xref: /linux/lib/stackdepot.c (revision e724e7aaf9ca794670a4d4931af7a7e24e37fec3)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Stack depot - a stack trace storage that avoids duplication.
4  *
5  * Internally, stack depot maintains a hash table of unique stacktraces. The
6  * stack traces themselves are stored contiguously one after another in a set
7  * of separate page allocations.
8  *
9  * Author: Alexander Potapenko <glider@google.com>
10  * Copyright (C) 2016 Google, Inc.
11  *
12  * Based on the code by Dmitry Chernenkov.
13  */
14 
15 #define pr_fmt(fmt) "stackdepot: " fmt
16 
17 #include <linux/gfp.h>
18 #include <linux/jhash.h>
19 #include <linux/kernel.h>
20 #include <linux/kmsan.h>
21 #include <linux/mm.h>
22 #include <linux/mutex.h>
23 #include <linux/percpu.h>
24 #include <linux/printk.h>
25 #include <linux/slab.h>
26 #include <linux/stacktrace.h>
27 #include <linux/stackdepot.h>
28 #include <linux/string.h>
29 #include <linux/types.h>
30 #include <linux/memblock.h>
31 #include <linux/kasan-enabled.h>
32 
33 #define DEPOT_HANDLE_BITS (sizeof(depot_stack_handle_t) * 8)
34 
35 #define DEPOT_VALID_BITS 1
36 #define DEPOT_POOL_ORDER 2 /* Pool size order, 4 pages */
37 #define DEPOT_POOL_SIZE (1LL << (PAGE_SHIFT + DEPOT_POOL_ORDER))
38 #define DEPOT_STACK_ALIGN 4
39 #define DEPOT_OFFSET_BITS (DEPOT_POOL_ORDER + PAGE_SHIFT - DEPOT_STACK_ALIGN)
40 #define DEPOT_POOL_INDEX_BITS (DEPOT_HANDLE_BITS - DEPOT_VALID_BITS - \
41 			       DEPOT_OFFSET_BITS - STACK_DEPOT_EXTRA_BITS)
42 #define DEPOT_POOLS_CAP 8192
43 #define DEPOT_MAX_POOLS \
44 	(((1LL << (DEPOT_POOL_INDEX_BITS)) < DEPOT_POOLS_CAP) ? \
45 	 (1LL << (DEPOT_POOL_INDEX_BITS)) : DEPOT_POOLS_CAP)
46 
47 /* Compact structure that stores a reference to a stack. */
48 union handle_parts {
49 	depot_stack_handle_t handle;
50 	struct {
51 		u32 pool_index	: DEPOT_POOL_INDEX_BITS;
52 		u32 offset	: DEPOT_OFFSET_BITS;
53 		u32 valid	: DEPOT_VALID_BITS;
54 		u32 extra	: STACK_DEPOT_EXTRA_BITS;
55 	};
56 };
57 
58 struct stack_record {
59 	struct stack_record *next;	/* Link in the hash table */
60 	u32 hash;			/* Hash in the hash table */
61 	u32 size;			/* Number of stored frames */
62 	union handle_parts handle;
63 	unsigned long entries[];	/* Variable-sized array of frames */
64 };
65 
66 static bool stack_depot_disabled;
67 static bool __stack_depot_early_init_requested __initdata = IS_ENABLED(CONFIG_STACKDEPOT_ALWAYS_INIT);
68 static bool __stack_depot_early_init_passed __initdata;
69 
70 /* Use one hash table bucket per 16 KB of memory. */
71 #define STACK_HASH_TABLE_SCALE 14
72 /* Limit the number of buckets between 4K and 1M. */
73 #define STACK_BUCKET_NUMBER_ORDER_MIN 12
74 #define STACK_BUCKET_NUMBER_ORDER_MAX 20
75 /* Initial seed for jhash2. */
76 #define STACK_HASH_SEED 0x9747b28c
77 
78 /* Hash table of pointers to stored stack traces. */
79 static struct stack_record **stack_table;
80 /* Fixed order of the number of table buckets. Used when KASAN is enabled. */
81 static unsigned int stack_bucket_number_order;
82 /* Hash mask for indexing the table. */
83 static unsigned int stack_hash_mask;
84 
85 /* Array of memory regions that store stack traces. */
86 static void *stack_pools[DEPOT_MAX_POOLS];
87 /* Currently used pool in stack_pools. */
88 static int pool_index;
89 /* Offset to the unused space in the currently used pool. */
90 static size_t pool_offset;
91 /* Lock that protects the variables above. */
92 static DEFINE_RAW_SPINLOCK(pool_lock);
93 /*
94  * Stack depot tries to keep an extra pool allocated even before it runs out
95  * of space in the currently used pool.
96  * This flag marks that this next extra pool needs to be allocated and
97  * initialized. It has the value 0 when either the next pool is not yet
98  * initialized or the limit on the number of pools is reached.
99  */
100 static int next_pool_required = 1;
101 
102 static int __init disable_stack_depot(char *str)
103 {
104 	int ret;
105 
106 	ret = kstrtobool(str, &stack_depot_disabled);
107 	if (!ret && stack_depot_disabled) {
108 		pr_info("disabled\n");
109 		stack_table = NULL;
110 	}
111 	return 0;
112 }
113 early_param("stack_depot_disable", disable_stack_depot);
114 
115 void __init stack_depot_request_early_init(void)
116 {
117 	/* Too late to request early init now. */
118 	WARN_ON(__stack_depot_early_init_passed);
119 
120 	__stack_depot_early_init_requested = true;
121 }
122 
123 /* Allocates a hash table via memblock. Can only be used during early boot. */
124 int __init stack_depot_early_init(void)
125 {
126 	unsigned long entries = 0;
127 
128 	/* This function must be called only once, from mm_init(). */
129 	if (WARN_ON(__stack_depot_early_init_passed))
130 		return 0;
131 	__stack_depot_early_init_passed = true;
132 
133 	/*
134 	 * If KASAN is enabled, use the maximum order: KASAN is frequently used
135 	 * in fuzzing scenarios, which leads to a large number of different
136 	 * stack traces being stored in stack depot.
137 	 */
138 	if (kasan_enabled() && !stack_bucket_number_order)
139 		stack_bucket_number_order = STACK_BUCKET_NUMBER_ORDER_MAX;
140 
141 	if (!__stack_depot_early_init_requested || stack_depot_disabled)
142 		return 0;
143 
144 	/*
145 	 * If stack_bucket_number_order is not set, leave entries as 0 to rely
146 	 * on the automatic calculations performed by alloc_large_system_hash.
147 	 */
148 	if (stack_bucket_number_order)
149 		entries = 1UL << stack_bucket_number_order;
150 	pr_info("allocating hash table via alloc_large_system_hash\n");
151 	stack_table = alloc_large_system_hash("stackdepot",
152 						sizeof(struct stack_record *),
153 						entries,
154 						STACK_HASH_TABLE_SCALE,
155 						HASH_EARLY | HASH_ZERO,
156 						NULL,
157 						&stack_hash_mask,
158 						1UL << STACK_BUCKET_NUMBER_ORDER_MIN,
159 						1UL << STACK_BUCKET_NUMBER_ORDER_MAX);
160 	if (!stack_table) {
161 		pr_err("hash table allocation failed, disabling\n");
162 		stack_depot_disabled = true;
163 		return -ENOMEM;
164 	}
165 
166 	return 0;
167 }
168 
169 /* Allocates a hash table via kvcalloc. Can be used after boot. */
170 int stack_depot_init(void)
171 {
172 	static DEFINE_MUTEX(stack_depot_init_mutex);
173 	unsigned long entries;
174 	int ret = 0;
175 
176 	mutex_lock(&stack_depot_init_mutex);
177 
178 	if (stack_depot_disabled || stack_table)
179 		goto out_unlock;
180 
181 	/*
182 	 * Similarly to stack_depot_early_init, use stack_bucket_number_order
183 	 * if assigned, and rely on automatic scaling otherwise.
184 	 */
185 	if (stack_bucket_number_order) {
186 		entries = 1UL << stack_bucket_number_order;
187 	} else {
188 		int scale = STACK_HASH_TABLE_SCALE;
189 
190 		entries = nr_free_buffer_pages();
191 		entries = roundup_pow_of_two(entries);
192 
193 		if (scale > PAGE_SHIFT)
194 			entries >>= (scale - PAGE_SHIFT);
195 		else
196 			entries <<= (PAGE_SHIFT - scale);
197 	}
198 
199 	if (entries < 1UL << STACK_BUCKET_NUMBER_ORDER_MIN)
200 		entries = 1UL << STACK_BUCKET_NUMBER_ORDER_MIN;
201 	if (entries > 1UL << STACK_BUCKET_NUMBER_ORDER_MAX)
202 		entries = 1UL << STACK_BUCKET_NUMBER_ORDER_MAX;
203 
204 	pr_info("allocating hash table of %lu entries via kvcalloc\n", entries);
205 	stack_table = kvcalloc(entries, sizeof(struct stack_record *), GFP_KERNEL);
206 	if (!stack_table) {
207 		pr_err("hash table allocation failed, disabling\n");
208 		stack_depot_disabled = true;
209 		ret = -ENOMEM;
210 		goto out_unlock;
211 	}
212 	stack_hash_mask = entries - 1;
213 
214 out_unlock:
215 	mutex_unlock(&stack_depot_init_mutex);
216 
217 	return ret;
218 }
219 EXPORT_SYMBOL_GPL(stack_depot_init);
220 
221 /* Uses preallocated memory to initialize a new stack depot pool. */
222 static void depot_init_pool(void **prealloc)
223 {
224 	/*
225 	 * If the next pool is already initialized or the maximum number of
226 	 * pools is reached, do not use the preallocated memory.
227 	 * smp_load_acquire() here pairs with smp_store_release() below and
228 	 * in depot_alloc_stack().
229 	 */
230 	if (!smp_load_acquire(&next_pool_required))
231 		return;
232 
233 	/* Check if the current pool is not yet allocated. */
234 	if (stack_pools[pool_index] == NULL) {
235 		/* Use the preallocated memory for the current pool. */
236 		stack_pools[pool_index] = *prealloc;
237 		*prealloc = NULL;
238 	} else {
239 		/*
240 		 * Otherwise, use the preallocated memory for the next pool
241 		 * as long as we do not exceed the maximum number of pools.
242 		 */
243 		if (pool_index + 1 < DEPOT_MAX_POOLS) {
244 			stack_pools[pool_index + 1] = *prealloc;
245 			*prealloc = NULL;
246 		}
247 		/*
248 		 * At this point, either the next pool is initialized or the
249 		 * maximum number of pools is reached. In either case, take
250 		 * note that initializing another pool is not required.
251 		 * This smp_store_release pairs with smp_load_acquire() above
252 		 * and in stack_depot_save().
253 		 */
254 		smp_store_release(&next_pool_required, 0);
255 	}
256 }
257 
258 /* Allocates a new stack in a stack depot pool. */
259 static struct stack_record *
260 depot_alloc_stack(unsigned long *entries, int size, u32 hash, void **prealloc)
261 {
262 	struct stack_record *stack;
263 	size_t required_size = struct_size(stack, entries, size);
264 
265 	required_size = ALIGN(required_size, 1 << DEPOT_STACK_ALIGN);
266 
267 	/* Check if there is not enough space in the current pool. */
268 	if (unlikely(pool_offset + required_size > DEPOT_POOL_SIZE)) {
269 		/* Bail out if we reached the pool limit. */
270 		if (unlikely(pool_index + 1 >= DEPOT_MAX_POOLS)) {
271 			WARN_ONCE(1, "Stack depot reached limit capacity");
272 			return NULL;
273 		}
274 
275 		/*
276 		 * Move on to the next pool.
277 		 * WRITE_ONCE pairs with potential concurrent read in
278 		 * stack_depot_fetch().
279 		 */
280 		WRITE_ONCE(pool_index, pool_index + 1);
281 		pool_offset = 0;
282 		/*
283 		 * If the maximum number of pools is not reached, take note
284 		 * that the next pool needs to initialized.
285 		 * smp_store_release() here pairs with smp_load_acquire() in
286 		 * stack_depot_save() and depot_init_pool().
287 		 */
288 		if (pool_index + 1 < DEPOT_MAX_POOLS)
289 			smp_store_release(&next_pool_required, 1);
290 	}
291 
292 	/* Assign the preallocated memory to a pool if required. */
293 	if (*prealloc)
294 		depot_init_pool(prealloc);
295 
296 	/* Check if we have a pool to save the stack trace. */
297 	if (stack_pools[pool_index] == NULL)
298 		return NULL;
299 
300 	/* Save the stack trace. */
301 	stack = stack_pools[pool_index] + pool_offset;
302 	stack->hash = hash;
303 	stack->size = size;
304 	stack->handle.pool_index = pool_index;
305 	stack->handle.offset = pool_offset >> DEPOT_STACK_ALIGN;
306 	stack->handle.valid = 1;
307 	stack->handle.extra = 0;
308 	memcpy(stack->entries, entries, flex_array_size(stack, entries, size));
309 	pool_offset += required_size;
310 	/*
311 	 * Let KMSAN know the stored stack record is initialized. This shall
312 	 * prevent false positive reports if instrumented code accesses it.
313 	 */
314 	kmsan_unpoison_memory(stack, required_size);
315 
316 	return stack;
317 }
318 
319 /* Calculates the hash for a stack. */
320 static inline u32 hash_stack(unsigned long *entries, unsigned int size)
321 {
322 	return jhash2((u32 *)entries,
323 		      array_size(size,  sizeof(*entries)) / sizeof(u32),
324 		      STACK_HASH_SEED);
325 }
326 
327 /*
328  * Non-instrumented version of memcmp().
329  * Does not check the lexicographical order, only the equality.
330  */
331 static inline
332 int stackdepot_memcmp(const unsigned long *u1, const unsigned long *u2,
333 			unsigned int n)
334 {
335 	for ( ; n-- ; u1++, u2++) {
336 		if (*u1 != *u2)
337 			return 1;
338 	}
339 	return 0;
340 }
341 
342 /* Finds a stack in a bucket of the hash table. */
343 static inline struct stack_record *find_stack(struct stack_record *bucket,
344 					     unsigned long *entries, int size,
345 					     u32 hash)
346 {
347 	struct stack_record *found;
348 
349 	for (found = bucket; found; found = found->next) {
350 		if (found->hash == hash &&
351 		    found->size == size &&
352 		    !stackdepot_memcmp(entries, found->entries, size))
353 			return found;
354 	}
355 	return NULL;
356 }
357 
358 depot_stack_handle_t __stack_depot_save(unsigned long *entries,
359 					unsigned int nr_entries,
360 					gfp_t alloc_flags, bool can_alloc)
361 {
362 	struct stack_record *found = NULL, **bucket;
363 	union handle_parts retval = { .handle = 0 };
364 	struct page *page = NULL;
365 	void *prealloc = NULL;
366 	unsigned long flags;
367 	u32 hash;
368 
369 	/*
370 	 * If this stack trace is from an interrupt, including anything before
371 	 * interrupt entry usually leads to unbounded stack depot growth.
372 	 *
373 	 * Since use of filter_irq_stacks() is a requirement to ensure stack
374 	 * depot can efficiently deduplicate interrupt stacks, always
375 	 * filter_irq_stacks() to simplify all callers' use of stack depot.
376 	 */
377 	nr_entries = filter_irq_stacks(entries, nr_entries);
378 
379 	if (unlikely(nr_entries == 0) || stack_depot_disabled)
380 		goto fast_exit;
381 
382 	hash = hash_stack(entries, nr_entries);
383 	bucket = &stack_table[hash & stack_hash_mask];
384 
385 	/*
386 	 * Fast path: look the stack trace up without locking.
387 	 * The smp_load_acquire() here pairs with smp_store_release() to
388 	 * |bucket| below.
389 	 */
390 	found = find_stack(smp_load_acquire(bucket), entries, nr_entries, hash);
391 	if (found)
392 		goto exit;
393 
394 	/*
395 	 * Check if another stack pool needs to be initialized. If so, allocate
396 	 * the memory now - we won't be able to do that under the lock.
397 	 *
398 	 * The smp_load_acquire() here pairs with smp_store_release() to
399 	 * |next_pool_inited| in depot_alloc_stack() and depot_init_pool().
400 	 */
401 	if (unlikely(can_alloc && smp_load_acquire(&next_pool_required))) {
402 		/*
403 		 * Zero out zone modifiers, as we don't have specific zone
404 		 * requirements. Keep the flags related to allocation in atomic
405 		 * contexts and I/O.
406 		 */
407 		alloc_flags &= ~GFP_ZONEMASK;
408 		alloc_flags &= (GFP_ATOMIC | GFP_KERNEL);
409 		alloc_flags |= __GFP_NOWARN;
410 		page = alloc_pages(alloc_flags, DEPOT_POOL_ORDER);
411 		if (page)
412 			prealloc = page_address(page);
413 	}
414 
415 	raw_spin_lock_irqsave(&pool_lock, flags);
416 
417 	found = find_stack(*bucket, entries, nr_entries, hash);
418 	if (!found) {
419 		struct stack_record *new =
420 			depot_alloc_stack(entries, nr_entries, hash, &prealloc);
421 
422 		if (new) {
423 			new->next = *bucket;
424 			/*
425 			 * This smp_store_release() pairs with
426 			 * smp_load_acquire() from |bucket| above.
427 			 */
428 			smp_store_release(bucket, new);
429 			found = new;
430 		}
431 	} else if (prealloc) {
432 		/*
433 		 * Stack depot already contains this stack trace, but let's
434 		 * keep the preallocated memory for the future.
435 		 */
436 		depot_init_pool(&prealloc);
437 	}
438 
439 	raw_spin_unlock_irqrestore(&pool_lock, flags);
440 exit:
441 	if (prealloc) {
442 		/* Stack depot didn't use this memory, free it. */
443 		free_pages((unsigned long)prealloc, DEPOT_POOL_ORDER);
444 	}
445 	if (found)
446 		retval.handle = found->handle.handle;
447 fast_exit:
448 	return retval.handle;
449 }
450 EXPORT_SYMBOL_GPL(__stack_depot_save);
451 
452 depot_stack_handle_t stack_depot_save(unsigned long *entries,
453 				      unsigned int nr_entries,
454 				      gfp_t alloc_flags)
455 {
456 	return __stack_depot_save(entries, nr_entries, alloc_flags, true);
457 }
458 EXPORT_SYMBOL_GPL(stack_depot_save);
459 
460 unsigned int stack_depot_fetch(depot_stack_handle_t handle,
461 			       unsigned long **entries)
462 {
463 	union handle_parts parts = { .handle = handle };
464 	/*
465 	 * READ_ONCE pairs with potential concurrent write in
466 	 * depot_alloc_stack.
467 	 */
468 	int pool_index_cached = READ_ONCE(pool_index);
469 	void *pool;
470 	size_t offset = parts.offset << DEPOT_STACK_ALIGN;
471 	struct stack_record *stack;
472 
473 	*entries = NULL;
474 	/*
475 	 * Let KMSAN know *entries is initialized. This shall prevent false
476 	 * positive reports if instrumented code accesses it.
477 	 */
478 	kmsan_unpoison_memory(entries, sizeof(*entries));
479 
480 	if (!handle)
481 		return 0;
482 
483 	if (parts.pool_index > pool_index_cached) {
484 		WARN(1, "pool index %d out of bounds (%d) for stack id %08x\n",
485 			parts.pool_index, pool_index_cached, handle);
486 		return 0;
487 	}
488 	pool = stack_pools[parts.pool_index];
489 	if (!pool)
490 		return 0;
491 	stack = pool + offset;
492 
493 	*entries = stack->entries;
494 	return stack->size;
495 }
496 EXPORT_SYMBOL_GPL(stack_depot_fetch);
497 
498 void stack_depot_print(depot_stack_handle_t stack)
499 {
500 	unsigned long *entries;
501 	unsigned int nr_entries;
502 
503 	nr_entries = stack_depot_fetch(stack, &entries);
504 	if (nr_entries > 0)
505 		stack_trace_print(entries, nr_entries, 0);
506 }
507 EXPORT_SYMBOL_GPL(stack_depot_print);
508 
509 int stack_depot_snprint(depot_stack_handle_t handle, char *buf, size_t size,
510 		       int spaces)
511 {
512 	unsigned long *entries;
513 	unsigned int nr_entries;
514 
515 	nr_entries = stack_depot_fetch(handle, &entries);
516 	return nr_entries ? stack_trace_snprint(buf, size, entries, nr_entries,
517 						spaces) : 0;
518 }
519 EXPORT_SYMBOL_GPL(stack_depot_snprint);
520 
521 depot_stack_handle_t __must_check stack_depot_set_extra_bits(
522 			depot_stack_handle_t handle, unsigned int extra_bits)
523 {
524 	union handle_parts parts = { .handle = handle };
525 
526 	/* Don't set extra bits on empty handles. */
527 	if (!handle)
528 		return 0;
529 
530 	parts.extra = extra_bits;
531 	return parts.handle;
532 }
533 EXPORT_SYMBOL(stack_depot_set_extra_bits);
534 
535 unsigned int stack_depot_get_extra_bits(depot_stack_handle_t handle)
536 {
537 	union handle_parts parts = { .handle = handle };
538 
539 	return parts.extra;
540 }
541 EXPORT_SYMBOL(stack_depot_get_extra_bits);
542