1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * KFENCE guarded object allocator and fault handling.
4 *
5 * Copyright (C) 2020, Google LLC.
6 */
7
8 #define pr_fmt(fmt) "kfence: " fmt
9
10 #include <linux/atomic.h>
11 #include <linux/bug.h>
12 #include <linux/debugfs.h>
13 #include <linux/hash.h>
14 #include <linux/irq_work.h>
15 #include <linux/jhash.h>
16 #include <linux/kasan-enabled.h>
17 #include <linux/kcsan-checks.h>
18 #include <linux/kfence.h>
19 #include <linux/kmemleak.h>
20 #include <linux/list.h>
21 #include <linux/lockdep.h>
22 #include <linux/log2.h>
23 #include <linux/memblock.h>
24 #include <linux/moduleparam.h>
25 #include <linux/nodemask.h>
26 #include <linux/notifier.h>
27 #include <linux/panic_notifier.h>
28 #include <linux/random.h>
29 #include <linux/rcupdate.h>
30 #include <linux/reboot.h>
31 #include <linux/sched/clock.h>
32 #include <linux/seq_file.h>
33 #include <linux/slab.h>
34 #include <linux/spinlock.h>
35 #include <linux/string.h>
36
37 #include <asm/kfence.h>
38
39 #include "kfence.h"
40
41 /* Disables KFENCE on the first warning assuming an irrecoverable error. */
42 #define KFENCE_WARN_ON(cond) \
43 ({ \
44 const bool __cond = WARN_ON(cond); \
45 if (unlikely(__cond)) { \
46 WRITE_ONCE(kfence_enabled, false); \
47 disabled_by_warn = true; \
48 } \
49 __cond; \
50 })
51
52 /* === Data ================================================================= */
53
54 bool kfence_enabled __read_mostly;
55 static bool disabled_by_warn __read_mostly;
56
57 unsigned long kfence_sample_interval __read_mostly = CONFIG_KFENCE_SAMPLE_INTERVAL;
58 EXPORT_SYMBOL_GPL(kfence_sample_interval); /* Export for test modules. */
59
60 #ifdef MODULE_PARAM_PREFIX
61 #undef MODULE_PARAM_PREFIX
62 #endif
63 #define MODULE_PARAM_PREFIX "kfence."
64
65 static int kfence_enable_late(void);
param_set_sample_interval(const char * val,const struct kernel_param * kp)66 static int param_set_sample_interval(const char *val, const struct kernel_param *kp)
67 {
68 unsigned long num;
69 int ret = kstrtoul(val, 0, &num);
70
71 if (ret < 0)
72 return ret;
73
74 /* Using 0 to indicate KFENCE is disabled. */
75 if (!num && READ_ONCE(kfence_enabled)) {
76 pr_info("disabled\n");
77 WRITE_ONCE(kfence_enabled, false);
78 }
79
80 if (num && kasan_hw_tags_enabled()) {
81 pr_info("disabled as KASAN HW tags are enabled\n");
82 return -EINVAL;
83 }
84
85 *((unsigned long *)kp->arg) = num;
86
87 if (num && !READ_ONCE(kfence_enabled) && system_state != SYSTEM_BOOTING)
88 return disabled_by_warn ? -EINVAL : kfence_enable_late();
89 return 0;
90 }
91
param_get_sample_interval(char * buffer,const struct kernel_param * kp)92 static int param_get_sample_interval(char *buffer, const struct kernel_param *kp)
93 {
94 if (!READ_ONCE(kfence_enabled))
95 return sprintf(buffer, "0\n");
96
97 return param_get_ulong(buffer, kp);
98 }
99
100 static const struct kernel_param_ops sample_interval_param_ops = {
101 .set = param_set_sample_interval,
102 .get = param_get_sample_interval,
103 };
104 module_param_cb(sample_interval, &sample_interval_param_ops, &kfence_sample_interval, 0600);
105
106 /* Pool usage% threshold when currently covered allocations are skipped. */
107 static unsigned long kfence_skip_covered_thresh __read_mostly = 75;
108 module_param_named(skip_covered_thresh, kfence_skip_covered_thresh, ulong, 0644);
109
110 /* Allocation burst count: number of excess KFENCE allocations per sample. */
111 static unsigned int kfence_burst __read_mostly;
112 module_param_named(burst, kfence_burst, uint, 0644);
113
114 /* If true, use a deferrable timer. */
115 static bool kfence_deferrable __read_mostly = IS_ENABLED(CONFIG_KFENCE_DEFERRABLE);
116 module_param_named(deferrable, kfence_deferrable, bool, 0444);
117
118 /* If true, check all canary bytes on panic. */
119 static bool kfence_check_on_panic __read_mostly;
120 module_param_named(check_on_panic, kfence_check_on_panic, bool, 0444);
121
122 /* The pool of pages used for guard pages and objects. */
123 char *__kfence_pool __read_mostly;
124 EXPORT_SYMBOL(__kfence_pool); /* Export for test modules. */
125
126 /*
127 * Per-object metadata, with one-to-one mapping of object metadata to
128 * backing pages (in __kfence_pool).
129 */
130 static_assert(CONFIG_KFENCE_NUM_OBJECTS > 0);
131 struct kfence_metadata *kfence_metadata __read_mostly;
132
133 /*
134 * If kfence_metadata is not NULL, it may be accessed by kfence_shutdown_cache().
135 * So introduce kfence_metadata_init to initialize metadata, and then make
136 * kfence_metadata visible after initialization is successful. This prevents
137 * potential UAF or access to uninitialized metadata.
138 */
139 static struct kfence_metadata *kfence_metadata_init __read_mostly;
140
141 /* Freelist with available objects. */
142 DEFINE_RAW_SPINLOCK(kfence_freelist_lock); /* Lock protecting freelist. */
143 static struct list_head kfence_freelist __guarded_by(&kfence_freelist_lock) = LIST_HEAD_INIT(kfence_freelist);
144
145 /*
146 * The static key to set up a KFENCE allocation; or if static keys are not used
147 * to gate allocations, to avoid a load and compare if KFENCE is disabled.
148 */
149 DEFINE_STATIC_KEY_FALSE(kfence_allocation_key);
150
151 /* Gates the allocation, ensuring only one succeeds in a given period. */
152 atomic_t kfence_allocation_gate = ATOMIC_INIT(1);
153
154 /*
155 * A Counting Bloom filter of allocation coverage: limits currently covered
156 * allocations of the same source filling up the pool.
157 *
158 * Assuming a range of 15%-85% unique allocations in the pool at any point in
159 * time, the below parameters provide a probablity of 0.02-0.33 for false
160 * positive hits respectively:
161 *
162 * P(alloc_traces) = (1 - e^(-HNUM * (alloc_traces / SIZE)) ^ HNUM
163 */
164 #define ALLOC_COVERED_HNUM 2
165 #define ALLOC_COVERED_ORDER (const_ilog2(CONFIG_KFENCE_NUM_OBJECTS) + 2)
166 #define ALLOC_COVERED_SIZE (1 << ALLOC_COVERED_ORDER)
167 #define ALLOC_COVERED_HNEXT(h) hash_32(h, ALLOC_COVERED_ORDER)
168 #define ALLOC_COVERED_MASK (ALLOC_COVERED_SIZE - 1)
169 static atomic_t alloc_covered[ALLOC_COVERED_SIZE];
170
171 /* Stack depth used to determine uniqueness of an allocation. */
172 #define UNIQUE_ALLOC_STACK_DEPTH ((size_t)8)
173
174 /*
175 * Randomness for stack hashes, making the same collisions across reboots and
176 * different machines less likely.
177 */
178 static u32 stack_hash_seed __ro_after_init;
179
180 /* Statistics counters for debugfs. */
181 enum kfence_counter_id {
182 KFENCE_COUNTER_ALLOCATED,
183 KFENCE_COUNTER_ALLOCS,
184 KFENCE_COUNTER_FREES,
185 KFENCE_COUNTER_ZOMBIES,
186 KFENCE_COUNTER_BUGS,
187 KFENCE_COUNTER_SKIP_INCOMPAT,
188 KFENCE_COUNTER_SKIP_CAPACITY,
189 KFENCE_COUNTER_SKIP_COVERED,
190 KFENCE_COUNTER_COUNT,
191 };
192 static atomic_long_t counters[KFENCE_COUNTER_COUNT];
193 static const char *const counter_names[] = {
194 [KFENCE_COUNTER_ALLOCATED] = "currently allocated",
195 [KFENCE_COUNTER_ALLOCS] = "total allocations",
196 [KFENCE_COUNTER_FREES] = "total frees",
197 [KFENCE_COUNTER_ZOMBIES] = "zombie allocations",
198 [KFENCE_COUNTER_BUGS] = "total bugs",
199 [KFENCE_COUNTER_SKIP_INCOMPAT] = "skipped allocations (incompatible)",
200 [KFENCE_COUNTER_SKIP_CAPACITY] = "skipped allocations (capacity)",
201 [KFENCE_COUNTER_SKIP_COVERED] = "skipped allocations (covered)",
202 };
203 static_assert(ARRAY_SIZE(counter_names) == KFENCE_COUNTER_COUNT);
204
205 /* === Internals ============================================================ */
206
should_skip_covered(void)207 static inline bool should_skip_covered(void)
208 {
209 unsigned long thresh = (CONFIG_KFENCE_NUM_OBJECTS * kfence_skip_covered_thresh) / 100;
210
211 return atomic_long_read(&counters[KFENCE_COUNTER_ALLOCATED]) > thresh;
212 }
213
get_alloc_stack_hash(unsigned long * stack_entries,size_t num_entries)214 static u32 get_alloc_stack_hash(unsigned long *stack_entries, size_t num_entries)
215 {
216 num_entries = min(num_entries, UNIQUE_ALLOC_STACK_DEPTH);
217 num_entries = filter_irq_stacks(stack_entries, num_entries);
218 return jhash(stack_entries, num_entries * sizeof(stack_entries[0]), stack_hash_seed);
219 }
220
221 /*
222 * Adds (or subtracts) count @val for allocation stack trace hash
223 * @alloc_stack_hash from Counting Bloom filter.
224 */
alloc_covered_add(u32 alloc_stack_hash,int val)225 static void alloc_covered_add(u32 alloc_stack_hash, int val)
226 {
227 int i;
228
229 for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
230 atomic_add(val, &alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]);
231 alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
232 }
233 }
234
235 /*
236 * Returns true if the allocation stack trace hash @alloc_stack_hash is
237 * currently contained (non-zero count) in Counting Bloom filter.
238 */
alloc_covered_contains(u32 alloc_stack_hash)239 static bool alloc_covered_contains(u32 alloc_stack_hash)
240 {
241 int i;
242
243 for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
244 if (!atomic_read(&alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]))
245 return false;
246 alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
247 }
248
249 return true;
250 }
251
kfence_protect(unsigned long addr)252 static bool kfence_protect(unsigned long addr)
253 {
254 return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), true));
255 }
256
kfence_unprotect(unsigned long addr)257 static bool kfence_unprotect(unsigned long addr)
258 {
259 return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), false));
260 }
261
metadata_to_pageaddr(const struct kfence_metadata * meta)262 static inline unsigned long metadata_to_pageaddr(const struct kfence_metadata *meta)
263 __must_hold(&meta->lock)
264 {
265 unsigned long offset = (meta - kfence_metadata + 1) * PAGE_SIZE * 2;
266 unsigned long pageaddr = (unsigned long)&__kfence_pool[offset];
267
268 /* The checks do not affect performance; only called from slow-paths. */
269
270 /* Only call with a pointer into kfence_metadata. */
271 if (KFENCE_WARN_ON(meta < kfence_metadata ||
272 meta >= kfence_metadata + CONFIG_KFENCE_NUM_OBJECTS))
273 return 0;
274
275 /*
276 * This metadata object only ever maps to 1 page; verify that the stored
277 * address is in the expected range.
278 */
279 if (KFENCE_WARN_ON(ALIGN_DOWN(meta->addr, PAGE_SIZE) != pageaddr))
280 return 0;
281
282 return pageaddr;
283 }
284
kfence_obj_allocated(const struct kfence_metadata * meta)285 static inline bool kfence_obj_allocated(const struct kfence_metadata *meta)
286 {
287 enum kfence_object_state state = READ_ONCE(meta->state);
288
289 return state == KFENCE_OBJECT_ALLOCATED || state == KFENCE_OBJECT_RCU_FREEING;
290 }
291
292 /*
293 * Update the object's metadata state, including updating the alloc/free stacks
294 * depending on the state transition.
295 */
296 static noinline void
metadata_update_state(struct kfence_metadata * meta,enum kfence_object_state next,unsigned long * stack_entries,size_t num_stack_entries)297 metadata_update_state(struct kfence_metadata *meta, enum kfence_object_state next,
298 unsigned long *stack_entries, size_t num_stack_entries)
299 __must_hold(&meta->lock)
300 {
301 struct kfence_track *track =
302 next == KFENCE_OBJECT_ALLOCATED ? &meta->alloc_track : &meta->free_track;
303
304 lockdep_assert_held(&meta->lock);
305
306 /* Stack has been saved when calling rcu, skip. */
307 if (READ_ONCE(meta->state) == KFENCE_OBJECT_RCU_FREEING)
308 goto out;
309
310 if (stack_entries) {
311 memcpy(track->stack_entries, stack_entries,
312 num_stack_entries * sizeof(stack_entries[0]));
313 } else {
314 /*
315 * Skip over 1 (this) functions; noinline ensures we do not
316 * accidentally skip over the caller by never inlining.
317 */
318 num_stack_entries = stack_trace_save(track->stack_entries, KFENCE_STACK_DEPTH, 1);
319 }
320 track->num_stack_entries = num_stack_entries;
321 track->pid = task_pid_nr(current);
322 track->cpu = raw_smp_processor_id();
323 track->ts_nsec = local_clock(); /* Same source as printk timestamps. */
324
325 out:
326 /*
327 * Pairs with READ_ONCE() in
328 * kfence_shutdown_cache(),
329 * kfence_handle_page_fault().
330 */
331 WRITE_ONCE(meta->state, next);
332 }
333
334 #ifdef CONFIG_KMSAN
335 #define check_canary_attributes noinline __no_kmsan_checks
336 #else
337 #define check_canary_attributes inline
338 #endif
339
340 /* Check canary byte at @addr. */
check_canary_byte(u8 * addr)341 static check_canary_attributes bool check_canary_byte(u8 *addr)
342 {
343 struct kfence_metadata *meta;
344 enum kfence_fault fault;
345 unsigned long flags;
346
347 if (likely(*addr == KFENCE_CANARY_PATTERN_U8(addr)))
348 return true;
349
350 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
351
352 meta = addr_to_metadata((unsigned long)addr);
353 raw_spin_lock_irqsave(&meta->lock, flags);
354 fault = kfence_report_error((unsigned long)addr, false, NULL, meta, KFENCE_ERROR_CORRUPTION);
355 raw_spin_unlock_irqrestore(&meta->lock, flags);
356 kfence_handle_fault(fault);
357
358 return false;
359 }
360
set_canary(const struct kfence_metadata * meta)361 static inline void set_canary(const struct kfence_metadata *meta)
362 {
363 const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
364 unsigned long addr = pageaddr;
365
366 /*
367 * The canary may be written to part of the object memory, but it does
368 * not affect it. The user should initialize the object before using it.
369 */
370 for (; addr < meta->addr; addr += sizeof(u64))
371 *((u64 *)addr) = KFENCE_CANARY_PATTERN_U64;
372
373 addr = ALIGN_DOWN(meta->addr + meta->size, sizeof(u64));
374 for (; addr - pageaddr < PAGE_SIZE; addr += sizeof(u64))
375 *((u64 *)addr) = KFENCE_CANARY_PATTERN_U64;
376 }
377
378 static check_canary_attributes void
check_canary(const struct kfence_metadata * meta)379 check_canary(const struct kfence_metadata *meta)
380 {
381 const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
382 unsigned long addr = pageaddr;
383
384 /*
385 * We'll iterate over each canary byte per-side until a corrupted byte
386 * is found. However, we'll still iterate over the canary bytes to the
387 * right of the object even if there was an error in the canary bytes to
388 * the left of the object. Specifically, if check_canary_byte()
389 * generates an error, showing both sides might give more clues as to
390 * what the error is about when displaying which bytes were corrupted.
391 */
392
393 /* Apply to left of object. */
394 for (; meta->addr - addr >= sizeof(u64); addr += sizeof(u64)) {
395 if (unlikely(*((u64 *)addr) != KFENCE_CANARY_PATTERN_U64))
396 break;
397 }
398
399 /*
400 * If the canary is corrupted in a certain 64 bytes, or the canary
401 * memory cannot be completely covered by multiple consecutive 64 bytes,
402 * it needs to be checked one by one.
403 */
404 for (; addr < meta->addr; addr++) {
405 if (unlikely(!check_canary_byte((u8 *)addr)))
406 break;
407 }
408
409 /* Apply to right of object. */
410 for (addr = meta->addr + meta->size; addr % sizeof(u64) != 0; addr++) {
411 if (unlikely(!check_canary_byte((u8 *)addr)))
412 return;
413 }
414 for (; addr - pageaddr < PAGE_SIZE; addr += sizeof(u64)) {
415 if (unlikely(*((u64 *)addr) != KFENCE_CANARY_PATTERN_U64)) {
416
417 for (; addr - pageaddr < PAGE_SIZE; addr++) {
418 if (!check_canary_byte((u8 *)addr))
419 return;
420 }
421 }
422 }
423 }
424
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)425 static void *kfence_guarded_alloc(struct kmem_cache *cache, size_t size, gfp_t gfp,
426 unsigned long *stack_entries, size_t num_stack_entries,
427 u32 alloc_stack_hash)
428 {
429 struct kfence_metadata *meta = NULL;
430 unsigned long flags;
431 struct slab *slab;
432 void *addr;
433 const bool random_right_allocate = get_random_u32_below(2);
434 const bool random_fault = CONFIG_KFENCE_STRESS_TEST_FAULTS &&
435 !get_random_u32_below(CONFIG_KFENCE_STRESS_TEST_FAULTS);
436
437 /* Try to obtain a free object. */
438 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
439 if (!list_empty(&kfence_freelist)) {
440 meta = list_entry(kfence_freelist.next, struct kfence_metadata, list);
441 list_del_init(&meta->list);
442 }
443 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
444 if (!meta) {
445 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_CAPACITY]);
446 return NULL;
447 }
448
449 if (unlikely(!raw_spin_trylock_irqsave(&meta->lock, flags))) {
450 /*
451 * This is extremely unlikely -- we are reporting on a
452 * use-after-free, which locked meta->lock, and the reporting
453 * code via printk calls kmalloc() which ends up in
454 * kfence_alloc() and tries to grab the same object that we're
455 * reporting on. While it has never been observed, lockdep does
456 * report that there is a possibility of deadlock. Fix it by
457 * using trylock and bailing out gracefully.
458 */
459 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
460 /* Put the object back on the freelist. */
461 list_add_tail(&meta->list, &kfence_freelist);
462 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
463
464 return NULL;
465 }
466
467 meta->addr = metadata_to_pageaddr(meta);
468 /* Unprotect if we're reusing this page. */
469 if (meta->state == KFENCE_OBJECT_FREED)
470 kfence_unprotect(meta->addr);
471
472 /*
473 * Note: for allocations made before RNG initialization, will always
474 * return zero. We still benefit from enabling KFENCE as early as
475 * possible, even when the RNG is not yet available, as this will allow
476 * KFENCE to detect bugs due to earlier allocations. The only downside
477 * is that the out-of-bounds accesses detected are deterministic for
478 * such allocations.
479 */
480 if (random_right_allocate) {
481 /* Allocate on the "right" side, re-calculate address. */
482 meta->addr += PAGE_SIZE - size;
483 meta->addr = ALIGN_DOWN(meta->addr, cache->align);
484 }
485
486 addr = (void *)meta->addr;
487
488 /* Update remaining metadata. */
489 metadata_update_state(meta, KFENCE_OBJECT_ALLOCATED, stack_entries, num_stack_entries);
490 /* Pairs with READ_ONCE() in kfence_shutdown_cache(). */
491 WRITE_ONCE(meta->cache, cache);
492 meta->size = size;
493 meta->alloc_stack_hash = alloc_stack_hash;
494 raw_spin_unlock_irqrestore(&meta->lock, flags);
495
496 alloc_covered_add(alloc_stack_hash, 1);
497
498 /* Set required slab fields. */
499 slab = virt_to_slab(addr);
500 slab->slab_cache = cache;
501 slab->objects = 1;
502
503 /* Memory initialization. */
504 set_canary(meta);
505
506 /*
507 * We check slab_want_init_on_alloc() ourselves, rather than letting
508 * slab do the initialization, as otherwise it might overwrite KFENCE's
509 * redzone.
510 */
511 if (unlikely(slab_want_init_on_alloc(gfp, cache)))
512 memzero_explicit(addr, size);
513 if (cache->ctor)
514 cache->ctor(addr);
515
516 if (random_fault)
517 kfence_protect(meta->addr); /* Random "faults" by protecting the object. */
518
519 atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCATED]);
520 atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCS]);
521
522 return addr;
523 }
524
kfence_guarded_free(void * addr,struct kfence_metadata * meta,bool zombie)525 static void kfence_guarded_free(void *addr, struct kfence_metadata *meta, bool zombie)
526 {
527 struct kcsan_scoped_access assert_page_exclusive;
528 u32 alloc_stack_hash;
529 unsigned long flags;
530 bool init;
531
532 raw_spin_lock_irqsave(&meta->lock, flags);
533
534 if (!kfence_obj_allocated(meta) || meta->addr != (unsigned long)addr) {
535 enum kfence_fault fault;
536
537 /* Invalid or double-free, bail out. */
538 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
539 fault = kfence_report_error((unsigned long)addr, false, NULL, meta,
540 KFENCE_ERROR_INVALID_FREE);
541 raw_spin_unlock_irqrestore(&meta->lock, flags);
542 kfence_handle_fault(fault);
543 return;
544 }
545
546 /* Detect racy use-after-free, or incorrect reallocation of this page by KFENCE. */
547 kcsan_begin_scoped_access((void *)ALIGN_DOWN((unsigned long)addr, PAGE_SIZE), PAGE_SIZE,
548 KCSAN_ACCESS_SCOPED | KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT,
549 &assert_page_exclusive);
550
551 if (CONFIG_KFENCE_STRESS_TEST_FAULTS)
552 kfence_unprotect((unsigned long)addr); /* To check canary bytes. */
553
554 /* Restore page protection if there was an OOB access. */
555 if (meta->unprotected_page) {
556 memzero_explicit((void *)ALIGN_DOWN(meta->unprotected_page, PAGE_SIZE), PAGE_SIZE);
557 kfence_protect(meta->unprotected_page);
558 meta->unprotected_page = 0;
559 }
560
561 /* Mark the object as freed. */
562 metadata_update_state(meta, KFENCE_OBJECT_FREED, NULL, 0);
563 init = slab_want_init_on_free(meta->cache);
564 alloc_stack_hash = meta->alloc_stack_hash;
565 raw_spin_unlock_irqrestore(&meta->lock, flags);
566
567 alloc_covered_add(alloc_stack_hash, -1);
568
569 /* Check canary bytes for memory corruption. */
570 check_canary(meta);
571
572 /*
573 * Clear memory if init-on-free is set. While we protect the page, the
574 * data is still there, and after a use-after-free is detected, we
575 * unprotect the page, so the data is still accessible.
576 */
577 if (!zombie && unlikely(init))
578 memzero_explicit(addr, meta->size);
579
580 /* Protect to detect use-after-frees. */
581 kfence_protect((unsigned long)addr);
582
583 kcsan_end_scoped_access(&assert_page_exclusive);
584 if (!zombie) {
585 /* Add it to the tail of the freelist for reuse. */
586 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
587 KFENCE_WARN_ON(!list_empty(&meta->list));
588 list_add_tail(&meta->list, &kfence_freelist);
589 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
590
591 atomic_long_dec(&counters[KFENCE_COUNTER_ALLOCATED]);
592 atomic_long_inc(&counters[KFENCE_COUNTER_FREES]);
593 } else {
594 /* See kfence_shutdown_cache(). */
595 atomic_long_inc(&counters[KFENCE_COUNTER_ZOMBIES]);
596 }
597 }
598
rcu_guarded_free(struct rcu_head * h)599 static void rcu_guarded_free(struct rcu_head *h)
600 {
601 struct kfence_metadata *meta = container_of(h, struct kfence_metadata, rcu_head);
602
603 kfence_guarded_free((void *)meta->addr, meta, false);
604 }
605
606 /*
607 * Initialization of the KFENCE pool after its allocation.
608 * Returns 0 on success; otherwise returns the address up to
609 * which partial initialization succeeded.
610 */
kfence_init_pool(void)611 static unsigned long kfence_init_pool(void)
612 __context_unsafe(/* constructor */)
613 {
614 unsigned long addr, start_pfn;
615 int i, rand;
616
617 if (!arch_kfence_init_pool())
618 return (unsigned long)__kfence_pool;
619
620 addr = (unsigned long)__kfence_pool;
621 start_pfn = PHYS_PFN(virt_to_phys(__kfence_pool));
622
623 /*
624 * Set up object pages: they must have PGTY_slab set to avoid freeing
625 * them as real pages.
626 *
627 * We also want to avoid inserting kfence_free() in the kfree()
628 * fast-path in SLUB, and therefore need to ensure kfree() correctly
629 * enters __slab_free() slow-path.
630 */
631 for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
632 struct page *page;
633
634 if (!i || (i % 2))
635 continue;
636
637 page = pfn_to_page(start_pfn + i);
638 __SetPageSlab(page);
639 }
640
641 /*
642 * Protect the first 2 pages. The first page is mostly unnecessary, and
643 * merely serves as an extended guard page. However, adding one
644 * additional page in the beginning gives us an even number of pages,
645 * which simplifies the mapping of address to metadata index.
646 */
647 for (i = 0; i < 2; i++) {
648 if (unlikely(!kfence_protect(addr)))
649 return addr;
650
651 addr += PAGE_SIZE;
652 }
653
654 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
655 struct kfence_metadata *meta = &kfence_metadata_init[i];
656
657 /* Initialize metadata. */
658 INIT_LIST_HEAD(&meta->list);
659 raw_spin_lock_init(&meta->lock);
660 meta->state = KFENCE_OBJECT_UNUSED;
661 /* Use addr to randomize the freelist. */
662 meta->addr = i;
663
664 /* Protect the right redzone. */
665 if (unlikely(!kfence_protect(addr + 2 * i * PAGE_SIZE + PAGE_SIZE)))
666 goto reset_slab;
667 }
668
669 for (i = CONFIG_KFENCE_NUM_OBJECTS; i > 0; i--) {
670 rand = get_random_u32_below(i);
671 swap(kfence_metadata_init[i - 1].addr, kfence_metadata_init[rand].addr);
672 }
673
674 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
675 struct kfence_metadata *meta_1 = &kfence_metadata_init[i];
676 struct kfence_metadata *meta_2 = &kfence_metadata_init[meta_1->addr];
677
678 list_add_tail(&meta_2->list, &kfence_freelist);
679 }
680 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
681 kfence_metadata_init[i].addr = addr;
682 addr += 2 * PAGE_SIZE;
683 }
684
685 /*
686 * Make kfence_metadata visible only when initialization is successful.
687 * Otherwise, if the initialization fails and kfence_metadata is freed,
688 * it may cause UAF in kfence_shutdown_cache().
689 */
690 smp_store_release(&kfence_metadata, kfence_metadata_init);
691 return 0;
692
693 reset_slab:
694 addr += 2 * i * PAGE_SIZE;
695 for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
696 struct page *page;
697
698 if (!i || (i % 2))
699 continue;
700
701 page = pfn_to_page(start_pfn + i);
702 __ClearPageSlab(page);
703 }
704
705 return addr;
706 }
707
kfence_init_pool_early(void)708 static bool __init kfence_init_pool_early(void)
709 {
710 unsigned long addr;
711
712 if (!__kfence_pool)
713 return false;
714
715 addr = kfence_init_pool();
716
717 if (!addr) {
718 /*
719 * The pool is live and will never be deallocated from this point on.
720 * Ignore the pool object from the kmemleak phys object tree, as it would
721 * otherwise overlap with allocations returned by kfence_alloc(), which
722 * are registered with kmemleak through the slab post-alloc hook.
723 */
724 kmemleak_ignore_phys(__pa(__kfence_pool));
725 return true;
726 }
727
728 /*
729 * Only release unprotected pages, and do not try to go back and change
730 * page attributes due to risk of failing to do so as well. If changing
731 * page attributes for some pages fails, it is very likely that it also
732 * fails for the first page, and therefore expect addr==__kfence_pool in
733 * most failure cases.
734 */
735 memblock_free((void *)addr, KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool));
736 __kfence_pool = NULL;
737
738 memblock_free(kfence_metadata_init, KFENCE_METADATA_SIZE);
739 kfence_metadata_init = NULL;
740
741 return false;
742 }
743
744 /* === DebugFS Interface ==================================================== */
745
stats_show(struct seq_file * seq,void * v)746 static int stats_show(struct seq_file *seq, void *v)
747 {
748 int i;
749
750 seq_printf(seq, "enabled: %i\n", READ_ONCE(kfence_enabled));
751 for (i = 0; i < KFENCE_COUNTER_COUNT; i++)
752 seq_printf(seq, "%s: %ld\n", counter_names[i], atomic_long_read(&counters[i]));
753
754 return 0;
755 }
756 DEFINE_SHOW_ATTRIBUTE(stats);
757
758 /*
759 * debugfs seq_file operations for /sys/kernel/debug/kfence/objects.
760 * start_object() and next_object() return the object index + 1, because NULL is used
761 * to stop iteration.
762 */
start_object(struct seq_file * seq,loff_t * pos)763 static void *start_object(struct seq_file *seq, loff_t *pos)
764 {
765 if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
766 return (void *)((long)*pos + 1);
767 return NULL;
768 }
769
stop_object(struct seq_file * seq,void * v)770 static void stop_object(struct seq_file *seq, void *v)
771 {
772 }
773
next_object(struct seq_file * seq,void * v,loff_t * pos)774 static void *next_object(struct seq_file *seq, void *v, loff_t *pos)
775 {
776 ++*pos;
777 if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
778 return (void *)((long)*pos + 1);
779 return NULL;
780 }
781
show_object(struct seq_file * seq,void * v)782 static int show_object(struct seq_file *seq, void *v)
783 {
784 struct kfence_metadata *meta = &kfence_metadata[(long)v - 1];
785 unsigned long flags;
786
787 raw_spin_lock_irqsave(&meta->lock, flags);
788 kfence_print_object(seq, meta);
789 raw_spin_unlock_irqrestore(&meta->lock, flags);
790 seq_puts(seq, "---------------------------------\n");
791
792 return 0;
793 }
794
795 static const struct seq_operations objects_sops = {
796 .start = start_object,
797 .next = next_object,
798 .stop = stop_object,
799 .show = show_object,
800 };
801 DEFINE_SEQ_ATTRIBUTE(objects);
802
kfence_debugfs_init(void)803 static int kfence_debugfs_init(void)
804 {
805 struct dentry *kfence_dir;
806
807 if (!READ_ONCE(kfence_enabled))
808 return 0;
809
810 kfence_dir = debugfs_create_dir("kfence", NULL);
811 debugfs_create_file("stats", 0444, kfence_dir, NULL, &stats_fops);
812 debugfs_create_file("objects", 0400, kfence_dir, NULL, &objects_fops);
813 return 0;
814 }
815
816 late_initcall(kfence_debugfs_init);
817
818 /* === Panic Notifier ====================================================== */
819
kfence_check_all_canary(void)820 static void kfence_check_all_canary(void)
821 {
822 int i;
823
824 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
825 struct kfence_metadata *meta = &kfence_metadata[i];
826
827 if (kfence_obj_allocated(meta))
828 check_canary(meta);
829 }
830 }
831
kfence_check_canary_callback(struct notifier_block * nb,unsigned long reason,void * arg)832 static int kfence_check_canary_callback(struct notifier_block *nb,
833 unsigned long reason, void *arg)
834 {
835 if (READ_ONCE(kfence_enabled))
836 kfence_check_all_canary();
837 return NOTIFY_OK;
838 }
839
840 static struct notifier_block kfence_check_canary_notifier = {
841 .notifier_call = kfence_check_canary_callback,
842 };
843
844 /* === Allocation Gate Timer ================================================ */
845
846 static struct delayed_work kfence_timer;
847
848 #ifdef CONFIG_KFENCE_STATIC_KEYS
849 /* Wait queue to wake up allocation-gate timer task. */
850 static DECLARE_WAIT_QUEUE_HEAD(allocation_wait);
851
kfence_reboot_callback(struct notifier_block * nb,unsigned long action,void * data)852 static int kfence_reboot_callback(struct notifier_block *nb,
853 unsigned long action, void *data)
854 {
855 /*
856 * Disable kfence to avoid static keys IPI synchronization during
857 * late shutdown/kexec
858 */
859 WRITE_ONCE(kfence_enabled, false);
860 /* Cancel any pending timer work */
861 cancel_delayed_work(&kfence_timer);
862 /*
863 * Wake up any blocked toggle_allocation_gate() so it can complete
864 * early while the system is still able to handle IPIs.
865 */
866 wake_up(&allocation_wait);
867
868 return NOTIFY_OK;
869 }
870
871 static struct notifier_block kfence_reboot_notifier = {
872 .notifier_call = kfence_reboot_callback,
873 .priority = INT_MAX, /* Run early to stop timers ASAP */
874 };
875
wake_up_kfence_timer(struct irq_work * work)876 static void wake_up_kfence_timer(struct irq_work *work)
877 {
878 wake_up(&allocation_wait);
879 }
880 static DEFINE_IRQ_WORK(wake_up_kfence_timer_work, wake_up_kfence_timer);
881 #endif
882
883 /*
884 * Set up delayed work, which will enable and disable the static key. We need to
885 * use a work queue (rather than a simple timer), since enabling and disabling a
886 * static key cannot be done from an interrupt.
887 *
888 * Note: Toggling a static branch currently causes IPIs, and here we'll end up
889 * with a total of 2 IPIs to all CPUs. If this ends up a problem in future (with
890 * more aggressive sampling intervals), we could get away with a variant that
891 * avoids IPIs, at the cost of not immediately capturing allocations if the
892 * instructions remain cached.
893 */
toggle_allocation_gate(struct work_struct * work)894 static void toggle_allocation_gate(struct work_struct *work)
895 {
896 if (!READ_ONCE(kfence_enabled))
897 return;
898
899 atomic_set(&kfence_allocation_gate, -kfence_burst);
900 #ifdef CONFIG_KFENCE_STATIC_KEYS
901 /* Enable static key, and await allocation to happen. */
902 static_branch_enable(&kfence_allocation_key);
903
904 wait_event_idle(allocation_wait,
905 atomic_read(&kfence_allocation_gate) > 0 ||
906 !READ_ONCE(kfence_enabled));
907
908 /* Disable static key and reset timer. */
909 static_branch_disable(&kfence_allocation_key);
910 #endif
911 queue_delayed_work(system_dfl_wq, &kfence_timer,
912 msecs_to_jiffies(kfence_sample_interval));
913 }
914
915 /* === Public interface ===================================================== */
916
kfence_alloc_pool_and_metadata(void)917 void __init kfence_alloc_pool_and_metadata(void)
918 {
919 if (!kfence_sample_interval)
920 return;
921
922 /*
923 * If KASAN hardware tags are enabled, disable KFENCE, because it
924 * does not support MTE yet.
925 */
926 if (kasan_hw_tags_enabled()) {
927 pr_info("disabled as KASAN HW tags are enabled\n");
928 if (__kfence_pool) {
929 memblock_free(__kfence_pool, KFENCE_POOL_SIZE);
930 __kfence_pool = NULL;
931 }
932 kfence_sample_interval = 0;
933 return;
934 }
935
936 /*
937 * If the pool has already been initialized by arch, there is no need to
938 * re-allocate the memory pool.
939 */
940 if (!__kfence_pool)
941 __kfence_pool = memblock_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
942
943 if (!__kfence_pool) {
944 pr_err("failed to allocate pool\n");
945 return;
946 }
947
948 /* The memory allocated by memblock has been zeroed out. */
949 kfence_metadata_init = memblock_alloc(KFENCE_METADATA_SIZE, PAGE_SIZE);
950 if (!kfence_metadata_init) {
951 pr_err("failed to allocate metadata\n");
952 memblock_free(__kfence_pool, KFENCE_POOL_SIZE);
953 __kfence_pool = NULL;
954 }
955 }
956
kfence_init_enable(void)957 static void kfence_init_enable(void)
958 {
959 if (!IS_ENABLED(CONFIG_KFENCE_STATIC_KEYS))
960 static_branch_enable(&kfence_allocation_key);
961
962 if (kfence_deferrable)
963 INIT_DEFERRABLE_WORK(&kfence_timer, toggle_allocation_gate);
964 else
965 INIT_DELAYED_WORK(&kfence_timer, toggle_allocation_gate);
966
967 if (kfence_check_on_panic)
968 atomic_notifier_chain_register(&panic_notifier_list, &kfence_check_canary_notifier);
969
970 #ifdef CONFIG_KFENCE_STATIC_KEYS
971 register_reboot_notifier(&kfence_reboot_notifier);
972 #endif
973
974 WRITE_ONCE(kfence_enabled, true);
975 queue_delayed_work(system_dfl_wq, &kfence_timer, 0);
976
977 pr_info("initialized - using %lu bytes for %d objects at 0x%p-0x%p\n", KFENCE_POOL_SIZE,
978 CONFIG_KFENCE_NUM_OBJECTS, (void *)__kfence_pool,
979 (void *)(__kfence_pool + KFENCE_POOL_SIZE));
980 }
981
kfence_init(void)982 void __init kfence_init(void)
983 {
984 stack_hash_seed = get_random_u32();
985
986 /* Setting kfence_sample_interval to 0 on boot disables KFENCE. */
987 if (!kfence_sample_interval)
988 return;
989
990 if (!kfence_init_pool_early()) {
991 pr_err("%s failed\n", __func__);
992 return;
993 }
994
995 kfence_init_enable();
996 }
997
kfence_init_late(void)998 static int kfence_init_late(void)
999 {
1000 const unsigned long nr_pages_pool = KFENCE_POOL_SIZE / PAGE_SIZE;
1001 const unsigned long nr_pages_meta = KFENCE_METADATA_SIZE / PAGE_SIZE;
1002 unsigned long addr = (unsigned long)__kfence_pool;
1003 unsigned long free_size = KFENCE_POOL_SIZE;
1004 int err = -ENOMEM;
1005
1006 #ifdef CONFIG_CONTIG_ALLOC
1007 struct page *pages;
1008
1009 pages = alloc_contig_pages(nr_pages_pool, GFP_KERNEL | __GFP_SKIP_KASAN,
1010 first_online_node, NULL);
1011 if (!pages)
1012 return -ENOMEM;
1013
1014 __kfence_pool = page_to_virt(pages);
1015 pages = alloc_contig_pages(nr_pages_meta, GFP_KERNEL | __GFP_SKIP_KASAN,
1016 first_online_node, NULL);
1017 if (pages)
1018 kfence_metadata_init = page_to_virt(pages);
1019 #else
1020 if (nr_pages_pool > MAX_ORDER_NR_PAGES ||
1021 nr_pages_meta > MAX_ORDER_NR_PAGES) {
1022 pr_warn("KFENCE_NUM_OBJECTS too large for buddy allocator\n");
1023 return -EINVAL;
1024 }
1025
1026 __kfence_pool = alloc_pages_exact(KFENCE_POOL_SIZE,
1027 GFP_KERNEL | __GFP_SKIP_KASAN);
1028 if (!__kfence_pool)
1029 return -ENOMEM;
1030
1031 kfence_metadata_init = alloc_pages_exact(KFENCE_METADATA_SIZE,
1032 GFP_KERNEL | __GFP_SKIP_KASAN);
1033 #endif
1034
1035 if (!kfence_metadata_init)
1036 goto free_pool;
1037
1038 memzero_explicit(kfence_metadata_init, KFENCE_METADATA_SIZE);
1039 addr = kfence_init_pool();
1040 if (!addr) {
1041 kfence_init_enable();
1042 kfence_debugfs_init();
1043 return 0;
1044 }
1045
1046 pr_err("%s failed\n", __func__);
1047 free_size = KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool);
1048 err = -EBUSY;
1049
1050 #ifdef CONFIG_CONTIG_ALLOC
1051 free_contig_range(page_to_pfn(virt_to_page((void *)kfence_metadata_init)),
1052 nr_pages_meta);
1053 free_pool:
1054 free_contig_range(page_to_pfn(virt_to_page((void *)addr)),
1055 free_size / PAGE_SIZE);
1056 #else
1057 free_pages_exact((void *)kfence_metadata_init, KFENCE_METADATA_SIZE);
1058 free_pool:
1059 free_pages_exact((void *)addr, free_size);
1060 #endif
1061
1062 kfence_metadata_init = NULL;
1063 __kfence_pool = NULL;
1064 return err;
1065 }
1066
kfence_enable_late(void)1067 static int kfence_enable_late(void)
1068 {
1069 if (!__kfence_pool)
1070 return kfence_init_late();
1071
1072 WRITE_ONCE(kfence_enabled, true);
1073 queue_delayed_work(system_dfl_wq, &kfence_timer, 0);
1074 pr_info("re-enabled\n");
1075 return 0;
1076 }
1077
kfence_shutdown_cache(struct kmem_cache * s)1078 void kfence_shutdown_cache(struct kmem_cache *s)
1079 {
1080 unsigned long flags;
1081 struct kfence_metadata *meta;
1082 int i;
1083
1084 /* Pairs with release in kfence_init_pool(). */
1085 if (!smp_load_acquire(&kfence_metadata))
1086 return;
1087
1088 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
1089 bool in_use;
1090
1091 meta = &kfence_metadata[i];
1092
1093 /*
1094 * If we observe some inconsistent cache and state pair where we
1095 * should have returned false here, cache destruction is racing
1096 * with either kmem_cache_alloc() or kmem_cache_free(). Taking
1097 * the lock will not help, as different critical section
1098 * serialization will have the same outcome.
1099 */
1100 if (READ_ONCE(meta->cache) != s || !kfence_obj_allocated(meta))
1101 continue;
1102
1103 raw_spin_lock_irqsave(&meta->lock, flags);
1104 in_use = meta->cache == s && kfence_obj_allocated(meta);
1105 raw_spin_unlock_irqrestore(&meta->lock, flags);
1106
1107 if (in_use) {
1108 /*
1109 * This cache still has allocations, and we should not
1110 * release them back into the freelist so they can still
1111 * safely be used and retain the kernel's default
1112 * behaviour of keeping the allocations alive (leak the
1113 * cache); however, they effectively become "zombie
1114 * allocations" as the KFENCE objects are the only ones
1115 * still in use and the owning cache is being destroyed.
1116 *
1117 * We mark them freed, so that any subsequent use shows
1118 * more useful error messages that will include stack
1119 * traces of the user of the object, the original
1120 * allocation, and caller to shutdown_cache().
1121 */
1122 kfence_guarded_free((void *)meta->addr, meta, /*zombie=*/true);
1123 }
1124 }
1125
1126 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
1127 meta = &kfence_metadata[i];
1128
1129 /* See above. */
1130 if (READ_ONCE(meta->cache) != s || READ_ONCE(meta->state) != KFENCE_OBJECT_FREED)
1131 continue;
1132
1133 raw_spin_lock_irqsave(&meta->lock, flags);
1134 if (meta->cache == s && meta->state == KFENCE_OBJECT_FREED)
1135 meta->cache = NULL;
1136 raw_spin_unlock_irqrestore(&meta->lock, flags);
1137 }
1138 }
1139
__kfence_alloc(struct kmem_cache * s,size_t size,gfp_t flags)1140 void *__kfence_alloc(struct kmem_cache *s, size_t size, gfp_t flags)
1141 {
1142 unsigned long stack_entries[KFENCE_STACK_DEPTH];
1143 size_t num_stack_entries;
1144 u32 alloc_stack_hash;
1145 int allocation_gate;
1146
1147 /*
1148 * Perform size check before switching kfence_allocation_gate, so that
1149 * we don't disable KFENCE without making an allocation.
1150 */
1151 if (size > PAGE_SIZE) {
1152 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
1153 return NULL;
1154 }
1155
1156 /*
1157 * Skip allocations from non-default zones, including DMA. We cannot
1158 * guarantee that pages in the KFENCE pool will have the requested
1159 * properties (e.g. reside in DMAable memory).
1160 */
1161 if ((flags & GFP_ZONEMASK) ||
1162 ((flags & __GFP_THISNODE) && num_online_nodes() > 1) ||
1163 (s->flags & (SLAB_CACHE_DMA | SLAB_CACHE_DMA32))) {
1164 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
1165 return NULL;
1166 }
1167
1168 /*
1169 * Skip allocations for this slab, if KFENCE has been disabled for
1170 * this slab.
1171 */
1172 if (s->flags & SLAB_SKIP_KFENCE)
1173 return NULL;
1174
1175 allocation_gate = atomic_inc_return(&kfence_allocation_gate);
1176 if (allocation_gate > 1)
1177 return NULL;
1178 #ifdef CONFIG_KFENCE_STATIC_KEYS
1179 /*
1180 * waitqueue_active() is fully ordered after the update of
1181 * kfence_allocation_gate per atomic_inc_return().
1182 */
1183 if (allocation_gate == 1 && waitqueue_active(&allocation_wait)) {
1184 /*
1185 * Calling wake_up() here may deadlock when allocations happen
1186 * from within timer code. Use an irq_work to defer it.
1187 */
1188 irq_work_queue(&wake_up_kfence_timer_work);
1189 }
1190 #endif
1191
1192 if (!READ_ONCE(kfence_enabled))
1193 return NULL;
1194
1195 num_stack_entries = stack_trace_save(stack_entries, KFENCE_STACK_DEPTH, 0);
1196
1197 /*
1198 * Do expensive check for coverage of allocation in slow-path after
1199 * allocation_gate has already become non-zero, even though it might
1200 * mean not making any allocation within a given sample interval.
1201 *
1202 * This ensures reasonable allocation coverage when the pool is almost
1203 * full, including avoiding long-lived allocations of the same source
1204 * filling up the pool (e.g. pagecache allocations).
1205 */
1206 alloc_stack_hash = get_alloc_stack_hash(stack_entries, num_stack_entries);
1207 if (should_skip_covered() && alloc_covered_contains(alloc_stack_hash)) {
1208 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_COVERED]);
1209 return NULL;
1210 }
1211
1212 return kfence_guarded_alloc(s, size, flags, stack_entries, num_stack_entries,
1213 alloc_stack_hash);
1214 }
1215
kfence_ksize(const void * addr)1216 size_t kfence_ksize(const void *addr)
1217 {
1218 const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1219
1220 /*
1221 * Read locklessly -- if there is a race with __kfence_alloc(), this is
1222 * either a use-after-free or invalid access.
1223 */
1224 return meta ? meta->size : 0;
1225 }
1226
kfence_object_start(const void * addr)1227 void *kfence_object_start(const void *addr)
1228 {
1229 const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1230
1231 /*
1232 * Read locklessly -- if there is a race with __kfence_alloc(), this is
1233 * either a use-after-free or invalid access.
1234 */
1235 return meta ? (void *)meta->addr : NULL;
1236 }
1237
__kfence_free(void * addr)1238 void __kfence_free(void *addr)
1239 {
1240 struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1241
1242 /*
1243 * If the objects of the cache are SLAB_TYPESAFE_BY_RCU, defer freeing
1244 * the object, as the object page may be recycled for other-typed
1245 * objects once it has been freed. meta->cache may be NULL if the cache
1246 * was destroyed.
1247 * Save the stack trace here so that reports show where the user freed
1248 * the object.
1249 */
1250 if (unlikely(meta->cache && (meta->cache->flags & SLAB_TYPESAFE_BY_RCU))) {
1251 unsigned long flags;
1252
1253 raw_spin_lock_irqsave(&meta->lock, flags);
1254 metadata_update_state(meta, KFENCE_OBJECT_RCU_FREEING, NULL, 0);
1255 raw_spin_unlock_irqrestore(&meta->lock, flags);
1256 call_rcu(&meta->rcu_head, rcu_guarded_free);
1257 } else {
1258 kfence_guarded_free(addr, meta, false);
1259 }
1260 }
1261
kfence_handle_page_fault(unsigned long addr,bool is_write,struct pt_regs * regs)1262 bool kfence_handle_page_fault(unsigned long addr, bool is_write, struct pt_regs *regs)
1263 {
1264 const int page_index = (addr - (unsigned long)__kfence_pool) / PAGE_SIZE;
1265 struct kfence_metadata *to_report = NULL;
1266 unsigned long unprotected_page = 0;
1267 enum kfence_error_type error_type;
1268 enum kfence_fault fault;
1269 unsigned long flags;
1270
1271 if (!is_kfence_address((void *)addr))
1272 return false;
1273
1274 if (!READ_ONCE(kfence_enabled)) /* If disabled at runtime ... */
1275 return kfence_unprotect(addr); /* ... unprotect and proceed. */
1276
1277 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
1278
1279 if (page_index % 2) {
1280 /* This is a redzone, report a buffer overflow. */
1281 struct kfence_metadata *meta;
1282 int distance = 0;
1283
1284 meta = addr_to_metadata(addr - PAGE_SIZE);
1285 if (meta && kfence_obj_allocated(meta)) {
1286 to_report = meta;
1287 /* Data race ok; distance calculation approximate. */
1288 distance = addr - data_race(meta->addr + meta->size);
1289 }
1290
1291 meta = addr_to_metadata(addr + PAGE_SIZE);
1292 if (meta && kfence_obj_allocated(meta)) {
1293 /* Data race ok; distance calculation approximate. */
1294 if (!to_report || distance > data_race(meta->addr) - addr)
1295 to_report = meta;
1296 }
1297
1298 if (!to_report)
1299 goto out;
1300
1301 error_type = KFENCE_ERROR_OOB;
1302 unprotected_page = addr;
1303
1304 /*
1305 * If the object was freed before we took the look we can still
1306 * report this as an OOB -- the report will simply show the
1307 * stacktrace of the free as well.
1308 */
1309 } else {
1310 to_report = addr_to_metadata(addr);
1311 if (!to_report)
1312 goto out;
1313
1314 error_type = KFENCE_ERROR_UAF;
1315 /*
1316 * We may race with __kfence_alloc(), and it is possible that a
1317 * freed object may be reallocated. We simply report this as a
1318 * use-after-free, with the stack trace showing the place where
1319 * the object was re-allocated.
1320 */
1321 }
1322
1323 out:
1324 if (to_report) {
1325 raw_spin_lock_irqsave(&to_report->lock, flags);
1326 to_report->unprotected_page = unprotected_page;
1327 fault = kfence_report_error(addr, is_write, regs, to_report, error_type);
1328 raw_spin_unlock_irqrestore(&to_report->lock, flags);
1329 } else {
1330 /* This may be a UAF or OOB access, but we can't be sure. */
1331 fault = kfence_report_error(addr, is_write, regs, NULL, KFENCE_ERROR_INVALID);
1332 }
1333
1334 kfence_handle_fault(fault);
1335
1336 return kfence_unprotect(addr); /* Unprotect and let access proceed. */
1337 }
1338