;
672 
673 		slab = page_slab(pfn_to_page(start_pfn + i));
674 #ifdef CONFIG_MEMCG
675 		slab->obj_exts = 0;
676 #endif
677 		__folio_clear_slab(slab_folio(slab));
678 	}
679 
680 	return addr;
681 }
682 
kfence_init_pool_early(void)683 static bool __init kfence_init_pool_early(void)
684 {
685 	unsigned long addr;
686 
687 	if (!__kfence_pool)
688 		return false;
689 
690 	addr = kfence_init_pool();
691 
692 	if (!addr) {
693 		/*
694 		 * The pool is live and will never be deallocated from this point on.
695 		 * Ignore the pool object from the kmemleak phys object tree, as it would
696 		 * otherwise overlap with allocations returned by kfence_alloc(), which
697 		 * are registered with kmemleak through the slab post-alloc hook.
698 		 */
699 		kmemleak_ignore_phys(__pa(__kfence_pool));
700 		return true;
701 	}
702 
703 	/*
704 	 * Only release unprotected pages, and do not try to go back and change
705 	 * page attributes due to risk of failing to do so as well. If changing
706 	 * page attributes for some pages fails, it is very likely that it also
707 	 * fails for the first page, and therefore expect addr==__kfence_pool in
708 	 * most failure cases.
709 	 */
710 	memblock_free_late(__pa(addr), KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool));
711 	__kfence_pool = NULL;
712 
713 	memblock_free_late(__pa(kfence_metadata_init), KFENCE_METADATA_SIZE);
714 	kfence_metadata_init = NULL;
715 
716 	return false;
717 }
718 
719 /* === DebugFS Interface ==================================================== */
720 
stats_show(struct seq_file * seq,void * v)721 static int stats_show(struct seq_file *seq, void *v)
722 {
723 	int i;
724 
725 	seq_printf(seq, "enabled: %i\n", READ_ONCE(kfence_enabled));
726 	for (i = 0; i < KFENCE_COUNTER_COUNT; i++)
727 		seq_printf(seq, "%s: %ld\n", counter_names[i], atomic_long_read(&counters[i]));
728 
729 	return 0;
730 }
731 DEFINE_SHOW_ATTRIBUTE(stats);
732 
733 /*
734  * debugfs seq_file operations for /sys/kernel/debug/kfence/objects.
735  * start_object() and next_object() return the object index + 1, because NULL is used
736  * to stop iteration.
737  */
start_object(struct seq_file * seq,loff_t * pos)738 static void *start_object(struct seq_file *seq, loff_t *pos)
739 {
740 	if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
741 		return (void *)((long)*pos + 1);
742 	return NULL;
743 }
744 
stop_object(struct seq_file * seq,void * v)745 static void stop_object(struct seq_file *seq, void *v)
746 {
747 }
748 
next_object(struct seq_file * seq,void * v,loff_t * pos)749 static void *next_object(struct seq_file *seq, void *v, loff_t *pos)
750 {
751 	++*pos;
752 	if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
753 		return (void *)((long)*pos + 1);
754 	return NULL;
755 }
756 
show_object(struct seq_file * seq,void * v)757 static int show_object(struct seq_file *seq, void *v)
758 {
759 	struct kfence_metadata *meta = &kfence_metadata[(long)v - 1];
760 	unsigned long flags;
761 
762 	raw_spin_lock_irqsave(&meta->lock, flags);
763 	kfence_print_object(seq, meta);
764 	raw_spin_unlock_irqrestore(&meta->lock, flags);
765 	seq_puts(seq, "---------------------------------\n");
766 
767 	return 0;
768 }
769 
770 static const struct seq_operations objects_sops = {
771 	.start = start_object,
772 	.next = next_object,
773 	.stop = stop_object,
774 	.show = show_object,
775 };
776 DEFINE_SEQ_ATTRIBUTE(objects);
777 
kfence_debugfs_init(void)778 static int kfence_debugfs_init(void)
779 {
780 	struct dentry *kfence_dir;
781 
782 	if (!READ_ONCE(kfence_enabled))
783 		return 0;
784 
785 	kfence_dir = debugfs_create_dir("kfence", NULL);
786 	debugfs_create_file("stats", 0444, kfence_dir, NULL, &stats_fops);
787 	debugfs_create_file("objects", 0400, kfence_dir, NULL, &objects_fops);
788 	return 0;
789 }
790 
791 late_initcall(kfence_debugfs_init);
792 
793 /* === Panic Notifier ====================================================== */
794 
kfence_check_all_canary(void)795 static void kfence_check_all_canary(void)
796 {
797 	int i;
798 
799 	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
800 		struct kfence_metadata *meta = &kfence_metadata[i];
801 
802 		if (kfence_obj_allocated(meta))
803 			check_canary(meta);
804 	}
805 }
806 
kfence_check_canary_callback(struct notifier_block * nb,unsigned long reason,void * arg)807 static int kfence_check_canary_callback(struct notifier_block *nb,
808 					unsigned long reason, void *arg)
809 {
810 	kfence_check_all_canary();
811 	return NOTIFY_OK;
812 }
813 
814 static struct notifier_block kfence_check_canary_notifier = {
815 	.notifier_call = kfence_check_canary_callback,
816 };
817 
818 /* === Allocation Gate Timer ================================================ */
819 
820 static struct delayed_work kfence_timer;
821 
822 #ifdef CONFIG_KFENCE_STATIC_KEYS
823 /* Wait queue to wake up allocation-gate timer task. */
824 static DECLARE_WAIT_QUEUE_HEAD(allocation_wait);
825 
wake_up_kfence_timer(struct irq_work * work)826 static void wake_up_kfence_timer(struct irq_work *work)
827 {
828 	wake_up(&allocation_wait);
829 }
830 static DEFINE_IRQ_WORK(wake_up_kfence_timer_work, wake_up_kfence_timer);
831 #endif
832 
833 /*
834  * Set up delayed work, which will enable and disable the static key. We need to
835  * use a work queue (rather than a simple timer), since enabling and disabling a
836  * static key cannot be done from an interrupt.
837  *
838  * Note: Toggling a static branch currently causes IPIs, and here we'll end up
839  * with a total of 2 IPIs to all CPUs. If this ends up a problem in future (with
840  * more aggressive sampling intervals), we could get away with a variant that
841  * avoids IPIs, at the cost of not immediately capturing allocations if the
842  * instructions remain cached.
843  */
toggle_allocation_gate(struct work_struct * work)844 static void toggle_allocation_gate(struct work_struct *work)
845 {
846 	if (!READ_ONCE(kfence_enabled))
847 		return;
848 
849 	atomic_set(&kfence_allocation_gate, -kfence_burst);
850 #ifdef CONFIG_KFENCE_STATIC_KEYS
851 	/* Enable static key, and await allocation to happen. */
852 	static_branch_enable(&kfence_allocation_key);
853 
854 	wait_event_idle(allocation_wait, atomic_read(&kfence_allocation_gate) > 0);
855 
856 	/* Disable static key and reset timer. */
857 	static_branch_disable(&kfence_allocation_key);
858 #endif
859 	queue_delayed_work(system_unbound_wq, &kfence_timer,
860 			   msecs_to_jiffies(kfence_sample_interval));
861 }
862 
863 /* === Public interface ===================================================== */
864 
kfence_alloc_pool_and_metadata(void)865 void __init kfence_alloc_pool_and_metadata(void)
866 {
867 	if (!kfence_sample_interval)
868 		return;
869 
870 	/*
871 	 * If the pool has already been initialized by arch, there is no need to
872 	 * re-allocate the memory pool.
873 	 */
874 	if (!__kfence_pool)
875 		__kfence_pool = memblock_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
876 
877 	if (!__kfence_pool) {
878 		pr_err("failed to allocate pool\n");
879 		return;
880 	}
881 
882 	/* The memory allocated by memblock has been zeroed out. */
883 	kfence_metadata_init = memblock_alloc(KFENCE_METADATA_SIZE, PAGE_SIZE);
884 	if (!kfence_metadata_init) {
885 		pr_err("failed to allocate metadata\n");
886 		memblock_free(__kfence_pool, KFENCE_POOL_SIZE);
887 		__kfence_pool = NULL;
888 	}
889 }
890 
kfence_init_enable(void)891 static void kfence_init_enable(void)
892 {
893 	if (!IS_ENABLED(CONFIG_KFENCE_STATIC_KEYS))
894 		static_branch_enable(&kfence_allocation_key);
895 
896 	if (kfence_deferrable)
897 		INIT_DEFERRABLE_WORK(&kfence_timer, toggle_allocation_gate);
898 	else
899 		INIT_DELAYED_WORK(&kfence_timer, toggle_allocation_gate);
900 
901 	if (kfence_check_on_panic)
902 		atomic_notifier_chain_register(&panic_notifier_list, &kfence_check_canary_notifier);
903 
904 	WRITE_ONCE(kfence_enabled, true);
905 	queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
906 
907 	pr_info("initialized - using %lu bytes for %d objects at 0x%p-0x%p\n", KFENCE_POOL_SIZE,
908 		CONFIG_KFENCE_NUM_OBJECTS, (void *)__kfence_pool,
909 		(void *)(__kfence_pool + KFENCE_POOL_SIZE));
910 }
911 
kfence_init(void)912 void __init kfence_init(void)
913 {
914 	stack_hash_seed = get_random_u32();
915 
916 	/* Setting kfence_sample_interval to 0 on boot disables KFENCE. */
917 	if (!kfence_sample_interval)
918 		return;
919 
920 	if (!kfence_init_pool_early()) {
921 		pr_err("%s failed\n", __func__);
922 		return;
923 	}
924 
925 	kfence_init_enable();
926 }
927 
kfence_init_late(void)928 static int kfence_init_late(void)
929 {
930 	const unsigned long nr_pages_pool = KFENCE_POOL_SIZE / PAGE_SIZE;
931 	const unsigned long nr_pages_meta = KFENCE_METADATA_SIZE / PAGE_SIZE;
932 	unsigned long addr = (unsigned long)__kfence_pool;
933 	unsigned long free_size = KFENCE_POOL_SIZE;
934 	int err = -ENOMEM;
935 
936 #ifdef CONFIG_CONTIG_ALLOC
937 	struct page *pages;
938 
939 	pages = alloc_contig_pages(nr_pages_pool, GFP_KERNEL, first_online_node,
940 				   NULL);
941 	if (!pages)
942 		return -ENOMEM;
943 
944 	__kfence_pool = page_to_virt(pages);
945 	pages = alloc_contig_pages(nr_pages_meta, GFP_KERNEL, first_online_node,
946 				   NULL);
947 	if (pages)
948 		kfence_metadata_init = page_to_virt(pages);
949 #else
950 	if (nr_pages_pool > MAX_ORDER_NR_PAGES ||
951 	    nr_pages_meta > MAX_ORDER_NR_PAGES) {
952 		pr_warn("KFENCE_NUM_OBJECTS too large for buddy allocator\n");
953 		return -EINVAL;
954 	}
955 
956 	__kfence_pool = alloc_pages_exact(KFENCE_POOL_SIZE, GFP_KERNEL);
957 	if (!__kfence_pool)
958 		return -ENOMEM;
959 
960 	kfence_metadata_init = alloc_pages_exact(KFENCE_METADATA_SIZE, GFP_KERNEL);
961 #endif
962 
963 	if (!kfence_metadata_init)
964 		goto free_pool;
965 
966 	memzero_explicit(kfence_metadata_init, KFENCE_METADATA_SIZE);
967 	addr = kfence_init_pool();
968 	if (!addr) {
969 		kfence_init_enable();
970 		kfence_debugfs_init();
971 		return 0;
972 	}
973 
974 	pr_err("%s failed\n", __func__);
975 	free_size = KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool);
976 	err = -EBUSY;
977 
978 #ifdef CONFIG_CONTIG_ALLOC
979 	free_contig_range(page_to_pfn(virt_to_page((void *)kfence_metadata_init)),
980 			  nr_pages_meta);
981 free_pool:
982 	free_contig_range(page_to_pfn(virt_to_page((void *)addr)),
983 			  free_size / PAGE_SIZE);
984 #else
985 	free_pages_exact((void *)kfence_metadata_init, KFENCE_METADATA_SIZE);
986 free_pool:
987 	free_pages_exact((void *)addr, free_size);
988 #endif
989 
990 	kfence_metadata_init = NULL;
991 	__kfence_pool = NULL;
992 	return err;
993 }
994 
kfence_enable_late(void)995 static int kfence_enable_late(void)
996 {
997 	if (!__kfence_pool)
998 		return kfence_init_late();
999 
1000 	WRITE_ONCE(kfence_enabled, true);
1001 	queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
1002 	pr_info("re-enabled\n");
1003 	return 0;
1004 }
1005 
kfence_shutdown_cache(struct kmem_cache * s)1006 void kfence_shutdown_cache(struct kmem_cache *s)
1007 {
1008 	unsigned long flags;
1009 	struct kfence_metadata *meta;
1010 	int i;
1011 
1012 	/* Pairs with release in kfence_init_pool(). */
1013 	if (!smp_load_acquire(&kfence_metadata))
1014 		return;
1015 
1016 	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
1017 		bool in_use;
1018 
1019 		meta = &kfence_metadata[i];
1020 
1021 		/*
1022 		 * If we observe some inconsistent cache and state pair where we
1023 		 * should have returned false here, cache destruction is racing
1024 		 * with either kmem_cache_alloc() or kmem_cache_free(). Taking
1025 		 * the lock will not help, as different critical section
1026 		 * serialization will have the same outcome.
1027 		 */
1028 		if (READ_ONCE(meta->cache) != s || !kfence_obj_allocated(meta))
1029 			continue;
1030 
1031 		raw_spin_lock_irqsave(&meta->lock, flags);
1032 		in_use = meta->cache == s && kfence_obj_allocated(meta);
1033 		raw_spin_unlock_irqrestore(&meta->lock, flags);
1034 
1035 		if (in_use) {
1036 			/*
1037 			 * This cache still has allocations, and we should not
1038 			 * release them back into the freelist so they can still
1039 			 * safely be used and retain the kernel's default
1040 			 * behaviour of keeping the allocations alive (leak the
1041 			 * cache); however, they effectively become "zombie
1042 			 * allocations" as the KFENCE objects are the only ones
1043 			 * still in use and the owning cache is being destroyed.
1044 			 *
1045 			 * We mark them freed, so that any subsequent use shows
1046 			 * more useful error messages that will include stack
1047 			 * traces of the user of the object, the original
1048 			 * allocation, and caller to shutdown_cache().
1049 			 */
1050 			kfence_guarded_free((void *)meta->addr, meta, /*zombie=*/true);
1051 		}
1052 	}
1053 
1054 	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
1055 		meta = &kfence_metadata[i];
1056 
1057 		/* See above. */
1058 		if (READ_ONCE(meta->cache) != s || READ_ONCE(meta->state) != KFENCE_OBJECT_FREED)
1059 			continue;
1060 
1061 		raw_spin_lock_irqsave(&meta->lock, flags);
1062 		if (meta->cache == s && meta->state == KFENCE_OBJECT_FREED)
1063 			meta->cache = NULL;
1064 		raw_spin_unlock_irqrestore(&meta->lock, flags);
1065 	}
1066 }
1067 
__kfence_alloc(struct kmem_cache * s,size_t size,gfp_t flags)1068 void *__kfence_alloc(struct kmem_cache *s, size_t size, gfp_t flags)
1069 {
1070 	unsigned long stack_entries[KFENCE_STACK_DEPTH];
1071 	size_t num_stack_entries;
1072 	u32 alloc_stack_hash;
1073 	int allocation_gate;
1074 
1075 	/*
1076 	 * Perform size check before switching kfence_allocation_gate, so that
1077 	 * we don't disable KFENCE without making an allocation.
1078 	 */
1079 	if (size > PAGE_SIZE) {
1080 		atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
1081 		return NULL;
1082 	}
1083 
1084 	/*
1085 	 * Skip allocations from non-default zones, including DMA. We cannot
1086 	 * guarantee that pages in the KFENCE pool will have the requested
1087 	 * properties (e.g. reside in DMAable memory).
1088 	 */
1089 	if ((flags & GFP_ZONEMASK) ||
1090 	    ((flags & __GFP_THISNODE) && num_online_nodes() > 1) ||
1091 	    (s->flags & (SLAB_CACHE_DMA | SLAB_CACHE_DMA32))) {
1092 		atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
1093 		return NULL;
1094 	}
1095 
1096 	/*
1097 	 * Skip allocations for this slab, if KFENCE has been disabled for
1098 	 * this slab.
1099 	 */
1100 	if (s->flags & SLAB_SKIP_KFENCE)
1101 		return NULL;
1102 
1103 	allocation_gate = atomic_inc_return(&kfence_allocation_gate);
1104 	if (allocation_gate > 1)
1105 		return NULL;
1106 #ifdef CONFIG_KFENCE_STATIC_KEYS
1107 	/*
1108 	 * waitqueue_active() is fully ordered after the update of
1109 	 * kfence_allocation_gate per atomic_inc_return().
1110 	 */
1111 	if (allocation_gate == 1 && waitqueue_active(&allocation_wait)) {
1112 		/*
1113 		 * Calling wake_up() here may deadlock when allocations happen
1114 		 * from within timer code. Use an irq_work to defer it.
1115 		 */
1116 		irq_work_queue(&wake_up_kfence_timer_work);
1117 	}
1118 #endif
1119 
1120 	if (!READ_ONCE(kfence_enabled))
1121 		return NULL;
1122 
1123 	num_stack_entries = stack_trace_save(stack_entries, KFENCE_STACK_DEPTH, 0);
1124 
1125 	/*
1126 	 * Do expensive check for coverage of allocation in slow-path after
1127 	 * allocation_gate has already become non-zero, even though it might
1128 	 * mean not making any allocation within a given sample interval.
1129 	 *
1130 	 * This ensures reasonable allocation coverage when the pool is almost
1131 	 * full, including avoiding long-lived allocations of the same source
1132 	 * filling up the pool (e.g. pagecache allocations).
1133 	 */
1134 	alloc_stack_hash = get_alloc_stack_hash(stack_entries, num_stack_entries);
1135 	if (should_skip_covered() && alloc_covered_contains(alloc_stack_hash)) {
1136 		atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_COVERED]);
1137 		return NULL;
1138 	}
1139 
1140 	return kfence_guarded_alloc(s, size, flags, stack_entries, num_stack_entries,
1141 				    alloc_stack_hash);
1142 }
1143 
kfence_ksize(const void * addr)1144 size_t kfence_ksize(const void *addr)
1145 {
1146 	const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1147 
1148 	/*
1149 	 * Read locklessly -- if there is a race with __kfence_alloc(), this is
1150 	 * either a use-after-free or invalid access.
1151 	 */
1152 	return meta ? meta->size : 0;
1153 }
1154 
kfence_object_start(const void * addr)1155 void *kfence_object_start(const void *addr)
1156 {
1157 	const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1158 
1159 	/*
1160 	 * Read locklessly -- if there is a race with __kfence_alloc(), this is
1161 	 * either a use-after-free or invalid access.
1162 	 */
1163 	return meta ? (void *)meta->addr : NULL;
1164 }
1165 
__kfence_free(void * addr)1166 void __kfence_free(void *addr)
1167 {
1168 	struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1169 
1170 #ifdef CONFIG_MEMCG
1171 	KFENCE_WARN_ON(meta->obj_exts.objcg);
1172 #endif
1173 	/*
1174 	 * If the objects of the cache are SLAB_TYPESAFE_BY_RCU, defer freeing
1175 	 * the object, as the object page may be recycled for other-typed
1176 	 * objects once it has been freed. meta->cache may be NULL if the cache
1177 	 * was destroyed.
1178 	 * Save the stack trace here so that reports show where the user freed
1179 	 * the object.
1180 	 */
1181 	if (unlikely(meta->cache && (meta->cache->flags & SLAB_TYPESAFE_BY_RCU))) {
1182 		unsigned long flags;
1183 
1184 		raw_spin_lock_irqsave(&meta->lock, flags);
1185 		metadata_update_state(meta, KFENCE_OBJECT_RCU_FREEING, NULL, 0);
1186 		raw_spin_unlock_irqrestore(&meta->lock, flags);
1187 		call_rcu(&meta->rcu_head, rcu_guarded_free);
1188 	} else {
1189 		kfence_guarded_free(addr, meta, false);
1190 	}
1191 }
1192 
kfence_handle_page_fault(unsigned long addr,bool is_write,struct pt_regs * regs)1193 bool kfence_handle_page_fault(unsigned long addr, bool is_write, struct pt_regs *regs)
1194 {
1195 	const int page_index = (addr - (unsigned long)__kfence_pool) / PAGE_SIZE;
1196 	struct kfence_metadata *to_report = NULL;
1197 	enum kfence_error_type error_type;
1198 	unsigned long flags;
1199 
1200 	if (!is_kfence_address((void *)addr))
1201 		return false;
1202 
1203 	if (!READ_ONCE(kfence_enabled)) /* If disabled at runtime ... */
1204 		return kfence_unprotect(addr); /* ... unprotect and proceed. */
1205 
1206 	atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
1207 
1208 	if (page_index % 2) {
1209 		/* This is a redzone, report a buffer overflow. */
1210 		struct kfence_metadata *meta;
1211 		int distance = 0;
1212 
1213 		meta = addr_to_metadata(addr - PAGE_SIZE);
1214 		if (meta && kfence_obj_allocated(meta)) {
1215 			to_report = meta;
1216 			/* Data race ok; distance calculation approximate. */
1217 			distance = addr - data_race(meta->addr + meta->size);
1218 		}
1219 
1220 		meta = addr_to_metadata(addr + PAGE_SIZE);
1221 		if (meta && kfence_obj_allocated(meta)) {
1222 			/* Data race ok; distance calculation approximate. */
1223 			if (!to_report || distance > data_race(meta->addr) - addr)
1224 				to_report = meta;
1225 		}
1226 
1227 		if (!to_report)
1228 			goto out;
1229 
1230 		raw_spin_lock_irqsave(&to_report->lock, flags);
1231 		to_report->unprotected_page = addr;
1232 		error_type = KFENCE_ERROR_OOB;
1233 
1234 		/*
1235 		 * If the object was freed before we took the look we can still
1236 		 * report this as an OOB -- the report will simply show the
1237 		 * stacktrace of the free as well.
1238 		 */
1239 	} else {
1240 		to_report = addr_to_metadata(addr);
1241 		if (!to_report)
1242 			goto out;
1243 
1244 		raw_spin_lock_irqsave(&to_report->lock, flags);
1245 		error_type = KFENCE_ERROR_UAF;
1246 		/*
1247 		 * We may race with __kfence_alloc(), and it is possible that a
1248 		 * freed object may be reallocated. We simply report this as a
1249 		 * use-after-free, with the stack trace showing the place where
1250 		 * the object was re-allocated.
1251 		 */
1252 	}
1253 
1254 out:
1255 	if (to_report) {
1256 		kfence_report_error(addr, is_write, regs, to_report, error_type);
1257 		raw_spin_unlock_irqrestore(&to_report->lock, flags);
1258 	} else {
1259 		/* This may be a UAF or OOB access, but we can't be sure. */
1260 		kfence_report_error(addr, is_write, regs, NULL, KFENCE_ERROR_INVALID);
1261 	}
1262 
1263 	return kfence_unprotect(addr); /* Unprotect and let access proceed. */
1264 }
1265