4 * Copyright (C) 2008 ARM Limited
5 * Written by Catalin Marinas <catalin.marinas@arm.com>
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
21 * For more information on the algorithm and kmemleak usage, please see
22 * Documentation/kmemleak.txt.
27 * The following locks and mutexes are used by kmemleak:
29 * - kmemleak_lock (rwlock): protects the object_list modifications and
30 * accesses to the object_tree_root. The object_list is the main list
31 * holding the metadata (struct kmemleak_object) for the allocated memory
32 * blocks. The object_tree_root is a priority search tree used to look-up
33 * metadata based on a pointer to the corresponding memory block. The
34 * kmemleak_object structures are added to the object_list and
35 * object_tree_root in the create_object() function called from the
36 * kmemleak_alloc() callback and removed in delete_object() called from the
37 * kmemleak_free() callback
38 * - kmemleak_object.lock (spinlock): protects a kmemleak_object. Accesses to
39 * the metadata (e.g. count) are protected by this lock. Note that some
40 * members of this structure may be protected by other means (atomic or
41 * kmemleak_lock). This lock is also held when scanning the corresponding
42 * memory block to avoid the kernel freeing it via the kmemleak_free()
43 * callback. This is less heavyweight than holding a global lock like
44 * kmemleak_lock during scanning
45 * - scan_mutex (mutex): ensures that only one thread may scan the memory for
46 * unreferenced objects at a time. The gray_list contains the objects which
47 * are already referenced or marked as false positives and need to be
48 * scanned. This list is only modified during a scanning episode when the
49 * scan_mutex is held. At the end of a scan, the gray_list is always empty.
50 * Note that the kmemleak_object.use_count is incremented when an object is
51 * added to the gray_list and therefore cannot be freed. This mutex also
52 * prevents multiple users of the "kmemleak" debugfs file together with
53 * modifications to the memory scanning parameters including the scan_thread
56 * The kmemleak_object structures have a use_count incremented or decremented
57 * using the get_object()/put_object() functions. When the use_count becomes
58 * 0, this count can no longer be incremented and put_object() schedules the
59 * kmemleak_object freeing via an RCU callback. All calls to the get_object()
60 * function must be protected by rcu_read_lock() to avoid accessing a freed
64 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
66 #include <linux/init.h>
67 #include <linux/kernel.h>
68 #include <linux/list.h>
69 #include <linux/sched.h>
70 #include <linux/jiffies.h>
71 #include <linux/delay.h>
72 #include <linux/module.h>
73 #include <linux/kthread.h>
74 #include <linux/prio_tree.h>
75 #include <linux/gfp.h>
77 #include <linux/debugfs.h>
78 #include <linux/seq_file.h>
79 #include <linux/cpumask.h>
80 #include <linux/spinlock.h>
81 #include <linux/mutex.h>
82 #include <linux/rcupdate.h>
83 #include <linux/stacktrace.h>
84 #include <linux/cache.h>
85 #include <linux/percpu.h>
86 #include <linux/hardirq.h>
87 #include <linux/mmzone.h>
88 #include <linux/slab.h>
89 #include <linux/thread_info.h>
90 #include <linux/err.h>
91 #include <linux/uaccess.h>
92 #include <linux/string.h>
93 #include <linux/nodemask.h>
95 #include <linux/workqueue.h>
96 #include <linux/crc32.h>
98 #include <asm/sections.h>
99 #include <asm/processor.h>
100 #include <asm/atomic.h>
102 #include <linux/kmemcheck.h>
103 #include <linux/kmemleak.h>
106 * Kmemleak configuration and common defines.
108 #define MAX_TRACE 16 /* stack trace length */
109 #define MSECS_MIN_AGE 5000 /* minimum object age for reporting */
110 #define SECS_FIRST_SCAN 60 /* delay before the first scan */
111 #define SECS_SCAN_WAIT 600 /* subsequent auto scanning delay */
112 #define MAX_SCAN_SIZE 4096 /* maximum size of a scanned block */
114 #define BYTES_PER_POINTER sizeof(void *)
116 /* GFP bitmask for kmemleak internal allocations */
117 #define GFP_KMEMLEAK_MASK (GFP_KERNEL | GFP_ATOMIC)
119 /* scanning area inside a memory block */
120 struct kmemleak_scan_area {
121 struct hlist_node node;
126 #define KMEMLEAK_GREY 0
127 #define KMEMLEAK_BLACK -1
130 * Structure holding the metadata for each allocated memory block.
131 * Modifications to such objects should be made while holding the
132 * object->lock. Insertions or deletions from object_list, gray_list or
133 * tree_node are already protected by the corresponding locks or mutex (see
134 * the notes on locking above). These objects are reference-counted
135 * (use_count) and freed using the RCU mechanism.
137 struct kmemleak_object {
139 unsigned long flags; /* object status flags */
140 struct list_head object_list;
141 struct list_head gray_list;
142 struct prio_tree_node tree_node;
143 struct rcu_head rcu; /* object_list lockless traversal */
144 /* object usage count; object freed when use_count == 0 */
146 unsigned long pointer;
148 /* minimum number of a pointers found before it is considered leak */
150 /* the total number of pointers found pointing to this object */
152 /* checksum for detecting modified objects */
154 /* memory ranges to be scanned inside an object (empty for all) */
155 struct hlist_head area_list;
156 unsigned long trace[MAX_TRACE];
157 unsigned int trace_len;
158 unsigned long jiffies; /* creation timestamp */
159 pid_t pid; /* pid of the current task */
160 char comm[TASK_COMM_LEN]; /* executable name */
163 /* flag representing the memory block allocation status */
164 #define OBJECT_ALLOCATED (1 << 0)
165 /* flag set after the first reporting of an unreference object */
166 #define OBJECT_REPORTED (1 << 1)
167 /* flag set to not scan the object */
168 #define OBJECT_NO_SCAN (1 << 2)
170 /* number of bytes to print per line; must be 16 or 32 */
171 #define HEX_ROW_SIZE 16
172 /* number of bytes to print at a time (1, 2, 4, 8) */
173 #define HEX_GROUP_SIZE 1
174 /* include ASCII after the hex output */
176 /* max number of lines to be printed */
177 #define HEX_MAX_LINES 2
179 /* the list of all allocated objects */
180 static LIST_HEAD(object_list);
181 /* the list of gray-colored objects (see color_gray comment below) */
182 static LIST_HEAD(gray_list);
183 /* prio search tree for object boundaries */
184 static struct prio_tree_root object_tree_root;
185 /* rw_lock protecting the access to object_list and prio_tree_root */
186 static DEFINE_RWLOCK(kmemleak_lock);
188 /* allocation caches for kmemleak internal data */
189 static struct kmem_cache *object_cache;
190 static struct kmem_cache *scan_area_cache;
192 /* set if tracing memory operations is enabled */
193 static atomic_t kmemleak_enabled = ATOMIC_INIT(0);
194 /* set in the late_initcall if there were no errors */
195 static atomic_t kmemleak_initialized = ATOMIC_INIT(0);
196 /* enables or disables early logging of the memory operations */
197 static atomic_t kmemleak_early_log = ATOMIC_INIT(1);
198 /* set if a fata kmemleak error has occurred */
199 static atomic_t kmemleak_error = ATOMIC_INIT(0);
201 /* minimum and maximum address that may be valid pointers */
202 static unsigned long min_addr = ULONG_MAX;
203 static unsigned long max_addr;
205 static struct task_struct *scan_thread;
206 /* used to avoid reporting of recently allocated objects */
207 static unsigned long jiffies_min_age;
208 static unsigned long jiffies_last_scan;
209 /* delay between automatic memory scannings */
210 static signed long jiffies_scan_wait;
211 /* enables or disables the task stacks scanning */
212 static int kmemleak_stack_scan = 1;
213 /* protects the memory scanning, parameters and debug/kmemleak file access */
214 static DEFINE_MUTEX(scan_mutex);
217 * Early object allocation/freeing logging. Kmemleak is initialized after the
218 * kernel allocator. However, both the kernel allocator and kmemleak may
219 * allocate memory blocks which need to be tracked. Kmemleak defines an
220 * arbitrary buffer to hold the allocation/freeing information before it is
224 /* kmemleak operation type for early logging */
236 * Structure holding the information passed to kmemleak callbacks during the
240 int op_type; /* kmemleak operation type */
241 const void *ptr; /* allocated/freed memory block */
242 size_t size; /* memory block size */
243 int min_count; /* minimum reference count */
244 unsigned long trace[MAX_TRACE]; /* stack trace */
245 unsigned int trace_len; /* stack trace length */
248 /* early logging buffer and current position */
249 static struct early_log
250 early_log[CONFIG_DEBUG_KMEMLEAK_EARLY_LOG_SIZE] __initdata;
251 static int crt_early_log __initdata;
253 static void kmemleak_disable(void);
256 * Print a warning and dump the stack trace.
258 #define kmemleak_warn(x...) do { \
264 * Macro invoked when a serious kmemleak condition occured and cannot be
265 * recovered from. Kmemleak will be disabled and further allocation/freeing
266 * tracing no longer available.
268 #define kmemleak_stop(x...) do { \
270 kmemleak_disable(); \
274 * Printing of the objects hex dump to the seq file. The number of lines to be
275 * printed is limited to HEX_MAX_LINES to prevent seq file spamming. The
276 * actual number of printed bytes depends on HEX_ROW_SIZE. It must be called
277 * with the object->lock held.
279 static void hex_dump_object(struct seq_file *seq,
280 struct kmemleak_object *object)
282 const u8 *ptr = (const u8 *)object->pointer;
283 int i, len, remaining;
284 unsigned char linebuf[HEX_ROW_SIZE * 5];
286 /* limit the number of lines to HEX_MAX_LINES */
288 min(object->size, (size_t)(HEX_MAX_LINES * HEX_ROW_SIZE));
290 seq_printf(seq, " hex dump (first %d bytes):\n", len);
291 for (i = 0; i < len; i += HEX_ROW_SIZE) {
292 int linelen = min(remaining, HEX_ROW_SIZE);
294 remaining -= HEX_ROW_SIZE;
295 hex_dump_to_buffer(ptr + i, linelen, HEX_ROW_SIZE,
296 HEX_GROUP_SIZE, linebuf, sizeof(linebuf),
298 seq_printf(seq, " %s\n", linebuf);
303 * Object colors, encoded with count and min_count:
304 * - white - orphan object, not enough references to it (count < min_count)
305 * - gray - not orphan, not marked as false positive (min_count == 0) or
306 * sufficient references to it (count >= min_count)
307 * - black - ignore, it doesn't contain references (e.g. text section)
308 * (min_count == -1). No function defined for this color.
309 * Newly created objects don't have any color assigned (object->count == -1)
310 * before the next memory scan when they become white.
312 static bool color_white(const struct kmemleak_object *object)
314 return object->count != KMEMLEAK_BLACK &&
315 object->count < object->min_count;
318 static bool color_gray(const struct kmemleak_object *object)
320 return object->min_count != KMEMLEAK_BLACK &&
321 object->count >= object->min_count;
325 * Objects are considered unreferenced only if their color is white, they have
326 * not be deleted and have a minimum age to avoid false positives caused by
327 * pointers temporarily stored in CPU registers.
329 static bool unreferenced_object(struct kmemleak_object *object)
331 return (color_white(object) && object->flags & OBJECT_ALLOCATED) &&
332 time_before_eq(object->jiffies + jiffies_min_age,
337 * Printing of the unreferenced objects information to the seq file. The
338 * print_unreferenced function must be called with the object->lock held.
340 static void print_unreferenced(struct seq_file *seq,
341 struct kmemleak_object *object)
344 unsigned int msecs_age = jiffies_to_msecs(jiffies - object->jiffies);
346 seq_printf(seq, "unreferenced object 0x%08lx (size %zu):\n",
347 object->pointer, object->size);
348 seq_printf(seq, " comm \"%s\", pid %d, jiffies %lu (age %d.%03ds)\n",
349 object->comm, object->pid, object->jiffies,
350 msecs_age / 1000, msecs_age % 1000);
351 hex_dump_object(seq, object);
352 seq_printf(seq, " backtrace:\n");
354 for (i = 0; i < object->trace_len; i++) {
355 void *ptr = (void *)object->trace[i];
356 seq_printf(seq, " [<%p>] %pS\n", ptr, ptr);
361 * Print the kmemleak_object information. This function is used mainly for
362 * debugging special cases when kmemleak operations. It must be called with
363 * the object->lock held.
365 static void dump_object_info(struct kmemleak_object *object)
367 struct stack_trace trace;
369 trace.nr_entries = object->trace_len;
370 trace.entries = object->trace;
372 pr_notice("Object 0x%08lx (size %zu):\n",
373 object->tree_node.start, object->size);
374 pr_notice(" comm \"%s\", pid %d, jiffies %lu\n",
375 object->comm, object->pid, object->jiffies);
376 pr_notice(" min_count = %d\n", object->min_count);
377 pr_notice(" count = %d\n", object->count);
378 pr_notice(" flags = 0x%lx\n", object->flags);
379 pr_notice(" checksum = %d\n", object->checksum);
380 pr_notice(" backtrace:\n");
381 print_stack_trace(&trace, 4);
385 * Look-up a memory block metadata (kmemleak_object) in the priority search
386 * tree based on a pointer value. If alias is 0, only values pointing to the
387 * beginning of the memory block are allowed. The kmemleak_lock must be held
388 * when calling this function.
390 static struct kmemleak_object *lookup_object(unsigned long ptr, int alias)
392 struct prio_tree_node *node;
393 struct prio_tree_iter iter;
394 struct kmemleak_object *object;
396 prio_tree_iter_init(&iter, &object_tree_root, ptr, ptr);
397 node = prio_tree_next(&iter);
399 object = prio_tree_entry(node, struct kmemleak_object,
401 if (!alias && object->pointer != ptr) {
402 kmemleak_warn("Found object by alias");
412 * Increment the object use_count. Return 1 if successful or 0 otherwise. Note
413 * that once an object's use_count reached 0, the RCU freeing was already
414 * registered and the object should no longer be used. This function must be
415 * called under the protection of rcu_read_lock().
417 static int get_object(struct kmemleak_object *object)
419 return atomic_inc_not_zero(&object->use_count);
423 * RCU callback to free a kmemleak_object.
425 static void free_object_rcu(struct rcu_head *rcu)
427 struct hlist_node *elem, *tmp;
428 struct kmemleak_scan_area *area;
429 struct kmemleak_object *object =
430 container_of(rcu, struct kmemleak_object, rcu);
433 * Once use_count is 0 (guaranteed by put_object), there is no other
434 * code accessing this object, hence no need for locking.
436 hlist_for_each_entry_safe(area, elem, tmp, &object->area_list, node) {
438 kmem_cache_free(scan_area_cache, area);
440 kmem_cache_free(object_cache, object);
444 * Decrement the object use_count. Once the count is 0, free the object using
445 * an RCU callback. Since put_object() may be called via the kmemleak_free() ->
446 * delete_object() path, the delayed RCU freeing ensures that there is no
447 * recursive call to the kernel allocator. Lock-less RCU object_list traversal
450 static void put_object(struct kmemleak_object *object)
452 if (!atomic_dec_and_test(&object->use_count))
455 /* should only get here after delete_object was called */
456 WARN_ON(object->flags & OBJECT_ALLOCATED);
458 call_rcu(&object->rcu, free_object_rcu);
462 * Look up an object in the prio search tree and increase its use_count.
464 static struct kmemleak_object *find_and_get_object(unsigned long ptr, int alias)
467 struct kmemleak_object *object = NULL;
470 read_lock_irqsave(&kmemleak_lock, flags);
471 if (ptr >= min_addr && ptr < max_addr)
472 object = lookup_object(ptr, alias);
473 read_unlock_irqrestore(&kmemleak_lock, flags);
475 /* check whether the object is still available */
476 if (object && !get_object(object))
484 * Save stack trace to the given array of MAX_TRACE size.
486 static int __save_stack_trace(unsigned long *trace)
488 struct stack_trace stack_trace;
490 stack_trace.max_entries = MAX_TRACE;
491 stack_trace.nr_entries = 0;
492 stack_trace.entries = trace;
493 stack_trace.skip = 2;
494 save_stack_trace(&stack_trace);
496 return stack_trace.nr_entries;
500 * Create the metadata (struct kmemleak_object) corresponding to an allocated
501 * memory block and add it to the object_list and object_tree_root.
503 static struct kmemleak_object *create_object(unsigned long ptr, size_t size,
504 int min_count, gfp_t gfp)
507 struct kmemleak_object *object;
508 struct prio_tree_node *node;
510 object = kmem_cache_alloc(object_cache, gfp & GFP_KMEMLEAK_MASK);
512 kmemleak_stop("Cannot allocate a kmemleak_object structure\n");
516 INIT_LIST_HEAD(&object->object_list);
517 INIT_LIST_HEAD(&object->gray_list);
518 INIT_HLIST_HEAD(&object->area_list);
519 spin_lock_init(&object->lock);
520 atomic_set(&object->use_count, 1);
521 object->flags = OBJECT_ALLOCATED;
522 object->pointer = ptr;
524 object->min_count = min_count;
525 object->count = 0; /* white color initially */
526 object->jiffies = jiffies;
527 object->checksum = 0;
529 /* task information */
532 strncpy(object->comm, "hardirq", sizeof(object->comm));
533 } else if (in_softirq()) {
535 strncpy(object->comm, "softirq", sizeof(object->comm));
537 object->pid = current->pid;
539 * There is a small chance of a race with set_task_comm(),
540 * however using get_task_comm() here may cause locking
541 * dependency issues with current->alloc_lock. In the worst
542 * case, the command line is not correct.
544 strncpy(object->comm, current->comm, sizeof(object->comm));
547 /* kernel backtrace */
548 object->trace_len = __save_stack_trace(object->trace);
550 INIT_PRIO_TREE_NODE(&object->tree_node);
551 object->tree_node.start = ptr;
552 object->tree_node.last = ptr + size - 1;
554 write_lock_irqsave(&kmemleak_lock, flags);
556 min_addr = min(min_addr, ptr);
557 max_addr = max(max_addr, ptr + size);
558 node = prio_tree_insert(&object_tree_root, &object->tree_node);
560 * The code calling the kernel does not yet have the pointer to the
561 * memory block to be able to free it. However, we still hold the
562 * kmemleak_lock here in case parts of the kernel started freeing
563 * random memory blocks.
565 if (node != &object->tree_node) {
566 kmemleak_stop("Cannot insert 0x%lx into the object search tree "
567 "(already existing)\n", ptr);
568 object = lookup_object(ptr, 1);
569 spin_lock(&object->lock);
570 dump_object_info(object);
571 spin_unlock(&object->lock);
575 list_add_tail_rcu(&object->object_list, &object_list);
577 write_unlock_irqrestore(&kmemleak_lock, flags);
582 * Remove the metadata (struct kmemleak_object) for a memory block from the
583 * object_list and object_tree_root and decrement its use_count.
585 static void __delete_object(struct kmemleak_object *object)
589 write_lock_irqsave(&kmemleak_lock, flags);
590 prio_tree_remove(&object_tree_root, &object->tree_node);
591 list_del_rcu(&object->object_list);
592 write_unlock_irqrestore(&kmemleak_lock, flags);
594 WARN_ON(!(object->flags & OBJECT_ALLOCATED));
595 WARN_ON(atomic_read(&object->use_count) < 2);
598 * Locking here also ensures that the corresponding memory block
599 * cannot be freed when it is being scanned.
601 spin_lock_irqsave(&object->lock, flags);
602 object->flags &= ~OBJECT_ALLOCATED;
603 spin_unlock_irqrestore(&object->lock, flags);
608 * Look up the metadata (struct kmemleak_object) corresponding to ptr and
611 static void delete_object_full(unsigned long ptr)
613 struct kmemleak_object *object;
615 object = find_and_get_object(ptr, 0);
618 kmemleak_warn("Freeing unknown object at 0x%08lx\n",
623 __delete_object(object);
628 * Look up the metadata (struct kmemleak_object) corresponding to ptr and
629 * delete it. If the memory block is partially freed, the function may create
630 * additional metadata for the remaining parts of the block.
632 static void delete_object_part(unsigned long ptr, size_t size)
634 struct kmemleak_object *object;
635 unsigned long start, end;
637 object = find_and_get_object(ptr, 1);
640 kmemleak_warn("Partially freeing unknown object at 0x%08lx "
641 "(size %zu)\n", ptr, size);
645 __delete_object(object);
648 * Create one or two objects that may result from the memory block
649 * split. Note that partial freeing is only done by free_bootmem() and
650 * this happens before kmemleak_init() is called. The path below is
651 * only executed during early log recording in kmemleak_init(), so
652 * GFP_KERNEL is enough.
654 start = object->pointer;
655 end = object->pointer + object->size;
657 create_object(start, ptr - start, object->min_count,
659 if (ptr + size < end)
660 create_object(ptr + size, end - ptr - size, object->min_count,
666 static void __paint_it(struct kmemleak_object *object, int color)
668 object->min_count = color;
669 if (color == KMEMLEAK_BLACK)
670 object->flags |= OBJECT_NO_SCAN;
673 static void paint_it(struct kmemleak_object *object, int color)
677 spin_lock_irqsave(&object->lock, flags);
678 __paint_it(object, color);
679 spin_unlock_irqrestore(&object->lock, flags);
682 static void paint_ptr(unsigned long ptr, int color)
684 struct kmemleak_object *object;
686 object = find_and_get_object(ptr, 0);
688 kmemleak_warn("Trying to color unknown object "
689 "at 0x%08lx as %s\n", ptr,
690 (color == KMEMLEAK_GREY) ? "Grey" :
691 (color == KMEMLEAK_BLACK) ? "Black" : "Unknown");
694 paint_it(object, color);
699 * Make a object permanently as gray-colored so that it can no longer be
700 * reported as a leak. This is used in general to mark a false positive.
702 static void make_gray_object(unsigned long ptr)
704 paint_ptr(ptr, KMEMLEAK_GREY);
708 * Mark the object as black-colored so that it is ignored from scans and
711 static void make_black_object(unsigned long ptr)
713 paint_ptr(ptr, KMEMLEAK_BLACK);
717 * Add a scanning area to the object. If at least one such area is added,
718 * kmemleak will only scan these ranges rather than the whole memory block.
720 static void add_scan_area(unsigned long ptr, size_t size, gfp_t gfp)
723 struct kmemleak_object *object;
724 struct kmemleak_scan_area *area;
726 object = find_and_get_object(ptr, 1);
728 kmemleak_warn("Adding scan area to unknown object at 0x%08lx\n",
733 area = kmem_cache_alloc(scan_area_cache, gfp & GFP_KMEMLEAK_MASK);
735 kmemleak_warn("Cannot allocate a scan area\n");
739 spin_lock_irqsave(&object->lock, flags);
740 if (ptr + size > object->pointer + object->size) {
741 kmemleak_warn("Scan area larger than object 0x%08lx\n", ptr);
742 dump_object_info(object);
743 kmem_cache_free(scan_area_cache, area);
747 INIT_HLIST_NODE(&area->node);
751 hlist_add_head(&area->node, &object->area_list);
753 spin_unlock_irqrestore(&object->lock, flags);
759 * Set the OBJECT_NO_SCAN flag for the object corresponding to the give
760 * pointer. Such object will not be scanned by kmemleak but references to it
763 static void object_no_scan(unsigned long ptr)
766 struct kmemleak_object *object;
768 object = find_and_get_object(ptr, 0);
770 kmemleak_warn("Not scanning unknown object at 0x%08lx\n", ptr);
774 spin_lock_irqsave(&object->lock, flags);
775 object->flags |= OBJECT_NO_SCAN;
776 spin_unlock_irqrestore(&object->lock, flags);
781 * Log an early kmemleak_* call to the early_log buffer. These calls will be
782 * processed later once kmemleak is fully initialized.
784 static void __init log_early(int op_type, const void *ptr, size_t size,
788 struct early_log *log;
790 if (crt_early_log >= ARRAY_SIZE(early_log)) {
791 pr_warning("Early log buffer exceeded, "
792 "please increase DEBUG_KMEMLEAK_EARLY_LOG_SIZE\n");
798 * There is no need for locking since the kernel is still in UP mode
799 * at this stage. Disabling the IRQs is enough.
801 local_irq_save(flags);
802 log = &early_log[crt_early_log];
803 log->op_type = op_type;
806 log->min_count = min_count;
807 if (op_type == KMEMLEAK_ALLOC)
808 log->trace_len = __save_stack_trace(log->trace);
810 local_irq_restore(flags);
814 * Log an early allocated block and populate the stack trace.
816 static void early_alloc(struct early_log *log)
818 struct kmemleak_object *object;
822 if (!atomic_read(&kmemleak_enabled) || !log->ptr || IS_ERR(log->ptr))
826 * RCU locking needed to ensure object is not freed via put_object().
829 object = create_object((unsigned long)log->ptr, log->size,
830 log->min_count, GFP_ATOMIC);
833 spin_lock_irqsave(&object->lock, flags);
834 for (i = 0; i < log->trace_len; i++)
835 object->trace[i] = log->trace[i];
836 object->trace_len = log->trace_len;
837 spin_unlock_irqrestore(&object->lock, flags);
843 * Memory allocation function callback. This function is called from the
844 * kernel allocators when a new block is allocated (kmem_cache_alloc, kmalloc,
847 void __ref kmemleak_alloc(const void *ptr, size_t size, int min_count,
850 pr_debug("%s(0x%p, %zu, %d)\n", __func__, ptr, size, min_count);
852 if (atomic_read(&kmemleak_enabled) && ptr && !IS_ERR(ptr))
853 create_object((unsigned long)ptr, size, min_count, gfp);
854 else if (atomic_read(&kmemleak_early_log))
855 log_early(KMEMLEAK_ALLOC, ptr, size, min_count);
857 EXPORT_SYMBOL_GPL(kmemleak_alloc);
860 * Memory freeing function callback. This function is called from the kernel
861 * allocators when a block is freed (kmem_cache_free, kfree, vfree etc.).
863 void __ref kmemleak_free(const void *ptr)
865 pr_debug("%s(0x%p)\n", __func__, ptr);
867 if (atomic_read(&kmemleak_enabled) && ptr && !IS_ERR(ptr))
868 delete_object_full((unsigned long)ptr);
869 else if (atomic_read(&kmemleak_early_log))
870 log_early(KMEMLEAK_FREE, ptr, 0, 0);
872 EXPORT_SYMBOL_GPL(kmemleak_free);
875 * Partial memory freeing function callback. This function is usually called
876 * from bootmem allocator when (part of) a memory block is freed.
878 void __ref kmemleak_free_part(const void *ptr, size_t size)
880 pr_debug("%s(0x%p)\n", __func__, ptr);
882 if (atomic_read(&kmemleak_enabled) && ptr && !IS_ERR(ptr))
883 delete_object_part((unsigned long)ptr, size);
884 else if (atomic_read(&kmemleak_early_log))
885 log_early(KMEMLEAK_FREE_PART, ptr, size, 0);
887 EXPORT_SYMBOL_GPL(kmemleak_free_part);
890 * Mark an already allocated memory block as a false positive. This will cause
891 * the block to no longer be reported as leak and always be scanned.
893 void __ref kmemleak_not_leak(const void *ptr)
895 pr_debug("%s(0x%p)\n", __func__, ptr);
897 if (atomic_read(&kmemleak_enabled) && ptr && !IS_ERR(ptr))
898 make_gray_object((unsigned long)ptr);
899 else if (atomic_read(&kmemleak_early_log))
900 log_early(KMEMLEAK_NOT_LEAK, ptr, 0, 0);
902 EXPORT_SYMBOL(kmemleak_not_leak);
905 * Ignore a memory block. This is usually done when it is known that the
906 * corresponding block is not a leak and does not contain any references to
907 * other allocated memory blocks.
909 void __ref kmemleak_ignore(const void *ptr)
911 pr_debug("%s(0x%p)\n", __func__, ptr);
913 if (atomic_read(&kmemleak_enabled) && ptr && !IS_ERR(ptr))
914 make_black_object((unsigned long)ptr);
915 else if (atomic_read(&kmemleak_early_log))
916 log_early(KMEMLEAK_IGNORE, ptr, 0, 0);
918 EXPORT_SYMBOL(kmemleak_ignore);
921 * Limit the range to be scanned in an allocated memory block.
923 void __ref kmemleak_scan_area(const void *ptr, size_t size, gfp_t gfp)
925 pr_debug("%s(0x%p)\n", __func__, ptr);
927 if (atomic_read(&kmemleak_enabled) && ptr && !IS_ERR(ptr))
928 add_scan_area((unsigned long)ptr, size, gfp);
929 else if (atomic_read(&kmemleak_early_log))
930 log_early(KMEMLEAK_SCAN_AREA, ptr, size, 0);
932 EXPORT_SYMBOL(kmemleak_scan_area);
935 * Inform kmemleak not to scan the given memory block.
937 void __ref kmemleak_no_scan(const void *ptr)
939 pr_debug("%s(0x%p)\n", __func__, ptr);
941 if (atomic_read(&kmemleak_enabled) && ptr && !IS_ERR(ptr))
942 object_no_scan((unsigned long)ptr);
943 else if (atomic_read(&kmemleak_early_log))
944 log_early(KMEMLEAK_NO_SCAN, ptr, 0, 0);
946 EXPORT_SYMBOL(kmemleak_no_scan);
949 * Update an object's checksum and return true if it was modified.
951 static bool update_checksum(struct kmemleak_object *object)
953 u32 old_csum = object->checksum;
955 if (!kmemcheck_is_obj_initialized(object->pointer, object->size))
958 object->checksum = crc32(0, (void *)object->pointer, object->size);
959 return object->checksum != old_csum;
963 * Memory scanning is a long process and it needs to be interruptable. This
964 * function checks whether such interrupt condition occured.
966 static int scan_should_stop(void)
968 if (!atomic_read(&kmemleak_enabled))
972 * This function may be called from either process or kthread context,
973 * hence the need to check for both stop conditions.
976 return signal_pending(current);
978 return kthread_should_stop();
984 * Scan a memory block (exclusive range) for valid pointers and add those
985 * found to the gray list.
987 static void scan_block(void *_start, void *_end,
988 struct kmemleak_object *scanned, int allow_resched)
991 unsigned long *start = PTR_ALIGN(_start, BYTES_PER_POINTER);
992 unsigned long *end = _end - (BYTES_PER_POINTER - 1);
994 for (ptr = start; ptr < end; ptr++) {
995 struct kmemleak_object *object;
997 unsigned long pointer;
1001 if (scan_should_stop())
1004 /* don't scan uninitialized memory */
1005 if (!kmemcheck_is_obj_initialized((unsigned long)ptr,
1011 object = find_and_get_object(pointer, 1);
1014 if (object == scanned) {
1015 /* self referenced, ignore */
1021 * Avoid the lockdep recursive warning on object->lock being
1022 * previously acquired in scan_object(). These locks are
1023 * enclosed by scan_mutex.
1025 spin_lock_irqsave_nested(&object->lock, flags,
1026 SINGLE_DEPTH_NESTING);
1027 if (!color_white(object)) {
1028 /* non-orphan, ignored or new */
1029 spin_unlock_irqrestore(&object->lock, flags);
1035 * Increase the object's reference count (number of pointers
1036 * to the memory block). If this count reaches the required
1037 * minimum, the object's color will become gray and it will be
1038 * added to the gray_list.
1041 if (color_gray(object)) {
1042 list_add_tail(&object->gray_list, &gray_list);
1043 spin_unlock_irqrestore(&object->lock, flags);
1047 spin_unlock_irqrestore(&object->lock, flags);
1053 * Scan a memory block corresponding to a kmemleak_object. A condition is
1054 * that object->use_count >= 1.
1056 static void scan_object(struct kmemleak_object *object)
1058 struct kmemleak_scan_area *area;
1059 struct hlist_node *elem;
1060 unsigned long flags;
1063 * Once the object->lock is acquired, the corresponding memory block
1064 * cannot be freed (the same lock is acquired in delete_object).
1066 spin_lock_irqsave(&object->lock, flags);
1067 if (object->flags & OBJECT_NO_SCAN)
1069 if (!(object->flags & OBJECT_ALLOCATED))
1070 /* already freed object */
1072 if (hlist_empty(&object->area_list)) {
1073 void *start = (void *)object->pointer;
1074 void *end = (void *)(object->pointer + object->size);
1076 while (start < end && (object->flags & OBJECT_ALLOCATED) &&
1077 !(object->flags & OBJECT_NO_SCAN)) {
1078 scan_block(start, min(start + MAX_SCAN_SIZE, end),
1080 start += MAX_SCAN_SIZE;
1082 spin_unlock_irqrestore(&object->lock, flags);
1084 spin_lock_irqsave(&object->lock, flags);
1087 hlist_for_each_entry(area, elem, &object->area_list, node)
1088 scan_block((void *)area->start,
1089 (void *)(area->start + area->size),
1092 spin_unlock_irqrestore(&object->lock, flags);
1096 * Scan the objects already referenced (gray objects). More objects will be
1097 * referenced and, if there are no memory leaks, all the objects are scanned.
1099 static void scan_gray_list(void)
1101 struct kmemleak_object *object, *tmp;
1104 * The list traversal is safe for both tail additions and removals
1105 * from inside the loop. The kmemleak objects cannot be freed from
1106 * outside the loop because their use_count was incremented.
1108 object = list_entry(gray_list.next, typeof(*object), gray_list);
1109 while (&object->gray_list != &gray_list) {
1112 /* may add new objects to the list */
1113 if (!scan_should_stop())
1114 scan_object(object);
1116 tmp = list_entry(object->gray_list.next, typeof(*object),
1119 /* remove the object from the list and release it */
1120 list_del(&object->gray_list);
1125 WARN_ON(!list_empty(&gray_list));
1129 * Scan data sections and all the referenced memory blocks allocated via the
1130 * kernel's standard allocators. This function must be called with the
1133 static void kmemleak_scan(void)
1135 unsigned long flags;
1136 struct kmemleak_object *object;
1140 jiffies_last_scan = jiffies;
1142 /* prepare the kmemleak_object's */
1144 list_for_each_entry_rcu(object, &object_list, object_list) {
1145 spin_lock_irqsave(&object->lock, flags);
1148 * With a few exceptions there should be a maximum of
1149 * 1 reference to any object at this point.
1151 if (atomic_read(&object->use_count) > 1) {
1152 pr_debug("object->use_count = %d\n",
1153 atomic_read(&object->use_count));
1154 dump_object_info(object);
1157 /* reset the reference count (whiten the object) */
1159 if (color_gray(object) && get_object(object))
1160 list_add_tail(&object->gray_list, &gray_list);
1162 spin_unlock_irqrestore(&object->lock, flags);
1166 /* data/bss scanning */
1167 scan_block(_sdata, _edata, NULL, 1);
1168 scan_block(__bss_start, __bss_stop, NULL, 1);
1171 /* per-cpu sections scanning */
1172 for_each_possible_cpu(i)
1173 scan_block(__per_cpu_start + per_cpu_offset(i),
1174 __per_cpu_end + per_cpu_offset(i), NULL, 1);
1178 * Struct page scanning for each node. The code below is not yet safe
1179 * with MEMORY_HOTPLUG.
1181 for_each_online_node(i) {
1182 pg_data_t *pgdat = NODE_DATA(i);
1183 unsigned long start_pfn = pgdat->node_start_pfn;
1184 unsigned long end_pfn = start_pfn + pgdat->node_spanned_pages;
1187 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1190 if (!pfn_valid(pfn))
1192 page = pfn_to_page(pfn);
1193 /* only scan if page is in use */
1194 if (page_count(page) == 0)
1196 scan_block(page, page + 1, NULL, 1);
1201 * Scanning the task stacks (may introduce false negatives).
1203 if (kmemleak_stack_scan) {
1204 struct task_struct *p, *g;
1206 read_lock(&tasklist_lock);
1207 do_each_thread(g, p) {
1208 scan_block(task_stack_page(p), task_stack_page(p) +
1209 THREAD_SIZE, NULL, 0);
1210 } while_each_thread(g, p);
1211 read_unlock(&tasklist_lock);
1215 * Scan the objects already referenced from the sections scanned
1221 * Check for new or unreferenced objects modified since the previous
1222 * scan and color them gray until the next scan.
1225 list_for_each_entry_rcu(object, &object_list, object_list) {
1226 spin_lock_irqsave(&object->lock, flags);
1227 if (color_white(object) && (object->flags & OBJECT_ALLOCATED)
1228 && update_checksum(object) && get_object(object)) {
1229 /* color it gray temporarily */
1230 object->count = object->min_count;
1231 list_add_tail(&object->gray_list, &gray_list);
1233 spin_unlock_irqrestore(&object->lock, flags);
1238 * Re-scan the gray list for modified unreferenced objects.
1243 * If scanning was stopped do not report any new unreferenced objects.
1245 if (scan_should_stop())
1249 * Scanning result reporting.
1252 list_for_each_entry_rcu(object, &object_list, object_list) {
1253 spin_lock_irqsave(&object->lock, flags);
1254 if (unreferenced_object(object) &&
1255 !(object->flags & OBJECT_REPORTED)) {
1256 object->flags |= OBJECT_REPORTED;
1259 spin_unlock_irqrestore(&object->lock, flags);
1264 pr_info("%d new suspected memory leaks (see "
1265 "/sys/kernel/debug/kmemleak)\n", new_leaks);
1270 * Thread function performing automatic memory scanning. Unreferenced objects
1271 * at the end of a memory scan are reported but only the first time.
1273 static int kmemleak_scan_thread(void *arg)
1275 static int first_run = 1;
1277 pr_info("Automatic memory scanning thread started\n");
1278 set_user_nice(current, 10);
1281 * Wait before the first scan to allow the system to fully initialize.
1285 ssleep(SECS_FIRST_SCAN);
1288 while (!kthread_should_stop()) {
1289 signed long timeout = jiffies_scan_wait;
1291 mutex_lock(&scan_mutex);
1293 mutex_unlock(&scan_mutex);
1295 /* wait before the next scan */
1296 while (timeout && !kthread_should_stop())
1297 timeout = schedule_timeout_interruptible(timeout);
1300 pr_info("Automatic memory scanning thread ended\n");
1306 * Start the automatic memory scanning thread. This function must be called
1307 * with the scan_mutex held.
1309 static void start_scan_thread(void)
1313 scan_thread = kthread_run(kmemleak_scan_thread, NULL, "kmemleak");
1314 if (IS_ERR(scan_thread)) {
1315 pr_warning("Failed to create the scan thread\n");
1321 * Stop the automatic memory scanning thread. This function must be called
1322 * with the scan_mutex held.
1324 static void stop_scan_thread(void)
1327 kthread_stop(scan_thread);
1333 * Iterate over the object_list and return the first valid object at or after
1334 * the required position with its use_count incremented. The function triggers
1335 * a memory scanning when the pos argument points to the first position.
1337 static void *kmemleak_seq_start(struct seq_file *seq, loff_t *pos)
1339 struct kmemleak_object *object;
1343 err = mutex_lock_interruptible(&scan_mutex);
1345 return ERR_PTR(err);
1348 list_for_each_entry_rcu(object, &object_list, object_list) {
1351 if (get_object(object))
1360 * Return the next object in the object_list. The function decrements the
1361 * use_count of the previous object and increases that of the next one.
1363 static void *kmemleak_seq_next(struct seq_file *seq, void *v, loff_t *pos)
1365 struct kmemleak_object *prev_obj = v;
1366 struct kmemleak_object *next_obj = NULL;
1367 struct list_head *n = &prev_obj->object_list;
1371 list_for_each_continue_rcu(n, &object_list) {
1372 next_obj = list_entry(n, struct kmemleak_object, object_list);
1373 if (get_object(next_obj))
1377 put_object(prev_obj);
1382 * Decrement the use_count of the last object required, if any.
1384 static void kmemleak_seq_stop(struct seq_file *seq, void *v)
1388 * kmemleak_seq_start may return ERR_PTR if the scan_mutex
1389 * waiting was interrupted, so only release it if !IS_ERR.
1392 mutex_unlock(&scan_mutex);
1399 * Print the information for an unreferenced object to the seq file.
1401 static int kmemleak_seq_show(struct seq_file *seq, void *v)
1403 struct kmemleak_object *object = v;
1404 unsigned long flags;
1406 spin_lock_irqsave(&object->lock, flags);
1407 if ((object->flags & OBJECT_REPORTED) && unreferenced_object(object))
1408 print_unreferenced(seq, object);
1409 spin_unlock_irqrestore(&object->lock, flags);
1413 static const struct seq_operations kmemleak_seq_ops = {
1414 .start = kmemleak_seq_start,
1415 .next = kmemleak_seq_next,
1416 .stop = kmemleak_seq_stop,
1417 .show = kmemleak_seq_show,
1420 static int kmemleak_open(struct inode *inode, struct file *file)
1422 if (!atomic_read(&kmemleak_enabled))
1425 return seq_open(file, &kmemleak_seq_ops);
1428 static int kmemleak_release(struct inode *inode, struct file *file)
1430 return seq_release(inode, file);
1433 static int dump_str_object_info(const char *str)
1435 unsigned long flags;
1436 struct kmemleak_object *object;
1439 addr= simple_strtoul(str, NULL, 0);
1440 object = find_and_get_object(addr, 0);
1442 pr_info("Unknown object at 0x%08lx\n", addr);
1446 spin_lock_irqsave(&object->lock, flags);
1447 dump_object_info(object);
1448 spin_unlock_irqrestore(&object->lock, flags);
1455 * We use grey instead of black to ensure we can do future scans on the same
1456 * objects. If we did not do future scans these black objects could
1457 * potentially contain references to newly allocated objects in the future and
1458 * we'd end up with false positives.
1460 static void kmemleak_clear(void)
1462 struct kmemleak_object *object;
1463 unsigned long flags;
1466 list_for_each_entry_rcu(object, &object_list, object_list) {
1467 spin_lock_irqsave(&object->lock, flags);
1468 if ((object->flags & OBJECT_REPORTED) &&
1469 unreferenced_object(object))
1470 __paint_it(object, KMEMLEAK_GREY);
1471 spin_unlock_irqrestore(&object->lock, flags);
1477 * File write operation to configure kmemleak at run-time. The following
1478 * commands can be written to the /sys/kernel/debug/kmemleak file:
1479 * off - disable kmemleak (irreversible)
1480 * stack=on - enable the task stacks scanning
1481 * stack=off - disable the tasks stacks scanning
1482 * scan=on - start the automatic memory scanning thread
1483 * scan=off - stop the automatic memory scanning thread
1484 * scan=... - set the automatic memory scanning period in seconds (0 to
1486 * scan - trigger a memory scan
1487 * clear - mark all current reported unreferenced kmemleak objects as
1488 * grey to ignore printing them
1489 * dump=... - dump information about the object found at the given address
1491 static ssize_t kmemleak_write(struct file *file, const char __user *user_buf,
1492 size_t size, loff_t *ppos)
1498 buf_size = min(size, (sizeof(buf) - 1));
1499 if (strncpy_from_user(buf, user_buf, buf_size) < 0)
1503 ret = mutex_lock_interruptible(&scan_mutex);
1507 if (strncmp(buf, "off", 3) == 0)
1509 else if (strncmp(buf, "stack=on", 8) == 0)
1510 kmemleak_stack_scan = 1;
1511 else if (strncmp(buf, "stack=off", 9) == 0)
1512 kmemleak_stack_scan = 0;
1513 else if (strncmp(buf, "scan=on", 7) == 0)
1514 start_scan_thread();
1515 else if (strncmp(buf, "scan=off", 8) == 0)
1517 else if (strncmp(buf, "scan=", 5) == 0) {
1520 ret = strict_strtoul(buf + 5, 0, &secs);
1525 jiffies_scan_wait = msecs_to_jiffies(secs * 1000);
1526 start_scan_thread();
1528 } else if (strncmp(buf, "scan", 4) == 0)
1530 else if (strncmp(buf, "clear", 5) == 0)
1532 else if (strncmp(buf, "dump=", 5) == 0)
1533 ret = dump_str_object_info(buf + 5);
1538 mutex_unlock(&scan_mutex);
1542 /* ignore the rest of the buffer, only one command at a time */
1547 static const struct file_operations kmemleak_fops = {
1548 .owner = THIS_MODULE,
1549 .open = kmemleak_open,
1551 .write = kmemleak_write,
1552 .llseek = seq_lseek,
1553 .release = kmemleak_release,
1557 * Perform the freeing of the kmemleak internal objects after waiting for any
1558 * current memory scan to complete.
1560 static void kmemleak_do_cleanup(struct work_struct *work)
1562 struct kmemleak_object *object;
1564 mutex_lock(&scan_mutex);
1568 list_for_each_entry_rcu(object, &object_list, object_list)
1569 delete_object_full(object->pointer);
1571 mutex_unlock(&scan_mutex);
1574 static DECLARE_WORK(cleanup_work, kmemleak_do_cleanup);
1577 * Disable kmemleak. No memory allocation/freeing will be traced once this
1578 * function is called. Disabling kmemleak is an irreversible operation.
1580 static void kmemleak_disable(void)
1582 /* atomically check whether it was already invoked */
1583 if (atomic_cmpxchg(&kmemleak_error, 0, 1))
1586 /* stop any memory operation tracing */
1587 atomic_set(&kmemleak_early_log, 0);
1588 atomic_set(&kmemleak_enabled, 0);
1590 /* check whether it is too early for a kernel thread */
1591 if (atomic_read(&kmemleak_initialized))
1592 schedule_work(&cleanup_work);
1594 pr_info("Kernel memory leak detector disabled\n");
1598 * Allow boot-time kmemleak disabling (enabled by default).
1600 static int kmemleak_boot_config(char *str)
1604 if (strcmp(str, "off") == 0)
1606 else if (strcmp(str, "on") != 0)
1610 early_param("kmemleak", kmemleak_boot_config);
1613 * Kmemleak initialization.
1615 void __init kmemleak_init(void)
1618 unsigned long flags;
1620 jiffies_min_age = msecs_to_jiffies(MSECS_MIN_AGE);
1621 jiffies_scan_wait = msecs_to_jiffies(SECS_SCAN_WAIT * 1000);
1623 object_cache = KMEM_CACHE(kmemleak_object, SLAB_NOLEAKTRACE);
1624 scan_area_cache = KMEM_CACHE(kmemleak_scan_area, SLAB_NOLEAKTRACE);
1625 INIT_PRIO_TREE_ROOT(&object_tree_root);
1627 /* the kernel is still in UP mode, so disabling the IRQs is enough */
1628 local_irq_save(flags);
1629 if (!atomic_read(&kmemleak_error)) {
1630 atomic_set(&kmemleak_enabled, 1);
1631 atomic_set(&kmemleak_early_log, 0);
1633 local_irq_restore(flags);
1636 * This is the point where tracking allocations is safe. Automatic
1637 * scanning is started during the late initcall. Add the early logged
1638 * callbacks to the kmemleak infrastructure.
1640 for (i = 0; i < crt_early_log; i++) {
1641 struct early_log *log = &early_log[i];
1643 switch (log->op_type) {
1644 case KMEMLEAK_ALLOC:
1648 kmemleak_free(log->ptr);
1650 case KMEMLEAK_FREE_PART:
1651 kmemleak_free_part(log->ptr, log->size);
1653 case KMEMLEAK_NOT_LEAK:
1654 kmemleak_not_leak(log->ptr);
1656 case KMEMLEAK_IGNORE:
1657 kmemleak_ignore(log->ptr);
1659 case KMEMLEAK_SCAN_AREA:
1660 kmemleak_scan_area(log->ptr, log->size, GFP_KERNEL);
1662 case KMEMLEAK_NO_SCAN:
1663 kmemleak_no_scan(log->ptr);
1672 * Late initialization function.
1674 static int __init kmemleak_late_init(void)
1676 struct dentry *dentry;
1678 atomic_set(&kmemleak_initialized, 1);
1680 if (atomic_read(&kmemleak_error)) {
1682 * Some error occured and kmemleak was disabled. There is a
1683 * small chance that kmemleak_disable() was called immediately
1684 * after setting kmemleak_initialized and we may end up with
1685 * two clean-up threads but serialized by scan_mutex.
1687 schedule_work(&cleanup_work);
1691 dentry = debugfs_create_file("kmemleak", S_IRUGO, NULL, NULL,
1694 pr_warning("Failed to create the debugfs kmemleak file\n");
1695 mutex_lock(&scan_mutex);
1696 start_scan_thread();
1697 mutex_unlock(&scan_mutex);
1699 pr_info("Kernel memory leak detector initialized\n");
1703 late_initcall(kmemleak_late_init);