Commit | Line | Data |
---|---|---|
1802d0be | 1 | // SPDX-License-Identifier: GPL-2.0-only |
cd11016e AP |
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 contiguos memory allocation. | |
15 | * | |
16 | * Author: Alexander Potapenko <glider@google.com> | |
17 | * Copyright (C) 2016 Google, Inc. | |
18 | * | |
19 | * Based on code by Dmitry Chernenkov. | |
cd11016e AP |
20 | */ |
21 | ||
22 | #include <linux/gfp.h> | |
505a0ef1 | 23 | #include <linux/interrupt.h> |
cd11016e AP |
24 | #include <linux/jhash.h> |
25 | #include <linux/kernel.h> | |
26 | #include <linux/mm.h> | |
27 | #include <linux/percpu.h> | |
28 | #include <linux/printk.h> | |
29 | #include <linux/slab.h> | |
30 | #include <linux/stacktrace.h> | |
31 | #include <linux/stackdepot.h> | |
32 | #include <linux/string.h> | |
33 | #include <linux/types.h> | |
34 | ||
35 | #define DEPOT_STACK_BITS (sizeof(depot_stack_handle_t) * 8) | |
36 | ||
7c31190b | 37 | #define STACK_ALLOC_NULL_PROTECTION_BITS 1 |
cd11016e AP |
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) | |
7c31190b JK |
43 | #define STACK_ALLOC_INDEX_BITS (DEPOT_STACK_BITS - \ |
44 | STACK_ALLOC_NULL_PROTECTION_BITS - STACK_ALLOC_OFFSET_BITS) | |
02754e0a | 45 | #define STACK_ALLOC_SLABS_CAP 8192 |
cd11016e AP |
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; | |
7c31190b | 56 | u32 valid : STACK_ALLOC_NULL_PROTECTION_BITS; |
cd11016e AP |
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[1]; /* 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_SPINLOCK(depot_lock); | |
74 | ||
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; | |
305e519c | 87 | *prealloc = NULL; |
cd11016e | 88 | } else { |
305e519c AP |
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 | } | |
cd11016e AP |
94 | /* |
95 | * This smp_store_release pairs with smp_load_acquire() from | |
ee050dc8 | 96 | * |next_slab_inited| above and in stack_depot_save(). |
cd11016e AP |
97 | */ |
98 | smp_store_release(&next_slab_inited, 1); | |
99 | } | |
cd11016e AP |
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, | |
105 | u32 hash, void **prealloc, gfp_t alloc_flags) | |
106 | { | |
107 | int required_size = offsetof(struct stack_record, entries) + | |
108 | sizeof(unsigned long) * size; | |
109 | struct stack_record *stack; | |
110 | ||
111 | required_size = ALIGN(required_size, 1 << STACK_ALLOC_ALIGN); | |
112 | ||
113 | if (unlikely(depot_offset + required_size > STACK_ALLOC_SIZE)) { | |
114 | if (unlikely(depot_index + 1 >= STACK_ALLOC_MAX_SLABS)) { | |
115 | WARN_ONCE(1, "Stack depot reached limit capacity"); | |
116 | return NULL; | |
117 | } | |
118 | depot_index++; | |
119 | depot_offset = 0; | |
120 | /* | |
121 | * smp_store_release() here pairs with smp_load_acquire() from | |
ee050dc8 | 122 | * |next_slab_inited| in stack_depot_save() and |
cd11016e AP |
123 | * init_stack_slab(). |
124 | */ | |
125 | if (depot_index + 1 < STACK_ALLOC_MAX_SLABS) | |
126 | smp_store_release(&next_slab_inited, 0); | |
127 | } | |
128 | init_stack_slab(prealloc); | |
129 | if (stack_slabs[depot_index] == NULL) | |
130 | return NULL; | |
131 | ||
132 | stack = stack_slabs[depot_index] + depot_offset; | |
133 | ||
134 | stack->hash = hash; | |
135 | stack->size = size; | |
136 | stack->handle.slabindex = depot_index; | |
137 | stack->handle.offset = depot_offset >> STACK_ALLOC_ALIGN; | |
7c31190b | 138 | stack->handle.valid = 1; |
cd11016e AP |
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 | ||
a571b272 AP |
162 | /* Use our own, non-instrumented version of memcmp(). |
163 | * | |
164 | * We actually don't care about the order, just the equality. | |
165 | */ | |
166 | static inline | |
167 | int stackdepot_memcmp(const unsigned long *u1, const unsigned long *u2, | |
168 | unsigned int n) | |
169 | { | |
170 | for ( ; n-- ; u1++, u2++) { | |
171 | if (*u1 != *u2) | |
172 | return 1; | |
173 | } | |
174 | return 0; | |
175 | } | |
176 | ||
cd11016e AP |
177 | /* Find a stack that is equal to the one stored in entries in the hash */ |
178 | static inline struct stack_record *find_stack(struct stack_record *bucket, | |
179 | unsigned long *entries, int size, | |
180 | u32 hash) | |
181 | { | |
182 | struct stack_record *found; | |
183 | ||
184 | for (found = bucket; found; found = found->next) { | |
185 | if (found->hash == hash && | |
186 | found->size == size && | |
a571b272 | 187 | !stackdepot_memcmp(entries, found->entries, size)) |
cd11016e | 188 | return found; |
cd11016e AP |
189 | } |
190 | return NULL; | |
191 | } | |
192 | ||
c0cfc337 TG |
193 | /** |
194 | * stack_depot_fetch - Fetch stack entries from a depot | |
195 | * | |
196 | * @handle: Stack depot handle which was returned from | |
197 | * stack_depot_save(). | |
198 | * @entries: Pointer to store the entries address | |
199 | * | |
200 | * Return: The number of trace entries for this depot. | |
201 | */ | |
202 | unsigned int stack_depot_fetch(depot_stack_handle_t handle, | |
203 | unsigned long **entries) | |
cd11016e AP |
204 | { |
205 | union handle_parts parts = { .handle = handle }; | |
69866e15 | 206 | void *slab; |
cd11016e | 207 | size_t offset = parts.offset << STACK_ALLOC_ALIGN; |
69866e15 AP |
208 | struct stack_record *stack; |
209 | ||
210 | *entries = NULL; | |
211 | if (parts.slabindex > depot_index) { | |
212 | WARN(1, "slab index %d out of bounds (%d) for stack id %08x\n", | |
213 | parts.slabindex, depot_index, handle); | |
214 | return 0; | |
215 | } | |
216 | slab = stack_slabs[parts.slabindex]; | |
217 | if (!slab) | |
218 | return 0; | |
219 | stack = slab + offset; | |
cd11016e | 220 | |
c0cfc337 TG |
221 | *entries = stack->entries; |
222 | return stack->size; | |
223 | } | |
224 | EXPORT_SYMBOL_GPL(stack_depot_fetch); | |
225 | ||
cd11016e | 226 | /** |
c0cfc337 TG |
227 | * stack_depot_save - Save a stack trace from an array |
228 | * | |
229 | * @entries: Pointer to storage array | |
230 | * @nr_entries: Size of the storage array | |
231 | * @alloc_flags: Allocation gfp flags | |
cd11016e | 232 | * |
c0cfc337 | 233 | * Return: The handle of the stack struct stored in depot |
cd11016e | 234 | */ |
c0cfc337 TG |
235 | depot_stack_handle_t stack_depot_save(unsigned long *entries, |
236 | unsigned int nr_entries, | |
237 | gfp_t alloc_flags) | |
cd11016e | 238 | { |
cd11016e | 239 | struct stack_record *found = NULL, **bucket; |
c0cfc337 | 240 | depot_stack_handle_t retval = 0; |
cd11016e AP |
241 | struct page *page = NULL; |
242 | void *prealloc = NULL; | |
c0cfc337 TG |
243 | unsigned long flags; |
244 | u32 hash; | |
cd11016e | 245 | |
c0cfc337 | 246 | if (unlikely(nr_entries == 0)) |
cd11016e AP |
247 | goto fast_exit; |
248 | ||
c0cfc337 | 249 | hash = hash_stack(entries, nr_entries); |
cd11016e AP |
250 | bucket = &stack_table[hash & STACK_HASH_MASK]; |
251 | ||
252 | /* | |
253 | * Fast path: look the stack trace up without locking. | |
254 | * The smp_load_acquire() here pairs with smp_store_release() to | |
255 | * |bucket| below. | |
256 | */ | |
c0cfc337 TG |
257 | found = find_stack(smp_load_acquire(bucket), entries, |
258 | nr_entries, hash); | |
cd11016e AP |
259 | if (found) |
260 | goto exit; | |
261 | ||
262 | /* | |
263 | * Check if the current or the next stack slab need to be initialized. | |
264 | * If so, allocate the memory - we won't be able to do that under the | |
265 | * lock. | |
266 | * | |
267 | * The smp_load_acquire() here pairs with smp_store_release() to | |
268 | * |next_slab_inited| in depot_alloc_stack() and init_stack_slab(). | |
269 | */ | |
270 | if (unlikely(!smp_load_acquire(&next_slab_inited))) { | |
271 | /* | |
272 | * Zero out zone modifiers, as we don't have specific zone | |
273 | * requirements. Keep the flags related to allocation in atomic | |
274 | * contexts and I/O. | |
275 | */ | |
276 | alloc_flags &= ~GFP_ZONEMASK; | |
277 | alloc_flags &= (GFP_ATOMIC | GFP_KERNEL); | |
87cc271d | 278 | alloc_flags |= __GFP_NOWARN; |
cd11016e AP |
279 | page = alloc_pages(alloc_flags, STACK_ALLOC_ORDER); |
280 | if (page) | |
281 | prealloc = page_address(page); | |
282 | } | |
283 | ||
284 | spin_lock_irqsave(&depot_lock, flags); | |
285 | ||
c0cfc337 | 286 | found = find_stack(*bucket, entries, nr_entries, hash); |
cd11016e AP |
287 | if (!found) { |
288 | struct stack_record *new = | |
c0cfc337 | 289 | depot_alloc_stack(entries, nr_entries, |
cd11016e AP |
290 | hash, &prealloc, alloc_flags); |
291 | if (new) { | |
292 | new->next = *bucket; | |
293 | /* | |
294 | * This smp_store_release() pairs with | |
295 | * smp_load_acquire() from |bucket| above. | |
296 | */ | |
297 | smp_store_release(bucket, new); | |
298 | found = new; | |
299 | } | |
300 | } else if (prealloc) { | |
301 | /* | |
302 | * We didn't need to store this stack trace, but let's keep | |
303 | * the preallocated memory for the future. | |
304 | */ | |
305 | WARN_ON(!init_stack_slab(&prealloc)); | |
306 | } | |
307 | ||
308 | spin_unlock_irqrestore(&depot_lock, flags); | |
309 | exit: | |
310 | if (prealloc) { | |
311 | /* Nobody used this memory, ok to free it. */ | |
312 | free_pages((unsigned long)prealloc, STACK_ALLOC_ORDER); | |
313 | } | |
314 | if (found) | |
315 | retval = found->handle.handle; | |
316 | fast_exit: | |
317 | return retval; | |
318 | } | |
c0cfc337 | 319 | EXPORT_SYMBOL_GPL(stack_depot_save); |
505a0ef1 AP |
320 | |
321 | static inline int in_irqentry_text(unsigned long ptr) | |
322 | { | |
323 | return (ptr >= (unsigned long)&__irqentry_text_start && | |
324 | ptr < (unsigned long)&__irqentry_text_end) || | |
325 | (ptr >= (unsigned long)&__softirqentry_text_start && | |
326 | ptr < (unsigned long)&__softirqentry_text_end); | |
327 | } | |
328 | ||
329 | unsigned int filter_irq_stacks(unsigned long *entries, | |
330 | unsigned int nr_entries) | |
331 | { | |
332 | unsigned int i; | |
333 | ||
334 | for (i = 0; i < nr_entries; i++) { | |
335 | if (in_irqentry_text(entries[i])) { | |
336 | /* Include the irqentry function into the stack. */ | |
337 | return i + 1; | |
338 | } | |
339 | } | |
340 | return nr_entries; | |
341 | } | |
342 | EXPORT_SYMBOL_GPL(filter_irq_stacks); |