• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Generic stack depot for storing stack traces.
4  *
5  * Some debugging tools need to save stack traces of certain events which can
6  * be later presented to the user. For example, KASAN needs to safe alloc and
7  * free stacks for each object, but storing two stack traces per object
8  * requires too much memory (e.g. SLUB_DEBUG needs 256 bytes per object for
9  * that).
10  *
11  * Instead, stack depot maintains a hashtable of unique stacktraces. Since alloc
12  * and free stacks repeat a lot, we save about 100x space.
13  * Stacks are never removed from depot, so we store them contiguously one after
14  * another in a contiguous memory allocation.
15  *
16  * Author: Alexander Potapenko <glider@google.com>
17  * Copyright (C) 2016 Google, Inc.
18  *
19  * Based on code by Dmitry Chernenkov.
20  */
21 
22 #include <linux/gfp.h>
23 #include <linux/jhash.h>
24 #include <linux/kernel.h>
25 #include <linux/mm.h>
26 #include <linux/percpu.h>
27 #include <linux/printk.h>
28 #include <linux/slab.h>
29 #include <linux/stacktrace.h>
30 #include <linux/stackdepot.h>
31 #include <linux/string.h>
32 #include <linux/types.h>
33 #include <linux/memblock.h>
34 
35 #define DEPOT_STACK_BITS (sizeof(depot_stack_handle_t) * 8)
36 
37 #define STACK_ALLOC_NULL_PROTECTION_BITS 1
38 #define STACK_ALLOC_ORDER 2 /* 'Slab' size order for stack depot, 4 pages */
39 #define STACK_ALLOC_SIZE (1LL << (PAGE_SHIFT + STACK_ALLOC_ORDER))
40 #define STACK_ALLOC_ALIGN 4
41 #define STACK_ALLOC_OFFSET_BITS (STACK_ALLOC_ORDER + PAGE_SHIFT - \
42 					STACK_ALLOC_ALIGN)
43 #define STACK_ALLOC_INDEX_BITS (DEPOT_STACK_BITS - \
44 		STACK_ALLOC_NULL_PROTECTION_BITS - STACK_ALLOC_OFFSET_BITS)
45 #define STACK_ALLOC_SLABS_CAP 8192
46 #define STACK_ALLOC_MAX_SLABS \
47 	(((1LL << (STACK_ALLOC_INDEX_BITS)) < STACK_ALLOC_SLABS_CAP) ? \
48 	 (1LL << (STACK_ALLOC_INDEX_BITS)) : STACK_ALLOC_SLABS_CAP)
49 
50 /* The compact structure to store the reference to stacks. */
51 union handle_parts {
52 	depot_stack_handle_t handle;
53 	struct {
54 		u32 slabindex : STACK_ALLOC_INDEX_BITS;
55 		u32 offset : STACK_ALLOC_OFFSET_BITS;
56 		u32 valid : STACK_ALLOC_NULL_PROTECTION_BITS;
57 	};
58 };
59 
60 struct stack_record {
61 	struct stack_record *next;	/* Link in the hashtable */
62 	u32 hash;			/* Hash in the hastable */
63 	u32 size;			/* Number of frames in the stack */
64 	union handle_parts handle;
65 	unsigned long entries[];	/* Variable-sized array of entries. */
66 };
67 
68 static void *stack_slabs[STACK_ALLOC_MAX_SLABS];
69 
70 static int depot_index;
71 static int next_slab_inited;
72 static size_t depot_offset;
73 static DEFINE_RAW_SPINLOCK(depot_lock);
74 
init_stack_slab(void ** prealloc)75 static bool init_stack_slab(void **prealloc)
76 {
77 	if (!*prealloc)
78 		return false;
79 	/*
80 	 * This smp_load_acquire() pairs with smp_store_release() to
81 	 * |next_slab_inited| below and in depot_alloc_stack().
82 	 */
83 	if (smp_load_acquire(&next_slab_inited))
84 		return true;
85 	if (stack_slabs[depot_index] == NULL) {
86 		stack_slabs[depot_index] = *prealloc;
87 		*prealloc = NULL;
88 	} else {
89 		/* If this is the last depot slab, do not touch the next one. */
90 		if (depot_index + 1 < STACK_ALLOC_MAX_SLABS) {
91 			stack_slabs[depot_index + 1] = *prealloc;
92 			*prealloc = NULL;
93 		}
94 		/*
95 		 * This smp_store_release pairs with smp_load_acquire() from
96 		 * |next_slab_inited| above and in stack_depot_save().
97 		 */
98 		smp_store_release(&next_slab_inited, 1);
99 	}
100 	return true;
101 }
102 
103 /* Allocation of a new stack in raw storage */
104 static struct stack_record *
depot_alloc_stack(unsigned long * entries,int size,u32 hash,void ** prealloc)105 depot_alloc_stack(unsigned long *entries, int size, u32 hash, void **prealloc)
106 {
107 	struct stack_record *stack;
108 	size_t required_size = struct_size(stack, entries, size);
109 
110 	required_size = ALIGN(required_size, 1 << STACK_ALLOC_ALIGN);
111 
112 	if (unlikely(depot_offset + required_size > STACK_ALLOC_SIZE)) {
113 		if (unlikely(depot_index + 1 >= STACK_ALLOC_MAX_SLABS)) {
114 			WARN_ONCE(1, "Stack depot reached limit capacity");
115 			return NULL;
116 		}
117 		depot_index++;
118 		depot_offset = 0;
119 		/*
120 		 * smp_store_release() here pairs with smp_load_acquire() from
121 		 * |next_slab_inited| in stack_depot_save() and
122 		 * init_stack_slab().
123 		 */
124 		if (depot_index + 1 < STACK_ALLOC_MAX_SLABS)
125 			smp_store_release(&next_slab_inited, 0);
126 	}
127 	init_stack_slab(prealloc);
128 	if (stack_slabs[depot_index] == NULL)
129 		return NULL;
130 
131 	stack = stack_slabs[depot_index] + depot_offset;
132 
133 	stack->hash = hash;
134 	stack->size = size;
135 	stack->handle.slabindex = depot_index;
136 	stack->handle.offset = depot_offset >> STACK_ALLOC_ALIGN;
137 	stack->handle.valid = 1;
138 	memcpy(stack->entries, entries, flex_array_size(stack, entries, size));
139 	depot_offset += required_size;
140 
141 	return stack;
142 }
143 
144 #define STACK_HASH_SIZE (1L << CONFIG_STACK_HASH_ORDER)
145 #define STACK_HASH_MASK (STACK_HASH_SIZE - 1)
146 #define STACK_HASH_SEED 0x9747b28c
147 
148 static bool stack_depot_disable;
149 static struct stack_record **stack_table;
150 
is_stack_depot_disabled(char * str)151 static int __init is_stack_depot_disabled(char *str)
152 {
153 	int ret;
154 
155 	ret = kstrtobool(str, &stack_depot_disable);
156 	if (!ret && stack_depot_disable) {
157 		pr_info("Stack Depot is disabled\n");
158 		stack_table = NULL;
159 	}
160 	return 0;
161 }
162 early_param("stack_depot_disable", is_stack_depot_disabled);
163 
stack_depot_init(void)164 int __init stack_depot_init(void)
165 {
166 	if (!stack_depot_disable) {
167 		size_t size = (STACK_HASH_SIZE * sizeof(struct stack_record *));
168 		int i;
169 
170 		stack_table = memblock_alloc(size, size);
171 		for (i = 0; i < STACK_HASH_SIZE;  i++)
172 			stack_table[i] = NULL;
173 	}
174 	return 0;
175 }
176 
177 /* Calculate hash for a stack */
hash_stack(unsigned long * entries,unsigned int size)178 static inline u32 hash_stack(unsigned long *entries, unsigned int size)
179 {
180 	return jhash2((u32 *)entries,
181 		      array_size(size,  sizeof(*entries)) / sizeof(u32),
182 		      STACK_HASH_SEED);
183 }
184 
185 /* Use our own, non-instrumented version of memcmp().
186  *
187  * We actually don't care about the order, just the equality.
188  */
189 static inline
stackdepot_memcmp(const unsigned long * u1,const unsigned long * u2,unsigned int n)190 int stackdepot_memcmp(const unsigned long *u1, const unsigned long *u2,
191 			unsigned int n)
192 {
193 	for ( ; n-- ; u1++, u2++) {
194 		if (*u1 != *u2)
195 			return 1;
196 	}
197 	return 0;
198 }
199 
200 /* Find a stack that is equal to the one stored in entries in the hash */
find_stack(struct stack_record * bucket,unsigned long * entries,int size,u32 hash)201 static inline struct stack_record *find_stack(struct stack_record *bucket,
202 					     unsigned long *entries, int size,
203 					     u32 hash)
204 {
205 	struct stack_record *found;
206 
207 	for (found = bucket; found; found = found->next) {
208 		if (found->hash == hash &&
209 		    found->size == size &&
210 		    !stackdepot_memcmp(entries, found->entries, size))
211 			return found;
212 	}
213 	return NULL;
214 }
215 
216 /**
217  * stack_depot_fetch - Fetch stack entries from a depot
218  *
219  * @handle:		Stack depot handle which was returned from
220  *			stack_depot_save().
221  * @entries:		Pointer to store the entries address
222  *
223  * Return: The number of trace entries for this depot.
224  */
stack_depot_fetch(depot_stack_handle_t handle,unsigned long ** entries)225 unsigned int stack_depot_fetch(depot_stack_handle_t handle,
226 			       unsigned long **entries)
227 {
228 	union handle_parts parts = { .handle = handle };
229 	void *slab;
230 	size_t offset = parts.offset << STACK_ALLOC_ALIGN;
231 	struct stack_record *stack;
232 
233 	*entries = NULL;
234 	if (parts.slabindex > depot_index) {
235 		WARN(1, "slab index %d out of bounds (%d) for stack id %08x\n",
236 			parts.slabindex, depot_index, handle);
237 		return 0;
238 	}
239 	slab = stack_slabs[parts.slabindex];
240 	if (!slab)
241 		return 0;
242 	stack = slab + offset;
243 
244 	*entries = stack->entries;
245 	return stack->size;
246 }
247 EXPORT_SYMBOL_GPL(stack_depot_fetch);
248 
249 /**
250  * __stack_depot_save - Save a stack trace from an array
251  *
252  * @entries:		Pointer to storage array
253  * @nr_entries:		Size of the storage array
254  * @alloc_flags:	Allocation gfp flags
255  * @can_alloc:		Allocate stack slabs (increased chance of failure if false)
256  *
257  * Saves a stack trace from @entries array of size @nr_entries. If @can_alloc is
258  * %true, is allowed to replenish the stack slab pool in case no space is left
259  * (allocates using GFP flags of @alloc_flags). If @can_alloc is %false, avoids
260  * any allocations and will fail if no space is left to store the stack trace.
261  *
262  * If the stack trace in @entries is from an interrupt, only the portion up to
263  * interrupt entry is saved.
264  *
265  * Context: Any context, but setting @can_alloc to %false is required if
266  *          alloc_pages() cannot be used from the current context. Currently
267  *          this is the case from contexts where neither %GFP_ATOMIC nor
268  *          %GFP_NOWAIT can be used (NMI, raw_spin_lock).
269  *
270  * Return: The handle of the stack struct stored in depot, 0 on failure.
271  */
__stack_depot_save(unsigned long * entries,unsigned int nr_entries,gfp_t alloc_flags,bool can_alloc)272 depot_stack_handle_t __stack_depot_save(unsigned long *entries,
273 					unsigned int nr_entries,
274 					gfp_t alloc_flags, bool can_alloc)
275 {
276 	struct stack_record *found = NULL, **bucket;
277 	depot_stack_handle_t retval = 0;
278 	struct page *page = NULL;
279 	void *prealloc = NULL;
280 	unsigned long flags;
281 	u32 hash;
282 
283 	/*
284 	 * If this stack trace is from an interrupt, including anything before
285 	 * interrupt entry usually leads to unbounded stackdepot growth.
286 	 *
287 	 * Because use of filter_irq_stacks() is a requirement to ensure
288 	 * stackdepot can efficiently deduplicate interrupt stacks, always
289 	 * filter_irq_stacks() to simplify all callers' use of stackdepot.
290 	 */
291 	nr_entries = filter_irq_stacks(entries, nr_entries);
292 
293 	if (unlikely(nr_entries == 0) || stack_depot_disable)
294 		goto fast_exit;
295 
296 	hash = hash_stack(entries, nr_entries);
297 	bucket = &stack_table[hash & STACK_HASH_MASK];
298 
299 	/*
300 	 * Fast path: look the stack trace up without locking.
301 	 * The smp_load_acquire() here pairs with smp_store_release() to
302 	 * |bucket| below.
303 	 */
304 	found = find_stack(smp_load_acquire(bucket), entries,
305 			   nr_entries, hash);
306 	if (found)
307 		goto exit;
308 
309 	/*
310 	 * Check if the current or the next stack slab need to be initialized.
311 	 * If so, allocate the memory - we won't be able to do that under the
312 	 * lock.
313 	 *
314 	 * The smp_load_acquire() here pairs with smp_store_release() to
315 	 * |next_slab_inited| in depot_alloc_stack() and init_stack_slab().
316 	 */
317 	if (unlikely(can_alloc && !smp_load_acquire(&next_slab_inited))) {
318 		/*
319 		 * Zero out zone modifiers, as we don't have specific zone
320 		 * requirements. Keep the flags related to allocation in atomic
321 		 * contexts and I/O.
322 		 */
323 		alloc_flags &= ~GFP_ZONEMASK;
324 		alloc_flags &= (GFP_ATOMIC | GFP_KERNEL);
325 		alloc_flags |= __GFP_NOWARN;
326 		page = alloc_pages(alloc_flags, STACK_ALLOC_ORDER);
327 		if (page)
328 			prealloc = page_address(page);
329 	}
330 
331 	raw_spin_lock_irqsave(&depot_lock, flags);
332 
333 	found = find_stack(*bucket, entries, nr_entries, hash);
334 	if (!found) {
335 		struct stack_record *new = depot_alloc_stack(entries, nr_entries, hash, &prealloc);
336 
337 		if (new) {
338 			new->next = *bucket;
339 			/*
340 			 * This smp_store_release() pairs with
341 			 * smp_load_acquire() from |bucket| above.
342 			 */
343 			smp_store_release(bucket, new);
344 			found = new;
345 		}
346 	} else if (prealloc) {
347 		/*
348 		 * We didn't need to store this stack trace, but let's keep
349 		 * the preallocated memory for the future.
350 		 */
351 		WARN_ON(!init_stack_slab(&prealloc));
352 	}
353 
354 	raw_spin_unlock_irqrestore(&depot_lock, flags);
355 exit:
356 	if (prealloc) {
357 		/* Nobody used this memory, ok to free it. */
358 		free_pages((unsigned long)prealloc, STACK_ALLOC_ORDER);
359 	}
360 	if (found)
361 		retval = found->handle.handle;
362 fast_exit:
363 	return retval;
364 }
365 EXPORT_SYMBOL_GPL(__stack_depot_save);
366 
367 /**
368  * stack_depot_save - Save a stack trace from an array
369  *
370  * @entries:		Pointer to storage array
371  * @nr_entries:		Size of the storage array
372  * @alloc_flags:	Allocation gfp flags
373  *
374  * Context: Contexts where allocations via alloc_pages() are allowed.
375  *          See __stack_depot_save() for more details.
376  *
377  * Return: The handle of the stack struct stored in depot, 0 on failure.
378  */
stack_depot_save(unsigned long * entries,unsigned int nr_entries,gfp_t alloc_flags)379 depot_stack_handle_t stack_depot_save(unsigned long *entries,
380 				      unsigned int nr_entries,
381 				      gfp_t alloc_flags)
382 {
383 	return __stack_depot_save(entries, nr_entries, alloc_flags, true);
384 }
385 EXPORT_SYMBOL_GPL(stack_depot_save);
386