GNU Linux-libre 6.9.1-gnu
[releases.git] / mm / kfence / core.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * KFENCE guarded object allocator and fault handling.
4  *
5  * Copyright (C) 2020, Google LLC.
6  */
7
8 #define pr_fmt(fmt) "kfence: " fmt
9
10 #include <linux/atomic.h>
11 #include <linux/bug.h>
12 #include <linux/debugfs.h>
13 #include <linux/hash.h>
14 #include <linux/irq_work.h>
15 #include <linux/jhash.h>
16 #include <linux/kcsan-checks.h>
17 #include <linux/kfence.h>
18 #include <linux/kmemleak.h>
19 #include <linux/list.h>
20 #include <linux/lockdep.h>
21 #include <linux/log2.h>
22 #include <linux/memblock.h>
23 #include <linux/moduleparam.h>
24 #include <linux/notifier.h>
25 #include <linux/panic_notifier.h>
26 #include <linux/random.h>
27 #include <linux/rcupdate.h>
28 #include <linux/sched/clock.h>
29 #include <linux/seq_file.h>
30 #include <linux/slab.h>
31 #include <linux/spinlock.h>
32 #include <linux/string.h>
33
34 #include <asm/kfence.h>
35
36 #include "kfence.h"
37
38 /* Disables KFENCE on the first warning assuming an irrecoverable error. */
39 #define KFENCE_WARN_ON(cond)                                                   \
40         ({                                                                     \
41                 const bool __cond = WARN_ON(cond);                             \
42                 if (unlikely(__cond)) {                                        \
43                         WRITE_ONCE(kfence_enabled, false);                     \
44                         disabled_by_warn = true;                               \
45                 }                                                              \
46                 __cond;                                                        \
47         })
48
49 /* === Data ================================================================= */
50
51 static bool kfence_enabled __read_mostly;
52 static bool disabled_by_warn __read_mostly;
53
54 unsigned long kfence_sample_interval __read_mostly = CONFIG_KFENCE_SAMPLE_INTERVAL;
55 EXPORT_SYMBOL_GPL(kfence_sample_interval); /* Export for test modules. */
56
57 #ifdef MODULE_PARAM_PREFIX
58 #undef MODULE_PARAM_PREFIX
59 #endif
60 #define MODULE_PARAM_PREFIX "kfence."
61
62 static int kfence_enable_late(void);
63 static int param_set_sample_interval(const char *val, const struct kernel_param *kp)
64 {
65         unsigned long num;
66         int ret = kstrtoul(val, 0, &num);
67
68         if (ret < 0)
69                 return ret;
70
71         /* Using 0 to indicate KFENCE is disabled. */
72         if (!num && READ_ONCE(kfence_enabled)) {
73                 pr_info("disabled\n");
74                 WRITE_ONCE(kfence_enabled, false);
75         }
76
77         *((unsigned long *)kp->arg) = num;
78
79         if (num && !READ_ONCE(kfence_enabled) && system_state != SYSTEM_BOOTING)
80                 return disabled_by_warn ? -EINVAL : kfence_enable_late();
81         return 0;
82 }
83
84 static int param_get_sample_interval(char *buffer, const struct kernel_param *kp)
85 {
86         if (!READ_ONCE(kfence_enabled))
87                 return sprintf(buffer, "0\n");
88
89         return param_get_ulong(buffer, kp);
90 }
91
92 static const struct kernel_param_ops sample_interval_param_ops = {
93         .set = param_set_sample_interval,
94         .get = param_get_sample_interval,
95 };
96 module_param_cb(sample_interval, &sample_interval_param_ops, &kfence_sample_interval, 0600);
97
98 /* Pool usage% threshold when currently covered allocations are skipped. */
99 static unsigned long kfence_skip_covered_thresh __read_mostly = 75;
100 module_param_named(skip_covered_thresh, kfence_skip_covered_thresh, ulong, 0644);
101
102 /* If true, use a deferrable timer. */
103 static bool kfence_deferrable __read_mostly = IS_ENABLED(CONFIG_KFENCE_DEFERRABLE);
104 module_param_named(deferrable, kfence_deferrable, bool, 0444);
105
106 /* If true, check all canary bytes on panic. */
107 static bool kfence_check_on_panic __read_mostly;
108 module_param_named(check_on_panic, kfence_check_on_panic, bool, 0444);
109
110 /* The pool of pages used for guard pages and objects. */
111 char *__kfence_pool __read_mostly;
112 EXPORT_SYMBOL(__kfence_pool); /* Export for test modules. */
113
114 /*
115  * Per-object metadata, with one-to-one mapping of object metadata to
116  * backing pages (in __kfence_pool).
117  */
118 static_assert(CONFIG_KFENCE_NUM_OBJECTS > 0);
119 struct kfence_metadata *kfence_metadata __read_mostly;
120
121 /*
122  * If kfence_metadata is not NULL, it may be accessed by kfence_shutdown_cache().
123  * So introduce kfence_metadata_init to initialize metadata, and then make
124  * kfence_metadata visible after initialization is successful. This prevents
125  * potential UAF or access to uninitialized metadata.
126  */
127 static struct kfence_metadata *kfence_metadata_init __read_mostly;
128
129 /* Freelist with available objects. */
130 static struct list_head kfence_freelist = LIST_HEAD_INIT(kfence_freelist);
131 static DEFINE_RAW_SPINLOCK(kfence_freelist_lock); /* Lock protecting freelist. */
132
133 /*
134  * The static key to set up a KFENCE allocation; or if static keys are not used
135  * to gate allocations, to avoid a load and compare if KFENCE is disabled.
136  */
137 DEFINE_STATIC_KEY_FALSE(kfence_allocation_key);
138
139 /* Gates the allocation, ensuring only one succeeds in a given period. */
140 atomic_t kfence_allocation_gate = ATOMIC_INIT(1);
141
142 /*
143  * A Counting Bloom filter of allocation coverage: limits currently covered
144  * allocations of the same source filling up the pool.
145  *
146  * Assuming a range of 15%-85% unique allocations in the pool at any point in
147  * time, the below parameters provide a probablity of 0.02-0.33 for false
148  * positive hits respectively:
149  *
150  *      P(alloc_traces) = (1 - e^(-HNUM * (alloc_traces / SIZE)) ^ HNUM
151  */
152 #define ALLOC_COVERED_HNUM      2
153 #define ALLOC_COVERED_ORDER     (const_ilog2(CONFIG_KFENCE_NUM_OBJECTS) + 2)
154 #define ALLOC_COVERED_SIZE      (1 << ALLOC_COVERED_ORDER)
155 #define ALLOC_COVERED_HNEXT(h)  hash_32(h, ALLOC_COVERED_ORDER)
156 #define ALLOC_COVERED_MASK      (ALLOC_COVERED_SIZE - 1)
157 static atomic_t alloc_covered[ALLOC_COVERED_SIZE];
158
159 /* Stack depth used to determine uniqueness of an allocation. */
160 #define UNIQUE_ALLOC_STACK_DEPTH ((size_t)8)
161
162 /*
163  * Randomness for stack hashes, making the same collisions across reboots and
164  * different machines less likely.
165  */
166 static u32 stack_hash_seed __ro_after_init;
167
168 /* Statistics counters for debugfs. */
169 enum kfence_counter_id {
170         KFENCE_COUNTER_ALLOCATED,
171         KFENCE_COUNTER_ALLOCS,
172         KFENCE_COUNTER_FREES,
173         KFENCE_COUNTER_ZOMBIES,
174         KFENCE_COUNTER_BUGS,
175         KFENCE_COUNTER_SKIP_INCOMPAT,
176         KFENCE_COUNTER_SKIP_CAPACITY,
177         KFENCE_COUNTER_SKIP_COVERED,
178         KFENCE_COUNTER_COUNT,
179 };
180 static atomic_long_t counters[KFENCE_COUNTER_COUNT];
181 static const char *const counter_names[] = {
182         [KFENCE_COUNTER_ALLOCATED]      = "currently allocated",
183         [KFENCE_COUNTER_ALLOCS]         = "total allocations",
184         [KFENCE_COUNTER_FREES]          = "total frees",
185         [KFENCE_COUNTER_ZOMBIES]        = "zombie allocations",
186         [KFENCE_COUNTER_BUGS]           = "total bugs",
187         [KFENCE_COUNTER_SKIP_INCOMPAT]  = "skipped allocations (incompatible)",
188         [KFENCE_COUNTER_SKIP_CAPACITY]  = "skipped allocations (capacity)",
189         [KFENCE_COUNTER_SKIP_COVERED]   = "skipped allocations (covered)",
190 };
191 static_assert(ARRAY_SIZE(counter_names) == KFENCE_COUNTER_COUNT);
192
193 /* === Internals ============================================================ */
194
195 static inline bool should_skip_covered(void)
196 {
197         unsigned long thresh = (CONFIG_KFENCE_NUM_OBJECTS * kfence_skip_covered_thresh) / 100;
198
199         return atomic_long_read(&counters[KFENCE_COUNTER_ALLOCATED]) > thresh;
200 }
201
202 static u32 get_alloc_stack_hash(unsigned long *stack_entries, size_t num_entries)
203 {
204         num_entries = min(num_entries, UNIQUE_ALLOC_STACK_DEPTH);
205         num_entries = filter_irq_stacks(stack_entries, num_entries);
206         return jhash(stack_entries, num_entries * sizeof(stack_entries[0]), stack_hash_seed);
207 }
208
209 /*
210  * Adds (or subtracts) count @val for allocation stack trace hash
211  * @alloc_stack_hash from Counting Bloom filter.
212  */
213 static void alloc_covered_add(u32 alloc_stack_hash, int val)
214 {
215         int i;
216
217         for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
218                 atomic_add(val, &alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]);
219                 alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
220         }
221 }
222
223 /*
224  * Returns true if the allocation stack trace hash @alloc_stack_hash is
225  * currently contained (non-zero count) in Counting Bloom filter.
226  */
227 static bool alloc_covered_contains(u32 alloc_stack_hash)
228 {
229         int i;
230
231         for (i = 0; i < ALLOC_COVERED_HNUM; i++) {
232                 if (!atomic_read(&alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]))
233                         return false;
234                 alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash);
235         }
236
237         return true;
238 }
239
240 static bool kfence_protect(unsigned long addr)
241 {
242         return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), true));
243 }
244
245 static bool kfence_unprotect(unsigned long addr)
246 {
247         return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), false));
248 }
249
250 static inline unsigned long metadata_to_pageaddr(const struct kfence_metadata *meta)
251 {
252         unsigned long offset = (meta - kfence_metadata + 1) * PAGE_SIZE * 2;
253         unsigned long pageaddr = (unsigned long)&__kfence_pool[offset];
254
255         /* The checks do not affect performance; only called from slow-paths. */
256
257         /* Only call with a pointer into kfence_metadata. */
258         if (KFENCE_WARN_ON(meta < kfence_metadata ||
259                            meta >= kfence_metadata + CONFIG_KFENCE_NUM_OBJECTS))
260                 return 0;
261
262         /*
263          * This metadata object only ever maps to 1 page; verify that the stored
264          * address is in the expected range.
265          */
266         if (KFENCE_WARN_ON(ALIGN_DOWN(meta->addr, PAGE_SIZE) != pageaddr))
267                 return 0;
268
269         return pageaddr;
270 }
271
272 /*
273  * Update the object's metadata state, including updating the alloc/free stacks
274  * depending on the state transition.
275  */
276 static noinline void
277 metadata_update_state(struct kfence_metadata *meta, enum kfence_object_state next,
278                       unsigned long *stack_entries, size_t num_stack_entries)
279 {
280         struct kfence_track *track =
281                 next == KFENCE_OBJECT_FREED ? &meta->free_track : &meta->alloc_track;
282
283         lockdep_assert_held(&meta->lock);
284
285         if (stack_entries) {
286                 memcpy(track->stack_entries, stack_entries,
287                        num_stack_entries * sizeof(stack_entries[0]));
288         } else {
289                 /*
290                  * Skip over 1 (this) functions; noinline ensures we do not
291                  * accidentally skip over the caller by never inlining.
292                  */
293                 num_stack_entries = stack_trace_save(track->stack_entries, KFENCE_STACK_DEPTH, 1);
294         }
295         track->num_stack_entries = num_stack_entries;
296         track->pid = task_pid_nr(current);
297         track->cpu = raw_smp_processor_id();
298         track->ts_nsec = local_clock(); /* Same source as printk timestamps. */
299
300         /*
301          * Pairs with READ_ONCE() in
302          *      kfence_shutdown_cache(),
303          *      kfence_handle_page_fault().
304          */
305         WRITE_ONCE(meta->state, next);
306 }
307
308 /* Check canary byte at @addr. */
309 static inline bool check_canary_byte(u8 *addr)
310 {
311         struct kfence_metadata *meta;
312         unsigned long flags;
313
314         if (likely(*addr == KFENCE_CANARY_PATTERN_U8(addr)))
315                 return true;
316
317         atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
318
319         meta = addr_to_metadata((unsigned long)addr);
320         raw_spin_lock_irqsave(&meta->lock, flags);
321         kfence_report_error((unsigned long)addr, false, NULL, meta, KFENCE_ERROR_CORRUPTION);
322         raw_spin_unlock_irqrestore(&meta->lock, flags);
323
324         return false;
325 }
326
327 static inline void set_canary(const struct kfence_metadata *meta)
328 {
329         const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
330         unsigned long addr = pageaddr;
331
332         /*
333          * The canary may be written to part of the object memory, but it does
334          * not affect it. The user should initialize the object before using it.
335          */
336         for (; addr < meta->addr; addr += sizeof(u64))
337                 *((u64 *)addr) = KFENCE_CANARY_PATTERN_U64;
338
339         addr = ALIGN_DOWN(meta->addr + meta->size, sizeof(u64));
340         for (; addr - pageaddr < PAGE_SIZE; addr += sizeof(u64))
341                 *((u64 *)addr) = KFENCE_CANARY_PATTERN_U64;
342 }
343
344 static inline void check_canary(const struct kfence_metadata *meta)
345 {
346         const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
347         unsigned long addr = pageaddr;
348
349         /*
350          * We'll iterate over each canary byte per-side until a corrupted byte
351          * is found. However, we'll still iterate over the canary bytes to the
352          * right of the object even if there was an error in the canary bytes to
353          * the left of the object. Specifically, if check_canary_byte()
354          * generates an error, showing both sides might give more clues as to
355          * what the error is about when displaying which bytes were corrupted.
356          */
357
358         /* Apply to left of object. */
359         for (; meta->addr - addr >= sizeof(u64); addr += sizeof(u64)) {
360                 if (unlikely(*((u64 *)addr) != KFENCE_CANARY_PATTERN_U64))
361                         break;
362         }
363
364         /*
365          * If the canary is corrupted in a certain 64 bytes, or the canary
366          * memory cannot be completely covered by multiple consecutive 64 bytes,
367          * it needs to be checked one by one.
368          */
369         for (; addr < meta->addr; addr++) {
370                 if (unlikely(!check_canary_byte((u8 *)addr)))
371                         break;
372         }
373
374         /* Apply to right of object. */
375         for (addr = meta->addr + meta->size; addr % sizeof(u64) != 0; addr++) {
376                 if (unlikely(!check_canary_byte((u8 *)addr)))
377                         return;
378         }
379         for (; addr - pageaddr < PAGE_SIZE; addr += sizeof(u64)) {
380                 if (unlikely(*((u64 *)addr) != KFENCE_CANARY_PATTERN_U64)) {
381
382                         for (; addr - pageaddr < PAGE_SIZE; addr++) {
383                                 if (!check_canary_byte((u8 *)addr))
384                                         return;
385                         }
386                 }
387         }
388 }
389
390 static void *kfence_guarded_alloc(struct kmem_cache *cache, size_t size, gfp_t gfp,
391                                   unsigned long *stack_entries, size_t num_stack_entries,
392                                   u32 alloc_stack_hash)
393 {
394         struct kfence_metadata *meta = NULL;
395         unsigned long flags;
396         struct slab *slab;
397         void *addr;
398         const bool random_right_allocate = get_random_u32_below(2);
399         const bool random_fault = CONFIG_KFENCE_STRESS_TEST_FAULTS &&
400                                   !get_random_u32_below(CONFIG_KFENCE_STRESS_TEST_FAULTS);
401
402         /* Try to obtain a free object. */
403         raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
404         if (!list_empty(&kfence_freelist)) {
405                 meta = list_entry(kfence_freelist.next, struct kfence_metadata, list);
406                 list_del_init(&meta->list);
407         }
408         raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
409         if (!meta) {
410                 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_CAPACITY]);
411                 return NULL;
412         }
413
414         if (unlikely(!raw_spin_trylock_irqsave(&meta->lock, flags))) {
415                 /*
416                  * This is extremely unlikely -- we are reporting on a
417                  * use-after-free, which locked meta->lock, and the reporting
418                  * code via printk calls kmalloc() which ends up in
419                  * kfence_alloc() and tries to grab the same object that we're
420                  * reporting on. While it has never been observed, lockdep does
421                  * report that there is a possibility of deadlock. Fix it by
422                  * using trylock and bailing out gracefully.
423                  */
424                 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
425                 /* Put the object back on the freelist. */
426                 list_add_tail(&meta->list, &kfence_freelist);
427                 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
428
429                 return NULL;
430         }
431
432         meta->addr = metadata_to_pageaddr(meta);
433         /* Unprotect if we're reusing this page. */
434         if (meta->state == KFENCE_OBJECT_FREED)
435                 kfence_unprotect(meta->addr);
436
437         /*
438          * Note: for allocations made before RNG initialization, will always
439          * return zero. We still benefit from enabling KFENCE as early as
440          * possible, even when the RNG is not yet available, as this will allow
441          * KFENCE to detect bugs due to earlier allocations. The only downside
442          * is that the out-of-bounds accesses detected are deterministic for
443          * such allocations.
444          */
445         if (random_right_allocate) {
446                 /* Allocate on the "right" side, re-calculate address. */
447                 meta->addr += PAGE_SIZE - size;
448                 meta->addr = ALIGN_DOWN(meta->addr, cache->align);
449         }
450
451         addr = (void *)meta->addr;
452
453         /* Update remaining metadata. */
454         metadata_update_state(meta, KFENCE_OBJECT_ALLOCATED, stack_entries, num_stack_entries);
455         /* Pairs with READ_ONCE() in kfence_shutdown_cache(). */
456         WRITE_ONCE(meta->cache, cache);
457         meta->size = size;
458         meta->alloc_stack_hash = alloc_stack_hash;
459         raw_spin_unlock_irqrestore(&meta->lock, flags);
460
461         alloc_covered_add(alloc_stack_hash, 1);
462
463         /* Set required slab fields. */
464         slab = virt_to_slab((void *)meta->addr);
465         slab->slab_cache = cache;
466         slab->objects = 1;
467
468         /* Memory initialization. */
469         set_canary(meta);
470
471         /*
472          * We check slab_want_init_on_alloc() ourselves, rather than letting
473          * SL*B do the initialization, as otherwise we might overwrite KFENCE's
474          * redzone.
475          */
476         if (unlikely(slab_want_init_on_alloc(gfp, cache)))
477                 memzero_explicit(addr, size);
478         if (cache->ctor)
479                 cache->ctor(addr);
480
481         if (random_fault)
482                 kfence_protect(meta->addr); /* Random "faults" by protecting the object. */
483
484         atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCATED]);
485         atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCS]);
486
487         return addr;
488 }
489
490 static void kfence_guarded_free(void *addr, struct kfence_metadata *meta, bool zombie)
491 {
492         struct kcsan_scoped_access assert_page_exclusive;
493         unsigned long flags;
494         bool init;
495
496         raw_spin_lock_irqsave(&meta->lock, flags);
497
498         if (meta->state != KFENCE_OBJECT_ALLOCATED || meta->addr != (unsigned long)addr) {
499                 /* Invalid or double-free, bail out. */
500                 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
501                 kfence_report_error((unsigned long)addr, false, NULL, meta,
502                                     KFENCE_ERROR_INVALID_FREE);
503                 raw_spin_unlock_irqrestore(&meta->lock, flags);
504                 return;
505         }
506
507         /* Detect racy use-after-free, or incorrect reallocation of this page by KFENCE. */
508         kcsan_begin_scoped_access((void *)ALIGN_DOWN((unsigned long)addr, PAGE_SIZE), PAGE_SIZE,
509                                   KCSAN_ACCESS_SCOPED | KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT,
510                                   &assert_page_exclusive);
511
512         if (CONFIG_KFENCE_STRESS_TEST_FAULTS)
513                 kfence_unprotect((unsigned long)addr); /* To check canary bytes. */
514
515         /* Restore page protection if there was an OOB access. */
516         if (meta->unprotected_page) {
517                 memzero_explicit((void *)ALIGN_DOWN(meta->unprotected_page, PAGE_SIZE), PAGE_SIZE);
518                 kfence_protect(meta->unprotected_page);
519                 meta->unprotected_page = 0;
520         }
521
522         /* Mark the object as freed. */
523         metadata_update_state(meta, KFENCE_OBJECT_FREED, NULL, 0);
524         init = slab_want_init_on_free(meta->cache);
525         raw_spin_unlock_irqrestore(&meta->lock, flags);
526
527         alloc_covered_add(meta->alloc_stack_hash, -1);
528
529         /* Check canary bytes for memory corruption. */
530         check_canary(meta);
531
532         /*
533          * Clear memory if init-on-free is set. While we protect the page, the
534          * data is still there, and after a use-after-free is detected, we
535          * unprotect the page, so the data is still accessible.
536          */
537         if (!zombie && unlikely(init))
538                 memzero_explicit(addr, meta->size);
539
540         /* Protect to detect use-after-frees. */
541         kfence_protect((unsigned long)addr);
542
543         kcsan_end_scoped_access(&assert_page_exclusive);
544         if (!zombie) {
545                 /* Add it to the tail of the freelist for reuse. */
546                 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
547                 KFENCE_WARN_ON(!list_empty(&meta->list));
548                 list_add_tail(&meta->list, &kfence_freelist);
549                 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
550
551                 atomic_long_dec(&counters[KFENCE_COUNTER_ALLOCATED]);
552                 atomic_long_inc(&counters[KFENCE_COUNTER_FREES]);
553         } else {
554                 /* See kfence_shutdown_cache(). */
555                 atomic_long_inc(&counters[KFENCE_COUNTER_ZOMBIES]);
556         }
557 }
558
559 static void rcu_guarded_free(struct rcu_head *h)
560 {
561         struct kfence_metadata *meta = container_of(h, struct kfence_metadata, rcu_head);
562
563         kfence_guarded_free((void *)meta->addr, meta, false);
564 }
565
566 /*
567  * Initialization of the KFENCE pool after its allocation.
568  * Returns 0 on success; otherwise returns the address up to
569  * which partial initialization succeeded.
570  */
571 static unsigned long kfence_init_pool(void)
572 {
573         unsigned long addr;
574         struct page *pages;
575         int i;
576
577         if (!arch_kfence_init_pool())
578                 return (unsigned long)__kfence_pool;
579
580         addr = (unsigned long)__kfence_pool;
581         pages = virt_to_page(__kfence_pool);
582
583         /*
584          * Set up object pages: they must have PG_slab set, to avoid freeing
585          * these as real pages.
586          *
587          * We also want to avoid inserting kfence_free() in the kfree()
588          * fast-path in SLUB, and therefore need to ensure kfree() correctly
589          * enters __slab_free() slow-path.
590          */
591         for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
592                 struct slab *slab = page_slab(nth_page(pages, i));
593
594                 if (!i || (i % 2))
595                         continue;
596
597                 __folio_set_slab(slab_folio(slab));
598 #ifdef CONFIG_MEMCG
599                 slab->memcg_data = (unsigned long)&kfence_metadata_init[i / 2 - 1].objcg |
600                                    MEMCG_DATA_OBJCGS;
601 #endif
602         }
603
604         /*
605          * Protect the first 2 pages. The first page is mostly unnecessary, and
606          * merely serves as an extended guard page. However, adding one
607          * additional page in the beginning gives us an even number of pages,
608          * which simplifies the mapping of address to metadata index.
609          */
610         for (i = 0; i < 2; i++) {
611                 if (unlikely(!kfence_protect(addr)))
612                         return addr;
613
614                 addr += PAGE_SIZE;
615         }
616
617         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
618                 struct kfence_metadata *meta = &kfence_metadata_init[i];
619
620                 /* Initialize metadata. */
621                 INIT_LIST_HEAD(&meta->list);
622                 raw_spin_lock_init(&meta->lock);
623                 meta->state = KFENCE_OBJECT_UNUSED;
624                 meta->addr = addr; /* Initialize for validation in metadata_to_pageaddr(). */
625                 list_add_tail(&meta->list, &kfence_freelist);
626
627                 /* Protect the right redzone. */
628                 if (unlikely(!kfence_protect(addr + PAGE_SIZE)))
629                         goto reset_slab;
630
631                 addr += 2 * PAGE_SIZE;
632         }
633
634         /*
635          * Make kfence_metadata visible only when initialization is successful.
636          * Otherwise, if the initialization fails and kfence_metadata is freed,
637          * it may cause UAF in kfence_shutdown_cache().
638          */
639         smp_store_release(&kfence_metadata, kfence_metadata_init);
640         return 0;
641
642 reset_slab:
643         for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
644                 struct slab *slab = page_slab(nth_page(pages, i));
645
646                 if (!i || (i % 2))
647                         continue;
648 #ifdef CONFIG_MEMCG
649                 slab->memcg_data = 0;
650 #endif
651                 __folio_clear_slab(slab_folio(slab));
652         }
653
654         return addr;
655 }
656
657 static bool __init kfence_init_pool_early(void)
658 {
659         unsigned long addr;
660
661         if (!__kfence_pool)
662                 return false;
663
664         addr = kfence_init_pool();
665
666         if (!addr) {
667                 /*
668                  * The pool is live and will never be deallocated from this point on.
669                  * Ignore the pool object from the kmemleak phys object tree, as it would
670                  * otherwise overlap with allocations returned by kfence_alloc(), which
671                  * are registered with kmemleak through the slab post-alloc hook.
672                  */
673                 kmemleak_ignore_phys(__pa(__kfence_pool));
674                 return true;
675         }
676
677         /*
678          * Only release unprotected pages, and do not try to go back and change
679          * page attributes due to risk of failing to do so as well. If changing
680          * page attributes for some pages fails, it is very likely that it also
681          * fails for the first page, and therefore expect addr==__kfence_pool in
682          * most failure cases.
683          */
684         memblock_free_late(__pa(addr), KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool));
685         __kfence_pool = NULL;
686
687         memblock_free_late(__pa(kfence_metadata_init), KFENCE_METADATA_SIZE);
688         kfence_metadata_init = NULL;
689
690         return false;
691 }
692
693 /* === DebugFS Interface ==================================================== */
694
695 static int stats_show(struct seq_file *seq, void *v)
696 {
697         int i;
698
699         seq_printf(seq, "enabled: %i\n", READ_ONCE(kfence_enabled));
700         for (i = 0; i < KFENCE_COUNTER_COUNT; i++)
701                 seq_printf(seq, "%s: %ld\n", counter_names[i], atomic_long_read(&counters[i]));
702
703         return 0;
704 }
705 DEFINE_SHOW_ATTRIBUTE(stats);
706
707 /*
708  * debugfs seq_file operations for /sys/kernel/debug/kfence/objects.
709  * start_object() and next_object() return the object index + 1, because NULL is used
710  * to stop iteration.
711  */
712 static void *start_object(struct seq_file *seq, loff_t *pos)
713 {
714         if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
715                 return (void *)((long)*pos + 1);
716         return NULL;
717 }
718
719 static void stop_object(struct seq_file *seq, void *v)
720 {
721 }
722
723 static void *next_object(struct seq_file *seq, void *v, loff_t *pos)
724 {
725         ++*pos;
726         if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
727                 return (void *)((long)*pos + 1);
728         return NULL;
729 }
730
731 static int show_object(struct seq_file *seq, void *v)
732 {
733         struct kfence_metadata *meta = &kfence_metadata[(long)v - 1];
734         unsigned long flags;
735
736         raw_spin_lock_irqsave(&meta->lock, flags);
737         kfence_print_object(seq, meta);
738         raw_spin_unlock_irqrestore(&meta->lock, flags);
739         seq_puts(seq, "---------------------------------\n");
740
741         return 0;
742 }
743
744 static const struct seq_operations objects_sops = {
745         .start = start_object,
746         .next = next_object,
747         .stop = stop_object,
748         .show = show_object,
749 };
750 DEFINE_SEQ_ATTRIBUTE(objects);
751
752 static int kfence_debugfs_init(void)
753 {
754         struct dentry *kfence_dir;
755
756         if (!READ_ONCE(kfence_enabled))
757                 return 0;
758
759         kfence_dir = debugfs_create_dir("kfence", NULL);
760         debugfs_create_file("stats", 0444, kfence_dir, NULL, &stats_fops);
761         debugfs_create_file("objects", 0400, kfence_dir, NULL, &objects_fops);
762         return 0;
763 }
764
765 late_initcall(kfence_debugfs_init);
766
767 /* === Panic Notifier ====================================================== */
768
769 static void kfence_check_all_canary(void)
770 {
771         int i;
772
773         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
774                 struct kfence_metadata *meta = &kfence_metadata[i];
775
776                 if (meta->state == KFENCE_OBJECT_ALLOCATED)
777                         check_canary(meta);
778         }
779 }
780
781 static int kfence_check_canary_callback(struct notifier_block *nb,
782                                         unsigned long reason, void *arg)
783 {
784         kfence_check_all_canary();
785         return NOTIFY_OK;
786 }
787
788 static struct notifier_block kfence_check_canary_notifier = {
789         .notifier_call = kfence_check_canary_callback,
790 };
791
792 /* === Allocation Gate Timer ================================================ */
793
794 static struct delayed_work kfence_timer;
795
796 #ifdef CONFIG_KFENCE_STATIC_KEYS
797 /* Wait queue to wake up allocation-gate timer task. */
798 static DECLARE_WAIT_QUEUE_HEAD(allocation_wait);
799
800 static void wake_up_kfence_timer(struct irq_work *work)
801 {
802         wake_up(&allocation_wait);
803 }
804 static DEFINE_IRQ_WORK(wake_up_kfence_timer_work, wake_up_kfence_timer);
805 #endif
806
807 /*
808  * Set up delayed work, which will enable and disable the static key. We need to
809  * use a work queue (rather than a simple timer), since enabling and disabling a
810  * static key cannot be done from an interrupt.
811  *
812  * Note: Toggling a static branch currently causes IPIs, and here we'll end up
813  * with a total of 2 IPIs to all CPUs. If this ends up a problem in future (with
814  * more aggressive sampling intervals), we could get away with a variant that
815  * avoids IPIs, at the cost of not immediately capturing allocations if the
816  * instructions remain cached.
817  */
818 static void toggle_allocation_gate(struct work_struct *work)
819 {
820         if (!READ_ONCE(kfence_enabled))
821                 return;
822
823         atomic_set(&kfence_allocation_gate, 0);
824 #ifdef CONFIG_KFENCE_STATIC_KEYS
825         /* Enable static key, and await allocation to happen. */
826         static_branch_enable(&kfence_allocation_key);
827
828         wait_event_idle(allocation_wait, atomic_read(&kfence_allocation_gate));
829
830         /* Disable static key and reset timer. */
831         static_branch_disable(&kfence_allocation_key);
832 #endif
833         queue_delayed_work(system_unbound_wq, &kfence_timer,
834                            msecs_to_jiffies(kfence_sample_interval));
835 }
836
837 /* === Public interface ===================================================== */
838
839 void __init kfence_alloc_pool_and_metadata(void)
840 {
841         if (!kfence_sample_interval)
842                 return;
843
844         /*
845          * If the pool has already been initialized by arch, there is no need to
846          * re-allocate the memory pool.
847          */
848         if (!__kfence_pool)
849                 __kfence_pool = memblock_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
850
851         if (!__kfence_pool) {
852                 pr_err("failed to allocate pool\n");
853                 return;
854         }
855
856         /* The memory allocated by memblock has been zeroed out. */
857         kfence_metadata_init = memblock_alloc(KFENCE_METADATA_SIZE, PAGE_SIZE);
858         if (!kfence_metadata_init) {
859                 pr_err("failed to allocate metadata\n");
860                 memblock_free(__kfence_pool, KFENCE_POOL_SIZE);
861                 __kfence_pool = NULL;
862         }
863 }
864
865 static void kfence_init_enable(void)
866 {
867         if (!IS_ENABLED(CONFIG_KFENCE_STATIC_KEYS))
868                 static_branch_enable(&kfence_allocation_key);
869
870         if (kfence_deferrable)
871                 INIT_DEFERRABLE_WORK(&kfence_timer, toggle_allocation_gate);
872         else
873                 INIT_DELAYED_WORK(&kfence_timer, toggle_allocation_gate);
874
875         if (kfence_check_on_panic)
876                 atomic_notifier_chain_register(&panic_notifier_list, &kfence_check_canary_notifier);
877
878         WRITE_ONCE(kfence_enabled, true);
879         queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
880
881         pr_info("initialized - using %lu bytes for %d objects at 0x%p-0x%p\n", KFENCE_POOL_SIZE,
882                 CONFIG_KFENCE_NUM_OBJECTS, (void *)__kfence_pool,
883                 (void *)(__kfence_pool + KFENCE_POOL_SIZE));
884 }
885
886 void __init kfence_init(void)
887 {
888         stack_hash_seed = get_random_u32();
889
890         /* Setting kfence_sample_interval to 0 on boot disables KFENCE. */
891         if (!kfence_sample_interval)
892                 return;
893
894         if (!kfence_init_pool_early()) {
895                 pr_err("%s failed\n", __func__);
896                 return;
897         }
898
899         kfence_init_enable();
900 }
901
902 static int kfence_init_late(void)
903 {
904         const unsigned long nr_pages_pool = KFENCE_POOL_SIZE / PAGE_SIZE;
905         const unsigned long nr_pages_meta = KFENCE_METADATA_SIZE / PAGE_SIZE;
906         unsigned long addr = (unsigned long)__kfence_pool;
907         unsigned long free_size = KFENCE_POOL_SIZE;
908         int err = -ENOMEM;
909
910 #ifdef CONFIG_CONTIG_ALLOC
911         struct page *pages;
912
913         pages = alloc_contig_pages(nr_pages_pool, GFP_KERNEL, first_online_node,
914                                    NULL);
915         if (!pages)
916                 return -ENOMEM;
917
918         __kfence_pool = page_to_virt(pages);
919         pages = alloc_contig_pages(nr_pages_meta, GFP_KERNEL, first_online_node,
920                                    NULL);
921         if (pages)
922                 kfence_metadata_init = page_to_virt(pages);
923 #else
924         if (nr_pages_pool > MAX_ORDER_NR_PAGES ||
925             nr_pages_meta > MAX_ORDER_NR_PAGES) {
926                 pr_warn("KFENCE_NUM_OBJECTS too large for buddy allocator\n");
927                 return -EINVAL;
928         }
929
930         __kfence_pool = alloc_pages_exact(KFENCE_POOL_SIZE, GFP_KERNEL);
931         if (!__kfence_pool)
932                 return -ENOMEM;
933
934         kfence_metadata_init = alloc_pages_exact(KFENCE_METADATA_SIZE, GFP_KERNEL);
935 #endif
936
937         if (!kfence_metadata_init)
938                 goto free_pool;
939
940         memzero_explicit(kfence_metadata_init, KFENCE_METADATA_SIZE);
941         addr = kfence_init_pool();
942         if (!addr) {
943                 kfence_init_enable();
944                 kfence_debugfs_init();
945                 return 0;
946         }
947
948         pr_err("%s failed\n", __func__);
949         free_size = KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool);
950         err = -EBUSY;
951
952 #ifdef CONFIG_CONTIG_ALLOC
953         free_contig_range(page_to_pfn(virt_to_page((void *)kfence_metadata_init)),
954                           nr_pages_meta);
955 free_pool:
956         free_contig_range(page_to_pfn(virt_to_page((void *)addr)),
957                           free_size / PAGE_SIZE);
958 #else
959         free_pages_exact((void *)kfence_metadata_init, KFENCE_METADATA_SIZE);
960 free_pool:
961         free_pages_exact((void *)addr, free_size);
962 #endif
963
964         kfence_metadata_init = NULL;
965         __kfence_pool = NULL;
966         return err;
967 }
968
969 static int kfence_enable_late(void)
970 {
971         if (!__kfence_pool)
972                 return kfence_init_late();
973
974         WRITE_ONCE(kfence_enabled, true);
975         queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
976         pr_info("re-enabled\n");
977         return 0;
978 }
979
980 void kfence_shutdown_cache(struct kmem_cache *s)
981 {
982         unsigned long flags;
983         struct kfence_metadata *meta;
984         int i;
985
986         /* Pairs with release in kfence_init_pool(). */
987         if (!smp_load_acquire(&kfence_metadata))
988                 return;
989
990         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
991                 bool in_use;
992
993                 meta = &kfence_metadata[i];
994
995                 /*
996                  * If we observe some inconsistent cache and state pair where we
997                  * should have returned false here, cache destruction is racing
998                  * with either kmem_cache_alloc() or kmem_cache_free(). Taking
999                  * the lock will not help, as different critical section
1000                  * serialization will have the same outcome.
1001                  */
1002                 if (READ_ONCE(meta->cache) != s ||
1003                     READ_ONCE(meta->state) != KFENCE_OBJECT_ALLOCATED)
1004                         continue;
1005
1006                 raw_spin_lock_irqsave(&meta->lock, flags);
1007                 in_use = meta->cache == s && meta->state == KFENCE_OBJECT_ALLOCATED;
1008                 raw_spin_unlock_irqrestore(&meta->lock, flags);
1009
1010                 if (in_use) {
1011                         /*
1012                          * This cache still has allocations, and we should not
1013                          * release them back into the freelist so they can still
1014                          * safely be used and retain the kernel's default
1015                          * behaviour of keeping the allocations alive (leak the
1016                          * cache); however, they effectively become "zombie
1017                          * allocations" as the KFENCE objects are the only ones
1018                          * still in use and the owning cache is being destroyed.
1019                          *
1020                          * We mark them freed, so that any subsequent use shows
1021                          * more useful error messages that will include stack
1022                          * traces of the user of the object, the original
1023                          * allocation, and caller to shutdown_cache().
1024                          */
1025                         kfence_guarded_free((void *)meta->addr, meta, /*zombie=*/true);
1026                 }
1027         }
1028
1029         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
1030                 meta = &kfence_metadata[i];
1031
1032                 /* See above. */
1033                 if (READ_ONCE(meta->cache) != s || READ_ONCE(meta->state) != KFENCE_OBJECT_FREED)
1034                         continue;
1035
1036                 raw_spin_lock_irqsave(&meta->lock, flags);
1037                 if (meta->cache == s && meta->state == KFENCE_OBJECT_FREED)
1038                         meta->cache = NULL;
1039                 raw_spin_unlock_irqrestore(&meta->lock, flags);
1040         }
1041 }
1042
1043 void *__kfence_alloc(struct kmem_cache *s, size_t size, gfp_t flags)
1044 {
1045         unsigned long stack_entries[KFENCE_STACK_DEPTH];
1046         size_t num_stack_entries;
1047         u32 alloc_stack_hash;
1048
1049         /*
1050          * Perform size check before switching kfence_allocation_gate, so that
1051          * we don't disable KFENCE without making an allocation.
1052          */
1053         if (size > PAGE_SIZE) {
1054                 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
1055                 return NULL;
1056         }
1057
1058         /*
1059          * Skip allocations from non-default zones, including DMA. We cannot
1060          * guarantee that pages in the KFENCE pool will have the requested
1061          * properties (e.g. reside in DMAable memory).
1062          */
1063         if ((flags & GFP_ZONEMASK) ||
1064             (s->flags & (SLAB_CACHE_DMA | SLAB_CACHE_DMA32))) {
1065                 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]);
1066                 return NULL;
1067         }
1068
1069         /*
1070          * Skip allocations for this slab, if KFENCE has been disabled for
1071          * this slab.
1072          */
1073         if (s->flags & SLAB_SKIP_KFENCE)
1074                 return NULL;
1075
1076         if (atomic_inc_return(&kfence_allocation_gate) > 1)
1077                 return NULL;
1078 #ifdef CONFIG_KFENCE_STATIC_KEYS
1079         /*
1080          * waitqueue_active() is fully ordered after the update of
1081          * kfence_allocation_gate per atomic_inc_return().
1082          */
1083         if (waitqueue_active(&allocation_wait)) {
1084                 /*
1085                  * Calling wake_up() here may deadlock when allocations happen
1086                  * from within timer code. Use an irq_work to defer it.
1087                  */
1088                 irq_work_queue(&wake_up_kfence_timer_work);
1089         }
1090 #endif
1091
1092         if (!READ_ONCE(kfence_enabled))
1093                 return NULL;
1094
1095         num_stack_entries = stack_trace_save(stack_entries, KFENCE_STACK_DEPTH, 0);
1096
1097         /*
1098          * Do expensive check for coverage of allocation in slow-path after
1099          * allocation_gate has already become non-zero, even though it might
1100          * mean not making any allocation within a given sample interval.
1101          *
1102          * This ensures reasonable allocation coverage when the pool is almost
1103          * full, including avoiding long-lived allocations of the same source
1104          * filling up the pool (e.g. pagecache allocations).
1105          */
1106         alloc_stack_hash = get_alloc_stack_hash(stack_entries, num_stack_entries);
1107         if (should_skip_covered() && alloc_covered_contains(alloc_stack_hash)) {
1108                 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_COVERED]);
1109                 return NULL;
1110         }
1111
1112         return kfence_guarded_alloc(s, size, flags, stack_entries, num_stack_entries,
1113                                     alloc_stack_hash);
1114 }
1115
1116 size_t kfence_ksize(const void *addr)
1117 {
1118         const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1119
1120         /*
1121          * Read locklessly -- if there is a race with __kfence_alloc(), this is
1122          * either a use-after-free or invalid access.
1123          */
1124         return meta ? meta->size : 0;
1125 }
1126
1127 void *kfence_object_start(const void *addr)
1128 {
1129         const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1130
1131         /*
1132          * Read locklessly -- if there is a race with __kfence_alloc(), this is
1133          * either a use-after-free or invalid access.
1134          */
1135         return meta ? (void *)meta->addr : NULL;
1136 }
1137
1138 void __kfence_free(void *addr)
1139 {
1140         struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
1141
1142 #ifdef CONFIG_MEMCG
1143         KFENCE_WARN_ON(meta->objcg);
1144 #endif
1145         /*
1146          * If the objects of the cache are SLAB_TYPESAFE_BY_RCU, defer freeing
1147          * the object, as the object page may be recycled for other-typed
1148          * objects once it has been freed. meta->cache may be NULL if the cache
1149          * was destroyed.
1150          */
1151         if (unlikely(meta->cache && (meta->cache->flags & SLAB_TYPESAFE_BY_RCU)))
1152                 call_rcu(&meta->rcu_head, rcu_guarded_free);
1153         else
1154                 kfence_guarded_free(addr, meta, false);
1155 }
1156
1157 bool kfence_handle_page_fault(unsigned long addr, bool is_write, struct pt_regs *regs)
1158 {
1159         const int page_index = (addr - (unsigned long)__kfence_pool) / PAGE_SIZE;
1160         struct kfence_metadata *to_report = NULL;
1161         enum kfence_error_type error_type;
1162         unsigned long flags;
1163
1164         if (!is_kfence_address((void *)addr))
1165                 return false;
1166
1167         if (!READ_ONCE(kfence_enabled)) /* If disabled at runtime ... */
1168                 return kfence_unprotect(addr); /* ... unprotect and proceed. */
1169
1170         atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
1171
1172         if (page_index % 2) {
1173                 /* This is a redzone, report a buffer overflow. */
1174                 struct kfence_metadata *meta;
1175                 int distance = 0;
1176
1177                 meta = addr_to_metadata(addr - PAGE_SIZE);
1178                 if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
1179                         to_report = meta;
1180                         /* Data race ok; distance calculation approximate. */
1181                         distance = addr - data_race(meta->addr + meta->size);
1182                 }
1183
1184                 meta = addr_to_metadata(addr + PAGE_SIZE);
1185                 if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
1186                         /* Data race ok; distance calculation approximate. */
1187                         if (!to_report || distance > data_race(meta->addr) - addr)
1188                                 to_report = meta;
1189                 }
1190
1191                 if (!to_report)
1192                         goto out;
1193
1194                 raw_spin_lock_irqsave(&to_report->lock, flags);
1195                 to_report->unprotected_page = addr;
1196                 error_type = KFENCE_ERROR_OOB;
1197
1198                 /*
1199                  * If the object was freed before we took the look we can still
1200                  * report this as an OOB -- the report will simply show the
1201                  * stacktrace of the free as well.
1202                  */
1203         } else {
1204                 to_report = addr_to_metadata(addr);
1205                 if (!to_report)
1206                         goto out;
1207
1208                 raw_spin_lock_irqsave(&to_report->lock, flags);
1209                 error_type = KFENCE_ERROR_UAF;
1210                 /*
1211                  * We may race with __kfence_alloc(), and it is possible that a
1212                  * freed object may be reallocated. We simply report this as a
1213                  * use-after-free, with the stack trace showing the place where
1214                  * the object was re-allocated.
1215                  */
1216         }
1217
1218 out:
1219         if (to_report) {
1220                 kfence_report_error(addr, is_write, regs, to_report, error_type);
1221                 raw_spin_unlock_irqrestore(&to_report->lock, flags);
1222         } else {
1223                 /* This may be a UAF or OOB access, but we can't be sure. */
1224                 kfence_report_error(addr, is_write, regs, NULL, KFENCE_ERROR_INVALID);
1225         }
1226
1227         return kfence_unprotect(addr); /* Unprotect and let access proceed. */
1228 }