1 // SPDX-License-Identifier: GPL-2.0-only
3 * Generic stack depot for storing stack traces.
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
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.
16 * Author: Alexander Potapenko <glider@google.com>
17 * Copyright (C) 2016 Google, Inc.
19 * Based on code by Dmitry Chernenkov.
22 #include <linux/gfp.h>
23 #include <linux/jhash.h>
24 #include <linux/kernel.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>
34 #define DEPOT_STACK_BITS (sizeof(depot_stack_handle_t) * 8)
36 #define STACK_ALLOC_NULL_PROTECTION_BITS 1
37 #define STACK_ALLOC_ORDER 2 /* 'Slab' size order for stack depot, 4 pages */
38 #define STACK_ALLOC_SIZE (1LL << (PAGE_SHIFT + STACK_ALLOC_ORDER))
39 #define STACK_ALLOC_ALIGN 4
40 #define STACK_ALLOC_OFFSET_BITS (STACK_ALLOC_ORDER + PAGE_SHIFT - \
42 #define STACK_ALLOC_INDEX_BITS (DEPOT_STACK_BITS - \
43 STACK_ALLOC_NULL_PROTECTION_BITS - STACK_ALLOC_OFFSET_BITS)
44 #define STACK_ALLOC_SLABS_CAP 8192
45 #define STACK_ALLOC_MAX_SLABS \
46 (((1LL << (STACK_ALLOC_INDEX_BITS)) < STACK_ALLOC_SLABS_CAP) ? \
47 (1LL << (STACK_ALLOC_INDEX_BITS)) : STACK_ALLOC_SLABS_CAP)
49 /* The compact structure to store the reference to stacks. */
51 depot_stack_handle_t handle;
53 u32 slabindex : STACK_ALLOC_INDEX_BITS;
54 u32 offset : STACK_ALLOC_OFFSET_BITS;
55 u32 valid : STACK_ALLOC_NULL_PROTECTION_BITS;
60 struct stack_record *next; /* Link in the hashtable */
61 u32 hash; /* Hash in the hastable */
62 u32 size; /* Number of frames in the stack */
63 union handle_parts handle;
64 unsigned long entries[1]; /* Variable-sized array of entries. */
67 static void *stack_slabs[STACK_ALLOC_MAX_SLABS];
69 static int depot_index;
70 static int next_slab_inited;
71 static size_t depot_offset;
72 static DEFINE_SPINLOCK(depot_lock);
74 static bool init_stack_slab(void **prealloc)
79 * This smp_load_acquire() pairs with smp_store_release() to
80 * |next_slab_inited| below and in depot_alloc_stack().
82 if (smp_load_acquire(&next_slab_inited))
84 if (stack_slabs[depot_index] == NULL) {
85 stack_slabs[depot_index] = *prealloc;
88 /* If this is the last depot slab, do not touch the next one. */
89 if (depot_index + 1 < STACK_ALLOC_MAX_SLABS) {
90 stack_slabs[depot_index + 1] = *prealloc;
94 * This smp_store_release pairs with smp_load_acquire() from
95 * |next_slab_inited| above and in stack_depot_save().
97 smp_store_release(&next_slab_inited, 1);
102 /* Allocation of a new stack in raw storage */
103 static struct stack_record *depot_alloc_stack(unsigned long *entries, int size,
104 u32 hash, void **prealloc, gfp_t alloc_flags)
106 int required_size = offsetof(struct stack_record, entries) +
107 sizeof(unsigned long) * size;
108 struct stack_record *stack;
110 required_size = ALIGN(required_size, 1 << STACK_ALLOC_ALIGN);
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");
120 * smp_store_release() here pairs with smp_load_acquire() from
121 * |next_slab_inited| in stack_depot_save() and
124 if (depot_index + 1 < STACK_ALLOC_MAX_SLABS)
125 smp_store_release(&next_slab_inited, 0);
127 init_stack_slab(prealloc);
128 if (stack_slabs[depot_index] == NULL)
131 stack = stack_slabs[depot_index] + depot_offset;
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, size * sizeof(unsigned long));
139 depot_offset += required_size;
144 #define STACK_HASH_ORDER 20
145 #define STACK_HASH_SIZE (1L << STACK_HASH_ORDER)
146 #define STACK_HASH_MASK (STACK_HASH_SIZE - 1)
147 #define STACK_HASH_SEED 0x9747b28c
149 static struct stack_record *stack_table[STACK_HASH_SIZE] = {
150 [0 ... STACK_HASH_SIZE - 1] = NULL
153 /* Calculate hash for a stack */
154 static inline u32 hash_stack(unsigned long *entries, unsigned int size)
156 return jhash2((u32 *)entries,
157 size * sizeof(unsigned long) / sizeof(u32),
161 /* Use our own, non-instrumented version of memcmp().
163 * We actually don't care about the order, just the equality.
166 int stackdepot_memcmp(const unsigned long *u1, const unsigned long *u2,
169 for ( ; n-- ; u1++, u2++) {
176 /* Find a stack that is equal to the one stored in entries in the hash */
177 static inline struct stack_record *find_stack(struct stack_record *bucket,
178 unsigned long *entries, int size,
181 struct stack_record *found;
183 for (found = bucket; found; found = found->next) {
184 if (found->hash == hash &&
185 found->size == size &&
186 !stackdepot_memcmp(entries, found->entries, size))
193 * stack_depot_fetch - Fetch stack entries from a depot
195 * @handle: Stack depot handle which was returned from
196 * stack_depot_save().
197 * @entries: Pointer to store the entries address
199 * Return: The number of trace entries for this depot.
201 unsigned int stack_depot_fetch(depot_stack_handle_t handle,
202 unsigned long **entries)
204 union handle_parts parts = { .handle = handle };
205 void *slab = stack_slabs[parts.slabindex];
206 size_t offset = parts.offset << STACK_ALLOC_ALIGN;
207 struct stack_record *stack = slab + offset;
209 *entries = stack->entries;
212 EXPORT_SYMBOL_GPL(stack_depot_fetch);
215 * stack_depot_save - Save a stack trace from an array
217 * @entries: Pointer to storage array
218 * @nr_entries: Size of the storage array
219 * @alloc_flags: Allocation gfp flags
221 * Return: The handle of the stack struct stored in depot
223 depot_stack_handle_t stack_depot_save(unsigned long *entries,
224 unsigned int nr_entries,
227 struct stack_record *found = NULL, **bucket;
228 depot_stack_handle_t retval = 0;
229 struct page *page = NULL;
230 void *prealloc = NULL;
234 if (unlikely(nr_entries == 0))
237 hash = hash_stack(entries, nr_entries);
238 bucket = &stack_table[hash & STACK_HASH_MASK];
241 * Fast path: look the stack trace up without locking.
242 * The smp_load_acquire() here pairs with smp_store_release() to
245 found = find_stack(smp_load_acquire(bucket), entries,
251 * Check if the current or the next stack slab need to be initialized.
252 * If so, allocate the memory - we won't be able to do that under the
255 * The smp_load_acquire() here pairs with smp_store_release() to
256 * |next_slab_inited| in depot_alloc_stack() and init_stack_slab().
258 if (unlikely(!smp_load_acquire(&next_slab_inited))) {
260 * Zero out zone modifiers, as we don't have specific zone
261 * requirements. Keep the flags related to allocation in atomic
264 alloc_flags &= ~GFP_ZONEMASK;
265 alloc_flags &= (GFP_ATOMIC | GFP_KERNEL);
266 alloc_flags |= __GFP_NOWARN;
267 page = alloc_pages(alloc_flags, STACK_ALLOC_ORDER);
269 prealloc = page_address(page);
272 spin_lock_irqsave(&depot_lock, flags);
274 found = find_stack(*bucket, entries, nr_entries, hash);
276 struct stack_record *new =
277 depot_alloc_stack(entries, nr_entries,
278 hash, &prealloc, alloc_flags);
282 * This smp_store_release() pairs with
283 * smp_load_acquire() from |bucket| above.
285 smp_store_release(bucket, new);
288 } else if (prealloc) {
290 * We didn't need to store this stack trace, but let's keep
291 * the preallocated memory for the future.
293 WARN_ON(!init_stack_slab(&prealloc));
296 spin_unlock_irqrestore(&depot_lock, flags);
299 /* Nobody used this memory, ok to free it. */
300 free_pages((unsigned long)prealloc, STACK_ALLOC_ORDER);
303 retval = found->handle.handle;
307 EXPORT_SYMBOL_GPL(stack_depot_save);