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