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/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>
32
33 #include <asm/kfence.h>
34
35 #include "kfence.h"
36
37 /* Disables KFENCE on the first warning assuming an irrecoverable error. */
38 #define KFENCE_WARN_ON(cond) \
39 ({ \
40 const bool __cond = WARN_ON(cond); \
41 if (unlikely(__cond)) \
42 WRITE_ONCE(kfence_enabled, false); \
43 __cond; \
44 })
45
46 /* === Data ================================================================= */
47
48 static bool kfence_enabled __read_mostly;
49
50 static unsigned long kfence_sample_interval __read_mostly = CONFIG_KFENCE_SAMPLE_INTERVAL;
51
52 #ifdef MODULE_PARAM_PREFIX
53 #undef MODULE_PARAM_PREFIX
54 #endif
55 #define MODULE_PARAM_PREFIX "kfence."
56
param_set_sample_interval(const char * val,const struct kernel_param * kp)57 static int param_set_sample_interval(const char *val, const struct kernel_param *kp)
58 {
59 unsigned long num;
60 int ret = kstrtoul(val, 0, &num);
61
62 if (ret < 0)
63 return ret;
64
65 if (!num) /* Using 0 to indicate KFENCE is disabled. */
66 WRITE_ONCE(kfence_enabled, false);
67 else if (!READ_ONCE(kfence_enabled) && system_state != SYSTEM_BOOTING)
68 return -EINVAL; /* Cannot (re-)enable KFENCE on-the-fly. */
69
70 *((unsigned long *)kp->arg) = num;
71 return 0;
72 }
73
param_get_sample_interval(char * buffer,const struct kernel_param * kp)74 static int param_get_sample_interval(char *buffer, const struct kernel_param *kp)
75 {
76 if (!READ_ONCE(kfence_enabled))
77 return sprintf(buffer, "0\n");
78
79 return param_get_ulong(buffer, kp);
80 }
81
82 static const struct kernel_param_ops sample_interval_param_ops = {
83 .set = param_set_sample_interval,
84 .get = param_get_sample_interval,
85 };
86 module_param_cb(sample_interval, &sample_interval_param_ops, &kfence_sample_interval, 0600);
87
88 /* Pool usage% threshold when currently covered allocations are skipped. */
89 static unsigned long kfence_skip_covered_thresh __read_mostly = 75;
90 module_param_named(skip_covered_thresh, kfence_skip_covered_thresh, ulong, 0644);
91
92 /* The pool of pages used for guard pages and objects. */
93 char *__kfence_pool __ro_after_init;
94 EXPORT_SYMBOL(__kfence_pool); /* Export for test modules. */
95
96 /*
97 * Per-object metadata, with one-to-one mapping of object metadata to
98 * backing pages (in __kfence_pool).
99 */
100 static_assert(CONFIG_KFENCE_NUM_OBJECTS > 0);
101 struct kfence_metadata kfence_metadata[CONFIG_KFENCE_NUM_OBJECTS];
102
103 /* Freelist with available objects. */
104 static struct list_head kfence_freelist = LIST_HEAD_INIT(kfence_freelist);
105 static DEFINE_RAW_SPINLOCK(kfence_freelist_lock); /* Lock protecting freelist. */
106
107 /*
108 * The static key to set up a KFENCE allocation; or if static keys are not used
109 * to gate allocations, to avoid a load and compare if KFENCE is disabled.
110 */
111 DEFINE_STATIC_KEY_FALSE(kfence_allocation_key);
112
113 /* Gates the allocation, ensuring only one succeeds in a given period. */
114 atomic_t kfence_allocation_gate = ATOMIC_INIT(1);
115
116 /*
117 * A Counting Bloom filter of allocation coverage: limits currently covered
118 * allocations of the same source filling up the pool.
119 *
120 * Assuming a range of 15%-85% unique allocations in the pool at any point in
121 * time, the below parameters provide a probablity of 0.02-0.33 for false
122 * positive hits respectively:
123 *
124 * P(alloc_traces) = (1 - e^(-HNUM * (alloc_traces / SIZE)) ^ HNUM
125 */
126 #define ALLOC_COVERED_HNUM 2
127 #define ALLOC_COVERED_ORDER (const_ilog2(CONFIG_KFENCE_NUM_OBJECTS) + 2)
128 #define ALLOC_COVERED_SIZE (1 << ALLOC_COVERED_ORDER)
129 #define ALLOC_COVERED_HNEXT(h) hash_32(h, ALLOC_COVERED_ORDER)
130 #define ALLOC_COVERED_MASK (ALLOC_COVERED_SIZE - 1)
131 static atomic_t alloc_covered[ALLOC_COVERED_SIZE];
132
133 /* Stack depth used to determine uniqueness of an allocation. */
134 #define UNIQUE_ALLOC_STACK_DEPTH ((size_t)8)
135
136 /*
137 * Randomness for stack hashes, making the same collisions across reboots and
138 * different machines less likely.
139 */
140 static u32 stack_hash_seed __ro_after_init;
141
142 /* Statistics counters for debugfs. */
143 enum kfence_counter_id {
144 KFENCE_COUNTER_ALLOCATED,
145 KFENCE_COUNTER_ALLOCS,
146 KFENCE_COUNTER_FREES,
147 KFENCE_COUNTER_ZOMBIES,
148 KFENCE_COUNTER_BUGS,
149 KFENCE_COUNTER_SKIP_INCOMPAT,
150 KFENCE_COUNTER_SKIP_CAPACITY,
151 KFENCE_COUNTER_SKIP_COVERED,
152 KFENCE_COUNTER_COUNT,
153 };
154 static atomic_long_t counters[KFENCE_COUNTER_COUNT];
155 static const char *const counter_names[] = {
156 [KFENCE_COUNTER_ALLOCATED] = "currently allocated",
157 [KFENCE_COUNTER_ALLOCS] = "total allocations",
158 [KFENCE_COUNTER_FREES] = "total frees",
159 [KFENCE_COUNTER_ZOMBIES] = "zombie allocations",
160 [KFENCE_COUNTER_BUGS] = "total bugs",
161 [KFENCE_COUNTER_SKIP_INCOMPAT] = "skipped allocations (incompatible)",
162 [KFENCE_COUNTER_SKIP_CAPACITY] = "skipped allocations (capacity)",
163 [KFENCE_COUNTER_SKIP_COVERED] = "skipped allocations (covered)",
164 };
165 static_assert(ARRAY_SIZE(counter_names) == KFENCE_COUNTER_COUNT);
166
167 /* === Internals ============================================================ */
168
should_skip_covered(void)169 static inline bool should_skip_covered(void)
170 {
171 unsigned long thresh = (CONFIG_KFENCE_NUM_OBJECTS * kfence_skip_covered_thresh) / 100;
172
173 return atomic_long_read(&counters[KFENCE_COUNTER_ALLOCATED]) > thresh;
174 }
175
get_alloc_stack_hash(unsigned long * stack_entries,size_t num_entries)176 static u32 get_alloc_stack_hash(unsigned long *stack_entries, size_t num_entries)
177 {
178 num_entries = min(num_entries, UNIQUE_ALLOC_STACK_DEPTH);
179 num_entries = filter_irq_stacks(stack_entries, num_entries);
180 return jhash(stack_entries, num_entries * sizeof(stack_entries[0]), stack_hash_seed);
181 }
182
183 /*
184 * Adds (or subtracts) count @val for allocation stack trace hash
185 * @alloc_stack_hash from Counting Bloom filter.
186 */
alloc_covered_add(u32 alloc_stack_hash,int val)187 static void alloc_covered_add(u32 alloc_stack_hash, int val)
188 {
189 int i;
190
191 for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
192 atomic_add(val, &alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]);
193 alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
194 }
195 }
196
197 /*
198 * Returns true if the allocation stack trace hash @alloc_stack_hash is
199 * currently contained (non-zero count) in Counting Bloom filter.
200 */
alloc_covered_contains(u32 alloc_stack_hash)201 static bool alloc_covered_contains(u32 alloc_stack_hash)
202 {
203 int i;
204
205 for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
206 if (!atomic_read(&alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]))
207 return false;
208 alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
209 }
210
211 return true;
212 }
213
kfence_protect(unsigned long addr)214 static bool kfence_protect(unsigned long addr)
215 {
216 return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), true));
217 }
218
kfence_unprotect(unsigned long addr)219 static bool kfence_unprotect(unsigned long addr)
220 {
221 return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), false));
222 }
223
metadata_to_pageaddr(const struct kfence_metadata * meta)224 static inline unsigned long metadata_to_pageaddr(const struct kfence_metadata *meta)
225 {
226 unsigned long offset = (meta - kfence_metadata + 1) * PAGE_SIZE * 2;
227 unsigned long pageaddr = (unsigned long)&__kfence_pool[offset];
228
229 /* The checks do not affect performance; only called from slow-paths. */
230
231 /* Only call with a pointer into kfence_metadata. */
232 if (KFENCE_WARN_ON(meta < kfence_metadata ||
233 meta >= kfence_metadata + CONFIG_KFENCE_NUM_OBJECTS))
234 return 0;
235
236 /*
237 * This metadata object only ever maps to 1 page; verify that the stored
238 * address is in the expected range.
239 */
240 if (KFENCE_WARN_ON(ALIGN_DOWN(meta->addr, PAGE_SIZE) != pageaddr))
241 return 0;
242
243 return pageaddr;
244 }
245
246 /*
247 * Update the object's metadata state, including updating the alloc/free stacks
248 * depending on the state transition.
249 */
250 static noinline void
metadata_update_state(struct kfence_metadata * meta,enum kfence_object_state next,unsigned long * stack_entries,size_t num_stack_entries)251 metadata_update_state(struct kfence_metadata *meta, enum kfence_object_state next,
252 unsigned long *stack_entries, size_t num_stack_entries)
253 {
254 struct kfence_track *track =
255 next == KFENCE_OBJECT_FREED ? &meta->free_track : &meta->alloc_track;
256
257 lockdep_assert_held(&meta->lock);
258
259 if (stack_entries) {
260 memcpy(track->stack_entries, stack_entries,
261 num_stack_entries * sizeof(stack_entries[0]));
262 } else {
263 /*
264 * Skip over 1 (this) functions; noinline ensures we do not
265 * accidentally skip over the caller by never inlining.
266 */
267 num_stack_entries = stack_trace_save(track->stack_entries, KFENCE_STACK_DEPTH, 1);
268 }
269 track->num_stack_entries = num_stack_entries;
270 track->pid = task_pid_nr(current);
271 track->cpu = raw_smp_processor_id();
272 track->ts_nsec = local_clock(); /* Same source as printk timestamps. */
273
274 /*
275 * Pairs with READ_ONCE() in
276 * kfence_shutdown_cache(),
277 * kfence_handle_page_fault().
278 */
279 WRITE_ONCE(meta->state, next);
280 }
281
282 /* Write canary byte to @addr. */
set_canary_byte(u8 * addr)283 static inline bool set_canary_byte(u8 *addr)
284 {
285 *addr = KFENCE_CANARY_PATTERN(addr);
286 return true;
287 }
288
289 /* Check canary byte at @addr. */
check_canary_byte(u8 * addr)290 static inline bool check_canary_byte(u8 *addr)
291 {
292 if (likely(*addr == KFENCE_CANARY_PATTERN(addr)))
293 return true;
294
295 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
296 kfence_report_error((unsigned long)addr, false, NULL, addr_to_metadata((unsigned long)addr),
297 KFENCE_ERROR_CORRUPTION);
298 return false;
299 }
300
301 /* __always_inline this to ensure we won't do an indirect call to fn. */
for_each_canary(const struct kfence_metadata * meta,bool (* fn)(u8 *))302 static __always_inline void for_each_canary(const struct kfence_metadata *meta, bool (*fn)(u8 *))
303 {
304 const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
305 unsigned long addr;
306
307 lockdep_assert_held(&meta->lock);
308
309 /*
310 * We'll iterate over each canary byte per-side until fn() returns
311 * false. However, we'll still iterate over the canary bytes to the
312 * right of the object even if there was an error in the canary bytes to
313 * the left of the object. Specifically, if check_canary_byte()
314 * generates an error, showing both sides might give more clues as to
315 * what the error is about when displaying which bytes were corrupted.
316 */
317
318 /* Apply to left of object. */
319 for (addr = pageaddr; addr < meta->addr; addr++) {
320 if (!fn((u8 *)addr))
321 break;
322 }
323
324 /* Apply to right of object. */
325 for (addr = meta->addr + meta->size; addr < pageaddr + PAGE_SIZE; addr++) {
326 if (!fn((u8 *)addr))
327 break;
328 }
329 }
330
kfence_guarded_alloc(struct kmem_cache * cache,size_t size,gfp_t gfp,unsigned long * stack_entries,size_t num_stack_entries,u32 alloc_stack_hash)331 static void *kfence_guarded_alloc(struct kmem_cache *cache, size_t size, gfp_t gfp,
332 unsigned long *stack_entries, size_t num_stack_entries,
333 u32 alloc_stack_hash)
334 {
335 struct kfence_metadata *meta = NULL;
336 unsigned long flags;
337 struct page *page;
338 void *addr;
339
340 /* Try to obtain a free object. */
341 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
342 if (!list_empty(&kfence_freelist)) {
343 meta = list_entry(kfence_freelist.next, struct kfence_metadata, list);
344 list_del_init(&meta->list);
345 }
346 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
347 if (!meta) {
348 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_CAPACITY]);
349 return NULL;
350 }
351
352 if (unlikely(!raw_spin_trylock_irqsave(&meta->lock, flags))) {
353 /*
354 * This is extremely unlikely -- we are reporting on a
355 * use-after-free, which locked meta->lock, and the reporting
356 * code via printk calls kmalloc() which ends up in
357 * kfence_alloc() and tries to grab the same object that we're
358 * reporting on. While it has never been observed, lockdep does
359 * report that there is a possibility of deadlock. Fix it by
360 * using trylock and bailing out gracefully.
361 */
362 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
363 /* Put the object back on the freelist. */
364 list_add_tail(&meta->list, &kfence_freelist);
365 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
366
367 return NULL;
368 }
369
370 meta->addr = metadata_to_pageaddr(meta);
371 /* Unprotect if we're reusing this page. */
372 if (meta->state == KFENCE_OBJECT_FREED)
373 kfence_unprotect(meta->addr);
374
375 /*
376 * Note: for allocations made before RNG initialization, will always
377 * return zero. We still benefit from enabling KFENCE as early as
378 * possible, even when the RNG is not yet available, as this will allow
379 * KFENCE to detect bugs due to earlier allocations. The only downside
380 * is that the out-of-bounds accesses detected are deterministic for
381 * such allocations.
382 */
383 if (prandom_u32_max(2)) {
384 /* Allocate on the "right" side, re-calculate address. */
385 meta->addr += PAGE_SIZE - size;
386 meta->addr = ALIGN_DOWN(meta->addr, cache->align);
387 }
388
389 addr = (void *)meta->addr;
390
391 /* Update remaining metadata. */
392 metadata_update_state(meta, KFENCE_OBJECT_ALLOCATED, stack_entries, num_stack_entries);
393 /* Pairs with READ_ONCE() in kfence_shutdown_cache(). */
394 WRITE_ONCE(meta->cache, cache);
395 meta->size = size;
396 meta->alloc_stack_hash = alloc_stack_hash;
397
398 for_each_canary(meta, set_canary_byte);
399
400 /* Set required struct page fields. */
401 page = virt_to_page(meta->addr);
402 page->slab_cache = cache;
403 if (IS_ENABLED(CONFIG_SLUB))
404 page->objects = 1;
405 if (IS_ENABLED(CONFIG_SLAB))
406 page->s_mem = addr;
407
408 raw_spin_unlock_irqrestore(&meta->lock, flags);
409
410 alloc_covered_add(alloc_stack_hash, 1);
411
412 /* Memory initialization. */
413
414 /*
415 * We check slab_want_init_on_alloc() ourselves, rather than letting
416 * SL*B do the initialization, as otherwise we might overwrite KFENCE's
417 * redzone.
418 */
419 if (unlikely(slab_want_init_on_alloc(gfp, cache)))
420 memzero_explicit(addr, size);
421 if (cache->ctor)
422 cache->ctor(addr);
423
424 if (CONFIG_KFENCE_STRESS_TEST_FAULTS && !prandom_u32_max(CONFIG_KFENCE_STRESS_TEST_FAULTS))
425 kfence_protect(meta->addr); /* Random "faults" by protecting the object. */
426
427 atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCATED]);
428 atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCS]);
429
430 return addr;
431 }
432
kfence_guarded_free(void * addr,struct kfence_metadata * meta,bool zombie)433 static void kfence_guarded_free(void *addr, struct kfence_metadata *meta, bool zombie)
434 {
435 struct kcsan_scoped_access assert_page_exclusive;
436 unsigned long flags;
437
438 raw_spin_lock_irqsave(&meta->lock, flags);
439
440 if (meta->state != KFENCE_OBJECT_ALLOCATED || meta->addr != (unsigned long)addr) {
441 /* Invalid or double-free, bail out. */
442 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
443 kfence_report_error((unsigned long)addr, false, NULL, meta,
444 KFENCE_ERROR_INVALID_FREE);
445 raw_spin_unlock_irqrestore(&meta->lock, flags);
446 return;
447 }
448
449 /* Detect racy use-after-free, or incorrect reallocation of this page by KFENCE. */
450 kcsan_begin_scoped_access((void *)ALIGN_DOWN((unsigned long)addr, PAGE_SIZE), PAGE_SIZE,
451 KCSAN_ACCESS_SCOPED | KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT,
452 &assert_page_exclusive);
453
454 if (CONFIG_KFENCE_STRESS_TEST_FAULTS)
455 kfence_unprotect((unsigned long)addr); /* To check canary bytes. */
456
457 /* Restore page protection if there was an OOB access. */
458 if (meta->unprotected_page) {
459 memzero_explicit((void *)ALIGN_DOWN(meta->unprotected_page, PAGE_SIZE), PAGE_SIZE);
460 kfence_protect(meta->unprotected_page);
461 meta->unprotected_page = 0;
462 }
463
464 /* Check canary bytes for memory corruption. */
465 for_each_canary(meta, check_canary_byte);
466
467 /*
468 * Clear memory if init-on-free is set. While we protect the page, the
469 * data is still there, and after a use-after-free is detected, we
470 * unprotect the page, so the data is still accessible.
471 */
472 if (!zombie && unlikely(slab_want_init_on_free(meta->cache)))
473 memzero_explicit(addr, meta->size);
474
475 /* Mark the object as freed. */
476 metadata_update_state(meta, KFENCE_OBJECT_FREED, NULL, 0);
477
478 raw_spin_unlock_irqrestore(&meta->lock, flags);
479
480 alloc_covered_add(meta->alloc_stack_hash, -1);
481
482 /* Protect to detect use-after-frees. */
483 kfence_protect((unsigned long)addr);
484
485 kcsan_end_scoped_access(&assert_page_exclusive);
486 if (!zombie) {
487 /* Add it to the tail of the freelist for reuse. */
488 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
489 KFENCE_WARN_ON(!list_empty(&meta->list));
490 list_add_tail(&meta->list, &kfence_freelist);
491 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
492
493 atomic_long_dec(&counters[KFENCE_COUNTER_ALLOCATED]);
494 atomic_long_inc(&counters[KFENCE_COUNTER_FREES]);
495 } else {
496 /* See kfence_shutdown_cache(). */
497 atomic_long_inc(&counters[KFENCE_COUNTER_ZOMBIES]);
498 }
499 }
500
rcu_guarded_free(struct rcu_head * h)501 static void rcu_guarded_free(struct rcu_head *h)
502 {
503 struct kfence_metadata *meta = container_of(h, struct kfence_metadata, rcu_head);
504
505 kfence_guarded_free((void *)meta->addr, meta, false);
506 }
507
kfence_init_pool(void)508 static bool __init kfence_init_pool(void)
509 {
510 unsigned long addr = (unsigned long)__kfence_pool;
511 struct page *pages;
512 int i;
513 char *p;
514
515 if (!__kfence_pool)
516 return false;
517
518 if (!arch_kfence_init_pool())
519 goto err;
520
521 pages = virt_to_page(addr);
522
523 /*
524 * Set up object pages: they must have PG_slab set, to avoid freeing
525 * these as real pages.
526 *
527 * We also want to avoid inserting kfence_free() in the kfree()
528 * fast-path in SLUB, and therefore need to ensure kfree() correctly
529 * enters __slab_free() slow-path.
530 */
531 for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
532 struct page *page = &pages[i];
533
534 if (!i || (i % 2))
535 continue;
536
537 /* Verify we do not have a compound head page. */
538 if (WARN_ON(compound_head(&pages[i]) != &pages[i]))
539 goto err;
540
541 __SetPageSlab(page);
542 #ifdef CONFIG_MEMCG
543 page->memcg_data = (unsigned long)&kfence_metadata[i / 2 - 1].objcg |
544 MEMCG_DATA_OBJCGS;
545 #endif
546 }
547
548 /*
549 * Protect the first 2 pages. The first page is mostly unnecessary, and
550 * merely serves as an extended guard page. However, adding one
551 * additional page in the beginning gives us an even number of pages,
552 * which simplifies the mapping of address to metadata index.
553 */
554 for (i = 0; i < 2; i++) {
555 if (unlikely(!kfence_protect(addr)))
556 goto err;
557
558 addr += PAGE_SIZE;
559 }
560
561 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
562 struct kfence_metadata *meta = &kfence_metadata[i];
563
564 /* Initialize metadata. */
565 INIT_LIST_HEAD(&meta->list);
566 raw_spin_lock_init(&meta->lock);
567 meta->state = KFENCE_OBJECT_UNUSED;
568 meta->addr = addr; /* Initialize for validation in metadata_to_pageaddr(). */
569 list_add_tail(&meta->list, &kfence_freelist);
570
571 /* Protect the right redzone. */
572 if (unlikely(!kfence_protect(addr + PAGE_SIZE)))
573 goto err;
574
575 addr += 2 * PAGE_SIZE;
576 }
577
578 /*
579 * The pool is live and will never be deallocated from this point on.
580 * Remove the pool object from the kmemleak object tree, as it would
581 * otherwise overlap with allocations returned by kfence_alloc(), which
582 * are registered with kmemleak through the slab post-alloc hook.
583 */
584 kmemleak_free(__kfence_pool);
585
586 return true;
587
588 err:
589 /*
590 * Only release unprotected pages, and do not try to go back and change
591 * page attributes due to risk of failing to do so as well. If changing
592 * page attributes for some pages fails, it is very likely that it also
593 * fails for the first page, and therefore expect addr==__kfence_pool in
594 * most failure cases.
595 */
596 for (p = (char *)addr; p < __kfence_pool + KFENCE_POOL_SIZE; p += PAGE_SIZE) {
597 struct page *page = virt_to_page(p);
598
599 if (!PageSlab(page))
600 continue;
601 #ifdef CONFIG_MEMCG
602 page->memcg_data = 0;
603 #endif
604 __ClearPageSlab(page);
605 }
606 memblock_free_late(__pa(addr), KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool));
607 __kfence_pool = NULL;
608 return false;
609 }
610
611 /* === DebugFS Interface ==================================================== */
612
stats_show(struct seq_file * seq,void * v)613 static int stats_show(struct seq_file *seq, void *v)
614 {
615 int i;
616
617 seq_printf(seq, "enabled: %i\n", READ_ONCE(kfence_enabled));
618 for (i = 0; i < KFENCE_COUNTER_COUNT; i++)
619 seq_printf(seq, "%s: %ld\n", counter_names[i], atomic_long_read(&counters[i]));
620
621 return 0;
622 }
623 DEFINE_SHOW_ATTRIBUTE(stats);
624
625 /*
626 * debugfs seq_file operations for /sys/kernel/debug/kfence/objects.
627 * start_object() and next_object() return the object index + 1, because NULL is used
628 * to stop iteration.
629 */
start_object(struct seq_file * seq,loff_t * pos)630 static void *start_object(struct seq_file *seq, loff_t *pos)
631 {
632 if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
633 return (void *)((long)*pos + 1);
634 return NULL;
635 }
636
stop_object(struct seq_file * seq,void * v)637 static void stop_object(struct seq_file *seq, void *v)
638 {
639 }
640
next_object(struct seq_file * seq,void * v,loff_t * pos)641 static void *next_object(struct seq_file *seq, void *v, loff_t *pos)
642 {
643 ++*pos;
644 if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
645 return (void *)((long)*pos + 1);
646 return NULL;
647 }
648
show_object(struct seq_file * seq,void * v)649 static int show_object(struct seq_file *seq, void *v)
650 {
651 struct kfence_metadata *meta = &kfence_metadata[(long)v - 1];
652 unsigned long flags;
653
654 raw_spin_lock_irqsave(&meta->lock, flags);
655 kfence_print_object(seq, meta);
656 raw_spin_unlock_irqrestore(&meta->lock, flags);
657 seq_puts(seq, "---------------------------------\n");
658
659 return 0;
660 }
661
662 static const struct seq_operations object_seqops = {
663 .start = start_object,
664 .next = next_object,
665 .stop = stop_object,
666 .show = show_object,
667 };
668
open_objects(struct inode * inode,struct file * file)669 static int open_objects(struct inode *inode, struct file *file)
670 {
671 return seq_open(file, &object_seqops);
672 }
673
674 static const struct file_operations objects_fops = {
675 .open = open_objects,
676 .read = seq_read,
677 .llseek = seq_lseek,
678 .release = seq_release,
679 };
680
kfence_debugfs_init(void)681 static int kfence_debugfs_init(void)
682 {
683 struct dentry *kfence_dir;
684
685 if (!READ_ONCE(kfence_enabled))
686 return 0;
687
688 kfence_dir = debugfs_create_dir("kfence", NULL);
689 debugfs_create_file("stats", 0444, kfence_dir, NULL, &stats_fops);
690 debugfs_create_file("objects", 0400, kfence_dir, NULL, &objects_fops);
691 return 0;
692 }
693
694 late_initcall(kfence_debugfs_init);
695
696 /* === Allocation Gate Timer ================================================ */
697
698 #ifdef CONFIG_KFENCE_STATIC_KEYS
699 /* Wait queue to wake up allocation-gate timer task. */
700 static DECLARE_WAIT_QUEUE_HEAD(allocation_wait);
701
wake_up_kfence_timer(struct irq_work * work)702 static void wake_up_kfence_timer(struct irq_work *work)
703 {
704 wake_up(&allocation_wait);
705 }
706 static DEFINE_IRQ_WORK(wake_up_kfence_timer_work, wake_up_kfence_timer);
707 #endif
708
709 /*
710 * Set up delayed work, which will enable and disable the static key. We need to
711 * use a work queue (rather than a simple timer), since enabling and disabling a
712 * static key cannot be done from an interrupt.
713 *
714 * Note: Toggling a static branch currently causes IPIs, and here we'll end up
715 * with a total of 2 IPIs to all CPUs. If this ends up a problem in future (with
716 * more aggressive sampling intervals), we could get away with a variant that
717 * avoids IPIs, at the cost of not immediately capturing allocations if the
718 * instructions remain cached.
719 */
720 static struct delayed_work kfence_timer;
toggle_allocation_gate(struct work_struct * work)721 static void toggle_allocation_gate(struct work_struct *work)
722 {
723 if (!READ_ONCE(kfence_enabled))
724 return;
725
726 atomic_set(&kfence_allocation_gate, 0);
727 #ifdef CONFIG_KFENCE_STATIC_KEYS
728 /* Enable static key, and await allocation to happen. */
729 static_branch_enable(&kfence_allocation_key);
730
731 if (sysctl_hung_task_timeout_secs) {
732 /*
733 * During low activity with no allocations we might wait a
734 * while; let's avoid the hung task warning.
735 */
736 wait_event_idle_timeout(allocation_wait, atomic_read(&kfence_allocation_gate),
737 sysctl_hung_task_timeout_secs * HZ / 2);
738 } else {
739 wait_event_idle(allocation_wait, atomic_read(&kfence_allocation_gate));
740 }
741
742 /* Disable static key and reset timer. */
743 static_branch_disable(&kfence_allocation_key);
744 #endif
745 queue_delayed_work(system_unbound_wq, &kfence_timer,
746 msecs_to_jiffies(kfence_sample_interval));
747 }
748 static DECLARE_DELAYED_WORK(kfence_timer, toggle_allocation_gate);
749
750 /* === Public interface ===================================================== */
751
kfence_alloc_pool(void)752 void __init kfence_alloc_pool(void)
753 {
754 if (!kfence_sample_interval)
755 return;
756
757 __kfence_pool = memblock_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
758
759 if (!__kfence_pool)
760 pr_err("failed to allocate pool\n");
761 }
762
kfence_init(void)763 void __init kfence_init(void)
764 {
765 /* Setting kfence_sample_interval to 0 on boot disables KFENCE. */
766 if (!kfence_sample_interval)
767 return;
768
769 stack_hash_seed = (u32)random_get_entropy();
770 if (!kfence_init_pool()) {
771 pr_err("%s failed\n", __func__);
772 return;
773 }
774
775 if (!IS_ENABLED(CONFIG_KFENCE_STATIC_KEYS))
776 static_branch_enable(&kfence_allocation_key);
777 WRITE_ONCE(kfence_enabled, true);
778 queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
779 pr_info("initialized - using %lu bytes for %d objects at 0x%p-0x%p\n", KFENCE_POOL_SIZE,
780 CONFIG_KFENCE_NUM_OBJECTS, (void *)__kfence_pool,
781 (void *)(__kfence_pool + KFENCE_POOL_SIZE));
782 }
783
kfence_shutdown_cache(struct kmem_cache * s)784 void kfence_shutdown_cache(struct kmem_cache *s)
785 {
786 unsigned long flags;
787 struct kfence_metadata *meta;
788 int i;
789
790 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
791 bool in_use;
792
793 meta = &kfence_metadata[i];
794
795 /*
796 * If we observe some inconsistent cache and state pair where we
797 * should have returned false here, cache destruction is racing
798 * with either kmem_cache_alloc() or kmem_cache_free(). Taking
799 * the lock will not help, as different critical section
800 * serialization will have the same outcome.
801 */
802 if (READ_ONCE(meta->cache) != s ||
803 READ_ONCE(meta->state) != KFENCE_OBJECT_ALLOCATED)
804 continue;
805
806 raw_spin_lock_irqsave(&meta->lock, flags);
807 in_use = meta->cache == s && meta->state == KFENCE_OBJECT_ALLOCATED;
808 raw_spin_unlock_irqrestore(&meta->lock, flags);
809
810 if (in_use) {
811 /*
812 * This cache still has allocations, and we should not
813 * release them back into the freelist so they can still
814 * safely be used and retain the kernel's default
815 * behaviour of keeping the allocations alive (leak the
816 * cache); however, they effectively become "zombie
817 * allocations" as the KFENCE objects are the only ones
818 * still in use and the owning cache is being destroyed.
819 *
820 * We mark them freed, so that any subsequent use shows
821 * more useful error messages that will include stack
822 * traces of the user of the object, the original
823 * allocation, and caller to shutdown_cache().
824 */
825 kfence_guarded_free((void *)meta->addr, meta, /*zombie=*/true);
826 }
827 }
828
829 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
830 meta = &kfence_metadata[i];
831
832 /* See above. */
833 if (READ_ONCE(meta->cache) != s || READ_ONCE(meta->state) != KFENCE_OBJECT_FREED)
834 continue;
835
836 raw_spin_lock_irqsave(&meta->lock, flags);
837 if (meta->cache == s && meta->state == KFENCE_OBJECT_FREED)
838 meta->cache = NULL;
839 raw_spin_unlock_irqrestore(&meta->lock, flags);
840 }
841 }
842
__kfence_alloc(struct kmem_cache * s,size_t size,gfp_t flags)843 void *__kfence_alloc(struct kmem_cache *s, size_t size, gfp_t flags)
844 {
845 unsigned long stack_entries[KFENCE_STACK_DEPTH];
846 size_t num_stack_entries;
847 u32 alloc_stack_hash;
848
849 /*
850 * Perform size check before switching kfence_allocation_gate, so that
851 * we don't disable KFENCE without making an allocation.
852 */
853 if (size > PAGE_SIZE) {
854 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
855 return NULL;
856 }
857
858 /*
859 * Skip allocations from non-default zones, including DMA. We cannot
860 * guarantee that pages in the KFENCE pool will have the requested
861 * properties (e.g. reside in DMAable memory).
862 */
863 if ((flags & GFP_ZONEMASK) ||
864 (s->flags & (SLAB_CACHE_DMA | SLAB_CACHE_DMA32))) {
865 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
866 return NULL;
867 }
868
869 if (atomic_inc_return(&kfence_allocation_gate) > 1)
870 return NULL;
871 #ifdef CONFIG_KFENCE_STATIC_KEYS
872 /*
873 * waitqueue_active() is fully ordered after the update of
874 * kfence_allocation_gate per atomic_inc_return().
875 */
876 if (waitqueue_active(&allocation_wait)) {
877 /*
878 * Calling wake_up() here may deadlock when allocations happen
879 * from within timer code. Use an irq_work to defer it.
880 */
881 irq_work_queue(&wake_up_kfence_timer_work);
882 }
883 #endif
884
885 if (!READ_ONCE(kfence_enabled))
886 return NULL;
887
888 num_stack_entries = stack_trace_save(stack_entries, KFENCE_STACK_DEPTH, 0);
889
890 /*
891 * Do expensive check for coverage of allocation in slow-path after
892 * allocation_gate has already become non-zero, even though it might
893 * mean not making any allocation within a given sample interval.
894 *
895 * This ensures reasonable allocation coverage when the pool is almost
896 * full, including avoiding long-lived allocations of the same source
897 * filling up the pool (e.g. pagecache allocations).
898 */
899 alloc_stack_hash = get_alloc_stack_hash(stack_entries, num_stack_entries);
900 if (should_skip_covered() && alloc_covered_contains(alloc_stack_hash)) {
901 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_COVERED]);
902 return NULL;
903 }
904
905 return kfence_guarded_alloc(s, size, flags, stack_entries, num_stack_entries,
906 alloc_stack_hash);
907 }
908
kfence_ksize(const void * addr)909 size_t kfence_ksize(const void *addr)
910 {
911 const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
912
913 /*
914 * Read locklessly -- if there is a race with __kfence_alloc(), this is
915 * either a use-after-free or invalid access.
916 */
917 return meta ? meta->size : 0;
918 }
919
kfence_object_start(const void * addr)920 void *kfence_object_start(const void *addr)
921 {
922 const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
923
924 /*
925 * Read locklessly -- if there is a race with __kfence_alloc(), this is
926 * either a use-after-free or invalid access.
927 */
928 return meta ? (void *)meta->addr : NULL;
929 }
930
__kfence_free(void * addr)931 void __kfence_free(void *addr)
932 {
933 struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
934
935 #ifdef CONFIG_MEMCG
936 KFENCE_WARN_ON(meta->objcg);
937 #endif
938 /*
939 * If the objects of the cache are SLAB_TYPESAFE_BY_RCU, defer freeing
940 * the object, as the object page may be recycled for other-typed
941 * objects once it has been freed. meta->cache may be NULL if the cache
942 * was destroyed.
943 */
944 if (unlikely(meta->cache && (meta->cache->flags & SLAB_TYPESAFE_BY_RCU)))
945 call_rcu(&meta->rcu_head, rcu_guarded_free);
946 else
947 kfence_guarded_free(addr, meta, false);
948 }
949
kfence_handle_page_fault(unsigned long addr,bool is_write,struct pt_regs * regs)950 bool kfence_handle_page_fault(unsigned long addr, bool is_write, struct pt_regs *regs)
951 {
952 const int page_index = (addr - (unsigned long)__kfence_pool) / PAGE_SIZE;
953 struct kfence_metadata *to_report = NULL;
954 enum kfence_error_type error_type;
955 unsigned long flags;
956
957 if (!is_kfence_address((void *)addr))
958 return false;
959
960 if (!READ_ONCE(kfence_enabled)) /* If disabled at runtime ... */
961 return kfence_unprotect(addr); /* ... unprotect and proceed. */
962
963 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
964
965 if (page_index % 2) {
966 /* This is a redzone, report a buffer overflow. */
967 struct kfence_metadata *meta;
968 int distance = 0;
969
970 meta = addr_to_metadata(addr - PAGE_SIZE);
971 if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
972 to_report = meta;
973 /* Data race ok; distance calculation approximate. */
974 distance = addr - data_race(meta->addr + meta->size);
975 }
976
977 meta = addr_to_metadata(addr + PAGE_SIZE);
978 if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
979 /* Data race ok; distance calculation approximate. */
980 if (!to_report || distance > data_race(meta->addr) - addr)
981 to_report = meta;
982 }
983
984 if (!to_report)
985 goto out;
986
987 raw_spin_lock_irqsave(&to_report->lock, flags);
988 to_report->unprotected_page = addr;
989 error_type = KFENCE_ERROR_OOB;
990
991 /*
992 * If the object was freed before we took the look we can still
993 * report this as an OOB -- the report will simply show the
994 * stacktrace of the free as well.
995 */
996 } else {
997 to_report = addr_to_metadata(addr);
998 if (!to_report)
999 goto out;
1000
1001 raw_spin_lock_irqsave(&to_report->lock, flags);
1002 error_type = KFENCE_ERROR_UAF;
1003 /*
1004 * We may race with __kfence_alloc(), and it is possible that a
1005 * freed object may be reallocated. We simply report this as a
1006 * use-after-free, with the stack trace showing the place where
1007 * the object was re-allocated.
1008 */
1009 }
1010
1011 out:
1012 if (to_report) {
1013 kfence_report_error(addr, is_write, regs, to_report, error_type);
1014 raw_spin_unlock_irqrestore(&to_report->lock, flags);
1015 } else {
1016 /* This may be a UAF or OOB access, but we can't be sure. */
1017 kfence_report_error(addr, is_write, regs, NULL, KFENCE_ERROR_INVALID);
1018 }
1019
1020 return kfence_unprotect(addr); /* Unprotect and let access proceed. */
1021 }
1022