1 // SPDX-License-Identifier: GPL-2.0
3 * KFENCE guarded object allocator and fault handling.
5 * Copyright (C) 2020, Google LLC.
8 #define pr_fmt(fmt) "kfence: " fmt
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/random.h>
25 #include <linux/rcupdate.h>
26 #include <linux/sched/clock.h>
27 #include <linux/sched/sysctl.h>
28 #include <linux/seq_file.h>
29 #include <linux/slab.h>
30 #include <linux/spinlock.h>
31 #include <linux/string.h>
33 #include <asm/kfence.h>
37 /* Disables KFENCE on the first warning assuming an irrecoverable error. */
38 #define KFENCE_WARN_ON(cond) \
40 const bool __cond = WARN_ON(cond); \
41 if (unlikely(__cond)) { \
42 WRITE_ONCE(kfence_enabled, false); \
43 disabled_by_warn = true; \
48 /* === Data ================================================================= */
50 static bool kfence_enabled __read_mostly;
51 static bool disabled_by_warn __read_mostly;
53 unsigned long kfence_sample_interval __read_mostly = CONFIG_KFENCE_SAMPLE_INTERVAL;
54 EXPORT_SYMBOL_GPL(kfence_sample_interval); /* Export for test modules. */
56 #ifdef MODULE_PARAM_PREFIX
57 #undef MODULE_PARAM_PREFIX
59 #define MODULE_PARAM_PREFIX "kfence."
61 static int kfence_enable_late(void);
62 static int param_set_sample_interval(const char *val, const struct kernel_param *kp)
65 int ret = kstrtoul(val, 0, &num);
70 if (!num) /* Using 0 to indicate KFENCE is disabled. */
71 WRITE_ONCE(kfence_enabled, false);
73 *((unsigned long *)kp->arg) = num;
75 if (num && !READ_ONCE(kfence_enabled) && system_state != SYSTEM_BOOTING)
76 return disabled_by_warn ? -EINVAL : kfence_enable_late();
80 static int param_get_sample_interval(char *buffer, const struct kernel_param *kp)
82 if (!READ_ONCE(kfence_enabled))
83 return sprintf(buffer, "0\n");
85 return param_get_ulong(buffer, kp);
88 static const struct kernel_param_ops sample_interval_param_ops = {
89 .set = param_set_sample_interval,
90 .get = param_get_sample_interval,
92 module_param_cb(sample_interval, &sample_interval_param_ops, &kfence_sample_interval, 0600);
94 /* Pool usage% threshold when currently covered allocations are skipped. */
95 static unsigned long kfence_skip_covered_thresh __read_mostly = 75;
96 module_param_named(skip_covered_thresh, kfence_skip_covered_thresh, ulong, 0644);
98 /* If true, use a deferrable timer. */
99 static bool kfence_deferrable __read_mostly = IS_ENABLED(CONFIG_KFENCE_DEFERRABLE);
100 module_param_named(deferrable, kfence_deferrable, bool, 0444);
102 /* The pool of pages used for guard pages and objects. */
103 char *__kfence_pool __read_mostly;
104 EXPORT_SYMBOL(__kfence_pool); /* Export for test modules. */
107 * Per-object metadata, with one-to-one mapping of object metadata to
108 * backing pages (in __kfence_pool).
110 static_assert(CONFIG_KFENCE_NUM_OBJECTS > 0);
111 struct kfence_metadata kfence_metadata[CONFIG_KFENCE_NUM_OBJECTS];
113 /* Freelist with available objects. */
114 static struct list_head kfence_freelist = LIST_HEAD_INIT(kfence_freelist);
115 static DEFINE_RAW_SPINLOCK(kfence_freelist_lock); /* Lock protecting freelist. */
118 * The static key to set up a KFENCE allocation; or if static keys are not used
119 * to gate allocations, to avoid a load and compare if KFENCE is disabled.
121 DEFINE_STATIC_KEY_FALSE(kfence_allocation_key);
123 /* Gates the allocation, ensuring only one succeeds in a given period. */
124 atomic_t kfence_allocation_gate = ATOMIC_INIT(1);
127 * A Counting Bloom filter of allocation coverage: limits currently covered
128 * allocations of the same source filling up the pool.
130 * Assuming a range of 15%-85% unique allocations in the pool at any point in
131 * time, the below parameters provide a probablity of 0.02-0.33 for false
132 * positive hits respectively:
134 * P(alloc_traces) = (1 - e^(-HNUM * (alloc_traces / SIZE)) ^ HNUM
136 #define ALLOC_COVERED_HNUM 2
137 #define ALLOC_COVERED_ORDER (const_ilog2(CONFIG_KFENCE_NUM_OBJECTS) + 2)
138 #define ALLOC_COVERED_SIZE (1 << ALLOC_COVERED_ORDER)
139 #define ALLOC_COVERED_HNEXT(h) hash_32(h, ALLOC_COVERED_ORDER)
140 #define ALLOC_COVERED_MASK (ALLOC_COVERED_SIZE - 1)
141 static atomic_t alloc_covered[ALLOC_COVERED_SIZE];
143 /* Stack depth used to determine uniqueness of an allocation. */
144 #define UNIQUE_ALLOC_STACK_DEPTH ((size_t)8)
147 * Randomness for stack hashes, making the same collisions across reboots and
148 * different machines less likely.
150 static u32 stack_hash_seed __ro_after_init;
152 /* Statistics counters for debugfs. */
153 enum kfence_counter_id {
154 KFENCE_COUNTER_ALLOCATED,
155 KFENCE_COUNTER_ALLOCS,
156 KFENCE_COUNTER_FREES,
157 KFENCE_COUNTER_ZOMBIES,
159 KFENCE_COUNTER_SKIP_INCOMPAT,
160 KFENCE_COUNTER_SKIP_CAPACITY,
161 KFENCE_COUNTER_SKIP_COVERED,
162 KFENCE_COUNTER_COUNT,
164 static atomic_long_t counters[KFENCE_COUNTER_COUNT];
165 static const char *const counter_names[] = {
166 [KFENCE_COUNTER_ALLOCATED] = "currently allocated",
167 [KFENCE_COUNTER_ALLOCS] = "total allocations",
168 [KFENCE_COUNTER_FREES] = "total frees",
169 [KFENCE_COUNTER_ZOMBIES] = "zombie allocations",
170 [KFENCE_COUNTER_BUGS] = "total bugs",
171 [KFENCE_COUNTER_SKIP_INCOMPAT] = "skipped allocations (incompatible)",
172 [KFENCE_COUNTER_SKIP_CAPACITY] = "skipped allocations (capacity)",
173 [KFENCE_COUNTER_SKIP_COVERED] = "skipped allocations (covered)",
175 static_assert(ARRAY_SIZE(counter_names) == KFENCE_COUNTER_COUNT);
177 /* === Internals ============================================================ */
179 static inline bool should_skip_covered(void)
181 unsigned long thresh = (CONFIG_KFENCE_NUM_OBJECTS * kfence_skip_covered_thresh) / 100;
183 return atomic_long_read(&counters[KFENCE_COUNTER_ALLOCATED]) > thresh;
186 static u32 get_alloc_stack_hash(unsigned long *stack_entries, size_t num_entries)
188 num_entries = min(num_entries, UNIQUE_ALLOC_STACK_DEPTH);
189 num_entries = filter_irq_stacks(stack_entries, num_entries);
190 return jhash(stack_entries, num_entries * sizeof(stack_entries[0]), stack_hash_seed);
194 * Adds (or subtracts) count @val for allocation stack trace hash
195 * @alloc_stack_hash from Counting Bloom filter.
197 static void alloc_covered_add(u32 alloc_stack_hash, int val)
201 for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
202 atomic_add(val, &alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]);
203 alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
208 * Returns true if the allocation stack trace hash @alloc_stack_hash is
209 * currently contained (non-zero count) in Counting Bloom filter.
211 static bool alloc_covered_contains(u32 alloc_stack_hash)
215 for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
216 if (!atomic_read(&alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]))
218 alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
224 static bool kfence_protect(unsigned long addr)
226 return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), true));
229 static bool kfence_unprotect(unsigned long addr)
231 return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), false));
234 static inline unsigned long metadata_to_pageaddr(const struct kfence_metadata *meta)
236 unsigned long offset = (meta - kfence_metadata + 1) * PAGE_SIZE * 2;
237 unsigned long pageaddr = (unsigned long)&__kfence_pool[offset];
239 /* The checks do not affect performance; only called from slow-paths. */
241 /* Only call with a pointer into kfence_metadata. */
242 if (KFENCE_WARN_ON(meta < kfence_metadata ||
243 meta >= kfence_metadata + CONFIG_KFENCE_NUM_OBJECTS))
247 * This metadata object only ever maps to 1 page; verify that the stored
248 * address is in the expected range.
250 if (KFENCE_WARN_ON(ALIGN_DOWN(meta->addr, PAGE_SIZE) != pageaddr))
257 * Update the object's metadata state, including updating the alloc/free stacks
258 * depending on the state transition.
261 metadata_update_state(struct kfence_metadata *meta, enum kfence_object_state next,
262 unsigned long *stack_entries, size_t num_stack_entries)
264 struct kfence_track *track =
265 next == KFENCE_OBJECT_FREED ? &meta->free_track : &meta->alloc_track;
267 lockdep_assert_held(&meta->lock);
270 memcpy(track->stack_entries, stack_entries,
271 num_stack_entries * sizeof(stack_entries[0]));
274 * Skip over 1 (this) functions; noinline ensures we do not
275 * accidentally skip over the caller by never inlining.
277 num_stack_entries = stack_trace_save(track->stack_entries, KFENCE_STACK_DEPTH, 1);
279 track->num_stack_entries = num_stack_entries;
280 track->pid = task_pid_nr(current);
281 track->cpu = raw_smp_processor_id();
282 track->ts_nsec = local_clock(); /* Same source as printk timestamps. */
285 * Pairs with READ_ONCE() in
286 * kfence_shutdown_cache(),
287 * kfence_handle_page_fault().
289 WRITE_ONCE(meta->state, next);
292 /* Write canary byte to @addr. */
293 static inline bool set_canary_byte(u8 *addr)
295 *addr = KFENCE_CANARY_PATTERN(addr);
299 /* Check canary byte at @addr. */
300 static inline bool check_canary_byte(u8 *addr)
302 struct kfence_metadata *meta;
305 if (likely(*addr == KFENCE_CANARY_PATTERN(addr)))
308 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
310 meta = addr_to_metadata((unsigned long)addr);
311 raw_spin_lock_irqsave(&meta->lock, flags);
312 kfence_report_error((unsigned long)addr, false, NULL, meta, KFENCE_ERROR_CORRUPTION);
313 raw_spin_unlock_irqrestore(&meta->lock, flags);
318 /* __always_inline this to ensure we won't do an indirect call to fn. */
319 static __always_inline void for_each_canary(const struct kfence_metadata *meta, bool (*fn)(u8 *))
321 const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
325 * We'll iterate over each canary byte per-side until fn() returns
326 * false. However, we'll still iterate over the canary bytes to the
327 * right of the object even if there was an error in the canary bytes to
328 * the left of the object. Specifically, if check_canary_byte()
329 * generates an error, showing both sides might give more clues as to
330 * what the error is about when displaying which bytes were corrupted.
333 /* Apply to left of object. */
334 for (addr = pageaddr; addr < meta->addr; addr++) {
339 /* Apply to right of object. */
340 for (addr = meta->addr + meta->size; addr < pageaddr + PAGE_SIZE; addr++) {
346 static void *kfence_guarded_alloc(struct kmem_cache *cache, size_t size, gfp_t gfp,
347 unsigned long *stack_entries, size_t num_stack_entries,
348 u32 alloc_stack_hash)
350 struct kfence_metadata *meta = NULL;
355 /* Try to obtain a free object. */
356 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
357 if (!list_empty(&kfence_freelist)) {
358 meta = list_entry(kfence_freelist.next, struct kfence_metadata, list);
359 list_del_init(&meta->list);
361 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
363 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_CAPACITY]);
367 if (unlikely(!raw_spin_trylock_irqsave(&meta->lock, flags))) {
369 * This is extremely unlikely -- we are reporting on a
370 * use-after-free, which locked meta->lock, and the reporting
371 * code via printk calls kmalloc() which ends up in
372 * kfence_alloc() and tries to grab the same object that we're
373 * reporting on. While it has never been observed, lockdep does
374 * report that there is a possibility of deadlock. Fix it by
375 * using trylock and bailing out gracefully.
377 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
378 /* Put the object back on the freelist. */
379 list_add_tail(&meta->list, &kfence_freelist);
380 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
385 meta->addr = metadata_to_pageaddr(meta);
386 /* Unprotect if we're reusing this page. */
387 if (meta->state == KFENCE_OBJECT_FREED)
388 kfence_unprotect(meta->addr);
391 * Note: for allocations made before RNG initialization, will always
392 * return zero. We still benefit from enabling KFENCE as early as
393 * possible, even when the RNG is not yet available, as this will allow
394 * KFENCE to detect bugs due to earlier allocations. The only downside
395 * is that the out-of-bounds accesses detected are deterministic for
398 if (prandom_u32_max(2)) {
399 /* Allocate on the "right" side, re-calculate address. */
400 meta->addr += PAGE_SIZE - size;
401 meta->addr = ALIGN_DOWN(meta->addr, cache->align);
404 addr = (void *)meta->addr;
406 /* Update remaining metadata. */
407 metadata_update_state(meta, KFENCE_OBJECT_ALLOCATED, stack_entries, num_stack_entries);
408 /* Pairs with READ_ONCE() in kfence_shutdown_cache(). */
409 WRITE_ONCE(meta->cache, cache);
411 meta->alloc_stack_hash = alloc_stack_hash;
412 raw_spin_unlock_irqrestore(&meta->lock, flags);
414 alloc_covered_add(alloc_stack_hash, 1);
416 /* Set required slab fields. */
417 slab = virt_to_slab((void *)meta->addr);
418 slab->slab_cache = cache;
419 #if defined(CONFIG_SLUB)
421 #elif defined(CONFIG_SLAB)
425 /* Memory initialization. */
426 for_each_canary(meta, set_canary_byte);
429 * We check slab_want_init_on_alloc() ourselves, rather than letting
430 * SL*B do the initialization, as otherwise we might overwrite KFENCE's
433 if (unlikely(slab_want_init_on_alloc(gfp, cache)))
434 memzero_explicit(addr, size);
438 if (CONFIG_KFENCE_STRESS_TEST_FAULTS && !prandom_u32_max(CONFIG_KFENCE_STRESS_TEST_FAULTS))
439 kfence_protect(meta->addr); /* Random "faults" by protecting the object. */
441 atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCATED]);
442 atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCS]);
447 static void kfence_guarded_free(void *addr, struct kfence_metadata *meta, bool zombie)
449 struct kcsan_scoped_access assert_page_exclusive;
453 raw_spin_lock_irqsave(&meta->lock, flags);
455 if (meta->state != KFENCE_OBJECT_ALLOCATED || meta->addr != (unsigned long)addr) {
456 /* Invalid or double-free, bail out. */
457 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
458 kfence_report_error((unsigned long)addr, false, NULL, meta,
459 KFENCE_ERROR_INVALID_FREE);
460 raw_spin_unlock_irqrestore(&meta->lock, flags);
464 /* Detect racy use-after-free, or incorrect reallocation of this page by KFENCE. */
465 kcsan_begin_scoped_access((void *)ALIGN_DOWN((unsigned long)addr, PAGE_SIZE), PAGE_SIZE,
466 KCSAN_ACCESS_SCOPED | KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT,
467 &assert_page_exclusive);
469 if (CONFIG_KFENCE_STRESS_TEST_FAULTS)
470 kfence_unprotect((unsigned long)addr); /* To check canary bytes. */
472 /* Restore page protection if there was an OOB access. */
473 if (meta->unprotected_page) {
474 memzero_explicit((void *)ALIGN_DOWN(meta->unprotected_page, PAGE_SIZE), PAGE_SIZE);
475 kfence_protect(meta->unprotected_page);
476 meta->unprotected_page = 0;
479 /* Mark the object as freed. */
480 metadata_update_state(meta, KFENCE_OBJECT_FREED, NULL, 0);
481 init = slab_want_init_on_free(meta->cache);
482 raw_spin_unlock_irqrestore(&meta->lock, flags);
484 alloc_covered_add(meta->alloc_stack_hash, -1);
486 /* Check canary bytes for memory corruption. */
487 for_each_canary(meta, check_canary_byte);
490 * Clear memory if init-on-free is set. While we protect the page, the
491 * data is still there, and after a use-after-free is detected, we
492 * unprotect the page, so the data is still accessible.
494 if (!zombie && unlikely(init))
495 memzero_explicit(addr, meta->size);
497 /* Protect to detect use-after-frees. */
498 kfence_protect((unsigned long)addr);
500 kcsan_end_scoped_access(&assert_page_exclusive);
502 /* Add it to the tail of the freelist for reuse. */
503 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
504 KFENCE_WARN_ON(!list_empty(&meta->list));
505 list_add_tail(&meta->list, &kfence_freelist);
506 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
508 atomic_long_dec(&counters[KFENCE_COUNTER_ALLOCATED]);
509 atomic_long_inc(&counters[KFENCE_COUNTER_FREES]);
511 /* See kfence_shutdown_cache(). */
512 atomic_long_inc(&counters[KFENCE_COUNTER_ZOMBIES]);
516 static void rcu_guarded_free(struct rcu_head *h)
518 struct kfence_metadata *meta = container_of(h, struct kfence_metadata, rcu_head);
520 kfence_guarded_free((void *)meta->addr, meta, false);
524 * Initialization of the KFENCE pool after its allocation.
525 * Returns 0 on success; otherwise returns the address up to
526 * which partial initialization succeeded.
528 static unsigned long kfence_init_pool(void)
530 unsigned long addr = (unsigned long)__kfence_pool;
534 if (!arch_kfence_init_pool())
537 pages = virt_to_page(addr);
540 * Set up object pages: they must have PG_slab set, to avoid freeing
541 * these as real pages.
543 * We also want to avoid inserting kfence_free() in the kfree()
544 * fast-path in SLUB, and therefore need to ensure kfree() correctly
545 * enters __slab_free() slow-path.
547 for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
548 struct slab *slab = page_slab(&pages[i]);
553 /* Verify we do not have a compound head page. */
554 if (WARN_ON(compound_head(&pages[i]) != &pages[i]))
557 __folio_set_slab(slab_folio(slab));
559 slab->memcg_data = (unsigned long)&kfence_metadata[i / 2 - 1].objcg |
565 * Protect the first 2 pages. The first page is mostly unnecessary, and
566 * merely serves as an extended guard page. However, adding one
567 * additional page in the beginning gives us an even number of pages,
568 * which simplifies the mapping of address to metadata index.
570 for (i = 0; i < 2; i++) {
571 if (unlikely(!kfence_protect(addr)))
577 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
578 struct kfence_metadata *meta = &kfence_metadata[i];
580 /* Initialize metadata. */
581 INIT_LIST_HEAD(&meta->list);
582 raw_spin_lock_init(&meta->lock);
583 meta->state = KFENCE_OBJECT_UNUSED;
584 meta->addr = addr; /* Initialize for validation in metadata_to_pageaddr(). */
585 list_add_tail(&meta->list, &kfence_freelist);
587 /* Protect the right redzone. */
588 if (unlikely(!kfence_protect(addr + PAGE_SIZE)))
591 addr += 2 * PAGE_SIZE;
595 * The pool is live and will never be deallocated from this point on.
596 * Remove the pool object from the kmemleak object tree, as it would
597 * otherwise overlap with allocations returned by kfence_alloc(), which
598 * are registered with kmemleak through the slab post-alloc hook.
600 kmemleak_free(__kfence_pool);
605 static bool __init kfence_init_pool_early(void)
612 addr = kfence_init_pool();
618 * Only release unprotected pages, and do not try to go back and change
619 * page attributes due to risk of failing to do so as well. If changing
620 * page attributes for some pages fails, it is very likely that it also
621 * fails for the first page, and therefore expect addr==__kfence_pool in
622 * most failure cases.
624 for (char *p = (char *)addr; p < __kfence_pool + KFENCE_POOL_SIZE; p += PAGE_SIZE) {
625 struct slab *slab = virt_to_slab(p);
630 slab->memcg_data = 0;
632 __folio_clear_slab(slab_folio(slab));
634 memblock_free_late(__pa(addr), KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool));
635 __kfence_pool = NULL;
639 static bool kfence_init_pool_late(void)
641 unsigned long addr, free_size;
643 addr = kfence_init_pool();
649 free_size = KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool);
650 #ifdef CONFIG_CONTIG_ALLOC
651 free_contig_range(page_to_pfn(virt_to_page(addr)), free_size / PAGE_SIZE);
653 free_pages_exact((void *)addr, free_size);
655 __kfence_pool = NULL;
659 /* === DebugFS Interface ==================================================== */
661 static int stats_show(struct seq_file *seq, void *v)
665 seq_printf(seq, "enabled: %i\n", READ_ONCE(kfence_enabled));
666 for (i = 0; i < KFENCE_COUNTER_COUNT; i++)
667 seq_printf(seq, "%s: %ld\n", counter_names[i], atomic_long_read(&counters[i]));
671 DEFINE_SHOW_ATTRIBUTE(stats);
674 * debugfs seq_file operations for /sys/kernel/debug/kfence/objects.
675 * start_object() and next_object() return the object index + 1, because NULL is used
678 static void *start_object(struct seq_file *seq, loff_t *pos)
680 if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
681 return (void *)((long)*pos + 1);
685 static void stop_object(struct seq_file *seq, void *v)
689 static void *next_object(struct seq_file *seq, void *v, loff_t *pos)
692 if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
693 return (void *)((long)*pos + 1);
697 static int show_object(struct seq_file *seq, void *v)
699 struct kfence_metadata *meta = &kfence_metadata[(long)v - 1];
702 raw_spin_lock_irqsave(&meta->lock, flags);
703 kfence_print_object(seq, meta);
704 raw_spin_unlock_irqrestore(&meta->lock, flags);
705 seq_puts(seq, "---------------------------------\n");
710 static const struct seq_operations object_seqops = {
711 .start = start_object,
717 static int open_objects(struct inode *inode, struct file *file)
719 return seq_open(file, &object_seqops);
722 static const struct file_operations objects_fops = {
723 .open = open_objects,
726 .release = seq_release,
729 static int __init kfence_debugfs_init(void)
731 struct dentry *kfence_dir = debugfs_create_dir("kfence", NULL);
733 debugfs_create_file("stats", 0444, kfence_dir, NULL, &stats_fops);
734 debugfs_create_file("objects", 0400, kfence_dir, NULL, &objects_fops);
738 late_initcall(kfence_debugfs_init);
740 /* === Allocation Gate Timer ================================================ */
742 static struct delayed_work kfence_timer;
744 #ifdef CONFIG_KFENCE_STATIC_KEYS
745 /* Wait queue to wake up allocation-gate timer task. */
746 static DECLARE_WAIT_QUEUE_HEAD(allocation_wait);
748 static void wake_up_kfence_timer(struct irq_work *work)
750 wake_up(&allocation_wait);
752 static DEFINE_IRQ_WORK(wake_up_kfence_timer_work, wake_up_kfence_timer);
756 * Set up delayed work, which will enable and disable the static key. We need to
757 * use a work queue (rather than a simple timer), since enabling and disabling a
758 * static key cannot be done from an interrupt.
760 * Note: Toggling a static branch currently causes IPIs, and here we'll end up
761 * with a total of 2 IPIs to all CPUs. If this ends up a problem in future (with
762 * more aggressive sampling intervals), we could get away with a variant that
763 * avoids IPIs, at the cost of not immediately capturing allocations if the
764 * instructions remain cached.
766 static void toggle_allocation_gate(struct work_struct *work)
768 if (!READ_ONCE(kfence_enabled))
771 atomic_set(&kfence_allocation_gate, 0);
772 #ifdef CONFIG_KFENCE_STATIC_KEYS
773 /* Enable static key, and await allocation to happen. */
774 static_branch_enable(&kfence_allocation_key);
776 if (sysctl_hung_task_timeout_secs) {
778 * During low activity with no allocations we might wait a
779 * while; let's avoid the hung task warning.
781 wait_event_idle_timeout(allocation_wait, atomic_read(&kfence_allocation_gate),
782 sysctl_hung_task_timeout_secs * HZ / 2);
784 wait_event_idle(allocation_wait, atomic_read(&kfence_allocation_gate));
787 /* Disable static key and reset timer. */
788 static_branch_disable(&kfence_allocation_key);
790 queue_delayed_work(system_unbound_wq, &kfence_timer,
791 msecs_to_jiffies(kfence_sample_interval));
794 /* === Public interface ===================================================== */
796 void __init kfence_alloc_pool(void)
798 if (!kfence_sample_interval)
801 __kfence_pool = memblock_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
804 pr_err("failed to allocate pool\n");
807 static void kfence_init_enable(void)
809 if (!IS_ENABLED(CONFIG_KFENCE_STATIC_KEYS))
810 static_branch_enable(&kfence_allocation_key);
812 if (kfence_deferrable)
813 INIT_DEFERRABLE_WORK(&kfence_timer, toggle_allocation_gate);
815 INIT_DELAYED_WORK(&kfence_timer, toggle_allocation_gate);
817 WRITE_ONCE(kfence_enabled, true);
818 queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
820 pr_info("initialized - using %lu bytes for %d objects at 0x%p-0x%p\n", KFENCE_POOL_SIZE,
821 CONFIG_KFENCE_NUM_OBJECTS, (void *)__kfence_pool,
822 (void *)(__kfence_pool + KFENCE_POOL_SIZE));
825 void __init kfence_init(void)
827 stack_hash_seed = (u32)random_get_entropy();
829 /* Setting kfence_sample_interval to 0 on boot disables KFENCE. */
830 if (!kfence_sample_interval)
833 if (!kfence_init_pool_early()) {
834 pr_err("%s failed\n", __func__);
838 kfence_init_enable();
841 static int kfence_init_late(void)
843 const unsigned long nr_pages = KFENCE_POOL_SIZE / PAGE_SIZE;
844 #ifdef CONFIG_CONTIG_ALLOC
847 pages = alloc_contig_pages(nr_pages, GFP_KERNEL, first_online_node, NULL);
850 __kfence_pool = page_to_virt(pages);
852 if (nr_pages > MAX_ORDER_NR_PAGES) {
853 pr_warn("KFENCE_NUM_OBJECTS too large for buddy allocator\n");
856 __kfence_pool = alloc_pages_exact(KFENCE_POOL_SIZE, GFP_KERNEL);
861 if (!kfence_init_pool_late()) {
862 pr_err("%s failed\n", __func__);
866 kfence_init_enable();
870 static int kfence_enable_late(void)
873 return kfence_init_late();
875 WRITE_ONCE(kfence_enabled, true);
876 queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
880 void kfence_shutdown_cache(struct kmem_cache *s)
883 struct kfence_metadata *meta;
886 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
889 meta = &kfence_metadata[i];
892 * If we observe some inconsistent cache and state pair where we
893 * should have returned false here, cache destruction is racing
894 * with either kmem_cache_alloc() or kmem_cache_free(). Taking
895 * the lock will not help, as different critical section
896 * serialization will have the same outcome.
898 if (READ_ONCE(meta->cache) != s ||
899 READ_ONCE(meta->state) != KFENCE_OBJECT_ALLOCATED)
902 raw_spin_lock_irqsave(&meta->lock, flags);
903 in_use = meta->cache == s && meta->state == KFENCE_OBJECT_ALLOCATED;
904 raw_spin_unlock_irqrestore(&meta->lock, flags);
908 * This cache still has allocations, and we should not
909 * release them back into the freelist so they can still
910 * safely be used and retain the kernel's default
911 * behaviour of keeping the allocations alive (leak the
912 * cache); however, they effectively become "zombie
913 * allocations" as the KFENCE objects are the only ones
914 * still in use and the owning cache is being destroyed.
916 * We mark them freed, so that any subsequent use shows
917 * more useful error messages that will include stack
918 * traces of the user of the object, the original
919 * allocation, and caller to shutdown_cache().
921 kfence_guarded_free((void *)meta->addr, meta, /*zombie=*/true);
925 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
926 meta = &kfence_metadata[i];
929 if (READ_ONCE(meta->cache) != s || READ_ONCE(meta->state) != KFENCE_OBJECT_FREED)
932 raw_spin_lock_irqsave(&meta->lock, flags);
933 if (meta->cache == s && meta->state == KFENCE_OBJECT_FREED)
935 raw_spin_unlock_irqrestore(&meta->lock, flags);
939 void *__kfence_alloc(struct kmem_cache *s, size_t size, gfp_t flags)
941 unsigned long stack_entries[KFENCE_STACK_DEPTH];
942 size_t num_stack_entries;
943 u32 alloc_stack_hash;
946 * Perform size check before switching kfence_allocation_gate, so that
947 * we don't disable KFENCE without making an allocation.
949 if (size > PAGE_SIZE) {
950 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
955 * Skip allocations from non-default zones, including DMA. We cannot
956 * guarantee that pages in the KFENCE pool will have the requested
957 * properties (e.g. reside in DMAable memory).
959 if ((flags & GFP_ZONEMASK) ||
960 (s->flags & (SLAB_CACHE_DMA | SLAB_CACHE_DMA32))) {
961 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
965 if (atomic_inc_return(&kfence_allocation_gate) > 1)
967 #ifdef CONFIG_KFENCE_STATIC_KEYS
969 * waitqueue_active() is fully ordered after the update of
970 * kfence_allocation_gate per atomic_inc_return().
972 if (waitqueue_active(&allocation_wait)) {
974 * Calling wake_up() here may deadlock when allocations happen
975 * from within timer code. Use an irq_work to defer it.
977 irq_work_queue(&wake_up_kfence_timer_work);
981 if (!READ_ONCE(kfence_enabled))
984 num_stack_entries = stack_trace_save(stack_entries, KFENCE_STACK_DEPTH, 0);
987 * Do expensive check for coverage of allocation in slow-path after
988 * allocation_gate has already become non-zero, even though it might
989 * mean not making any allocation within a given sample interval.
991 * This ensures reasonable allocation coverage when the pool is almost
992 * full, including avoiding long-lived allocations of the same source
993 * filling up the pool (e.g. pagecache allocations).
995 alloc_stack_hash = get_alloc_stack_hash(stack_entries, num_stack_entries);
996 if (should_skip_covered() && alloc_covered_contains(alloc_stack_hash)) {
997 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_COVERED]);
1001 return kfence_guarded_alloc(s, size, flags, stack_entries, num_stack_entries,
1005 size_t kfence_ksize(const void *addr)
1007 const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1010 * Read locklessly -- if there is a race with __kfence_alloc(), this is
1011 * either a use-after-free or invalid access.
1013 return meta ? meta->size : 0;
1016 void *kfence_object_start(const void *addr)
1018 const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1021 * Read locklessly -- if there is a race with __kfence_alloc(), this is
1022 * either a use-after-free or invalid access.
1024 return meta ? (void *)meta->addr : NULL;
1027 void __kfence_free(void *addr)
1029 struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1032 KFENCE_WARN_ON(meta->objcg);
1035 * If the objects of the cache are SLAB_TYPESAFE_BY_RCU, defer freeing
1036 * the object, as the object page may be recycled for other-typed
1037 * objects once it has been freed. meta->cache may be NULL if the cache
1040 if (unlikely(meta->cache && (meta->cache->flags & SLAB_TYPESAFE_BY_RCU)))
1041 call_rcu(&meta->rcu_head, rcu_guarded_free);
1043 kfence_guarded_free(addr, meta, false);
1046 bool kfence_handle_page_fault(unsigned long addr, bool is_write, struct pt_regs *regs)
1048 const int page_index = (addr - (unsigned long)__kfence_pool) / PAGE_SIZE;
1049 struct kfence_metadata *to_report = NULL;
1050 enum kfence_error_type error_type;
1051 unsigned long flags;
1053 if (!is_kfence_address((void *)addr))
1056 if (!READ_ONCE(kfence_enabled)) /* If disabled at runtime ... */
1057 return kfence_unprotect(addr); /* ... unprotect and proceed. */
1059 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
1061 if (page_index % 2) {
1062 /* This is a redzone, report a buffer overflow. */
1063 struct kfence_metadata *meta;
1066 meta = addr_to_metadata(addr - PAGE_SIZE);
1067 if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
1069 /* Data race ok; distance calculation approximate. */
1070 distance = addr - data_race(meta->addr + meta->size);
1073 meta = addr_to_metadata(addr + PAGE_SIZE);
1074 if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
1075 /* Data race ok; distance calculation approximate. */
1076 if (!to_report || distance > data_race(meta->addr) - addr)
1083 raw_spin_lock_irqsave(&to_report->lock, flags);
1084 to_report->unprotected_page = addr;
1085 error_type = KFENCE_ERROR_OOB;
1088 * If the object was freed before we took the look we can still
1089 * report this as an OOB -- the report will simply show the
1090 * stacktrace of the free as well.
1093 to_report = addr_to_metadata(addr);
1097 raw_spin_lock_irqsave(&to_report->lock, flags);
1098 error_type = KFENCE_ERROR_UAF;
1100 * We may race with __kfence_alloc(), and it is possible that a
1101 * freed object may be reallocated. We simply report this as a
1102 * use-after-free, with the stack trace showing the place where
1103 * the object was re-allocated.
1109 kfence_report_error(addr, is_write, regs, to_report, error_type);
1110 raw_spin_unlock_irqrestore(&to_report->lock, flags);
1112 /* This may be a UAF or OOB access, but we can't be sure. */
1113 kfence_report_error(addr, is_write, regs, NULL, KFENCE_ERROR_INVALID);
1116 return kfence_unprotect(addr); /* Unprotect and let access proceed. */