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 contiguous 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>
33 #include <linux/memblock.h>
35 #define DEPOT_STACK_BITS (sizeof(depot_stack_handle_t) * 8)
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 - \
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)
50 /* The compact structure to store the reference to stacks. */
52 depot_stack_handle_t handle;
54 u32 slabindex : STACK_ALLOC_INDEX_BITS;
55 u32 offset : STACK_ALLOC_OFFSET_BITS;
56 u32 valid : STACK_ALLOC_NULL_PROTECTION_BITS;
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. */
68 static void *stack_slabs[STACK_ALLOC_MAX_SLABS];
70 static int depot_index;
71 static int next_slab_inited;
72 static size_t depot_offset;
73 static DEFINE_RAW_SPINLOCK(depot_lock);
75 static bool init_stack_slab(void **prealloc)
80 * This smp_load_acquire() pairs with smp_store_release() to
81 * |next_slab_inited| below and in depot_alloc_stack().
83 if (smp_load_acquire(&next_slab_inited))
85 if (stack_slabs[depot_index] == NULL) {
86 stack_slabs[depot_index] = *prealloc;
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;
95 * This smp_store_release pairs with smp_load_acquire() from
96 * |next_slab_inited| above and in stack_depot_save().
98 smp_store_release(&next_slab_inited, 1);
103 /* Allocation of a new stack in raw storage */
104 static struct stack_record *
105 depot_alloc_stack(unsigned long *entries, int size, u32 hash, void **prealloc)
107 struct stack_record *stack;
108 size_t required_size = struct_size(stack, entries, size);
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, flex_array_size(stack, entries, size));
139 depot_offset += required_size;
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
148 static bool stack_depot_disable;
149 static struct stack_record **stack_table;
151 static int __init is_stack_depot_disabled(char *str)
155 ret = kstrtobool(str, &stack_depot_disable);
156 if (!ret && stack_depot_disable) {
157 pr_info("Stack Depot is disabled\n");
162 early_param("stack_depot_disable", is_stack_depot_disabled);
164 int __init stack_depot_init(void)
166 if (!stack_depot_disable) {
167 size_t size = (STACK_HASH_SIZE * sizeof(struct stack_record *));
170 stack_table = memblock_alloc(size, size);
171 for (i = 0; i < STACK_HASH_SIZE; i++)
172 stack_table[i] = NULL;
177 /* Calculate hash for a stack */
178 static inline u32 hash_stack(unsigned long *entries, unsigned int size)
180 return jhash2((u32 *)entries,
181 array_size(size, sizeof(*entries)) / sizeof(u32),
185 /* Use our own, non-instrumented version of memcmp().
187 * We actually don't care about the order, just the equality.
190 int stackdepot_memcmp(const unsigned long *u1, const unsigned long *u2,
193 for ( ; n-- ; u1++, u2++) {
200 /* Find a stack that is equal to the one stored in entries in the hash */
201 static inline struct stack_record *find_stack(struct stack_record *bucket,
202 unsigned long *entries, int size,
205 struct stack_record *found;
207 for (found = bucket; found; found = found->next) {
208 if (found->hash == hash &&
209 found->size == size &&
210 !stackdepot_memcmp(entries, found->entries, size))
217 * stack_depot_snprint - print stack entries from a depot into a buffer
219 * @handle: Stack depot handle which was returned from
220 * stack_depot_save().
221 * @buf: Pointer to the print buffer
223 * @size: Size of the print buffer
225 * @spaces: Number of leading spaces to print
227 * Return: Number of bytes printed.
229 int stack_depot_snprint(depot_stack_handle_t handle, char *buf, size_t size,
232 unsigned long *entries;
233 unsigned int nr_entries;
235 nr_entries = stack_depot_fetch(handle, &entries);
236 return nr_entries ? stack_trace_snprint(buf, size, entries, nr_entries,
239 EXPORT_SYMBOL_GPL(stack_depot_snprint);
242 * stack_depot_print - print stack entries from a depot
244 * @stack: Stack depot handle which was returned from
245 * stack_depot_save().
248 void stack_depot_print(depot_stack_handle_t stack)
250 unsigned long *entries;
251 unsigned int nr_entries;
253 nr_entries = stack_depot_fetch(stack, &entries);
255 stack_trace_print(entries, nr_entries, 0);
257 EXPORT_SYMBOL_GPL(stack_depot_print);
260 * stack_depot_fetch - Fetch stack entries from a depot
262 * @handle: Stack depot handle which was returned from
263 * stack_depot_save().
264 * @entries: Pointer to store the entries address
266 * Return: The number of trace entries for this depot.
268 unsigned int stack_depot_fetch(depot_stack_handle_t handle,
269 unsigned long **entries)
271 union handle_parts parts = { .handle = handle };
273 size_t offset = parts.offset << STACK_ALLOC_ALIGN;
274 struct stack_record *stack;
280 if (parts.slabindex > depot_index) {
281 WARN(1, "slab index %d out of bounds (%d) for stack id %08x\n",
282 parts.slabindex, depot_index, handle);
285 slab = stack_slabs[parts.slabindex];
288 stack = slab + offset;
290 *entries = stack->entries;
293 EXPORT_SYMBOL_GPL(stack_depot_fetch);
296 * __stack_depot_save - Save a stack trace from an array
298 * @entries: Pointer to storage array
299 * @nr_entries: Size of the storage array
300 * @alloc_flags: Allocation gfp flags
301 * @can_alloc: Allocate stack slabs (increased chance of failure if false)
303 * Saves a stack trace from @entries array of size @nr_entries. If @can_alloc is
304 * %true, is allowed to replenish the stack slab pool in case no space is left
305 * (allocates using GFP flags of @alloc_flags). If @can_alloc is %false, avoids
306 * any allocations and will fail if no space is left to store the stack trace.
308 * Context: Any context, but setting @can_alloc to %false is required if
309 * alloc_pages() cannot be used from the current context. Currently
310 * this is the case from contexts where neither %GFP_ATOMIC nor
311 * %GFP_NOWAIT can be used (NMI, raw_spin_lock).
313 * Return: The handle of the stack struct stored in depot, 0 on failure.
315 depot_stack_handle_t __stack_depot_save(unsigned long *entries,
316 unsigned int nr_entries,
317 gfp_t alloc_flags, bool can_alloc)
319 struct stack_record *found = NULL, **bucket;
320 depot_stack_handle_t retval = 0;
321 struct page *page = NULL;
322 void *prealloc = NULL;
326 if (unlikely(nr_entries == 0) || stack_depot_disable)
329 hash = hash_stack(entries, nr_entries);
330 bucket = &stack_table[hash & STACK_HASH_MASK];
333 * Fast path: look the stack trace up without locking.
334 * The smp_load_acquire() here pairs with smp_store_release() to
337 found = find_stack(smp_load_acquire(bucket), entries,
343 * Check if the current or the next stack slab need to be initialized.
344 * If so, allocate the memory - we won't be able to do that under the
347 * The smp_load_acquire() here pairs with smp_store_release() to
348 * |next_slab_inited| in depot_alloc_stack() and init_stack_slab().
350 if (unlikely(can_alloc && !smp_load_acquire(&next_slab_inited))) {
352 * Zero out zone modifiers, as we don't have specific zone
353 * requirements. Keep the flags related to allocation in atomic
356 alloc_flags &= ~GFP_ZONEMASK;
357 alloc_flags &= (GFP_ATOMIC | GFP_KERNEL);
358 alloc_flags |= __GFP_NOWARN;
359 page = alloc_pages(alloc_flags, STACK_ALLOC_ORDER);
361 prealloc = page_address(page);
364 raw_spin_lock_irqsave(&depot_lock, flags);
366 found = find_stack(*bucket, entries, nr_entries, hash);
368 struct stack_record *new = depot_alloc_stack(entries, nr_entries, hash, &prealloc);
373 * This smp_store_release() pairs with
374 * smp_load_acquire() from |bucket| above.
376 smp_store_release(bucket, new);
379 } else if (prealloc) {
381 * We didn't need to store this stack trace, but let's keep
382 * the preallocated memory for the future.
384 WARN_ON(!init_stack_slab(&prealloc));
387 raw_spin_unlock_irqrestore(&depot_lock, flags);
390 /* Nobody used this memory, ok to free it. */
391 free_pages((unsigned long)prealloc, STACK_ALLOC_ORDER);
394 retval = found->handle.handle;
398 EXPORT_SYMBOL_GPL(__stack_depot_save);
401 * stack_depot_save - Save a stack trace from an array
403 * @entries: Pointer to storage array
404 * @nr_entries: Size of the storage array
405 * @alloc_flags: Allocation gfp flags
407 * Context: Contexts where allocations via alloc_pages() are allowed.
408 * See __stack_depot_save() for more details.
410 * Return: The handle of the stack struct stored in depot, 0 on failure.
412 depot_stack_handle_t stack_depot_save(unsigned long *entries,
413 unsigned int nr_entries,
416 return __stack_depot_save(entries, nr_entries, alloc_flags, true);
418 EXPORT_SYMBOL_GPL(stack_depot_save);