xref: /linux/lib/stackdepot.c (revision 0e9b70c1e3623fa110fb6be553e644524228ef60)
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/mm.h>
21 #include <linux/mutex.h>
22 #include <linux/percpu.h>
23 #include <linux/printk.h>
24 #include <linux/slab.h>
25 #include <linux/stacktrace.h>
26 #include <linux/stackdepot.h>
27 #include <linux/string.h>
28 #include <linux/types.h>
29 #include <linux/memblock.h>
30 #include <linux/kasan-enabled.h>
31 
32 #define DEPOT_HANDLE_BITS (sizeof(depot_stack_handle_t) * 8)
33 
34 #define DEPOT_VALID_BITS 1
35 #define DEPOT_POOL_ORDER 2 /* Pool size order, 4 pages */
36 #define DEPOT_POOL_SIZE (1LL << (PAGE_SHIFT + DEPOT_POOL_ORDER))
37 #define DEPOT_STACK_ALIGN 4
38 #define DEPOT_OFFSET_BITS (DEPOT_POOL_ORDER + PAGE_SHIFT - DEPOT_STACK_ALIGN)
39 #define DEPOT_POOL_INDEX_BITS (DEPOT_HANDLE_BITS - DEPOT_VALID_BITS - \
40 			       DEPOT_OFFSET_BITS - STACK_DEPOT_EXTRA_BITS)
41 #define DEPOT_POOLS_CAP 8192
42 #define DEPOT_MAX_POOLS \
43 	(((1LL << (DEPOT_POOL_INDEX_BITS)) < DEPOT_POOLS_CAP) ? \
44 	 (1LL << (DEPOT_POOL_INDEX_BITS)) : DEPOT_POOLS_CAP)
45 
46 /* Compact structure that stores a reference to a stack. */
47 union handle_parts {
48 	depot_stack_handle_t handle;
49 	struct {
50 		u32 pool_index	: DEPOT_POOL_INDEX_BITS;
51 		u32 offset	: DEPOT_OFFSET_BITS;
52 		u32 valid	: DEPOT_VALID_BITS;
53 		u32 extra	: STACK_DEPOT_EXTRA_BITS;
54 	};
55 };
56 
57 struct stack_record {
58 	struct stack_record *next;	/* Link in the hash table */
59 	u32 hash;			/* Hash in the hash table */
60 	u32 size;			/* Number of stored frames */
61 	union handle_parts handle;
62 	unsigned long entries[];	/* Variable-sized array of frames */
63 };
64 
65 static bool stack_depot_disabled;
66 static bool __stack_depot_early_init_requested __initdata = IS_ENABLED(CONFIG_STACKDEPOT_ALWAYS_INIT);
67 static bool __stack_depot_early_init_passed __initdata;
68 
69 /* Use one hash table bucket per 16 KB of memory. */
70 #define STACK_HASH_TABLE_SCALE 14
71 /* Limit the number of buckets between 4K and 1M. */
72 #define STACK_BUCKET_NUMBER_ORDER_MIN 12
73 #define STACK_BUCKET_NUMBER_ORDER_MAX 20
74 /* Initial seed for jhash2. */
75 #define STACK_HASH_SEED 0x9747b28c
76 
77 /* Hash table of pointers to stored stack traces. */
78 static struct stack_record **stack_table;
79 /* Fixed order of the number of table buckets. Used when KASAN is enabled. */
80 static unsigned int stack_bucket_number_order;
81 /* Hash mask for indexing the table. */
82 static unsigned int stack_hash_mask;
83 
84 /* Array of memory regions that store stack traces. */
85 static void *stack_pools[DEPOT_MAX_POOLS];
86 /* Currently used pool in stack_pools. */
87 static int pool_index;
88 /* Offset to the unused space in the currently used pool. */
89 static size_t pool_offset;
90 /* Lock that protects the variables above. */
91 static DEFINE_RAW_SPINLOCK(pool_lock);
92 /*
93  * Stack depot tries to keep an extra pool allocated even before it runs out
94  * of space in the currently used pool.
95  * This flag marks that this next extra pool needs to be allocated and
96  * initialized. It has the value 0 when either the next pool is not yet
97  * initialized or the limit on the number of pools is reached.
98  */
99 static int next_pool_required = 1;
100 
101 static int __init disable_stack_depot(char *str)
102 {
103 	int ret;
104 
105 	ret = kstrtobool(str, &stack_depot_disabled);
106 	if (!ret && stack_depot_disabled) {
107 		pr_info("disabled\n");
108 		stack_table = NULL;
109 	}
110 	return 0;
111 }
112 early_param("stack_depot_disable", disable_stack_depot);
113 
114 void __init stack_depot_request_early_init(void)
115 {
116 	/* Too late to request early init now. */
117 	WARN_ON(__stack_depot_early_init_passed);
118 
119 	__stack_depot_early_init_requested = true;
120 }
121 
122 /* Allocates a hash table via memblock. Can only be used during early boot. */
123 int __init stack_depot_early_init(void)
124 {
125 	unsigned long entries = 0;
126 
127 	/* This function must be called only once, from mm_init(). */
128 	if (WARN_ON(__stack_depot_early_init_passed))
129 		return 0;
130 	__stack_depot_early_init_passed = true;
131 
132 	/*
133 	 * If KASAN is enabled, use the maximum order: KASAN is frequently used
134 	 * in fuzzing scenarios, which leads to a large number of different
135 	 * stack traces being stored in stack depot.
136 	 */
137 	if (kasan_enabled() && !stack_bucket_number_order)
138 		stack_bucket_number_order = STACK_BUCKET_NUMBER_ORDER_MAX;
139 
140 	if (!__stack_depot_early_init_requested || stack_depot_disabled)
141 		return 0;
142 
143 	/*
144 	 * If stack_bucket_number_order is not set, leave entries as 0 to rely
145 	 * on the automatic calculations performed by alloc_large_system_hash.
146 	 */
147 	if (stack_bucket_number_order)
148 		entries = 1UL << stack_bucket_number_order;
149 	pr_info("allocating hash table via alloc_large_system_hash\n");
150 	stack_table = alloc_large_system_hash("stackdepot",
151 						sizeof(struct stack_record *),
152 						entries,
153 						STACK_HASH_TABLE_SCALE,
154 						HASH_EARLY | HASH_ZERO,
155 						NULL,
156 						&stack_hash_mask,
157 						1UL << STACK_BUCKET_NUMBER_ORDER_MIN,
158 						1UL << STACK_BUCKET_NUMBER_ORDER_MAX);
159 	if (!stack_table) {
160 		pr_err("hash table allocation failed, disabling\n");
161 		stack_depot_disabled = true;
162 		return -ENOMEM;
163 	}
164 
165 	return 0;
166 }
167 
168 /* Allocates a hash table via kvcalloc. Can be used after boot. */
169 int stack_depot_init(void)
170 {
171 	static DEFINE_MUTEX(stack_depot_init_mutex);
172 	unsigned long entries;
173 	int ret = 0;
174 
175 	mutex_lock(&stack_depot_init_mutex);
176 
177 	if (stack_depot_disabled || stack_table)
178 		goto out_unlock;
179 
180 	/*
181 	 * Similarly to stack_depot_early_init, use stack_bucket_number_order
182 	 * if assigned, and rely on automatic scaling otherwise.
183 	 */
184 	if (stack_bucket_number_order) {
185 		entries = 1UL << stack_bucket_number_order;
186 	} else {
187 		int scale = STACK_HASH_TABLE_SCALE;
188 
189 		entries = nr_free_buffer_pages();
190 		entries = roundup_pow_of_two(entries);
191 
192 		if (scale > PAGE_SHIFT)
193 			entries >>= (scale - PAGE_SHIFT);
194 		else
195 			entries <<= (PAGE_SHIFT - scale);
196 	}
197 
198 	if (entries < 1UL << STACK_BUCKET_NUMBER_ORDER_MIN)
199 		entries = 1UL << STACK_BUCKET_NUMBER_ORDER_MIN;
200 	if (entries > 1UL << STACK_BUCKET_NUMBER_ORDER_MAX)
201 		entries = 1UL << STACK_BUCKET_NUMBER_ORDER_MAX;
202 
203 	pr_info("allocating hash table of %lu entries via kvcalloc\n", entries);
204 	stack_table = kvcalloc(entries, sizeof(struct stack_record *), GFP_KERNEL);
205 	if (!stack_table) {
206 		pr_err("hash table allocation failed, disabling\n");
207 		stack_depot_disabled = true;
208 		ret = -ENOMEM;
209 		goto out_unlock;
210 	}
211 	stack_hash_mask = entries - 1;
212 
213 out_unlock:
214 	mutex_unlock(&stack_depot_init_mutex);
215 
216 	return ret;
217 }
218 EXPORT_SYMBOL_GPL(stack_depot_init);
219 
220 /* Uses preallocated memory to initialize a new stack depot pool. */
221 static void depot_init_pool(void **prealloc)
222 {
223 	/*
224 	 * If the next pool is already initialized or the maximum number of
225 	 * pools is reached, do not use the preallocated memory.
226 	 * smp_load_acquire() here pairs with smp_store_release() below and
227 	 * in depot_alloc_stack().
228 	 */
229 	if (!smp_load_acquire(&next_pool_required))
230 		return;
231 
232 	/* Check if the current pool is not yet allocated. */
233 	if (stack_pools[pool_index] == NULL) {
234 		/* Use the preallocated memory for the current pool. */
235 		stack_pools[pool_index] = *prealloc;
236 		*prealloc = NULL;
237 	} else {
238 		/*
239 		 * Otherwise, use the preallocated memory for the next pool
240 		 * as long as we do not exceed the maximum number of pools.
241 		 */
242 		if (pool_index + 1 < DEPOT_MAX_POOLS) {
243 			stack_pools[pool_index + 1] = *prealloc;
244 			*prealloc = NULL;
245 		}
246 		/*
247 		 * At this point, either the next pool is initialized or the
248 		 * maximum number of pools is reached. In either case, take
249 		 * note that initializing another pool is not required.
250 		 * This smp_store_release pairs with smp_load_acquire() above
251 		 * and in stack_depot_save().
252 		 */
253 		smp_store_release(&next_pool_required, 0);
254 	}
255 }
256 
257 /* Allocates a new stack in a stack depot pool. */
258 static struct stack_record *
259 depot_alloc_stack(unsigned long *entries, int size, u32 hash, void **prealloc)
260 {
261 	struct stack_record *stack;
262 	size_t required_size = struct_size(stack, entries, size);
263 
264 	required_size = ALIGN(required_size, 1 << DEPOT_STACK_ALIGN);
265 
266 	/* Check if there is not enough space in the current pool. */
267 	if (unlikely(pool_offset + required_size > DEPOT_POOL_SIZE)) {
268 		/* Bail out if we reached the pool limit. */
269 		if (unlikely(pool_index + 1 >= DEPOT_MAX_POOLS)) {
270 			WARN_ONCE(1, "Stack depot reached limit capacity");
271 			return NULL;
272 		}
273 
274 		/*
275 		 * Move on to the next pool.
276 		 * WRITE_ONCE pairs with potential concurrent read in
277 		 * stack_depot_fetch().
278 		 */
279 		WRITE_ONCE(pool_index, pool_index + 1);
280 		pool_offset = 0;
281 		/*
282 		 * If the maximum number of pools is not reached, take note
283 		 * that the next pool needs to initialized.
284 		 * smp_store_release() here pairs with smp_load_acquire() in
285 		 * stack_depot_save() and depot_init_pool().
286 		 */
287 		if (pool_index + 1 < DEPOT_MAX_POOLS)
288 			smp_store_release(&next_pool_required, 1);
289 	}
290 
291 	/* Assign the preallocated memory to a pool if required. */
292 	if (*prealloc)
293 		depot_init_pool(prealloc);
294 
295 	/* Check if we have a pool to save the stack trace. */
296 	if (stack_pools[pool_index] == NULL)
297 		return NULL;
298 
299 	/* Save the stack trace. */
300 	stack = stack_pools[pool_index] + pool_offset;
301 	stack->hash = hash;
302 	stack->size = size;
303 	stack->handle.pool_index = pool_index;
304 	stack->handle.offset = pool_offset >> DEPOT_STACK_ALIGN;
305 	stack->handle.valid = 1;
306 	stack->handle.extra = 0;
307 	memcpy(stack->entries, entries, flex_array_size(stack, entries, size));
308 	pool_offset += required_size;
309 
310 	return stack;
311 }
312 
313 /* Calculates the hash for a stack. */
314 static inline u32 hash_stack(unsigned long *entries, unsigned int size)
315 {
316 	return jhash2((u32 *)entries,
317 		      array_size(size,  sizeof(*entries)) / sizeof(u32),
318 		      STACK_HASH_SEED);
319 }
320 
321 /*
322  * Non-instrumented version of memcmp().
323  * Does not check the lexicographical order, only the equality.
324  */
325 static inline
326 int stackdepot_memcmp(const unsigned long *u1, const unsigned long *u2,
327 			unsigned int n)
328 {
329 	for ( ; n-- ; u1++, u2++) {
330 		if (*u1 != *u2)
331 			return 1;
332 	}
333 	return 0;
334 }
335 
336 /* Finds a stack in a bucket of the hash table. */
337 static inline struct stack_record *find_stack(struct stack_record *bucket,
338 					     unsigned long *entries, int size,
339 					     u32 hash)
340 {
341 	struct stack_record *found;
342 
343 	for (found = bucket; found; found = found->next) {
344 		if (found->hash == hash &&
345 		    found->size == size &&
346 		    !stackdepot_memcmp(entries, found->entries, size))
347 			return found;
348 	}
349 	return NULL;
350 }
351 
352 depot_stack_handle_t __stack_depot_save(unsigned long *entries,
353 					unsigned int nr_entries,
354 					gfp_t alloc_flags, bool can_alloc)
355 {
356 	struct stack_record *found = NULL, **bucket;
357 	union handle_parts retval = { .handle = 0 };
358 	struct page *page = NULL;
359 	void *prealloc = NULL;
360 	unsigned long flags;
361 	u32 hash;
362 
363 	/*
364 	 * If this stack trace is from an interrupt, including anything before
365 	 * interrupt entry usually leads to unbounded stack depot growth.
366 	 *
367 	 * Since use of filter_irq_stacks() is a requirement to ensure stack
368 	 * depot can efficiently deduplicate interrupt stacks, always
369 	 * filter_irq_stacks() to simplify all callers' use of stack depot.
370 	 */
371 	nr_entries = filter_irq_stacks(entries, nr_entries);
372 
373 	if (unlikely(nr_entries == 0) || stack_depot_disabled)
374 		goto fast_exit;
375 
376 	hash = hash_stack(entries, nr_entries);
377 	bucket = &stack_table[hash & stack_hash_mask];
378 
379 	/*
380 	 * Fast path: look the stack trace up without locking.
381 	 * The smp_load_acquire() here pairs with smp_store_release() to
382 	 * |bucket| below.
383 	 */
384 	found = find_stack(smp_load_acquire(bucket), entries, nr_entries, hash);
385 	if (found)
386 		goto exit;
387 
388 	/*
389 	 * Check if another stack pool needs to be initialized. If so, allocate
390 	 * the memory now - we won't be able to do that under the lock.
391 	 *
392 	 * The smp_load_acquire() here pairs with smp_store_release() to
393 	 * |next_pool_inited| in depot_alloc_stack() and depot_init_pool().
394 	 */
395 	if (unlikely(can_alloc && smp_load_acquire(&next_pool_required))) {
396 		/*
397 		 * Zero out zone modifiers, as we don't have specific zone
398 		 * requirements. Keep the flags related to allocation in atomic
399 		 * contexts and I/O.
400 		 */
401 		alloc_flags &= ~GFP_ZONEMASK;
402 		alloc_flags &= (GFP_ATOMIC | GFP_KERNEL);
403 		alloc_flags |= __GFP_NOWARN;
404 		page = alloc_pages(alloc_flags, DEPOT_POOL_ORDER);
405 		if (page)
406 			prealloc = page_address(page);
407 	}
408 
409 	raw_spin_lock_irqsave(&pool_lock, flags);
410 
411 	found = find_stack(*bucket, entries, nr_entries, hash);
412 	if (!found) {
413 		struct stack_record *new =
414 			depot_alloc_stack(entries, nr_entries, hash, &prealloc);
415 
416 		if (new) {
417 			new->next = *bucket;
418 			/*
419 			 * This smp_store_release() pairs with
420 			 * smp_load_acquire() from |bucket| above.
421 			 */
422 			smp_store_release(bucket, new);
423 			found = new;
424 		}
425 	} else if (prealloc) {
426 		/*
427 		 * Stack depot already contains this stack trace, but let's
428 		 * keep the preallocated memory for the future.
429 		 */
430 		depot_init_pool(&prealloc);
431 	}
432 
433 	raw_spin_unlock_irqrestore(&pool_lock, flags);
434 exit:
435 	if (prealloc) {
436 		/* Stack depot didn't use this memory, free it. */
437 		free_pages((unsigned long)prealloc, DEPOT_POOL_ORDER);
438 	}
439 	if (found)
440 		retval.handle = found->handle.handle;
441 fast_exit:
442 	return retval.handle;
443 }
444 EXPORT_SYMBOL_GPL(__stack_depot_save);
445 
446 depot_stack_handle_t stack_depot_save(unsigned long *entries,
447 				      unsigned int nr_entries,
448 				      gfp_t alloc_flags)
449 {
450 	return __stack_depot_save(entries, nr_entries, alloc_flags, true);
451 }
452 EXPORT_SYMBOL_GPL(stack_depot_save);
453 
454 unsigned int stack_depot_fetch(depot_stack_handle_t handle,
455 			       unsigned long **entries)
456 {
457 	union handle_parts parts = { .handle = handle };
458 	/*
459 	 * READ_ONCE pairs with potential concurrent write in
460 	 * depot_alloc_stack.
461 	 */
462 	int pool_index_cached = READ_ONCE(pool_index);
463 	void *pool;
464 	size_t offset = parts.offset << DEPOT_STACK_ALIGN;
465 	struct stack_record *stack;
466 
467 	*entries = NULL;
468 	if (!handle)
469 		return 0;
470 
471 	if (parts.pool_index > pool_index_cached) {
472 		WARN(1, "pool index %d out of bounds (%d) for stack id %08x\n",
473 			parts.pool_index, pool_index_cached, handle);
474 		return 0;
475 	}
476 	pool = stack_pools[parts.pool_index];
477 	if (!pool)
478 		return 0;
479 	stack = pool + offset;
480 
481 	*entries = stack->entries;
482 	return stack->size;
483 }
484 EXPORT_SYMBOL_GPL(stack_depot_fetch);
485 
486 void stack_depot_print(depot_stack_handle_t stack)
487 {
488 	unsigned long *entries;
489 	unsigned int nr_entries;
490 
491 	nr_entries = stack_depot_fetch(stack, &entries);
492 	if (nr_entries > 0)
493 		stack_trace_print(entries, nr_entries, 0);
494 }
495 EXPORT_SYMBOL_GPL(stack_depot_print);
496 
497 int stack_depot_snprint(depot_stack_handle_t handle, char *buf, size_t size,
498 		       int spaces)
499 {
500 	unsigned long *entries;
501 	unsigned int nr_entries;
502 
503 	nr_entries = stack_depot_fetch(handle, &entries);
504 	return nr_entries ? stack_trace_snprint(buf, size, entries, nr_entries,
505 						spaces) : 0;
506 }
507 EXPORT_SYMBOL_GPL(stack_depot_snprint);
508 
509 depot_stack_handle_t __must_check stack_depot_set_extra_bits(
510 			depot_stack_handle_t handle, unsigned int extra_bits)
511 {
512 	union handle_parts parts = { .handle = handle };
513 
514 	/* Don't set extra bits on empty handles. */
515 	if (!handle)
516 		return 0;
517 
518 	parts.extra = extra_bits;
519 	return parts.handle;
520 }
521 EXPORT_SYMBOL(stack_depot_set_extra_bits);
522 
523 unsigned int stack_depot_get_extra_bits(depot_stack_handle_t handle)
524 {
525 	union handle_parts parts = { .handle = handle };
526 
527 	return parts.extra;
528 }
529 EXPORT_SYMBOL(stack_depot_get_extra_bits);
530