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/kcsan-checks.h>
14 #include <linux/kfence.h>
15 #include <linux/list.h>
16 #include <linux/lockdep.h>
17 #include <linux/memblock.h>
18 #include <linux/moduleparam.h>
19 #include <linux/random.h>
20 #include <linux/rcupdate.h>
21 #include <linux/seq_file.h>
22 #include <linux/slab.h>
23 #include <linux/spinlock.h>
24 #include <linux/string.h>
26 #include <asm/kfence.h>
30 /* Disables KFENCE on the first warning assuming an irrecoverable error. */
31 #define KFENCE_WARN_ON(cond) \
33 const bool __cond = WARN_ON(cond); \
34 if (unlikely(__cond)) \
35 WRITE_ONCE(kfence_enabled, false); \
39 /* === Data ================================================================= */
41 static bool kfence_enabled __read_mostly;
43 static unsigned long kfence_sample_interval __read_mostly = CONFIG_KFENCE_SAMPLE_INTERVAL;
45 #ifdef MODULE_PARAM_PREFIX
46 #undef MODULE_PARAM_PREFIX
48 #define MODULE_PARAM_PREFIX "kfence."
50 static int param_set_sample_interval(const char *val, const struct kernel_param *kp)
53 int ret = kstrtoul(val, 0, &num);
58 if (!num) /* Using 0 to indicate KFENCE is disabled. */
59 WRITE_ONCE(kfence_enabled, false);
60 else if (!READ_ONCE(kfence_enabled) && system_state != SYSTEM_BOOTING)
61 return -EINVAL; /* Cannot (re-)enable KFENCE on-the-fly. */
63 *((unsigned long *)kp->arg) = num;
67 static int param_get_sample_interval(char *buffer, const struct kernel_param *kp)
69 if (!READ_ONCE(kfence_enabled))
70 return sprintf(buffer, "0\n");
72 return param_get_ulong(buffer, kp);
75 static const struct kernel_param_ops sample_interval_param_ops = {
76 .set = param_set_sample_interval,
77 .get = param_get_sample_interval,
79 module_param_cb(sample_interval, &sample_interval_param_ops, &kfence_sample_interval, 0600);
81 /* The pool of pages used for guard pages and objects. */
82 char *__kfence_pool __ro_after_init;
83 EXPORT_SYMBOL(__kfence_pool); /* Export for test modules. */
86 * Per-object metadata, with one-to-one mapping of object metadata to
87 * backing pages (in __kfence_pool).
89 static_assert(CONFIG_KFENCE_NUM_OBJECTS > 0);
90 struct kfence_metadata kfence_metadata[CONFIG_KFENCE_NUM_OBJECTS];
92 /* Freelist with available objects. */
93 static struct list_head kfence_freelist = LIST_HEAD_INIT(kfence_freelist);
94 static DEFINE_RAW_SPINLOCK(kfence_freelist_lock); /* Lock protecting freelist. */
96 #ifdef CONFIG_KFENCE_STATIC_KEYS
97 /* The static key to set up a KFENCE allocation. */
98 DEFINE_STATIC_KEY_FALSE(kfence_allocation_key);
101 /* Gates the allocation, ensuring only one succeeds in a given period. */
102 atomic_t kfence_allocation_gate = ATOMIC_INIT(1);
104 /* Statistics counters for debugfs. */
105 enum kfence_counter_id {
106 KFENCE_COUNTER_ALLOCATED,
107 KFENCE_COUNTER_ALLOCS,
108 KFENCE_COUNTER_FREES,
109 KFENCE_COUNTER_ZOMBIES,
111 KFENCE_COUNTER_COUNT,
113 static atomic_long_t counters[KFENCE_COUNTER_COUNT];
114 static const char *const counter_names[] = {
115 [KFENCE_COUNTER_ALLOCATED] = "currently allocated",
116 [KFENCE_COUNTER_ALLOCS] = "total allocations",
117 [KFENCE_COUNTER_FREES] = "total frees",
118 [KFENCE_COUNTER_ZOMBIES] = "zombie allocations",
119 [KFENCE_COUNTER_BUGS] = "total bugs",
121 static_assert(ARRAY_SIZE(counter_names) == KFENCE_COUNTER_COUNT);
123 /* === Internals ============================================================ */
125 static bool kfence_protect(unsigned long addr)
127 return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), true));
130 static bool kfence_unprotect(unsigned long addr)
132 return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), false));
135 static inline struct kfence_metadata *addr_to_metadata(unsigned long addr)
139 /* The checks do not affect performance; only called from slow-paths. */
141 if (!is_kfence_address((void *)addr))
145 * May be an invalid index if called with an address at the edge of
146 * __kfence_pool, in which case we would report an "invalid access"
149 index = (addr - (unsigned long)__kfence_pool) / (PAGE_SIZE * 2) - 1;
150 if (index < 0 || index >= CONFIG_KFENCE_NUM_OBJECTS)
153 return &kfence_metadata[index];
156 static inline unsigned long metadata_to_pageaddr(const struct kfence_metadata *meta)
158 unsigned long offset = (meta - kfence_metadata + 1) * PAGE_SIZE * 2;
159 unsigned long pageaddr = (unsigned long)&__kfence_pool[offset];
161 /* The checks do not affect performance; only called from slow-paths. */
163 /* Only call with a pointer into kfence_metadata. */
164 if (KFENCE_WARN_ON(meta < kfence_metadata ||
165 meta >= kfence_metadata + CONFIG_KFENCE_NUM_OBJECTS))
169 * This metadata object only ever maps to 1 page; verify that the stored
170 * address is in the expected range.
172 if (KFENCE_WARN_ON(ALIGN_DOWN(meta->addr, PAGE_SIZE) != pageaddr))
179 * Update the object's metadata state, including updating the alloc/free stacks
180 * depending on the state transition.
182 static noinline void metadata_update_state(struct kfence_metadata *meta,
183 enum kfence_object_state next)
185 struct kfence_track *track =
186 next == KFENCE_OBJECT_FREED ? &meta->free_track : &meta->alloc_track;
188 lockdep_assert_held(&meta->lock);
191 * Skip over 1 (this) functions; noinline ensures we do not accidentally
192 * skip over the caller by never inlining.
194 track->num_stack_entries = stack_trace_save(track->stack_entries, KFENCE_STACK_DEPTH, 1);
195 track->pid = task_pid_nr(current);
198 * Pairs with READ_ONCE() in
199 * kfence_shutdown_cache(),
200 * kfence_handle_page_fault().
202 WRITE_ONCE(meta->state, next);
205 /* Write canary byte to @addr. */
206 static inline bool set_canary_byte(u8 *addr)
208 *addr = KFENCE_CANARY_PATTERN(addr);
212 /* Check canary byte at @addr. */
213 static inline bool check_canary_byte(u8 *addr)
215 if (likely(*addr == KFENCE_CANARY_PATTERN(addr)))
218 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
219 kfence_report_error((unsigned long)addr, false, NULL, addr_to_metadata((unsigned long)addr),
220 KFENCE_ERROR_CORRUPTION);
224 /* __always_inline this to ensure we won't do an indirect call to fn. */
225 static __always_inline void for_each_canary(const struct kfence_metadata *meta, bool (*fn)(u8 *))
227 const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
230 lockdep_assert_held(&meta->lock);
233 * We'll iterate over each canary byte per-side until fn() returns
234 * false. However, we'll still iterate over the canary bytes to the
235 * right of the object even if there was an error in the canary bytes to
236 * the left of the object. Specifically, if check_canary_byte()
237 * generates an error, showing both sides might give more clues as to
238 * what the error is about when displaying which bytes were corrupted.
241 /* Apply to left of object. */
242 for (addr = pageaddr; addr < meta->addr; addr++) {
247 /* Apply to right of object. */
248 for (addr = meta->addr + meta->size; addr < pageaddr + PAGE_SIZE; addr++) {
254 static void *kfence_guarded_alloc(struct kmem_cache *cache, size_t size, gfp_t gfp)
256 struct kfence_metadata *meta = NULL;
261 /* Try to obtain a free object. */
262 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
263 if (!list_empty(&kfence_freelist)) {
264 meta = list_entry(kfence_freelist.next, struct kfence_metadata, list);
265 list_del_init(&meta->list);
267 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
271 if (unlikely(!raw_spin_trylock_irqsave(&meta->lock, flags))) {
273 * This is extremely unlikely -- we are reporting on a
274 * use-after-free, which locked meta->lock, and the reporting
275 * code via printk calls kmalloc() which ends up in
276 * kfence_alloc() and tries to grab the same object that we're
277 * reporting on. While it has never been observed, lockdep does
278 * report that there is a possibility of deadlock. Fix it by
279 * using trylock and bailing out gracefully.
281 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
282 /* Put the object back on the freelist. */
283 list_add_tail(&meta->list, &kfence_freelist);
284 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
289 meta->addr = metadata_to_pageaddr(meta);
290 /* Unprotect if we're reusing this page. */
291 if (meta->state == KFENCE_OBJECT_FREED)
292 kfence_unprotect(meta->addr);
295 * Note: for allocations made before RNG initialization, will always
296 * return zero. We still benefit from enabling KFENCE as early as
297 * possible, even when the RNG is not yet available, as this will allow
298 * KFENCE to detect bugs due to earlier allocations. The only downside
299 * is that the out-of-bounds accesses detected are deterministic for
302 if (prandom_u32_max(2)) {
303 /* Allocate on the "right" side, re-calculate address. */
304 meta->addr += PAGE_SIZE - size;
305 meta->addr = ALIGN_DOWN(meta->addr, cache->align);
308 addr = (void *)meta->addr;
310 /* Update remaining metadata. */
311 metadata_update_state(meta, KFENCE_OBJECT_ALLOCATED);
312 /* Pairs with READ_ONCE() in kfence_shutdown_cache(). */
313 WRITE_ONCE(meta->cache, cache);
315 for_each_canary(meta, set_canary_byte);
317 /* Set required struct page fields. */
318 page = virt_to_page(meta->addr);
319 page->slab_cache = cache;
320 if (IS_ENABLED(CONFIG_SLUB))
322 if (IS_ENABLED(CONFIG_SLAB))
325 raw_spin_unlock_irqrestore(&meta->lock, flags);
327 /* Memory initialization. */
330 * We check slab_want_init_on_alloc() ourselves, rather than letting
331 * SL*B do the initialization, as otherwise we might overwrite KFENCE's
334 if (unlikely(slab_want_init_on_alloc(gfp, cache)))
335 memzero_explicit(addr, size);
339 if (CONFIG_KFENCE_STRESS_TEST_FAULTS && !prandom_u32_max(CONFIG_KFENCE_STRESS_TEST_FAULTS))
340 kfence_protect(meta->addr); /* Random "faults" by protecting the object. */
342 atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCATED]);
343 atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCS]);
348 static void kfence_guarded_free(void *addr, struct kfence_metadata *meta, bool zombie)
350 struct kcsan_scoped_access assert_page_exclusive;
353 raw_spin_lock_irqsave(&meta->lock, flags);
355 if (meta->state != KFENCE_OBJECT_ALLOCATED || meta->addr != (unsigned long)addr) {
356 /* Invalid or double-free, bail out. */
357 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
358 kfence_report_error((unsigned long)addr, false, NULL, meta,
359 KFENCE_ERROR_INVALID_FREE);
360 raw_spin_unlock_irqrestore(&meta->lock, flags);
364 /* Detect racy use-after-free, or incorrect reallocation of this page by KFENCE. */
365 kcsan_begin_scoped_access((void *)ALIGN_DOWN((unsigned long)addr, PAGE_SIZE), PAGE_SIZE,
366 KCSAN_ACCESS_SCOPED | KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT,
367 &assert_page_exclusive);
369 if (CONFIG_KFENCE_STRESS_TEST_FAULTS)
370 kfence_unprotect((unsigned long)addr); /* To check canary bytes. */
372 /* Restore page protection if there was an OOB access. */
373 if (meta->unprotected_page) {
374 kfence_protect(meta->unprotected_page);
375 meta->unprotected_page = 0;
378 /* Check canary bytes for memory corruption. */
379 for_each_canary(meta, check_canary_byte);
382 * Clear memory if init-on-free is set. While we protect the page, the
383 * data is still there, and after a use-after-free is detected, we
384 * unprotect the page, so the data is still accessible.
386 if (!zombie && unlikely(slab_want_init_on_free(meta->cache)))
387 memzero_explicit(addr, meta->size);
389 /* Mark the object as freed. */
390 metadata_update_state(meta, KFENCE_OBJECT_FREED);
392 raw_spin_unlock_irqrestore(&meta->lock, flags);
394 /* Protect to detect use-after-frees. */
395 kfence_protect((unsigned long)addr);
397 kcsan_end_scoped_access(&assert_page_exclusive);
399 /* Add it to the tail of the freelist for reuse. */
400 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
401 KFENCE_WARN_ON(!list_empty(&meta->list));
402 list_add_tail(&meta->list, &kfence_freelist);
403 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
405 atomic_long_dec(&counters[KFENCE_COUNTER_ALLOCATED]);
406 atomic_long_inc(&counters[KFENCE_COUNTER_FREES]);
408 /* See kfence_shutdown_cache(). */
409 atomic_long_inc(&counters[KFENCE_COUNTER_ZOMBIES]);
413 static void rcu_guarded_free(struct rcu_head *h)
415 struct kfence_metadata *meta = container_of(h, struct kfence_metadata, rcu_head);
417 kfence_guarded_free((void *)meta->addr, meta, false);
420 static bool __init kfence_init_pool(void)
422 unsigned long addr = (unsigned long)__kfence_pool;
429 if (!arch_kfence_init_pool())
432 pages = virt_to_page(addr);
435 * Set up object pages: they must have PG_slab set, to avoid freeing
436 * these as real pages.
438 * We also want to avoid inserting kfence_free() in the kfree()
439 * fast-path in SLUB, and therefore need to ensure kfree() correctly
440 * enters __slab_free() slow-path.
442 for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
446 /* Verify we do not have a compound head page. */
447 if (WARN_ON(compound_head(&pages[i]) != &pages[i]))
450 __SetPageSlab(&pages[i]);
454 * Protect the first 2 pages. The first page is mostly unnecessary, and
455 * merely serves as an extended guard page. However, adding one
456 * additional page in the beginning gives us an even number of pages,
457 * which simplifies the mapping of address to metadata index.
459 for (i = 0; i < 2; i++) {
460 if (unlikely(!kfence_protect(addr)))
466 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
467 struct kfence_metadata *meta = &kfence_metadata[i];
469 /* Initialize metadata. */
470 INIT_LIST_HEAD(&meta->list);
471 raw_spin_lock_init(&meta->lock);
472 meta->state = KFENCE_OBJECT_UNUSED;
473 meta->addr = addr; /* Initialize for validation in metadata_to_pageaddr(). */
474 list_add_tail(&meta->list, &kfence_freelist);
476 /* Protect the right redzone. */
477 if (unlikely(!kfence_protect(addr + PAGE_SIZE)))
480 addr += 2 * PAGE_SIZE;
487 * Only release unprotected pages, and do not try to go back and change
488 * page attributes due to risk of failing to do so as well. If changing
489 * page attributes for some pages fails, it is very likely that it also
490 * fails for the first page, and therefore expect addr==__kfence_pool in
491 * most failure cases.
493 memblock_free_late(__pa(addr), KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool));
494 __kfence_pool = NULL;
498 /* === DebugFS Interface ==================================================== */
500 static int stats_show(struct seq_file *seq, void *v)
504 seq_printf(seq, "enabled: %i\n", READ_ONCE(kfence_enabled));
505 for (i = 0; i < KFENCE_COUNTER_COUNT; i++)
506 seq_printf(seq, "%s: %ld\n", counter_names[i], atomic_long_read(&counters[i]));
510 DEFINE_SHOW_ATTRIBUTE(stats);
513 * debugfs seq_file operations for /sys/kernel/debug/kfence/objects.
514 * start_object() and next_object() return the object index + 1, because NULL is used
517 static void *start_object(struct seq_file *seq, loff_t *pos)
519 if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
520 return (void *)((long)*pos + 1);
524 static void stop_object(struct seq_file *seq, void *v)
528 static void *next_object(struct seq_file *seq, void *v, loff_t *pos)
531 if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
532 return (void *)((long)*pos + 1);
536 static int show_object(struct seq_file *seq, void *v)
538 struct kfence_metadata *meta = &kfence_metadata[(long)v - 1];
541 raw_spin_lock_irqsave(&meta->lock, flags);
542 kfence_print_object(seq, meta);
543 raw_spin_unlock_irqrestore(&meta->lock, flags);
544 seq_puts(seq, "---------------------------------\n");
549 static const struct seq_operations object_seqops = {
550 .start = start_object,
556 static int open_objects(struct inode *inode, struct file *file)
558 return seq_open(file, &object_seqops);
561 static const struct file_operations objects_fops = {
562 .open = open_objects,
567 static int __init kfence_debugfs_init(void)
569 struct dentry *kfence_dir = debugfs_create_dir("kfence", NULL);
571 debugfs_create_file("stats", 0444, kfence_dir, NULL, &stats_fops);
572 debugfs_create_file("objects", 0400, kfence_dir, NULL, &objects_fops);
576 late_initcall(kfence_debugfs_init);
578 /* === Allocation Gate Timer ================================================ */
581 * Set up delayed work, which will enable and disable the static key. We need to
582 * use a work queue (rather than a simple timer), since enabling and disabling a
583 * static key cannot be done from an interrupt.
585 * Note: Toggling a static branch currently causes IPIs, and here we'll end up
586 * with a total of 2 IPIs to all CPUs. If this ends up a problem in future (with
587 * more aggressive sampling intervals), we could get away with a variant that
588 * avoids IPIs, at the cost of not immediately capturing allocations if the
589 * instructions remain cached.
591 static struct delayed_work kfence_timer;
592 static void toggle_allocation_gate(struct work_struct *work)
594 if (!READ_ONCE(kfence_enabled))
597 /* Enable static key, and await allocation to happen. */
598 atomic_set(&kfence_allocation_gate, 0);
599 #ifdef CONFIG_KFENCE_STATIC_KEYS
600 static_branch_enable(&kfence_allocation_key);
602 * Await an allocation. Timeout after 1 second, in case the kernel stops
603 * doing allocations, to avoid stalling this worker task for too long.
606 unsigned long end_wait = jiffies + HZ;
609 set_current_state(TASK_UNINTERRUPTIBLE);
610 if (atomic_read(&kfence_allocation_gate) != 0)
613 } while (time_before(jiffies, end_wait));
614 __set_current_state(TASK_RUNNING);
616 /* Disable static key and reset timer. */
617 static_branch_disable(&kfence_allocation_key);
619 schedule_delayed_work(&kfence_timer, msecs_to_jiffies(kfence_sample_interval));
621 static DECLARE_DELAYED_WORK(kfence_timer, toggle_allocation_gate);
623 /* === Public interface ===================================================== */
625 void __init kfence_alloc_pool(void)
627 if (!kfence_sample_interval)
630 __kfence_pool = memblock_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
633 pr_err("failed to allocate pool\n");
636 void __init kfence_init(void)
638 /* Setting kfence_sample_interval to 0 on boot disables KFENCE. */
639 if (!kfence_sample_interval)
642 if (!kfence_init_pool()) {
643 pr_err("%s failed\n", __func__);
647 WRITE_ONCE(kfence_enabled, true);
648 schedule_delayed_work(&kfence_timer, 0);
649 pr_info("initialized - using %lu bytes for %d objects at 0x%p-0x%p\n", KFENCE_POOL_SIZE,
650 CONFIG_KFENCE_NUM_OBJECTS, (void *)__kfence_pool,
651 (void *)(__kfence_pool + KFENCE_POOL_SIZE));
654 void kfence_shutdown_cache(struct kmem_cache *s)
657 struct kfence_metadata *meta;
660 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
663 meta = &kfence_metadata[i];
666 * If we observe some inconsistent cache and state pair where we
667 * should have returned false here, cache destruction is racing
668 * with either kmem_cache_alloc() or kmem_cache_free(). Taking
669 * the lock will not help, as different critical section
670 * serialization will have the same outcome.
672 if (READ_ONCE(meta->cache) != s ||
673 READ_ONCE(meta->state) != KFENCE_OBJECT_ALLOCATED)
676 raw_spin_lock_irqsave(&meta->lock, flags);
677 in_use = meta->cache == s && meta->state == KFENCE_OBJECT_ALLOCATED;
678 raw_spin_unlock_irqrestore(&meta->lock, flags);
682 * This cache still has allocations, and we should not
683 * release them back into the freelist so they can still
684 * safely be used and retain the kernel's default
685 * behaviour of keeping the allocations alive (leak the
686 * cache); however, they effectively become "zombie
687 * allocations" as the KFENCE objects are the only ones
688 * still in use and the owning cache is being destroyed.
690 * We mark them freed, so that any subsequent use shows
691 * more useful error messages that will include stack
692 * traces of the user of the object, the original
693 * allocation, and caller to shutdown_cache().
695 kfence_guarded_free((void *)meta->addr, meta, /*zombie=*/true);
699 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
700 meta = &kfence_metadata[i];
703 if (READ_ONCE(meta->cache) != s || READ_ONCE(meta->state) != KFENCE_OBJECT_FREED)
706 raw_spin_lock_irqsave(&meta->lock, flags);
707 if (meta->cache == s && meta->state == KFENCE_OBJECT_FREED)
709 raw_spin_unlock_irqrestore(&meta->lock, flags);
713 void *__kfence_alloc(struct kmem_cache *s, size_t size, gfp_t flags)
716 * allocation_gate only needs to become non-zero, so it doesn't make
717 * sense to continue writing to it and pay the associated contention
718 * cost, in case we have a large number of concurrent allocations.
720 if (atomic_read(&kfence_allocation_gate) || atomic_inc_return(&kfence_allocation_gate) > 1)
723 if (!READ_ONCE(kfence_enabled))
726 if (size > PAGE_SIZE)
729 return kfence_guarded_alloc(s, size, flags);
732 size_t kfence_ksize(const void *addr)
734 const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
737 * Read locklessly -- if there is a race with __kfence_alloc(), this is
738 * either a use-after-free or invalid access.
740 return meta ? meta->size : 0;
743 void *kfence_object_start(const void *addr)
745 const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
748 * Read locklessly -- if there is a race with __kfence_alloc(), this is
749 * either a use-after-free or invalid access.
751 return meta ? (void *)meta->addr : NULL;
754 void __kfence_free(void *addr)
756 struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
759 * If the objects of the cache are SLAB_TYPESAFE_BY_RCU, defer freeing
760 * the object, as the object page may be recycled for other-typed
761 * objects once it has been freed. meta->cache may be NULL if the cache
764 if (unlikely(meta->cache && (meta->cache->flags & SLAB_TYPESAFE_BY_RCU)))
765 call_rcu(&meta->rcu_head, rcu_guarded_free);
767 kfence_guarded_free(addr, meta, false);
770 bool kfence_handle_page_fault(unsigned long addr, bool is_write, struct pt_regs *regs)
772 const int page_index = (addr - (unsigned long)__kfence_pool) / PAGE_SIZE;
773 struct kfence_metadata *to_report = NULL;
774 enum kfence_error_type error_type;
777 if (!is_kfence_address((void *)addr))
780 if (!READ_ONCE(kfence_enabled)) /* If disabled at runtime ... */
781 return kfence_unprotect(addr); /* ... unprotect and proceed. */
783 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
785 if (page_index % 2) {
786 /* This is a redzone, report a buffer overflow. */
787 struct kfence_metadata *meta;
790 meta = addr_to_metadata(addr - PAGE_SIZE);
791 if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
793 /* Data race ok; distance calculation approximate. */
794 distance = addr - data_race(meta->addr + meta->size);
797 meta = addr_to_metadata(addr + PAGE_SIZE);
798 if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
799 /* Data race ok; distance calculation approximate. */
800 if (!to_report || distance > data_race(meta->addr) - addr)
807 raw_spin_lock_irqsave(&to_report->lock, flags);
808 to_report->unprotected_page = addr;
809 error_type = KFENCE_ERROR_OOB;
812 * If the object was freed before we took the look we can still
813 * report this as an OOB -- the report will simply show the
814 * stacktrace of the free as well.
817 to_report = addr_to_metadata(addr);
821 raw_spin_lock_irqsave(&to_report->lock, flags);
822 error_type = KFENCE_ERROR_UAF;
824 * We may race with __kfence_alloc(), and it is possible that a
825 * freed object may be reallocated. We simply report this as a
826 * use-after-free, with the stack trace showing the place where
827 * the object was re-allocated.
833 kfence_report_error(addr, is_write, regs, to_report, error_type);
834 raw_spin_unlock_irqrestore(&to_report->lock, flags);
836 /* This may be a UAF or OOB access, but we can't be sure. */
837 kfence_report_error(addr, is_write, regs, NULL, KFENCE_ERROR_INVALID);
840 return kfence_unprotect(addr); /* Unprotect and let access proceed. */