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 struct kfence_metadata *addr_to_metadata(unsigned long addr)
238 /* The checks do not affect performance; only called from slow-paths. */
240 if (!is_kfence_address((void *)addr))
244 * May be an invalid index if called with an address at the edge of
245 * __kfence_pool, in which case we would report an "invalid access"
248 index = (addr - (unsigned long)__kfence_pool) / (PAGE_SIZE * 2) - 1;
249 if (index < 0 || index >= CONFIG_KFENCE_NUM_OBJECTS)
252 return &kfence_metadata[index];
255 static inline unsigned long metadata_to_pageaddr(const struct kfence_metadata *meta)
257 unsigned long offset = (meta - kfence_metadata + 1) * PAGE_SIZE * 2;
258 unsigned long pageaddr = (unsigned long)&__kfence_pool[offset];
260 /* The checks do not affect performance; only called from slow-paths. */
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))
268 * This metadata object only ever maps to 1 page; verify that the stored
269 * address is in the expected range.
271 if (KFENCE_WARN_ON(ALIGN_DOWN(meta->addr, PAGE_SIZE) != pageaddr))
278 * Update the object's metadata state, including updating the alloc/free stacks
279 * depending on the state transition.
282 metadata_update_state(struct kfence_metadata *meta, enum kfence_object_state next,
283 unsigned long *stack_entries, size_t num_stack_entries)
285 struct kfence_track *track =
286 next == KFENCE_OBJECT_FREED ? &meta->free_track : &meta->alloc_track;
288 lockdep_assert_held(&meta->lock);
291 memcpy(track->stack_entries, stack_entries,
292 num_stack_entries * sizeof(stack_entries[0]));
295 * Skip over 1 (this) functions; noinline ensures we do not
296 * accidentally skip over the caller by never inlining.
298 num_stack_entries = stack_trace_save(track->stack_entries, KFENCE_STACK_DEPTH, 1);
300 track->num_stack_entries = num_stack_entries;
301 track->pid = task_pid_nr(current);
302 track->cpu = raw_smp_processor_id();
303 track->ts_nsec = local_clock(); /* Same source as printk timestamps. */
306 * Pairs with READ_ONCE() in
307 * kfence_shutdown_cache(),
308 * kfence_handle_page_fault().
310 WRITE_ONCE(meta->state, next);
313 /* Write canary byte to @addr. */
314 static inline bool set_canary_byte(u8 *addr)
316 *addr = KFENCE_CANARY_PATTERN(addr);
320 /* Check canary byte at @addr. */
321 static inline bool check_canary_byte(u8 *addr)
323 struct kfence_metadata *meta;
326 if (likely(*addr == KFENCE_CANARY_PATTERN(addr)))
329 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
331 meta = addr_to_metadata((unsigned long)addr);
332 raw_spin_lock_irqsave(&meta->lock, flags);
333 kfence_report_error((unsigned long)addr, false, NULL, meta, KFENCE_ERROR_CORRUPTION);
334 raw_spin_unlock_irqrestore(&meta->lock, flags);
339 /* __always_inline this to ensure we won't do an indirect call to fn. */
340 static __always_inline void for_each_canary(const struct kfence_metadata *meta, bool (*fn)(u8 *))
342 const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
346 * We'll iterate over each canary byte per-side until fn() returns
347 * false. However, we'll still iterate over the canary bytes to the
348 * right of the object even if there was an error in the canary bytes to
349 * the left of the object. Specifically, if check_canary_byte()
350 * generates an error, showing both sides might give more clues as to
351 * what the error is about when displaying which bytes were corrupted.
354 /* Apply to left of object. */
355 for (addr = pageaddr; addr < meta->addr; addr++) {
360 /* Apply to right of object. */
361 for (addr = meta->addr + meta->size; addr < pageaddr + PAGE_SIZE; addr++) {
367 static void *kfence_guarded_alloc(struct kmem_cache *cache, size_t size, gfp_t gfp,
368 unsigned long *stack_entries, size_t num_stack_entries,
369 u32 alloc_stack_hash)
371 struct kfence_metadata *meta = NULL;
376 /* Try to obtain a free object. */
377 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
378 if (!list_empty(&kfence_freelist)) {
379 meta = list_entry(kfence_freelist.next, struct kfence_metadata, list);
380 list_del_init(&meta->list);
382 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
384 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_CAPACITY]);
388 if (unlikely(!raw_spin_trylock_irqsave(&meta->lock, flags))) {
390 * This is extremely unlikely -- we are reporting on a
391 * use-after-free, which locked meta->lock, and the reporting
392 * code via printk calls kmalloc() which ends up in
393 * kfence_alloc() and tries to grab the same object that we're
394 * reporting on. While it has never been observed, lockdep does
395 * report that there is a possibility of deadlock. Fix it by
396 * using trylock and bailing out gracefully.
398 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
399 /* Put the object back on the freelist. */
400 list_add_tail(&meta->list, &kfence_freelist);
401 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
406 meta->addr = metadata_to_pageaddr(meta);
407 /* Unprotect if we're reusing this page. */
408 if (meta->state == KFENCE_OBJECT_FREED)
409 kfence_unprotect(meta->addr);
412 * Note: for allocations made before RNG initialization, will always
413 * return zero. We still benefit from enabling KFENCE as early as
414 * possible, even when the RNG is not yet available, as this will allow
415 * KFENCE to detect bugs due to earlier allocations. The only downside
416 * is that the out-of-bounds accesses detected are deterministic for
419 if (prandom_u32_max(2)) {
420 /* Allocate on the "right" side, re-calculate address. */
421 meta->addr += PAGE_SIZE - size;
422 meta->addr = ALIGN_DOWN(meta->addr, cache->align);
425 addr = (void *)meta->addr;
427 /* Update remaining metadata. */
428 metadata_update_state(meta, KFENCE_OBJECT_ALLOCATED, stack_entries, num_stack_entries);
429 /* Pairs with READ_ONCE() in kfence_shutdown_cache(). */
430 WRITE_ONCE(meta->cache, cache);
432 meta->alloc_stack_hash = alloc_stack_hash;
433 raw_spin_unlock_irqrestore(&meta->lock, flags);
435 alloc_covered_add(alloc_stack_hash, 1);
437 /* Set required slab fields. */
438 slab = virt_to_slab((void *)meta->addr);
439 slab->slab_cache = cache;
440 #if defined(CONFIG_SLUB)
442 #elif defined(CONFIG_SLAB)
446 /* Memory initialization. */
447 for_each_canary(meta, set_canary_byte);
450 * We check slab_want_init_on_alloc() ourselves, rather than letting
451 * SL*B do the initialization, as otherwise we might overwrite KFENCE's
454 if (unlikely(slab_want_init_on_alloc(gfp, cache)))
455 memzero_explicit(addr, size);
459 if (CONFIG_KFENCE_STRESS_TEST_FAULTS && !prandom_u32_max(CONFIG_KFENCE_STRESS_TEST_FAULTS))
460 kfence_protect(meta->addr); /* Random "faults" by protecting the object. */
462 atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCATED]);
463 atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCS]);
468 static void kfence_guarded_free(void *addr, struct kfence_metadata *meta, bool zombie)
470 struct kcsan_scoped_access assert_page_exclusive;
474 raw_spin_lock_irqsave(&meta->lock, flags);
476 if (meta->state != KFENCE_OBJECT_ALLOCATED || meta->addr != (unsigned long)addr) {
477 /* Invalid or double-free, bail out. */
478 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
479 kfence_report_error((unsigned long)addr, false, NULL, meta,
480 KFENCE_ERROR_INVALID_FREE);
481 raw_spin_unlock_irqrestore(&meta->lock, flags);
485 /* Detect racy use-after-free, or incorrect reallocation of this page by KFENCE. */
486 kcsan_begin_scoped_access((void *)ALIGN_DOWN((unsigned long)addr, PAGE_SIZE), PAGE_SIZE,
487 KCSAN_ACCESS_SCOPED | KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT,
488 &assert_page_exclusive);
490 if (CONFIG_KFENCE_STRESS_TEST_FAULTS)
491 kfence_unprotect((unsigned long)addr); /* To check canary bytes. */
493 /* Restore page protection if there was an OOB access. */
494 if (meta->unprotected_page) {
495 memzero_explicit((void *)ALIGN_DOWN(meta->unprotected_page, PAGE_SIZE), PAGE_SIZE);
496 kfence_protect(meta->unprotected_page);
497 meta->unprotected_page = 0;
500 /* Mark the object as freed. */
501 metadata_update_state(meta, KFENCE_OBJECT_FREED, NULL, 0);
502 init = slab_want_init_on_free(meta->cache);
503 raw_spin_unlock_irqrestore(&meta->lock, flags);
505 alloc_covered_add(meta->alloc_stack_hash, -1);
507 /* Check canary bytes for memory corruption. */
508 for_each_canary(meta, check_canary_byte);
511 * Clear memory if init-on-free is set. While we protect the page, the
512 * data is still there, and after a use-after-free is detected, we
513 * unprotect the page, so the data is still accessible.
515 if (!zombie && unlikely(init))
516 memzero_explicit(addr, meta->size);
518 /* Protect to detect use-after-frees. */
519 kfence_protect((unsigned long)addr);
521 kcsan_end_scoped_access(&assert_page_exclusive);
523 /* Add it to the tail of the freelist for reuse. */
524 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
525 KFENCE_WARN_ON(!list_empty(&meta->list));
526 list_add_tail(&meta->list, &kfence_freelist);
527 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
529 atomic_long_dec(&counters[KFENCE_COUNTER_ALLOCATED]);
530 atomic_long_inc(&counters[KFENCE_COUNTER_FREES]);
532 /* See kfence_shutdown_cache(). */
533 atomic_long_inc(&counters[KFENCE_COUNTER_ZOMBIES]);
537 static void rcu_guarded_free(struct rcu_head *h)
539 struct kfence_metadata *meta = container_of(h, struct kfence_metadata, rcu_head);
541 kfence_guarded_free((void *)meta->addr, meta, false);
545 * Initialization of the KFENCE pool after its allocation.
546 * Returns 0 on success; otherwise returns the address up to
547 * which partial initialization succeeded.
549 static unsigned long kfence_init_pool(void)
551 unsigned long addr = (unsigned long)__kfence_pool;
555 if (!arch_kfence_init_pool())
558 pages = virt_to_page(addr);
561 * Set up object pages: they must have PG_slab set, to avoid freeing
562 * these as real pages.
564 * We also want to avoid inserting kfence_free() in the kfree()
565 * fast-path in SLUB, and therefore need to ensure kfree() correctly
566 * enters __slab_free() slow-path.
568 for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
572 /* Verify we do not have a compound head page. */
573 if (WARN_ON(compound_head(&pages[i]) != &pages[i]))
576 __SetPageSlab(&pages[i]);
580 * Protect the first 2 pages. The first page is mostly unnecessary, and
581 * merely serves as an extended guard page. However, adding one
582 * additional page in the beginning gives us an even number of pages,
583 * which simplifies the mapping of address to metadata index.
585 for (i = 0; i < 2; i++) {
586 if (unlikely(!kfence_protect(addr)))
592 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
593 struct kfence_metadata *meta = &kfence_metadata[i];
595 /* Initialize metadata. */
596 INIT_LIST_HEAD(&meta->list);
597 raw_spin_lock_init(&meta->lock);
598 meta->state = KFENCE_OBJECT_UNUSED;
599 meta->addr = addr; /* Initialize for validation in metadata_to_pageaddr(). */
600 list_add_tail(&meta->list, &kfence_freelist);
602 /* Protect the right redzone. */
603 if (unlikely(!kfence_protect(addr + PAGE_SIZE)))
606 addr += 2 * PAGE_SIZE;
610 * The pool is live and will never be deallocated from this point on.
611 * Remove the pool object from the kmemleak object tree, as it would
612 * otherwise overlap with allocations returned by kfence_alloc(), which
613 * are registered with kmemleak through the slab post-alloc hook.
615 kmemleak_free(__kfence_pool);
620 static bool __init kfence_init_pool_early(void)
627 addr = kfence_init_pool();
633 * Only release unprotected pages, and do not try to go back and change
634 * page attributes due to risk of failing to do so as well. If changing
635 * page attributes for some pages fails, it is very likely that it also
636 * fails for the first page, and therefore expect addr==__kfence_pool in
637 * most failure cases.
639 memblock_free_late(__pa(addr), KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool));
640 __kfence_pool = NULL;
644 static bool kfence_init_pool_late(void)
646 unsigned long addr, free_size;
648 addr = kfence_init_pool();
654 free_size = KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool);
655 #ifdef CONFIG_CONTIG_ALLOC
656 free_contig_range(page_to_pfn(virt_to_page(addr)), free_size / PAGE_SIZE);
658 free_pages_exact((void *)addr, free_size);
660 __kfence_pool = NULL;
664 /* === DebugFS Interface ==================================================== */
666 static int stats_show(struct seq_file *seq, void *v)
670 seq_printf(seq, "enabled: %i\n", READ_ONCE(kfence_enabled));
671 for (i = 0; i < KFENCE_COUNTER_COUNT; i++)
672 seq_printf(seq, "%s: %ld\n", counter_names[i], atomic_long_read(&counters[i]));
676 DEFINE_SHOW_ATTRIBUTE(stats);
679 * debugfs seq_file operations for /sys/kernel/debug/kfence/objects.
680 * start_object() and next_object() return the object index + 1, because NULL is used
683 static void *start_object(struct seq_file *seq, loff_t *pos)
685 if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
686 return (void *)((long)*pos + 1);
690 static void stop_object(struct seq_file *seq, void *v)
694 static void *next_object(struct seq_file *seq, void *v, loff_t *pos)
697 if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
698 return (void *)((long)*pos + 1);
702 static int show_object(struct seq_file *seq, void *v)
704 struct kfence_metadata *meta = &kfence_metadata[(long)v - 1];
707 raw_spin_lock_irqsave(&meta->lock, flags);
708 kfence_print_object(seq, meta);
709 raw_spin_unlock_irqrestore(&meta->lock, flags);
710 seq_puts(seq, "---------------------------------\n");
715 static const struct seq_operations object_seqops = {
716 .start = start_object,
722 static int open_objects(struct inode *inode, struct file *file)
724 return seq_open(file, &object_seqops);
727 static const struct file_operations objects_fops = {
728 .open = open_objects,
731 .release = seq_release,
734 static int __init kfence_debugfs_init(void)
736 struct dentry *kfence_dir = debugfs_create_dir("kfence", NULL);
738 debugfs_create_file("stats", 0444, kfence_dir, NULL, &stats_fops);
739 debugfs_create_file("objects", 0400, kfence_dir, NULL, &objects_fops);
743 late_initcall(kfence_debugfs_init);
745 /* === Allocation Gate Timer ================================================ */
747 static struct delayed_work kfence_timer;
749 #ifdef CONFIG_KFENCE_STATIC_KEYS
750 /* Wait queue to wake up allocation-gate timer task. */
751 static DECLARE_WAIT_QUEUE_HEAD(allocation_wait);
753 static void wake_up_kfence_timer(struct irq_work *work)
755 wake_up(&allocation_wait);
757 static DEFINE_IRQ_WORK(wake_up_kfence_timer_work, wake_up_kfence_timer);
761 * Set up delayed work, which will enable and disable the static key. We need to
762 * use a work queue (rather than a simple timer), since enabling and disabling a
763 * static key cannot be done from an interrupt.
765 * Note: Toggling a static branch currently causes IPIs, and here we'll end up
766 * with a total of 2 IPIs to all CPUs. If this ends up a problem in future (with
767 * more aggressive sampling intervals), we could get away with a variant that
768 * avoids IPIs, at the cost of not immediately capturing allocations if the
769 * instructions remain cached.
771 static void toggle_allocation_gate(struct work_struct *work)
773 if (!READ_ONCE(kfence_enabled))
776 atomic_set(&kfence_allocation_gate, 0);
777 #ifdef CONFIG_KFENCE_STATIC_KEYS
778 /* Enable static key, and await allocation to happen. */
779 static_branch_enable(&kfence_allocation_key);
781 if (sysctl_hung_task_timeout_secs) {
783 * During low activity with no allocations we might wait a
784 * while; let's avoid the hung task warning.
786 wait_event_idle_timeout(allocation_wait, atomic_read(&kfence_allocation_gate),
787 sysctl_hung_task_timeout_secs * HZ / 2);
789 wait_event_idle(allocation_wait, atomic_read(&kfence_allocation_gate));
792 /* Disable static key and reset timer. */
793 static_branch_disable(&kfence_allocation_key);
795 queue_delayed_work(system_unbound_wq, &kfence_timer,
796 msecs_to_jiffies(kfence_sample_interval));
799 /* === Public interface ===================================================== */
801 void __init kfence_alloc_pool(void)
803 if (!kfence_sample_interval)
806 __kfence_pool = memblock_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
809 pr_err("failed to allocate pool\n");
812 static void kfence_init_enable(void)
814 if (!IS_ENABLED(CONFIG_KFENCE_STATIC_KEYS))
815 static_branch_enable(&kfence_allocation_key);
817 if (kfence_deferrable)
818 INIT_DEFERRABLE_WORK(&kfence_timer, toggle_allocation_gate);
820 INIT_DELAYED_WORK(&kfence_timer, toggle_allocation_gate);
822 WRITE_ONCE(kfence_enabled, true);
823 queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
825 pr_info("initialized - using %lu bytes for %d objects at 0x%p-0x%p\n", KFENCE_POOL_SIZE,
826 CONFIG_KFENCE_NUM_OBJECTS, (void *)__kfence_pool,
827 (void *)(__kfence_pool + KFENCE_POOL_SIZE));
830 void __init kfence_init(void)
832 stack_hash_seed = (u32)random_get_entropy();
834 /* Setting kfence_sample_interval to 0 on boot disables KFENCE. */
835 if (!kfence_sample_interval)
838 if (!kfence_init_pool_early()) {
839 pr_err("%s failed\n", __func__);
843 kfence_init_enable();
846 static int kfence_init_late(void)
848 const unsigned long nr_pages = KFENCE_POOL_SIZE / PAGE_SIZE;
849 #ifdef CONFIG_CONTIG_ALLOC
852 pages = alloc_contig_pages(nr_pages, GFP_KERNEL, first_online_node, NULL);
855 __kfence_pool = page_to_virt(pages);
857 if (nr_pages > MAX_ORDER_NR_PAGES) {
858 pr_warn("KFENCE_NUM_OBJECTS too large for buddy allocator\n");
861 __kfence_pool = alloc_pages_exact(KFENCE_POOL_SIZE, GFP_KERNEL);
866 if (!kfence_init_pool_late()) {
867 pr_err("%s failed\n", __func__);
871 kfence_init_enable();
875 static int kfence_enable_late(void)
878 return kfence_init_late();
880 WRITE_ONCE(kfence_enabled, true);
881 queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
885 void kfence_shutdown_cache(struct kmem_cache *s)
888 struct kfence_metadata *meta;
891 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
894 meta = &kfence_metadata[i];
897 * If we observe some inconsistent cache and state pair where we
898 * should have returned false here, cache destruction is racing
899 * with either kmem_cache_alloc() or kmem_cache_free(). Taking
900 * the lock will not help, as different critical section
901 * serialization will have the same outcome.
903 if (READ_ONCE(meta->cache) != s ||
904 READ_ONCE(meta->state) != KFENCE_OBJECT_ALLOCATED)
907 raw_spin_lock_irqsave(&meta->lock, flags);
908 in_use = meta->cache == s && meta->state == KFENCE_OBJECT_ALLOCATED;
909 raw_spin_unlock_irqrestore(&meta->lock, flags);
913 * This cache still has allocations, and we should not
914 * release them back into the freelist so they can still
915 * safely be used and retain the kernel's default
916 * behaviour of keeping the allocations alive (leak the
917 * cache); however, they effectively become "zombie
918 * allocations" as the KFENCE objects are the only ones
919 * still in use and the owning cache is being destroyed.
921 * We mark them freed, so that any subsequent use shows
922 * more useful error messages that will include stack
923 * traces of the user of the object, the original
924 * allocation, and caller to shutdown_cache().
926 kfence_guarded_free((void *)meta->addr, meta, /*zombie=*/true);
930 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
931 meta = &kfence_metadata[i];
934 if (READ_ONCE(meta->cache) != s || READ_ONCE(meta->state) != KFENCE_OBJECT_FREED)
937 raw_spin_lock_irqsave(&meta->lock, flags);
938 if (meta->cache == s && meta->state == KFENCE_OBJECT_FREED)
940 raw_spin_unlock_irqrestore(&meta->lock, flags);
944 void *__kfence_alloc(struct kmem_cache *s, size_t size, gfp_t flags)
946 unsigned long stack_entries[KFENCE_STACK_DEPTH];
947 size_t num_stack_entries;
948 u32 alloc_stack_hash;
951 * Perform size check before switching kfence_allocation_gate, so that
952 * we don't disable KFENCE without making an allocation.
954 if (size > PAGE_SIZE) {
955 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
960 * Skip allocations from non-default zones, including DMA. We cannot
961 * guarantee that pages in the KFENCE pool will have the requested
962 * properties (e.g. reside in DMAable memory).
964 if ((flags & GFP_ZONEMASK) ||
965 (s->flags & (SLAB_CACHE_DMA | SLAB_CACHE_DMA32))) {
966 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
970 if (atomic_inc_return(&kfence_allocation_gate) > 1)
972 #ifdef CONFIG_KFENCE_STATIC_KEYS
974 * waitqueue_active() is fully ordered after the update of
975 * kfence_allocation_gate per atomic_inc_return().
977 if (waitqueue_active(&allocation_wait)) {
979 * Calling wake_up() here may deadlock when allocations happen
980 * from within timer code. Use an irq_work to defer it.
982 irq_work_queue(&wake_up_kfence_timer_work);
986 if (!READ_ONCE(kfence_enabled))
989 num_stack_entries = stack_trace_save(stack_entries, KFENCE_STACK_DEPTH, 0);
992 * Do expensive check for coverage of allocation in slow-path after
993 * allocation_gate has already become non-zero, even though it might
994 * mean not making any allocation within a given sample interval.
996 * This ensures reasonable allocation coverage when the pool is almost
997 * full, including avoiding long-lived allocations of the same source
998 * filling up the pool (e.g. pagecache allocations).
1000 alloc_stack_hash = get_alloc_stack_hash(stack_entries, num_stack_entries);
1001 if (should_skip_covered() && alloc_covered_contains(alloc_stack_hash)) {
1002 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_COVERED]);
1006 return kfence_guarded_alloc(s, size, flags, stack_entries, num_stack_entries,
1010 size_t kfence_ksize(const void *addr)
1012 const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1015 * Read locklessly -- if there is a race with __kfence_alloc(), this is
1016 * either a use-after-free or invalid access.
1018 return meta ? meta->size : 0;
1021 void *kfence_object_start(const void *addr)
1023 const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1026 * Read locklessly -- if there is a race with __kfence_alloc(), this is
1027 * either a use-after-free or invalid access.
1029 return meta ? (void *)meta->addr : NULL;
1032 void __kfence_free(void *addr)
1034 struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1037 * If the objects of the cache are SLAB_TYPESAFE_BY_RCU, defer freeing
1038 * the object, as the object page may be recycled for other-typed
1039 * objects once it has been freed. meta->cache may be NULL if the cache
1042 if (unlikely(meta->cache && (meta->cache->flags & SLAB_TYPESAFE_BY_RCU)))
1043 call_rcu(&meta->rcu_head, rcu_guarded_free);
1045 kfence_guarded_free(addr, meta, false);
1048 bool kfence_handle_page_fault(unsigned long addr, bool is_write, struct pt_regs *regs)
1050 const int page_index = (addr - (unsigned long)__kfence_pool) / PAGE_SIZE;
1051 struct kfence_metadata *to_report = NULL;
1052 enum kfence_error_type error_type;
1053 unsigned long flags;
1055 if (!is_kfence_address((void *)addr))
1058 if (!READ_ONCE(kfence_enabled)) /* If disabled at runtime ... */
1059 return kfence_unprotect(addr); /* ... unprotect and proceed. */
1061 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
1063 if (page_index % 2) {
1064 /* This is a redzone, report a buffer overflow. */
1065 struct kfence_metadata *meta;
1068 meta = addr_to_metadata(addr - PAGE_SIZE);
1069 if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
1071 /* Data race ok; distance calculation approximate. */
1072 distance = addr - data_race(meta->addr + meta->size);
1075 meta = addr_to_metadata(addr + PAGE_SIZE);
1076 if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
1077 /* Data race ok; distance calculation approximate. */
1078 if (!to_report || distance > data_race(meta->addr) - addr)
1085 raw_spin_lock_irqsave(&to_report->lock, flags);
1086 to_report->unprotected_page = addr;
1087 error_type = KFENCE_ERROR_OOB;
1090 * If the object was freed before we took the look we can still
1091 * report this as an OOB -- the report will simply show the
1092 * stacktrace of the free as well.
1095 to_report = addr_to_metadata(addr);
1099 raw_spin_lock_irqsave(&to_report->lock, flags);
1100 error_type = KFENCE_ERROR_UAF;
1102 * We may race with __kfence_alloc(), and it is possible that a
1103 * freed object may be reallocated. We simply report this as a
1104 * use-after-free, with the stack trace showing the place where
1105 * the object was re-allocated.
1111 kfence_report_error(addr, is_write, regs, to_report, error_type);
1112 raw_spin_unlock_irqrestore(&to_report->lock, flags);
1114 /* This may be a UAF or OOB access, but we can't be sure. */
1115 kfence_report_error(addr, is_write, regs, NULL, KFENCE_ERROR_INVALID);
1118 return kfence_unprotect(addr); /* Unprotect and let access proceed. */