1// SPDX-License-Identifier: GPL-2.0-only
  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.
 20 */
 21
 22#include <linux/gfp.h>
 23#include <linux/jhash.h>
 24#include <linux/kernel.h>
 25#include <linux/mm.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
 34#define DEPOT_STACK_BITS (sizeof(depot_stack_handle_t) * 8)
 35
 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 - \
 41					STACK_ALLOC_ALIGN)
 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)
 48
 49/* The compact structure to store the reference to stacks. */
 50union handle_parts {
 51	depot_stack_handle_t handle;
 52	struct {
 53		u32 slabindex : STACK_ALLOC_INDEX_BITS;
 54		u32 offset : STACK_ALLOC_OFFSET_BITS;
 55		u32 valid : STACK_ALLOC_NULL_PROTECTION_BITS;
 56	};
 57};
 58
 59struct stack_record {
 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. */
 65};
 66
 67static void *stack_slabs[STACK_ALLOC_MAX_SLABS];
 68
 69static int depot_index;
 70static int next_slab_inited;
 71static size_t depot_offset;
 72static DEFINE_SPINLOCK(depot_lock);
 73
 74static bool init_stack_slab(void **prealloc)
 75{
 76	if (!*prealloc)
 77		return false;
 78	/*
 79	 * This smp_load_acquire() pairs with smp_store_release() to
 80	 * |next_slab_inited| below and in depot_alloc_stack().
 81	 */
 82	if (smp_load_acquire(&next_slab_inited))
 83		return true;
 84	if (stack_slabs[depot_index] == NULL) {
 85		stack_slabs[depot_index] = *prealloc;
 86	} else {
 87		stack_slabs[depot_index + 1] = *prealloc;
 88		/*
 89		 * This smp_store_release pairs with smp_load_acquire() from
 90		 * |next_slab_inited| above and in stack_depot_save().
 91		 */
 92		smp_store_release(&next_slab_inited, 1);
 93	}
 94	*prealloc = NULL;
 95	return true;
 96}
 97
 98/* Allocation of a new stack in raw storage */
 99static struct stack_record *depot_alloc_stack(unsigned long *entries, int size,
100		u32 hash, void **prealloc, gfp_t alloc_flags)
101{
102	int required_size = offsetof(struct stack_record, entries) +
103		sizeof(unsigned long) * size;
104	struct stack_record *stack;
105
106	required_size = ALIGN(required_size, 1 << STACK_ALLOC_ALIGN);
107
108	if (unlikely(depot_offset + required_size > STACK_ALLOC_SIZE)) {
109		if (unlikely(depot_index + 1 >= STACK_ALLOC_MAX_SLABS)) {
110			WARN_ONCE(1, "Stack depot reached limit capacity");
111			return NULL;
112		}
113		depot_index++;
114		depot_offset = 0;
115		/*
116		 * smp_store_release() here pairs with smp_load_acquire() from
117		 * |next_slab_inited| in stack_depot_save() and
118		 * init_stack_slab().
119		 */
120		if (depot_index + 1 < STACK_ALLOC_MAX_SLABS)
121			smp_store_release(&next_slab_inited, 0);
122	}
123	init_stack_slab(prealloc);
124	if (stack_slabs[depot_index] == NULL)
125		return NULL;
126
127	stack = stack_slabs[depot_index] + depot_offset;
128
129	stack->hash = hash;
130	stack->size = size;
131	stack->handle.slabindex = depot_index;
132	stack->handle.offset = depot_offset >> STACK_ALLOC_ALIGN;
133	stack->handle.valid = 1;
134	memcpy(stack->entries, entries, size * sizeof(unsigned long));
135	depot_offset += required_size;
136
137	return stack;
138}
139
140#define STACK_HASH_ORDER 20
141#define STACK_HASH_SIZE (1L << STACK_HASH_ORDER)
142#define STACK_HASH_MASK (STACK_HASH_SIZE - 1)
143#define STACK_HASH_SEED 0x9747b28c
144
145static struct stack_record *stack_table[STACK_HASH_SIZE] = {
146	[0 ...	STACK_HASH_SIZE - 1] = NULL
147};
148
149/* Calculate hash for a stack */
150static inline u32 hash_stack(unsigned long *entries, unsigned int size)
151{
152	return jhash2((u32 *)entries,
153			       size * sizeof(unsigned long) / sizeof(u32),
154			       STACK_HASH_SEED);
155}
156
157/* Use our own, non-instrumented version of memcmp().
158 *
159 * We actually don't care about the order, just the equality.
160 */
161static inline
162int stackdepot_memcmp(const unsigned long *u1, const unsigned long *u2,
163			unsigned int n)
164{
165	for ( ; n-- ; u1++, u2++) {
166		if (*u1 != *u2)
167			return 1;
168	}
169	return 0;
170}
171
172/* Find a stack that is equal to the one stored in entries in the hash */
173static inline struct stack_record *find_stack(struct stack_record *bucket,
174					     unsigned long *entries, int size,
175					     u32 hash)
176{
177	struct stack_record *found;
178
179	for (found = bucket; found; found = found->next) {
180		if (found->hash == hash &&
181		    found->size == size &&
182		    !stackdepot_memcmp(entries, found->entries, size))
183			return found;
184	}
185	return NULL;
186}
187
188/**
189 * stack_depot_fetch - Fetch stack entries from a depot
190 *
191 * @handle:		Stack depot handle which was returned from
192 *			stack_depot_save().
193 * @entries:		Pointer to store the entries address
194 *
195 * Return: The number of trace entries for this depot.
196 */
197unsigned int stack_depot_fetch(depot_stack_handle_t handle,
198			       unsigned long **entries)
199{
200	union handle_parts parts = { .handle = handle };
201	void *slab = stack_slabs[parts.slabindex];
202	size_t offset = parts.offset << STACK_ALLOC_ALIGN;
203	struct stack_record *stack = slab + offset;
204
205	*entries = stack->entries;
206	return stack->size;
207}
208EXPORT_SYMBOL_GPL(stack_depot_fetch);
209
210/**
211 * stack_depot_save - Save a stack trace from an array
212 *
213 * @entries:		Pointer to storage array
214 * @nr_entries:		Size of the storage array
215 * @alloc_flags:	Allocation gfp flags
216 *
217 * Return: The handle of the stack struct stored in depot
218 */
219depot_stack_handle_t stack_depot_save(unsigned long *entries,
220				      unsigned int nr_entries,
221				      gfp_t alloc_flags)
222{
223	struct stack_record *found = NULL, **bucket;
224	depot_stack_handle_t retval = 0;
225	struct page *page = NULL;
226	void *prealloc = NULL;
227	unsigned long flags;
228	u32 hash;
229
230	if (unlikely(nr_entries == 0))
231		goto fast_exit;
232
233	hash = hash_stack(entries, nr_entries);
234	bucket = &stack_table[hash & STACK_HASH_MASK];
235
236	/*
237	 * Fast path: look the stack trace up without locking.
238	 * The smp_load_acquire() here pairs with smp_store_release() to
239	 * |bucket| below.
240	 */
241	found = find_stack(smp_load_acquire(bucket), entries,
242			   nr_entries, hash);
243	if (found)
244		goto exit;
245
246	/*
247	 * Check if the current or the next stack slab need to be initialized.
248	 * If so, allocate the memory - we won't be able to do that under the
249	 * lock.
250	 *
251	 * The smp_load_acquire() here pairs with smp_store_release() to
252	 * |next_slab_inited| in depot_alloc_stack() and init_stack_slab().
253	 */
254	if (unlikely(!smp_load_acquire(&next_slab_inited))) {
255		/*
256		 * Zero out zone modifiers, as we don't have specific zone
257		 * requirements. Keep the flags related to allocation in atomic
258		 * contexts and I/O.
259		 */
260		alloc_flags &= ~GFP_ZONEMASK;
261		alloc_flags &= (GFP_ATOMIC | GFP_KERNEL);
262		alloc_flags |= __GFP_NOWARN;
263		page = alloc_pages(alloc_flags, STACK_ALLOC_ORDER);
264		if (page)
265			prealloc = page_address(page);
266	}
267
268	spin_lock_irqsave(&depot_lock, flags);
269
270	found = find_stack(*bucket, entries, nr_entries, hash);
271	if (!found) {
272		struct stack_record *new =
273			depot_alloc_stack(entries, nr_entries,
274					  hash, &prealloc, alloc_flags);
275		if (new) {
276			new->next = *bucket;
277			/*
278			 * This smp_store_release() pairs with
279			 * smp_load_acquire() from |bucket| above.
280			 */
281			smp_store_release(bucket, new);
282			found = new;
283		}
284	} else if (prealloc) {
285		/*
286		 * We didn't need to store this stack trace, but let's keep
287		 * the preallocated memory for the future.
288		 */
289		WARN_ON(!init_stack_slab(&prealloc));
290	}
291
292	spin_unlock_irqrestore(&depot_lock, flags);
293exit:
294	if (prealloc) {
295		/* Nobody used this memory, ok to free it. */
296		free_pages((unsigned long)prealloc, STACK_ALLOC_ORDER);
297	}
298	if (found)
299		retval = found->handle.handle;
300fast_exit:
301	return retval;
302}
303EXPORT_SYMBOL_GPL(stack_depot_save);