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