xref: /openbmc/linux/lib/stackdepot.c (revision 3e26a691)
1 /*
2  * Generic stack depot for storing stack traces.
3  *
4  * Some debugging tools need to save stack traces of certain events which can
5  * be later presented to the user. For example, KASAN needs to safe alloc and
6  * free stacks for each object, but storing two stack traces per object
7  * requires too much memory (e.g. SLUB_DEBUG needs 256 bytes per object for
8  * that).
9  *
10  * Instead, stack depot maintains a hashtable of unique stacktraces. Since alloc
11  * and free stacks repeat a lot, we save about 100x space.
12  * Stacks are never removed from depot, so we store them contiguously one after
13  * another in a contiguos memory allocation.
14  *
15  * Author: Alexander Potapenko <glider@google.com>
16  * Copyright (C) 2016 Google, Inc.
17  *
18  * Based on code by Dmitry Chernenkov.
19  *
20  * This program is free software; you can redistribute it and/or
21  * modify it under the terms of the GNU General Public License
22  * version 2 as published by the Free Software Foundation.
23  *
24  * This program is distributed in the hope that it will be useful, but
25  * WITHOUT ANY WARRANTY; without even the implied warranty of
26  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
27  * General Public License for more details.
28  *
29  */
30 
31 #include <linux/gfp.h>
32 #include <linux/jhash.h>
33 #include <linux/kernel.h>
34 #include <linux/mm.h>
35 #include <linux/percpu.h>
36 #include <linux/printk.h>
37 #include <linux/slab.h>
38 #include <linux/stacktrace.h>
39 #include <linux/stackdepot.h>
40 #include <linux/string.h>
41 #include <linux/types.h>
42 
43 #define DEPOT_STACK_BITS (sizeof(depot_stack_handle_t) * 8)
44 
45 #define STACK_ALLOC_ORDER 2 /* 'Slab' size order for stack depot, 4 pages */
46 #define STACK_ALLOC_SIZE (1LL << (PAGE_SHIFT + STACK_ALLOC_ORDER))
47 #define STACK_ALLOC_ALIGN 4
48 #define STACK_ALLOC_OFFSET_BITS (STACK_ALLOC_ORDER + PAGE_SHIFT - \
49 					STACK_ALLOC_ALIGN)
50 #define STACK_ALLOC_INDEX_BITS (DEPOT_STACK_BITS - STACK_ALLOC_OFFSET_BITS)
51 #define STACK_ALLOC_SLABS_CAP 1024
52 #define STACK_ALLOC_MAX_SLABS \
53 	(((1LL << (STACK_ALLOC_INDEX_BITS)) < STACK_ALLOC_SLABS_CAP) ? \
54 	 (1LL << (STACK_ALLOC_INDEX_BITS)) : STACK_ALLOC_SLABS_CAP)
55 
56 /* The compact structure to store the reference to stacks. */
57 union handle_parts {
58 	depot_stack_handle_t handle;
59 	struct {
60 		u32 slabindex : STACK_ALLOC_INDEX_BITS;
61 		u32 offset : STACK_ALLOC_OFFSET_BITS;
62 	};
63 };
64 
65 struct stack_record {
66 	struct stack_record *next;	/* Link in the hashtable */
67 	u32 hash;			/* Hash in the hastable */
68 	u32 size;			/* Number of frames in the stack */
69 	union handle_parts handle;
70 	unsigned long entries[1];	/* Variable-sized array of entries. */
71 };
72 
73 static void *stack_slabs[STACK_ALLOC_MAX_SLABS];
74 
75 static int depot_index;
76 static int next_slab_inited;
77 static size_t depot_offset;
78 static DEFINE_SPINLOCK(depot_lock);
79 
80 static bool init_stack_slab(void **prealloc)
81 {
82 	if (!*prealloc)
83 		return false;
84 	/*
85 	 * This smp_load_acquire() pairs with smp_store_release() to
86 	 * |next_slab_inited| below and in depot_alloc_stack().
87 	 */
88 	if (smp_load_acquire(&next_slab_inited))
89 		return true;
90 	if (stack_slabs[depot_index] == NULL) {
91 		stack_slabs[depot_index] = *prealloc;
92 	} else {
93 		stack_slabs[depot_index + 1] = *prealloc;
94 		/*
95 		 * This smp_store_release pairs with smp_load_acquire() from
96 		 * |next_slab_inited| above and in depot_save_stack().
97 		 */
98 		smp_store_release(&next_slab_inited, 1);
99 	}
100 	*prealloc = NULL;
101 	return true;
102 }
103 
104 /* Allocation of a new stack in raw storage */
105 static struct stack_record *depot_alloc_stack(unsigned long *entries, int size,
106 		u32 hash, void **prealloc, gfp_t alloc_flags)
107 {
108 	int required_size = offsetof(struct stack_record, entries) +
109 		sizeof(unsigned long) * size;
110 	struct stack_record *stack;
111 
112 	required_size = ALIGN(required_size, 1 << STACK_ALLOC_ALIGN);
113 
114 	if (unlikely(depot_offset + required_size > STACK_ALLOC_SIZE)) {
115 		if (unlikely(depot_index + 1 >= STACK_ALLOC_MAX_SLABS)) {
116 			WARN_ONCE(1, "Stack depot reached limit capacity");
117 			return NULL;
118 		}
119 		depot_index++;
120 		depot_offset = 0;
121 		/*
122 		 * smp_store_release() here pairs with smp_load_acquire() from
123 		 * |next_slab_inited| in depot_save_stack() and
124 		 * init_stack_slab().
125 		 */
126 		if (depot_index + 1 < STACK_ALLOC_MAX_SLABS)
127 			smp_store_release(&next_slab_inited, 0);
128 	}
129 	init_stack_slab(prealloc);
130 	if (stack_slabs[depot_index] == NULL)
131 		return NULL;
132 
133 	stack = stack_slabs[depot_index] + depot_offset;
134 
135 	stack->hash = hash;
136 	stack->size = size;
137 	stack->handle.slabindex = depot_index;
138 	stack->handle.offset = depot_offset >> STACK_ALLOC_ALIGN;
139 	memcpy(stack->entries, entries, size * sizeof(unsigned long));
140 	depot_offset += required_size;
141 
142 	return stack;
143 }
144 
145 #define STACK_HASH_ORDER 20
146 #define STACK_HASH_SIZE (1L << STACK_HASH_ORDER)
147 #define STACK_HASH_MASK (STACK_HASH_SIZE - 1)
148 #define STACK_HASH_SEED 0x9747b28c
149 
150 static struct stack_record *stack_table[STACK_HASH_SIZE] = {
151 	[0 ...	STACK_HASH_SIZE - 1] = NULL
152 };
153 
154 /* Calculate hash for a stack */
155 static inline u32 hash_stack(unsigned long *entries, unsigned int size)
156 {
157 	return jhash2((u32 *)entries,
158 			       size * sizeof(unsigned long) / sizeof(u32),
159 			       STACK_HASH_SEED);
160 }
161 
162 /* Find a stack that is equal to the one stored in entries in the hash */
163 static inline struct stack_record *find_stack(struct stack_record *bucket,
164 					     unsigned long *entries, int size,
165 					     u32 hash)
166 {
167 	struct stack_record *found;
168 
169 	for (found = bucket; found; found = found->next) {
170 		if (found->hash == hash &&
171 		    found->size == size &&
172 		    !memcmp(entries, found->entries,
173 			    size * sizeof(unsigned long))) {
174 			return found;
175 		}
176 	}
177 	return NULL;
178 }
179 
180 void depot_fetch_stack(depot_stack_handle_t handle, struct stack_trace *trace)
181 {
182 	union handle_parts parts = { .handle = handle };
183 	void *slab = stack_slabs[parts.slabindex];
184 	size_t offset = parts.offset << STACK_ALLOC_ALIGN;
185 	struct stack_record *stack = slab + offset;
186 
187 	trace->nr_entries = trace->max_entries = stack->size;
188 	trace->entries = stack->entries;
189 	trace->skip = 0;
190 }
191 
192 /**
193  * depot_save_stack - save stack in a stack depot.
194  * @trace - the stacktrace to save.
195  * @alloc_flags - flags for allocating additional memory if required.
196  *
197  * Returns the handle of the stack struct stored in depot.
198  */
199 depot_stack_handle_t depot_save_stack(struct stack_trace *trace,
200 				    gfp_t alloc_flags)
201 {
202 	u32 hash;
203 	depot_stack_handle_t retval = 0;
204 	struct stack_record *found = NULL, **bucket;
205 	unsigned long flags;
206 	struct page *page = NULL;
207 	void *prealloc = NULL;
208 
209 	if (unlikely(trace->nr_entries == 0))
210 		goto fast_exit;
211 
212 	hash = hash_stack(trace->entries, trace->nr_entries);
213 	/* Bad luck, we won't store this stack. */
214 	if (hash == 0)
215 		goto exit;
216 
217 	bucket = &stack_table[hash & STACK_HASH_MASK];
218 
219 	/*
220 	 * Fast path: look the stack trace up without locking.
221 	 * The smp_load_acquire() here pairs with smp_store_release() to
222 	 * |bucket| below.
223 	 */
224 	found = find_stack(smp_load_acquire(bucket), trace->entries,
225 			   trace->nr_entries, hash);
226 	if (found)
227 		goto exit;
228 
229 	/*
230 	 * Check if the current or the next stack slab need to be initialized.
231 	 * If so, allocate the memory - we won't be able to do that under the
232 	 * lock.
233 	 *
234 	 * The smp_load_acquire() here pairs with smp_store_release() to
235 	 * |next_slab_inited| in depot_alloc_stack() and init_stack_slab().
236 	 */
237 	if (unlikely(!smp_load_acquire(&next_slab_inited))) {
238 		/*
239 		 * Zero out zone modifiers, as we don't have specific zone
240 		 * requirements. Keep the flags related to allocation in atomic
241 		 * contexts and I/O.
242 		 */
243 		alloc_flags &= ~GFP_ZONEMASK;
244 		alloc_flags &= (GFP_ATOMIC | GFP_KERNEL);
245 		page = alloc_pages(alloc_flags, STACK_ALLOC_ORDER);
246 		if (page)
247 			prealloc = page_address(page);
248 	}
249 
250 	spin_lock_irqsave(&depot_lock, flags);
251 
252 	found = find_stack(*bucket, trace->entries, trace->nr_entries, hash);
253 	if (!found) {
254 		struct stack_record *new =
255 			depot_alloc_stack(trace->entries, trace->nr_entries,
256 					  hash, &prealloc, alloc_flags);
257 		if (new) {
258 			new->next = *bucket;
259 			/*
260 			 * This smp_store_release() pairs with
261 			 * smp_load_acquire() from |bucket| above.
262 			 */
263 			smp_store_release(bucket, new);
264 			found = new;
265 		}
266 	} else if (prealloc) {
267 		/*
268 		 * We didn't need to store this stack trace, but let's keep
269 		 * the preallocated memory for the future.
270 		 */
271 		WARN_ON(!init_stack_slab(&prealloc));
272 	}
273 
274 	spin_unlock_irqrestore(&depot_lock, flags);
275 exit:
276 	if (prealloc) {
277 		/* Nobody used this memory, ok to free it. */
278 		free_pages((unsigned long)prealloc, STACK_ALLOC_ORDER);
279 	}
280 	if (found)
281 		retval = found->handle.handle;
282 fast_exit:
283 	return retval;
284 }
285