xref: /linux/mm/kfence/core.c (revision bcb6058a4b4596f12065276faeb9363dc4887ea9)
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/reboot.h>
30 #include <linux/sched/clock.h>
31 #include <linux/seq_file.h>
32 #include <linux/slab.h>
33 #include <linux/spinlock.h>
34 #include <linux/string.h>
35 
36 #include <asm/kfence.h>
37 
38 #include "kfence.h"
39 
40 /* Disables KFENCE on the first warning assuming an irrecoverable error. */
41 #define KFENCE_WARN_ON(cond)                                                   \
42 	({                                                                     \
43 		const bool __cond = WARN_ON(cond);                             \
44 		if (unlikely(__cond)) {                                        \
45 			WRITE_ONCE(kfence_enabled, false);                     \
46 			disabled_by_warn = true;                               \
47 		}                                                              \
48 		__cond;                                                        \
49 	})
50 
51 /* === Data ================================================================= */
52 
53 static bool kfence_enabled __read_mostly;
54 static bool disabled_by_warn __read_mostly;
55 
56 unsigned long kfence_sample_interval __read_mostly = CONFIG_KFENCE_SAMPLE_INTERVAL;
57 EXPORT_SYMBOL_GPL(kfence_sample_interval); /* Export for test modules. */
58 
59 #ifdef MODULE_PARAM_PREFIX
60 #undef MODULE_PARAM_PREFIX
61 #endif
62 #define MODULE_PARAM_PREFIX "kfence."
63 
64 static int kfence_enable_late(void);
param_set_sample_interval(const char * val,const struct kernel_param * kp)65 static int param_set_sample_interval(const char *val, const struct kernel_param *kp)
66 {
67 	unsigned long num;
68 	int ret = kstrtoul(val, 0, &num);
69 
70 	if (ret < 0)
71 		return ret;
72 
73 	/* Using 0 to indicate KFENCE is disabled. */
74 	if (!num && READ_ONCE(kfence_enabled)) {
75 		pr_info("disabled\n");
76 		WRITE_ONCE(kfence_enabled, false);
77 	}
78 
79 	*((unsigned long *)kp->arg) = num;
80 
81 	if (num && !READ_ONCE(kfence_enabled) && system_state != SYSTEM_BOOTING)
82 		return disabled_by_warn ? -EINVAL : kfence_enable_late();
83 	return 0;
84 }
85 
param_get_sample_interval(char * buffer,const struct kernel_param * kp)86 static int param_get_sample_interval(char *buffer, const struct kernel_param *kp)
87 {
88 	if (!READ_ONCE(kfence_enabled))
89 		return sprintf(buffer, "0\n");
90 
91 	return param_get_ulong(buffer, kp);
92 }
93 
94 static const struct kernel_param_ops sample_interval_param_ops = {
95 	.set = param_set_sample_interval,
96 	.get = param_get_sample_interval,
97 };
98 module_param_cb(sample_interval, &sample_interval_param_ops, &kfence_sample_interval, 0600);
99 
100 /* Pool usage% threshold when currently covered allocations are skipped. */
101 static unsigned long kfence_skip_covered_thresh __read_mostly = 75;
102 module_param_named(skip_covered_thresh, kfence_skip_covered_thresh, ulong, 0644);
103 
104 /* Allocation burst count: number of excess KFENCE allocations per sample. */
105 static unsigned int kfence_burst __read_mostly;
106 module_param_named(burst, kfence_burst, uint, 0644);
107 
108 /* If true, use a deferrable timer. */
109 static bool kfence_deferrable __read_mostly = IS_ENABLED(CONFIG_KFENCE_DEFERRABLE);
110 module_param_named(deferrable, kfence_deferrable, bool, 0444);
111 
112 /* If true, check all canary bytes on panic. */
113 static bool kfence_check_on_panic __read_mostly;
114 module_param_named(check_on_panic, kfence_check_on_panic, bool, 0444);
115 
116 /* The pool of pages used for guard pages and objects. */
117 char *__kfence_pool __read_mostly;
118 EXPORT_SYMBOL(__kfence_pool); /* Export for test modules. */
119 
120 /*
121  * Per-object metadata, with one-to-one mapping of object metadata to
122  * backing pages (in __kfence_pool).
123  */
124 static_assert(CONFIG_KFENCE_NUM_OBJECTS > 0);
125 struct kfence_metadata *kfence_metadata __read_mostly;
126 
127 /*
128  * If kfence_metadata is not NULL, it may be accessed by kfence_shutdown_cache().
129  * So introduce kfence_metadata_init to initialize metadata, and then make
130  * kfence_metadata visible after initialization is successful. This prevents
131  * potential UAF or access to uninitialized metadata.
132  */
133 static struct kfence_metadata *kfence_metadata_init __read_mostly;
134 
135 /* Freelist with available objects. */
136 static struct list_head kfence_freelist = LIST_HEAD_INIT(kfence_freelist);
137 static DEFINE_RAW_SPINLOCK(kfence_freelist_lock); /* Lock protecting freelist. */
138 
139 /*
140  * The static key to set up a KFENCE allocation; or if static keys are not used
141  * to gate allocations, to avoid a load and compare if KFENCE is disabled.
142  */
143 DEFINE_STATIC_KEY_FALSE(kfence_allocation_key);
144 
145 /* Gates the allocation, ensuring only one succeeds in a given period. */
146 atomic_t kfence_allocation_gate = ATOMIC_INIT(1);
147 
148 /*
149  * A Counting Bloom filter of allocation coverage: limits currently covered
150  * allocations of the same source filling up the pool.
151  *
152  * Assuming a range of 15%-85% unique allocations in the pool at any point in
153  * time, the below parameters provide a probablity of 0.02-0.33 for false
154  * positive hits respectively:
155  *
156  *	P(alloc_traces) = (1 - e^(-HNUM * (alloc_traces / SIZE)) ^ HNUM
157  */
158 #define ALLOC_COVERED_HNUM	2
159 #define ALLOC_COVERED_ORDER	(const_ilog2(CONFIG_KFENCE_NUM_OBJECTS) + 2)
160 #define ALLOC_COVERED_SIZE	(1 << ALLOC_COVERED_ORDER)
161 #define ALLOC_COVERED_HNEXT(h)	hash_32(h, ALLOC_COVERED_ORDER)
162 #define ALLOC_COVERED_MASK	(ALLOC_COVERED_SIZE - 1)
163 static atomic_t alloc_covered[ALLOC_COVERED_SIZE];
164 
165 /* Stack depth used to determine uniqueness of an allocation. */
166 #define UNIQUE_ALLOC_STACK_DEPTH ((size_t)8)
167 
168 /*
169  * Randomness for stack hashes, making the same collisions across reboots and
170  * different machines less likely.
171  */
172 static u32 stack_hash_seed __ro_after_init;
173 
174 /* Statistics counters for debugfs. */
175 enum kfence_counter_id {
176 	KFENCE_COUNTER_ALLOCATED,
177 	KFENCE_COUNTER_ALLOCS,
178 	KFENCE_COUNTER_FREES,
179 	KFENCE_COUNTER_ZOMBIES,
180 	KFENCE_COUNTER_BUGS,
181 	KFENCE_COUNTER_SKIP_INCOMPAT,
182 	KFENCE_COUNTER_SKIP_CAPACITY,
183 	KFENCE_COUNTER_SKIP_COVERED,
184 	KFENCE_COUNTER_COUNT,
185 };
186 static atomic_long_t counters[KFENCE_COUNTER_COUNT];
187 static const char *const counter_names[] = {
188 	[KFENCE_COUNTER_ALLOCATED]	= "currently allocated",
189 	[KFENCE_COUNTER_ALLOCS]		= "total allocations",
190 	[KFENCE_COUNTER_FREES]		= "total frees",
191 	[KFENCE_COUNTER_ZOMBIES]	= "zombie allocations",
192 	[KFENCE_COUNTER_BUGS]		= "total bugs",
193 	[KFENCE_COUNTER_SKIP_INCOMPAT]	= "skipped allocations (incompatible)",
194 	[KFENCE_COUNTER_SKIP_CAPACITY]	= "skipped allocations (capacity)",
195 	[KFENCE_COUNTER_SKIP_COVERED]	= "skipped allocations (covered)",
196 };
197 static_assert(ARRAY_SIZE(counter_names) == KFENCE_COUNTER_COUNT);
198 
199 /* === Internals ============================================================ */
200 
should_skip_covered(void)201 static inline bool should_skip_covered(void)
202 {
203 	unsigned long thresh = (CONFIG_KFENCE_NUM_OBJECTS * kfence_skip_covered_thresh) / 100;
204 
205 	return atomic_long_read(&counters[KFENCE_COUNTER_ALLOCATED]) > thresh;
206 }
207 
get_alloc_stack_hash(unsigned long * stack_entries,size_t num_entries)208 static u32 get_alloc_stack_hash(unsigned long *stack_entries, size_t num_entries)
209 {
210 	num_entries = min(num_entries, UNIQUE_ALLOC_STACK_DEPTH);
211 	num_entries = filter_irq_stacks(stack_entries, num_entries);
212 	return jhash(stack_entries, num_entries * sizeof(stack_entries[0]), stack_hash_seed);
213 }
214 
215 /*
216  * Adds (or subtracts) count @val for allocation stack trace hash
217  * @alloc_stack_hash from Counting Bloom filter.
218  */
alloc_covered_add(u32 alloc_stack_hash,int val)219 static void alloc_covered_add(u32 alloc_stack_hash, int val)
220 {
221 	int i;
222 
223 	for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
224 		atomic_add(val, &alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]);
225 		alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
226 	}
227 }
228 
229 /*
230  * Returns true if the allocation stack trace hash @alloc_stack_hash is
231  * currently contained (non-zero count) in Counting Bloom filter.
232  */
alloc_covered_contains(u32 alloc_stack_hash)233 static bool alloc_covered_contains(u32 alloc_stack_hash)
234 {
235 	int i;
236 
237 	for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
238 		if (!atomic_read(&alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]))
239 			return false;
240 		alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
241 	}
242 
243 	return true;
244 }
245 
kfence_protect(unsigned long addr)246 static bool kfence_protect(unsigned long addr)
247 {
248 	return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), true));
249 }
250 
kfence_unprotect(unsigned long addr)251 static bool kfence_unprotect(unsigned long addr)
252 {
253 	return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), false));
254 }
255 
metadata_to_pageaddr(const struct kfence_metadata * meta)256 static inline unsigned long metadata_to_pageaddr(const struct kfence_metadata *meta)
257 {
258 	unsigned long offset = (meta - kfence_metadata + 1) * PAGE_SIZE * 2;
259 	unsigned long pageaddr = (unsigned long)&__kfence_pool[offset];
260 
261 	/* The checks do not affect performance; only called from slow-paths. */
262 
263 	/* Only call with a pointer into kfence_metadata. */
264 	if (KFENCE_WARN_ON(meta < kfence_metadata ||
265 			   meta >= kfence_metadata + CONFIG_KFENCE_NUM_OBJECTS))
266 		return 0;
267 
268 	/*
269 	 * This metadata object only ever maps to 1 page; verify that the stored
270 	 * address is in the expected range.
271 	 */
272 	if (KFENCE_WARN_ON(ALIGN_DOWN(meta->addr, PAGE_SIZE) != pageaddr))
273 		return 0;
274 
275 	return pageaddr;
276 }
277 
kfence_obj_allocated(const struct kfence_metadata * meta)278 static inline bool kfence_obj_allocated(const struct kfence_metadata *meta)
279 {
280 	enum kfence_object_state state = READ_ONCE(meta->state);
281 
282 	return state == KFENCE_OBJECT_ALLOCATED || state == KFENCE_OBJECT_RCU_FREEING;
283 }
284 
285 /*
286  * Update the object's metadata state, including updating the alloc/free stacks
287  * depending on the state transition.
288  */
289 static noinline void
metadata_update_state(struct kfence_metadata * meta,enum kfence_object_state next,unsigned long * stack_entries,size_t num_stack_entries)290 metadata_update_state(struct kfence_metadata *meta, enum kfence_object_state next,
291 		      unsigned long *stack_entries, size_t num_stack_entries)
292 {
293 	struct kfence_track *track =
294 		next == KFENCE_OBJECT_ALLOCATED ? &meta->alloc_track : &meta->free_track;
295 
296 	lockdep_assert_held(&meta->lock);
297 
298 	/* Stack has been saved when calling rcu, skip. */
299 	if (READ_ONCE(meta->state) == KFENCE_OBJECT_RCU_FREEING)
300 		goto out;
301 
302 	if (stack_entries) {
303 		memcpy(track->stack_entries, stack_entries,
304 		       num_stack_entries * sizeof(stack_entries[0]));
305 	} else {
306 		/*
307 		 * Skip over 1 (this) functions; noinline ensures we do not
308 		 * accidentally skip over the caller by never inlining.
309 		 */
310 		num_stack_entries = stack_trace_save(track->stack_entries, KFENCE_STACK_DEPTH, 1);
311 	}
312 	track->num_stack_entries = num_stack_entries;
313 	track->pid = task_pid_nr(current);
314 	track->cpu = raw_smp_processor_id();
315 	track->ts_nsec = local_clock(); /* Same source as printk timestamps. */
316 
317 out:
318 	/*
319 	 * Pairs with READ_ONCE() in
320 	 *	kfence_shutdown_cache(),
321 	 *	kfence_handle_page_fault().
322 	 */
323 	WRITE_ONCE(meta->state, next);
324 }
325 
326 #ifdef CONFIG_KMSAN
327 #define check_canary_attributes noinline __no_kmsan_checks
328 #else
329 #define check_canary_attributes inline
330 #endif
331 
332 /* Check canary byte at @addr. */
check_canary_byte(u8 * addr)333 static check_canary_attributes bool check_canary_byte(u8 *addr)
334 {
335 	struct kfence_metadata *meta;
336 	unsigned long flags;
337 
338 	if (likely(*addr == KFENCE_CANARY_PATTERN_U8(addr)))
339 		return true;
340 
341 	atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
342 
343 	meta = addr_to_metadata((unsigned long)addr);
344 	raw_spin_lock_irqsave(&meta->lock, flags);
345 	kfence_report_error((unsigned long)addr, false, NULL, meta, KFENCE_ERROR_CORRUPTION);
346 	raw_spin_unlock_irqrestore(&meta->lock, flags);
347 
348 	return false;
349 }
350 
set_canary(const struct kfence_metadata * meta)351 static inline void set_canary(const struct kfence_metadata *meta)
352 {
353 	const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
354 	unsigned long addr = pageaddr;
355 
356 	/*
357 	 * The canary may be written to part of the object memory, but it does
358 	 * not affect it. The user should initialize the object before using it.
359 	 */
360 	for (; addr < meta->addr; addr += sizeof(u64))
361 		*((u64 *)addr) = KFENCE_CANARY_PATTERN_U64;
362 
363 	addr = ALIGN_DOWN(meta->addr + meta->size, sizeof(u64));
364 	for (; addr - pageaddr < PAGE_SIZE; addr += sizeof(u64))
365 		*((u64 *)addr) = KFENCE_CANARY_PATTERN_U64;
366 }
367 
368 static check_canary_attributes void
check_canary(const struct kfence_metadata * meta)369 check_canary(const struct kfence_metadata *meta)
370 {
371 	const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
372 	unsigned long addr = pageaddr;
373 
374 	/*
375 	 * We'll iterate over each canary byte per-side until a corrupted byte
376 	 * is found. However, we'll still iterate over the canary bytes to the
377 	 * right of the object even if there was an error in the canary bytes to
378 	 * the left of the object. Specifically, if check_canary_byte()
379 	 * generates an error, showing both sides might give more clues as to
380 	 * what the error is about when displaying which bytes were corrupted.
381 	 */
382 
383 	/* Apply to left of object. */
384 	for (; meta->addr - addr >= sizeof(u64); addr += sizeof(u64)) {
385 		if (unlikely(*((u64 *)addr) != KFENCE_CANARY_PATTERN_U64))
386 			break;
387 	}
388 
389 	/*
390 	 * If the canary is corrupted in a certain 64 bytes, or the canary
391 	 * memory cannot be completely covered by multiple consecutive 64 bytes,
392 	 * it needs to be checked one by one.
393 	 */
394 	for (; addr < meta->addr; addr++) {
395 		if (unlikely(!check_canary_byte((u8 *)addr)))
396 			break;
397 	}
398 
399 	/* Apply to right of object. */
400 	for (addr = meta->addr + meta->size; addr % sizeof(u64) != 0; addr++) {
401 		if (unlikely(!check_canary_byte((u8 *)addr)))
402 			return;
403 	}
404 	for (; addr - pageaddr < PAGE_SIZE; addr += sizeof(u64)) {
405 		if (unlikely(*((u64 *)addr) != KFENCE_CANARY_PATTERN_U64)) {
406 
407 			for (; addr - pageaddr < PAGE_SIZE; addr++) {
408 				if (!check_canary_byte((u8 *)addr))
409 					return;
410 			}
411 		}
412 	}
413 }
414 
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)415 static void *kfence_guarded_alloc(struct kmem_cache *cache, size_t size, gfp_t gfp,
416 				  unsigned long *stack_entries, size_t num_stack_entries,
417 				  u32 alloc_stack_hash)
418 {
419 	struct kfence_metadata *meta = NULL;
420 	unsigned long flags;
421 	struct slab *slab;
422 	void *addr;
423 	const bool random_right_allocate = get_random_u32_below(2);
424 	const bool random_fault = CONFIG_KFENCE_STRESS_TEST_FAULTS &&
425 				  !get_random_u32_below(CONFIG_KFENCE_STRESS_TEST_FAULTS);
426 
427 	/* Try to obtain a free object. */
428 	raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
429 	if (!list_empty(&kfence_freelist)) {
430 		meta = list_entry(kfence_freelist.next, struct kfence_metadata, list);
431 		list_del_init(&meta->list);
432 	}
433 	raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
434 	if (!meta) {
435 		atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_CAPACITY]);
436 		return NULL;
437 	}
438 
439 	if (unlikely(!raw_spin_trylock_irqsave(&meta->lock, flags))) {
440 		/*
441 		 * This is extremely unlikely -- we are reporting on a
442 		 * use-after-free, which locked meta->lock, and the reporting
443 		 * code via printk calls kmalloc() which ends up in
444 		 * kfence_alloc() and tries to grab the same object that we're
445 		 * reporting on. While it has never been observed, lockdep does
446 		 * report that there is a possibility of deadlock. Fix it by
447 		 * using trylock and bailing out gracefully.
448 		 */
449 		raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
450 		/* Put the object back on the freelist. */
451 		list_add_tail(&meta->list, &kfence_freelist);
452 		raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
453 
454 		return NULL;
455 	}
456 
457 	meta->addr = metadata_to_pageaddr(meta);
458 	/* Unprotect if we're reusing this page. */
459 	if (meta->state == KFENCE_OBJECT_FREED)
460 		kfence_unprotect(meta->addr);
461 
462 	/*
463 	 * Note: for allocations made before RNG initialization, will always
464 	 * return zero. We still benefit from enabling KFENCE as early as
465 	 * possible, even when the RNG is not yet available, as this will allow
466 	 * KFENCE to detect bugs due to earlier allocations. The only downside
467 	 * is that the out-of-bounds accesses detected are deterministic for
468 	 * such allocations.
469 	 */
470 	if (random_right_allocate) {
471 		/* Allocate on the "right" side, re-calculate address. */
472 		meta->addr += PAGE_SIZE - size;
473 		meta->addr = ALIGN_DOWN(meta->addr, cache->align);
474 	}
475 
476 	addr = (void *)meta->addr;
477 
478 	/* Update remaining metadata. */
479 	metadata_update_state(meta, KFENCE_OBJECT_ALLOCATED, stack_entries, num_stack_entries);
480 	/* Pairs with READ_ONCE() in kfence_shutdown_cache(). */
481 	WRITE_ONCE(meta->cache, cache);
482 	meta->size = size;
483 	meta->alloc_stack_hash = alloc_stack_hash;
484 	raw_spin_unlock_irqrestore(&meta->lock, flags);
485 
486 	alloc_covered_add(alloc_stack_hash, 1);
487 
488 	/* Set required slab fields. */
489 	slab = virt_to_slab((void *)meta->addr);
490 	slab->slab_cache = cache;
491 	slab->objects = 1;
492 
493 	/* Memory initialization. */
494 	set_canary(meta);
495 
496 	/*
497 	 * We check slab_want_init_on_alloc() ourselves, rather than letting
498 	 * SL*B do the initialization, as otherwise we might overwrite KFENCE's
499 	 * redzone.
500 	 */
501 	if (unlikely(slab_want_init_on_alloc(gfp, cache)))
502 		memzero_explicit(addr, size);
503 	if (cache->ctor)
504 		cache->ctor(addr);
505 
506 	if (random_fault)
507 		kfence_protect(meta->addr); /* Random "faults" by protecting the object. */
508 
509 	atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCATED]);
510 	atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCS]);
511 
512 	return addr;
513 }
514 
kfence_guarded_free(void * addr,struct kfence_metadata * meta,bool zombie)515 static void kfence_guarded_free(void *addr, struct kfence_metadata *meta, bool zombie)
516 {
517 	struct kcsan_scoped_access assert_page_exclusive;
518 	unsigned long flags;
519 	bool init;
520 
521 	raw_spin_lock_irqsave(&meta->lock, flags);
522 
523 	if (!kfence_obj_allocated(meta) || meta->addr != (unsigned long)addr) {
524 		/* Invalid or double-free, bail out. */
525 		atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
526 		kfence_report_error((unsigned long)addr, false, NULL, meta,
527 				    KFENCE_ERROR_INVALID_FREE);
528 		raw_spin_unlock_irqrestore(&meta->lock, flags);
529 		return;
530 	}
531 
532 	/* Detect racy use-after-free, or incorrect reallocation of this page by KFENCE. */
533 	kcsan_begin_scoped_access((void *)ALIGN_DOWN((unsigned long)addr, PAGE_SIZE), PAGE_SIZE,
534 				  KCSAN_ACCESS_SCOPED | KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT,
535 				  &assert_page_exclusive);
536 
537 	if (CONFIG_KFENCE_STRESS_TEST_FAULTS)
538 		kfence_unprotect((unsigned long)addr); /* To check canary bytes. */
539 
540 	/* Restore page protection if there was an OOB access. */
541 	if (meta->unprotected_page) {
542 		memzero_explicit((void *)ALIGN_DOWN(meta->unprotected_page, PAGE_SIZE), PAGE_SIZE);
543 		kfence_protect(meta->unprotected_page);
544 		meta->unprotected_page = 0;
545 	}
546 
547 	/* Mark the object as freed. */
548 	metadata_update_state(meta, KFENCE_OBJECT_FREED, NULL, 0);
549 	init = slab_want_init_on_free(meta->cache);
550 	raw_spin_unlock_irqrestore(&meta->lock, flags);
551 
552 	alloc_covered_add(meta->alloc_stack_hash, -1);
553 
554 	/* Check canary bytes for memory corruption. */
555 	check_canary(meta);
556 
557 	/*
558 	 * Clear memory if init-on-free is set. While we protect the page, the
559 	 * data is still there, and after a use-after-free is detected, we
560 	 * unprotect the page, so the data is still accessible.
561 	 */
562 	if (!zombie && unlikely(init))
563 		memzero_explicit(addr, meta->size);
564 
565 	/* Protect to detect use-after-frees. */
566 	kfence_protect((unsigned long)addr);
567 
568 	kcsan_end_scoped_access(&assert_page_exclusive);
569 	if (!zombie) {
570 		/* Add it to the tail of the freelist for reuse. */
571 		raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
572 		KFENCE_WARN_ON(!list_empty(&meta->list));
573 		list_add_tail(&meta->list, &kfence_freelist);
574 		raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
575 
576 		atomic_long_dec(&counters[KFENCE_COUNTER_ALLOCATED]);
577 		atomic_long_inc(&counters[KFENCE_COUNTER_FREES]);
578 	} else {
579 		/* See kfence_shutdown_cache(). */
580 		atomic_long_inc(&counters[KFENCE_COUNTER_ZOMBIES]);
581 	}
582 }
583 
rcu_guarded_free(struct rcu_head * h)584 static void rcu_guarded_free(struct rcu_head *h)
585 {
586 	struct kfence_metadata *meta = container_of(h, struct kfence_metadata, rcu_head);
587 
588 	kfence_guarded_free((void *)meta->addr, meta, false);
589 }
590 
591 /*
592  * Initialization of the KFENCE pool after its allocation.
593  * Returns 0 on success; otherwise returns the address up to
594  * which partial initialization succeeded.
595  */
kfence_init_pool(void)596 static unsigned long kfence_init_pool(void)
597 {
598 	unsigned long addr, start_pfn;
599 	int i, rand;
600 
601 	if (!arch_kfence_init_pool())
602 		return (unsigned long)__kfence_pool;
603 
604 	addr = (unsigned long)__kfence_pool;
605 	start_pfn = PHYS_PFN(virt_to_phys(__kfence_pool));
606 
607 	/*
608 	 * Set up object pages: they must have PGTY_slab set to avoid freeing
609 	 * them as real pages.
610 	 *
611 	 * We also want to avoid inserting kfence_free() in the kfree()
612 	 * fast-path in SLUB, and therefore need to ensure kfree() correctly
613 	 * enters __slab_free() slow-path.
614 	 */
615 	for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
616 		struct page *page;
617 
618 		if (!i || (i % 2))
619 			continue;
620 
621 		page = pfn_to_page(start_pfn + i);
622 		__SetPageSlab(page);
623 #ifdef CONFIG_MEMCG
624 		struct slab *slab = page_slab(page);
625 		slab->obj_exts = (unsigned long)&kfence_metadata_init[i / 2 - 1].obj_exts |
626 				 MEMCG_DATA_OBJEXTS;
627 #endif
628 	}
629 
630 	/*
631 	 * Protect the first 2 pages. The first page is mostly unnecessary, and
632 	 * merely serves as an extended guard page. However, adding one
633 	 * additional page in the beginning gives us an even number of pages,
634 	 * which simplifies the mapping of address to metadata index.
635 	 */
636 	for (i = 0; i < 2; i++) {
637 		if (unlikely(!kfence_protect(addr)))
638 			return addr;
639 
640 		addr += PAGE_SIZE;
641 	}
642 
643 	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
644 		struct kfence_metadata *meta = &kfence_metadata_init[i];
645 
646 		/* Initialize metadata. */
647 		INIT_LIST_HEAD(&meta->list);
648 		raw_spin_lock_init(&meta->lock);
649 		meta->state = KFENCE_OBJECT_UNUSED;
650 		/* Use addr to randomize the freelist. */
651 		meta->addr = i;
652 
653 		/* Protect the right redzone. */
654 		if (unlikely(!kfence_protect(addr + 2 * i * PAGE_SIZE + PAGE_SIZE)))
655 			goto reset_slab;
656 	}
657 
658 	for (i = CONFIG_KFENCE_NUM_OBJECTS; i > 0; i--) {
659 		rand = get_random_u32_below(i);
660 		swap(kfence_metadata_init[i - 1].addr, kfence_metadata_init[rand].addr);
661 	}
662 
663 	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
664 		struct kfence_metadata *meta_1 = &kfence_metadata_init[i];
665 		struct kfence_metadata *meta_2 = &kfence_metadata_init[meta_1->addr];
666 
667 		list_add_tail(&meta_2->list, &kfence_freelist);
668 	}
669 	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
670 		kfence_metadata_init[i].addr = addr;
671 		addr += 2 * PAGE_SIZE;
672 	}
673 
674 	/*
675 	 * Make kfence_metadata visible only when initialization is successful.
676 	 * Otherwise, if the initialization fails and kfence_metadata is freed,
677 	 * it may cause UAF in kfence_shutdown_cache().
678 	 */
679 	smp_store_release(&kfence_metadata, kfence_metadata_init);
680 	return 0;
681 
682 reset_slab:
683 	addr += 2 * i * PAGE_SIZE;
684 	for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
685 		struct page *page;
686 
687 		if (!i || (i % 2))
688 			continue;
689 
690 		page = pfn_to_page(start_pfn + i);
691 #ifdef CONFIG_MEMCG
692 		struct slab *slab = page_slab(page);
693 		slab->obj_exts = 0;
694 #endif
695 		__ClearPageSlab(page);
696 	}
697 
698 	return addr;
699 }
700 
kfence_init_pool_early(void)701 static bool __init kfence_init_pool_early(void)
702 {
703 	unsigned long addr;
704 
705 	if (!__kfence_pool)
706 		return false;
707 
708 	addr = kfence_init_pool();
709 
710 	if (!addr) {
711 		/*
712 		 * The pool is live and will never be deallocated from this point on.
713 		 * Ignore the pool object from the kmemleak phys object tree, as it would
714 		 * otherwise overlap with allocations returned by kfence_alloc(), which
715 		 * are registered with kmemleak through the slab post-alloc hook.
716 		 */
717 		kmemleak_ignore_phys(__pa(__kfence_pool));
718 		return true;
719 	}
720 
721 	/*
722 	 * Only release unprotected pages, and do not try to go back and change
723 	 * page attributes due to risk of failing to do so as well. If changing
724 	 * page attributes for some pages fails, it is very likely that it also
725 	 * fails for the first page, and therefore expect addr==__kfence_pool in
726 	 * most failure cases.
727 	 */
728 	memblock_free_late(__pa(addr), KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool));
729 	__kfence_pool = NULL;
730 
731 	memblock_free_late(__pa(kfence_metadata_init), KFENCE_METADATA_SIZE);
732 	kfence_metadata_init = NULL;
733 
734 	return false;
735 }
736 
737 /* === DebugFS Interface ==================================================== */
738 
stats_show(struct seq_file * seq,void * v)739 static int stats_show(struct seq_file *seq, void *v)
740 {
741 	int i;
742 
743 	seq_printf(seq, "enabled: %i\n", READ_ONCE(kfence_enabled));
744 	for (i = 0; i < KFENCE_COUNTER_COUNT; i++)
745 		seq_printf(seq, "%s: %ld\n", counter_names[i], atomic_long_read(&counters[i]));
746 
747 	return 0;
748 }
749 DEFINE_SHOW_ATTRIBUTE(stats);
750 
751 /*
752  * debugfs seq_file operations for /sys/kernel/debug/kfence/objects.
753  * start_object() and next_object() return the object index + 1, because NULL is used
754  * to stop iteration.
755  */
start_object(struct seq_file * seq,loff_t * pos)756 static void *start_object(struct seq_file *seq, loff_t *pos)
757 {
758 	if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
759 		return (void *)((long)*pos + 1);
760 	return NULL;
761 }
762 
stop_object(struct seq_file * seq,void * v)763 static void stop_object(struct seq_file *seq, void *v)
764 {
765 }
766 
next_object(struct seq_file * seq,void * v,loff_t * pos)767 static void *next_object(struct seq_file *seq, void *v, loff_t *pos)
768 {
769 	++*pos;
770 	if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
771 		return (void *)((long)*pos + 1);
772 	return NULL;
773 }
774 
show_object(struct seq_file * seq,void * v)775 static int show_object(struct seq_file *seq, void *v)
776 {
777 	struct kfence_metadata *meta = &kfence_metadata[(long)v - 1];
778 	unsigned long flags;
779 
780 	raw_spin_lock_irqsave(&meta->lock, flags);
781 	kfence_print_object(seq, meta);
782 	raw_spin_unlock_irqrestore(&meta->lock, flags);
783 	seq_puts(seq, "---------------------------------\n");
784 
785 	return 0;
786 }
787 
788 static const struct seq_operations objects_sops = {
789 	.start = start_object,
790 	.next = next_object,
791 	.stop = stop_object,
792 	.show = show_object,
793 };
794 DEFINE_SEQ_ATTRIBUTE(objects);
795 
kfence_debugfs_init(void)796 static int kfence_debugfs_init(void)
797 {
798 	struct dentry *kfence_dir;
799 
800 	if (!READ_ONCE(kfence_enabled))
801 		return 0;
802 
803 	kfence_dir = debugfs_create_dir("kfence", NULL);
804 	debugfs_create_file("stats", 0444, kfence_dir, NULL, &stats_fops);
805 	debugfs_create_file("objects", 0400, kfence_dir, NULL, &objects_fops);
806 	return 0;
807 }
808 
809 late_initcall(kfence_debugfs_init);
810 
811 /* === Panic Notifier ====================================================== */
812 
kfence_check_all_canary(void)813 static void kfence_check_all_canary(void)
814 {
815 	int i;
816 
817 	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
818 		struct kfence_metadata *meta = &kfence_metadata[i];
819 
820 		if (kfence_obj_allocated(meta))
821 			check_canary(meta);
822 	}
823 }
824 
kfence_check_canary_callback(struct notifier_block * nb,unsigned long reason,void * arg)825 static int kfence_check_canary_callback(struct notifier_block *nb,
826 					unsigned long reason, void *arg)
827 {
828 	kfence_check_all_canary();
829 	return NOTIFY_OK;
830 }
831 
832 static struct notifier_block kfence_check_canary_notifier = {
833 	.notifier_call = kfence_check_canary_callback,
834 };
835 
836 /* === Allocation Gate Timer ================================================ */
837 
838 static struct delayed_work kfence_timer;
839 
840 #ifdef CONFIG_KFENCE_STATIC_KEYS
841 /* Wait queue to wake up allocation-gate timer task. */
842 static DECLARE_WAIT_QUEUE_HEAD(allocation_wait);
843 
kfence_reboot_callback(struct notifier_block * nb,unsigned long action,void * data)844 static int kfence_reboot_callback(struct notifier_block *nb,
845 				  unsigned long action, void *data)
846 {
847 	/*
848 	 * Disable kfence to avoid static keys IPI synchronization during
849 	 * late shutdown/kexec
850 	 */
851 	WRITE_ONCE(kfence_enabled, false);
852 	/* Cancel any pending timer work */
853 	cancel_delayed_work(&kfence_timer);
854 	/*
855 	 * Wake up any blocked toggle_allocation_gate() so it can complete
856 	 * early while the system is still able to handle IPIs.
857 	 */
858 	wake_up(&allocation_wait);
859 
860 	return NOTIFY_OK;
861 }
862 
863 static struct notifier_block kfence_reboot_notifier = {
864 	.notifier_call = kfence_reboot_callback,
865 	.priority = INT_MAX, /* Run early to stop timers ASAP */
866 };
867 
wake_up_kfence_timer(struct irq_work * work)868 static void wake_up_kfence_timer(struct irq_work *work)
869 {
870 	wake_up(&allocation_wait);
871 }
872 static DEFINE_IRQ_WORK(wake_up_kfence_timer_work, wake_up_kfence_timer);
873 #endif
874 
875 /*
876  * Set up delayed work, which will enable and disable the static key. We need to
877  * use a work queue (rather than a simple timer), since enabling and disabling a
878  * static key cannot be done from an interrupt.
879  *
880  * Note: Toggling a static branch currently causes IPIs, and here we'll end up
881  * with a total of 2 IPIs to all CPUs. If this ends up a problem in future (with
882  * more aggressive sampling intervals), we could get away with a variant that
883  * avoids IPIs, at the cost of not immediately capturing allocations if the
884  * instructions remain cached.
885  */
toggle_allocation_gate(struct work_struct * work)886 static void toggle_allocation_gate(struct work_struct *work)
887 {
888 	if (!READ_ONCE(kfence_enabled))
889 		return;
890 
891 	atomic_set(&kfence_allocation_gate, -kfence_burst);
892 #ifdef CONFIG_KFENCE_STATIC_KEYS
893 	/* Enable static key, and await allocation to happen. */
894 	static_branch_enable(&kfence_allocation_key);
895 
896 	wait_event_idle(allocation_wait,
897 			atomic_read(&kfence_allocation_gate) > 0 ||
898 			!READ_ONCE(kfence_enabled));
899 
900 	/* Disable static key and reset timer. */
901 	static_branch_disable(&kfence_allocation_key);
902 #endif
903 	queue_delayed_work(system_unbound_wq, &kfence_timer,
904 			   msecs_to_jiffies(kfence_sample_interval));
905 }
906 
907 /* === Public interface ===================================================== */
908 
kfence_alloc_pool_and_metadata(void)909 void __init kfence_alloc_pool_and_metadata(void)
910 {
911 	if (!kfence_sample_interval)
912 		return;
913 
914 	/*
915 	 * If the pool has already been initialized by arch, there is no need to
916 	 * re-allocate the memory pool.
917 	 */
918 	if (!__kfence_pool)
919 		__kfence_pool = memblock_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
920 
921 	if (!__kfence_pool) {
922 		pr_err("failed to allocate pool\n");
923 		return;
924 	}
925 
926 	/* The memory allocated by memblock has been zeroed out. */
927 	kfence_metadata_init = memblock_alloc(KFENCE_METADATA_SIZE, PAGE_SIZE);
928 	if (!kfence_metadata_init) {
929 		pr_err("failed to allocate metadata\n");
930 		memblock_free(__kfence_pool, KFENCE_POOL_SIZE);
931 		__kfence_pool = NULL;
932 	}
933 }
934 
kfence_init_enable(void)935 static void kfence_init_enable(void)
936 {
937 	if (!IS_ENABLED(CONFIG_KFENCE_STATIC_KEYS))
938 		static_branch_enable(&kfence_allocation_key);
939 
940 	if (kfence_deferrable)
941 		INIT_DEFERRABLE_WORK(&kfence_timer, toggle_allocation_gate);
942 	else
943 		INIT_DELAYED_WORK(&kfence_timer, toggle_allocation_gate);
944 
945 	if (kfence_check_on_panic)
946 		atomic_notifier_chain_register(&panic_notifier_list, &kfence_check_canary_notifier);
947 
948 #ifdef CONFIG_KFENCE_STATIC_KEYS
949 	register_reboot_notifier(&kfence_reboot_notifier);
950 #endif
951 
952 	WRITE_ONCE(kfence_enabled, true);
953 	queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
954 
955 	pr_info("initialized - using %lu bytes for %d objects at 0x%p-0x%p\n", KFENCE_POOL_SIZE,
956 		CONFIG_KFENCE_NUM_OBJECTS, (void *)__kfence_pool,
957 		(void *)(__kfence_pool + KFENCE_POOL_SIZE));
958 }
959 
kfence_init(void)960 void __init kfence_init(void)
961 {
962 	stack_hash_seed = get_random_u32();
963 
964 	/* Setting kfence_sample_interval to 0 on boot disables KFENCE. */
965 	if (!kfence_sample_interval)
966 		return;
967 
968 	if (!kfence_init_pool_early()) {
969 		pr_err("%s failed\n", __func__);
970 		return;
971 	}
972 
973 	kfence_init_enable();
974 }
975 
kfence_init_late(void)976 static int kfence_init_late(void)
977 {
978 	const unsigned long nr_pages_pool = KFENCE_POOL_SIZE / PAGE_SIZE;
979 	const unsigned long nr_pages_meta = KFENCE_METADATA_SIZE / PAGE_SIZE;
980 	unsigned long addr = (unsigned long)__kfence_pool;
981 	unsigned long free_size = KFENCE_POOL_SIZE;
982 	int err = -ENOMEM;
983 
984 #ifdef CONFIG_CONTIG_ALLOC
985 	struct page *pages;
986 
987 	pages = alloc_contig_pages(nr_pages_pool, GFP_KERNEL, first_online_node,
988 				   NULL);
989 	if (!pages)
990 		return -ENOMEM;
991 
992 	__kfence_pool = page_to_virt(pages);
993 	pages = alloc_contig_pages(nr_pages_meta, GFP_KERNEL, first_online_node,
994 				   NULL);
995 	if (pages)
996 		kfence_metadata_init = page_to_virt(pages);
997 #else
998 	if (nr_pages_pool > MAX_ORDER_NR_PAGES ||
999 	    nr_pages_meta > MAX_ORDER_NR_PAGES) {
1000 		pr_warn("KFENCE_NUM_OBJECTS too large for buddy allocator\n");
1001 		return -EINVAL;
1002 	}
1003 
1004 	__kfence_pool = alloc_pages_exact(KFENCE_POOL_SIZE, GFP_KERNEL);
1005 	if (!__kfence_pool)
1006 		return -ENOMEM;
1007 
1008 	kfence_metadata_init = alloc_pages_exact(KFENCE_METADATA_SIZE, GFP_KERNEL);
1009 #endif
1010 
1011 	if (!kfence_metadata_init)
1012 		goto free_pool;
1013 
1014 	memzero_explicit(kfence_metadata_init, KFENCE_METADATA_SIZE);
1015 	addr = kfence_init_pool();
1016 	if (!addr) {
1017 		kfence_init_enable();
1018 		kfence_debugfs_init();
1019 		return 0;
1020 	}
1021 
1022 	pr_err("%s failed\n", __func__);
1023 	free_size = KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool);
1024 	err = -EBUSY;
1025 
1026 #ifdef CONFIG_CONTIG_ALLOC
1027 	free_contig_range(page_to_pfn(virt_to_page((void *)kfence_metadata_init)),
1028 			  nr_pages_meta);
1029 free_pool:
1030 	free_contig_range(page_to_pfn(virt_to_page((void *)addr)),
1031 			  free_size / PAGE_SIZE);
1032 #else
1033 	free_pages_exact((void *)kfence_metadata_init, KFENCE_METADATA_SIZE);
1034 free_pool:
1035 	free_pages_exact((void *)addr, free_size);
1036 #endif
1037 
1038 	kfence_metadata_init = NULL;
1039 	__kfence_pool = NULL;
1040 	return err;
1041 }
1042 
kfence_enable_late(void)1043 static int kfence_enable_late(void)
1044 {
1045 	if (!__kfence_pool)
1046 		return kfence_init_late();
1047 
1048 	WRITE_ONCE(kfence_enabled, true);
1049 	queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
1050 	pr_info("re-enabled\n");
1051 	return 0;
1052 }
1053 
kfence_shutdown_cache(struct kmem_cache * s)1054 void kfence_shutdown_cache(struct kmem_cache *s)
1055 {
1056 	unsigned long flags;
1057 	struct kfence_metadata *meta;
1058 	int i;
1059 
1060 	/* Pairs with release in kfence_init_pool(). */
1061 	if (!smp_load_acquire(&kfence_metadata))
1062 		return;
1063 
1064 	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
1065 		bool in_use;
1066 
1067 		meta = &kfence_metadata[i];
1068 
1069 		/*
1070 		 * If we observe some inconsistent cache and state pair where we
1071 		 * should have returned false here, cache destruction is racing
1072 		 * with either kmem_cache_alloc() or kmem_cache_free(). Taking
1073 		 * the lock will not help, as different critical section
1074 		 * serialization will have the same outcome.
1075 		 */
1076 		if (READ_ONCE(meta->cache) != s || !kfence_obj_allocated(meta))
1077 			continue;
1078 
1079 		raw_spin_lock_irqsave(&meta->lock, flags);
1080 		in_use = meta->cache == s && kfence_obj_allocated(meta);
1081 		raw_spin_unlock_irqrestore(&meta->lock, flags);
1082 
1083 		if (in_use) {
1084 			/*
1085 			 * This cache still has allocations, and we should not
1086 			 * release them back into the freelist so they can still
1087 			 * safely be used and retain the kernel's default
1088 			 * behaviour of keeping the allocations alive (leak the
1089 			 * cache); however, they effectively become "zombie
1090 			 * allocations" as the KFENCE objects are the only ones
1091 			 * still in use and the owning cache is being destroyed.
1092 			 *
1093 			 * We mark them freed, so that any subsequent use shows
1094 			 * more useful error messages that will include stack
1095 			 * traces of the user of the object, the original
1096 			 * allocation, and caller to shutdown_cache().
1097 			 */
1098 			kfence_guarded_free((void *)meta->addr, meta, /*zombie=*/true);
1099 		}
1100 	}
1101 
1102 	for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
1103 		meta = &kfence_metadata[i];
1104 
1105 		/* See above. */
1106 		if (READ_ONCE(meta->cache) != s || READ_ONCE(meta->state) != KFENCE_OBJECT_FREED)
1107 			continue;
1108 
1109 		raw_spin_lock_irqsave(&meta->lock, flags);
1110 		if (meta->cache == s && meta->state == KFENCE_OBJECT_FREED)
1111 			meta->cache = NULL;
1112 		raw_spin_unlock_irqrestore(&meta->lock, flags);
1113 	}
1114 }
1115 
__kfence_alloc(struct kmem_cache * s,size_t size,gfp_t flags)1116 void *__kfence_alloc(struct kmem_cache *s, size_t size, gfp_t flags)
1117 {
1118 	unsigned long stack_entries[KFENCE_STACK_DEPTH];
1119 	size_t num_stack_entries;
1120 	u32 alloc_stack_hash;
1121 	int allocation_gate;
1122 
1123 	/*
1124 	 * Perform size check before switching kfence_allocation_gate, so that
1125 	 * we don't disable KFENCE without making an allocation.
1126 	 */
1127 	if (size > PAGE_SIZE) {
1128 		atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
1129 		return NULL;
1130 	}
1131 
1132 	/*
1133 	 * Skip allocations from non-default zones, including DMA. We cannot
1134 	 * guarantee that pages in the KFENCE pool will have the requested
1135 	 * properties (e.g. reside in DMAable memory).
1136 	 */
1137 	if ((flags & GFP_ZONEMASK) ||
1138 	    ((flags & __GFP_THISNODE) && num_online_nodes() > 1) ||
1139 	    (s->flags & (SLAB_CACHE_DMA | SLAB_CACHE_DMA32))) {
1140 		atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
1141 		return NULL;
1142 	}
1143 
1144 	/*
1145 	 * Skip allocations for this slab, if KFENCE has been disabled for
1146 	 * this slab.
1147 	 */
1148 	if (s->flags & SLAB_SKIP_KFENCE)
1149 		return NULL;
1150 
1151 	allocation_gate = atomic_inc_return(&kfence_allocation_gate);
1152 	if (allocation_gate > 1)
1153 		return NULL;
1154 #ifdef CONFIG_KFENCE_STATIC_KEYS
1155 	/*
1156 	 * waitqueue_active() is fully ordered after the update of
1157 	 * kfence_allocation_gate per atomic_inc_return().
1158 	 */
1159 	if (allocation_gate == 1 && waitqueue_active(&allocation_wait)) {
1160 		/*
1161 		 * Calling wake_up() here may deadlock when allocations happen
1162 		 * from within timer code. Use an irq_work to defer it.
1163 		 */
1164 		irq_work_queue(&wake_up_kfence_timer_work);
1165 	}
1166 #endif
1167 
1168 	if (!READ_ONCE(kfence_enabled))
1169 		return NULL;
1170 
1171 	num_stack_entries = stack_trace_save(stack_entries, KFENCE_STACK_DEPTH, 0);
1172 
1173 	/*
1174 	 * Do expensive check for coverage of allocation in slow-path after
1175 	 * allocation_gate has already become non-zero, even though it might
1176 	 * mean not making any allocation within a given sample interval.
1177 	 *
1178 	 * This ensures reasonable allocation coverage when the pool is almost
1179 	 * full, including avoiding long-lived allocations of the same source
1180 	 * filling up the pool (e.g. pagecache allocations).
1181 	 */
1182 	alloc_stack_hash = get_alloc_stack_hash(stack_entries, num_stack_entries);
1183 	if (should_skip_covered() && alloc_covered_contains(alloc_stack_hash)) {
1184 		atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_COVERED]);
1185 		return NULL;
1186 	}
1187 
1188 	return kfence_guarded_alloc(s, size, flags, stack_entries, num_stack_entries,
1189 				    alloc_stack_hash);
1190 }
1191 
kfence_ksize(const void * addr)1192 size_t kfence_ksize(const void *addr)
1193 {
1194 	const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1195 
1196 	/*
1197 	 * Read locklessly -- if there is a race with __kfence_alloc(), this is
1198 	 * either a use-after-free or invalid access.
1199 	 */
1200 	return meta ? meta->size : 0;
1201 }
1202 
kfence_object_start(const void * addr)1203 void *kfence_object_start(const void *addr)
1204 {
1205 	const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1206 
1207 	/*
1208 	 * Read locklessly -- if there is a race with __kfence_alloc(), this is
1209 	 * either a use-after-free or invalid access.
1210 	 */
1211 	return meta ? (void *)meta->addr : NULL;
1212 }
1213 
__kfence_free(void * addr)1214 void __kfence_free(void *addr)
1215 {
1216 	struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1217 
1218 #ifdef CONFIG_MEMCG
1219 	KFENCE_WARN_ON(meta->obj_exts.objcg);
1220 #endif
1221 	/*
1222 	 * If the objects of the cache are SLAB_TYPESAFE_BY_RCU, defer freeing
1223 	 * the object, as the object page may be recycled for other-typed
1224 	 * objects once it has been freed. meta->cache may be NULL if the cache
1225 	 * was destroyed.
1226 	 * Save the stack trace here so that reports show where the user freed
1227 	 * the object.
1228 	 */
1229 	if (unlikely(meta->cache && (meta->cache->flags & SLAB_TYPESAFE_BY_RCU))) {
1230 		unsigned long flags;
1231 
1232 		raw_spin_lock_irqsave(&meta->lock, flags);
1233 		metadata_update_state(meta, KFENCE_OBJECT_RCU_FREEING, NULL, 0);
1234 		raw_spin_unlock_irqrestore(&meta->lock, flags);
1235 		call_rcu(&meta->rcu_head, rcu_guarded_free);
1236 	} else {
1237 		kfence_guarded_free(addr, meta, false);
1238 	}
1239 }
1240 
kfence_handle_page_fault(unsigned long addr,bool is_write,struct pt_regs * regs)1241 bool kfence_handle_page_fault(unsigned long addr, bool is_write, struct pt_regs *regs)
1242 {
1243 	const int page_index = (addr - (unsigned long)__kfence_pool) / PAGE_SIZE;
1244 	struct kfence_metadata *to_report = NULL;
1245 	enum kfence_error_type error_type;
1246 	unsigned long flags;
1247 
1248 	if (!is_kfence_address((void *)addr))
1249 		return false;
1250 
1251 	if (!READ_ONCE(kfence_enabled)) /* If disabled at runtime ... */
1252 		return kfence_unprotect(addr); /* ... unprotect and proceed. */
1253 
1254 	atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
1255 
1256 	if (page_index % 2) {
1257 		/* This is a redzone, report a buffer overflow. */
1258 		struct kfence_metadata *meta;
1259 		int distance = 0;
1260 
1261 		meta = addr_to_metadata(addr - PAGE_SIZE);
1262 		if (meta && kfence_obj_allocated(meta)) {
1263 			to_report = meta;
1264 			/* Data race ok; distance calculation approximate. */
1265 			distance = addr - data_race(meta->addr + meta->size);
1266 		}
1267 
1268 		meta = addr_to_metadata(addr + PAGE_SIZE);
1269 		if (meta && kfence_obj_allocated(meta)) {
1270 			/* Data race ok; distance calculation approximate. */
1271 			if (!to_report || distance > data_race(meta->addr) - addr)
1272 				to_report = meta;
1273 		}
1274 
1275 		if (!to_report)
1276 			goto out;
1277 
1278 		raw_spin_lock_irqsave(&to_report->lock, flags);
1279 		to_report->unprotected_page = addr;
1280 		error_type = KFENCE_ERROR_OOB;
1281 
1282 		/*
1283 		 * If the object was freed before we took the look we can still
1284 		 * report this as an OOB -- the report will simply show the
1285 		 * stacktrace of the free as well.
1286 		 */
1287 	} else {
1288 		to_report = addr_to_metadata(addr);
1289 		if (!to_report)
1290 			goto out;
1291 
1292 		raw_spin_lock_irqsave(&to_report->lock, flags);
1293 		error_type = KFENCE_ERROR_UAF;
1294 		/*
1295 		 * We may race with __kfence_alloc(), and it is possible that a
1296 		 * freed object may be reallocated. We simply report this as a
1297 		 * use-after-free, with the stack trace showing the place where
1298 		 * the object was re-allocated.
1299 		 */
1300 	}
1301 
1302 out:
1303 	if (to_report) {
1304 		kfence_report_error(addr, is_write, regs, to_report, error_type);
1305 		raw_spin_unlock_irqrestore(&to_report->lock, flags);
1306 	} else {
1307 		/* This may be a UAF or OOB access, but we can't be sure. */
1308 		kfence_report_error(addr, is_write, regs, NULL, KFENCE_ERROR_INVALID);
1309 	}
1310 
1311 	return kfence_unprotect(addr); /* Unprotect and let access proceed. */
1312 }
1313