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_NULL_PROTECTION_BITS 1
46 #define STACK_ALLOC_ORDER 2 /* 'Slab' size order for stack depot, 4 pages */
47 #define STACK_ALLOC_SIZE (1LL << (PAGE_SHIFT + STACK_ALLOC_ORDER))
48 #define STACK_ALLOC_ALIGN 4
49 #define STACK_ALLOC_OFFSET_BITS (STACK_ALLOC_ORDER + PAGE_SHIFT - \
50 					STACK_ALLOC_ALIGN)
51 #define STACK_ALLOC_INDEX_BITS (DEPOT_STACK_BITS - \
52 		STACK_ALLOC_NULL_PROTECTION_BITS - STACK_ALLOC_OFFSET_BITS)
53 #define STACK_ALLOC_SLABS_CAP 8192
54 #define STACK_ALLOC_MAX_SLABS \
55 	(((1LL << (STACK_ALLOC_INDEX_BITS)) < STACK_ALLOC_SLABS_CAP) ? \
56 	 (1LL << (STACK_ALLOC_INDEX_BITS)) : STACK_ALLOC_SLABS_CAP)
57 
58 /* The compact structure to store the reference to stacks. */
59 union handle_parts {
60 	depot_stack_handle_t handle;
61 	struct {
62 		u32 slabindex : STACK_ALLOC_INDEX_BITS;
63 		u32 offset : STACK_ALLOC_OFFSET_BITS;
64 		u32 valid : STACK_ALLOC_NULL_PROTECTION_BITS;
65 	};
66 };
67 
68 struct stack_record {
69 	struct stack_record *next;	/* Link in the hashtable */
70 	u32 hash;			/* Hash in the hastable */
71 	u32 size;			/* Number of frames in the stack */
72 	union handle_parts handle;
73 	unsigned long entries[1];	/* Variable-sized array of entries. */
74 };
75 
76 static void *stack_slabs[STACK_ALLOC_MAX_SLABS];
77 
78 static int depot_index;
79 static int next_slab_inited;
80 static size_t depot_offset;
81 static DEFINE_RAW_SPINLOCK(depot_lock);
82 
init_stack_slab(void ** prealloc)83 static bool init_stack_slab(void **prealloc)
84 {
85 	if (!*prealloc)
86 		return false;
87 	/*
88 	 * This smp_load_acquire() pairs with smp_store_release() to
89 	 * |next_slab_inited| below and in depot_alloc_stack().
90 	 */
91 	if (smp_load_acquire(&next_slab_inited))
92 		return true;
93 	if (stack_slabs[depot_index] == NULL) {
94 		stack_slabs[depot_index] = *prealloc;
95 		*prealloc = NULL;
96 	} else {
97 		/* If this is the last depot slab, do not touch the next one. */
98 		if (depot_index + 1 < STACK_ALLOC_MAX_SLABS) {
99 			stack_slabs[depot_index + 1] = *prealloc;
100 			*prealloc = NULL;
101 		}
102 		/*
103 		 * This smp_store_release pairs with smp_load_acquire() from
104 		 * |next_slab_inited| above and in depot_save_stack().
105 		 */
106 		smp_store_release(&next_slab_inited, 1);
107 	}
108 	return true;
109 }
110 
111 /* Allocation of a new stack in raw storage */
depot_alloc_stack(unsigned long * entries,int size,u32 hash,void ** prealloc,gfp_t alloc_flags)112 static struct stack_record *depot_alloc_stack(unsigned long *entries, int size,
113 		u32 hash, void **prealloc, gfp_t alloc_flags)
114 {
115 	int required_size = offsetof(struct stack_record, entries) +
116 		sizeof(unsigned long) * size;
117 	struct stack_record *stack;
118 
119 	required_size = ALIGN(required_size, 1 << STACK_ALLOC_ALIGN);
120 
121 	if (unlikely(depot_offset + required_size > STACK_ALLOC_SIZE)) {
122 		if (unlikely(depot_index + 1 >= STACK_ALLOC_MAX_SLABS)) {
123 			WARN_ONCE(1, "Stack depot reached limit capacity");
124 			return NULL;
125 		}
126 		depot_index++;
127 		depot_offset = 0;
128 		/*
129 		 * smp_store_release() here pairs with smp_load_acquire() from
130 		 * |next_slab_inited| in depot_save_stack() and
131 		 * init_stack_slab().
132 		 */
133 		if (depot_index + 1 < STACK_ALLOC_MAX_SLABS)
134 			smp_store_release(&next_slab_inited, 0);
135 	}
136 	init_stack_slab(prealloc);
137 	if (stack_slabs[depot_index] == NULL)
138 		return NULL;
139 
140 	stack = stack_slabs[depot_index] + depot_offset;
141 
142 	stack->hash = hash;
143 	stack->size = size;
144 	stack->handle.slabindex = depot_index;
145 	stack->handle.offset = depot_offset >> STACK_ALLOC_ALIGN;
146 	stack->handle.valid = 1;
147 	memcpy(stack->entries, entries, size * sizeof(unsigned long));
148 	depot_offset += required_size;
149 
150 	return stack;
151 }
152 
153 #define STACK_HASH_ORDER 20
154 #define STACK_HASH_SIZE (1L << STACK_HASH_ORDER)
155 #define STACK_HASH_MASK (STACK_HASH_SIZE - 1)
156 #define STACK_HASH_SEED 0x9747b28c
157 
158 static struct stack_record *stack_table[STACK_HASH_SIZE] = {
159 	[0 ...	STACK_HASH_SIZE - 1] = NULL
160 };
161 
162 /* Calculate hash for a stack */
hash_stack(unsigned long * entries,unsigned int size)163 static inline u32 hash_stack(unsigned long *entries, unsigned int size)
164 {
165 	return jhash2((u32 *)entries,
166 			       size * sizeof(unsigned long) / sizeof(u32),
167 			       STACK_HASH_SEED);
168 }
169 
170 /* Use our own, non-instrumented version of memcmp().
171  *
172  * We actually don't care about the order, just the equality.
173  */
174 static inline
stackdepot_memcmp(const unsigned long * u1,const unsigned long * u2,unsigned int n)175 int stackdepot_memcmp(const unsigned long *u1, const unsigned long *u2,
176 			unsigned int n)
177 {
178 	for ( ; n-- ; u1++, u2++) {
179 		if (*u1 != *u2)
180 			return 1;
181 	}
182 	return 0;
183 }
184 
185 /* 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)186 static inline struct stack_record *find_stack(struct stack_record *bucket,
187 					     unsigned long *entries, int size,
188 					     u32 hash)
189 {
190 	struct stack_record *found;
191 
192 	for (found = bucket; found; found = found->next) {
193 		if (found->hash == hash &&
194 		    found->size == size &&
195 		    !stackdepot_memcmp(entries, found->entries, size))
196 			return found;
197 	}
198 	return NULL;
199 }
200 
depot_fetch_stack(depot_stack_handle_t handle,struct stack_trace * trace)201 void depot_fetch_stack(depot_stack_handle_t handle, struct stack_trace *trace)
202 {
203 	union handle_parts parts = { .handle = handle };
204 	void *slab = stack_slabs[parts.slabindex];
205 	size_t offset = parts.offset << STACK_ALLOC_ALIGN;
206 	struct stack_record *stack = slab + offset;
207 
208 	trace->nr_entries = trace->max_entries = stack->size;
209 	trace->entries = stack->entries;
210 	trace->skip = 0;
211 }
212 EXPORT_SYMBOL_GPL(depot_fetch_stack);
213 
214 /**
215  * depot_save_stack - save stack in a stack depot.
216  * @trace - the stacktrace to save.
217  * @alloc_flags - flags for allocating additional memory if required.
218  *
219  * Returns the handle of the stack struct stored in depot.
220  */
depot_save_stack(struct stack_trace * trace,gfp_t alloc_flags)221 depot_stack_handle_t depot_save_stack(struct stack_trace *trace,
222 				    gfp_t alloc_flags)
223 {
224 	u32 hash;
225 	depot_stack_handle_t retval = 0;
226 	struct stack_record *found = NULL, **bucket;
227 	unsigned long flags;
228 	struct page *page = NULL;
229 	void *prealloc = NULL;
230 
231 	if (unlikely(trace->nr_entries == 0))
232 		goto fast_exit;
233 
234 	hash = hash_stack(trace->entries, trace->nr_entries);
235 	bucket = &stack_table[hash & STACK_HASH_MASK];
236 
237 	/*
238 	 * Fast path: look the stack trace up without locking.
239 	 * The smp_load_acquire() here pairs with smp_store_release() to
240 	 * |bucket| below.
241 	 */
242 	found = find_stack(smp_load_acquire(bucket), trace->entries,
243 			   trace->nr_entries, hash);
244 	if (found)
245 		goto exit;
246 
247 	/*
248 	 * Check if the current or the next stack slab need to be initialized.
249 	 * If so, allocate the memory - we won't be able to do that under the
250 	 * lock.
251 	 *
252 	 * The smp_load_acquire() here pairs with smp_store_release() to
253 	 * |next_slab_inited| in depot_alloc_stack() and init_stack_slab().
254 	 */
255 	if (unlikely(!smp_load_acquire(&next_slab_inited))) {
256 		/*
257 		 * Zero out zone modifiers, as we don't have specific zone
258 		 * requirements. Keep the flags related to allocation in atomic
259 		 * contexts and I/O.
260 		 */
261 		alloc_flags &= ~GFP_ZONEMASK;
262 		alloc_flags &= (GFP_ATOMIC | GFP_KERNEL);
263 		alloc_flags |= __GFP_NOWARN;
264 		page = alloc_pages(alloc_flags, STACK_ALLOC_ORDER);
265 		if (page)
266 			prealloc = page_address(page);
267 	}
268 
269 	raw_spin_lock_irqsave(&depot_lock, flags);
270 
271 	found = find_stack(*bucket, trace->entries, trace->nr_entries, hash);
272 	if (!found) {
273 		struct stack_record *new =
274 			depot_alloc_stack(trace->entries, trace->nr_entries,
275 					  hash, &prealloc, alloc_flags);
276 		if (new) {
277 			new->next = *bucket;
278 			/*
279 			 * This smp_store_release() pairs with
280 			 * smp_load_acquire() from |bucket| above.
281 			 */
282 			smp_store_release(bucket, new);
283 			found = new;
284 		}
285 	} else if (prealloc) {
286 		/*
287 		 * We didn't need to store this stack trace, but let's keep
288 		 * the preallocated memory for the future.
289 		 */
290 		WARN_ON(!init_stack_slab(&prealloc));
291 	}
292 
293 	raw_spin_unlock_irqrestore(&depot_lock, flags);
294 exit:
295 	if (prealloc) {
296 		/* Nobody used this memory, ok to free it. */
297 		free_pages((unsigned long)prealloc, STACK_ALLOC_ORDER);
298 	}
299 	if (found)
300 		retval = found->handle.handle;
301 fast_exit:
302 	return retval;
303 }
304 EXPORT_SYMBOL_GPL(depot_save_stack);
305