page allocator: do not setup zonelist cache when there is only one node
[linux-2.6.git] / mm / kmemleak.c
1 /*
2  * mm/kmemleak.c
3  *
4  * Copyright (C) 2008 ARM Limited
5  * Written by Catalin Marinas <catalin.marinas@arm.com>
6  *
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.
10  *
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.
15  *
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
19  *
20  *
21  * For more information on the algorithm and kmemleak usage, please see
22  * Documentation/kmemleak.txt.
23  *
24  * Notes on locking
25  * ----------------
26  *
27  * The following locks and mutexes are used by kmemleak:
28  *
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
52  * - kmemleak_mutex (mutex): prevents multiple users of the "kmemleak" debugfs
53  *   file together with modifications to the memory scanning parameters
54  *   including the scan_thread pointer
55  *
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
61  * structure.
62  */
63
64 #include <linux/init.h>
65 #include <linux/kernel.h>
66 #include <linux/list.h>
67 #include <linux/sched.h>
68 #include <linux/jiffies.h>
69 #include <linux/delay.h>
70 #include <linux/module.h>
71 #include <linux/kthread.h>
72 #include <linux/prio_tree.h>
73 #include <linux/gfp.h>
74 #include <linux/fs.h>
75 #include <linux/debugfs.h>
76 #include <linux/seq_file.h>
77 #include <linux/cpumask.h>
78 #include <linux/spinlock.h>
79 #include <linux/mutex.h>
80 #include <linux/rcupdate.h>
81 #include <linux/stacktrace.h>
82 #include <linux/cache.h>
83 #include <linux/percpu.h>
84 #include <linux/hardirq.h>
85 #include <linux/mmzone.h>
86 #include <linux/slab.h>
87 #include <linux/thread_info.h>
88 #include <linux/err.h>
89 #include <linux/uaccess.h>
90 #include <linux/string.h>
91 #include <linux/nodemask.h>
92 #include <linux/mm.h>
93
94 #include <asm/sections.h>
95 #include <asm/processor.h>
96 #include <asm/atomic.h>
97
98 #include <linux/kmemleak.h>
99
100 /*
101  * Kmemleak configuration and common defines.
102  */
103 #define MAX_TRACE               16      /* stack trace length */
104 #define REPORTS_NR              50      /* maximum number of reported leaks */
105 #define MSECS_MIN_AGE           5000    /* minimum object age for reporting */
106 #define MSECS_SCAN_YIELD        10      /* CPU yielding period */
107 #define SECS_FIRST_SCAN         60      /* delay before the first scan */
108 #define SECS_SCAN_WAIT          600     /* subsequent auto scanning delay */
109
110 #define BYTES_PER_POINTER       sizeof(void *)
111
112 /* scanning area inside a memory block */
113 struct kmemleak_scan_area {
114         struct hlist_node node;
115         unsigned long offset;
116         size_t length;
117 };
118
119 /*
120  * Structure holding the metadata for each allocated memory block.
121  * Modifications to such objects should be made while holding the
122  * object->lock. Insertions or deletions from object_list, gray_list or
123  * tree_node are already protected by the corresponding locks or mutex (see
124  * the notes on locking above). These objects are reference-counted
125  * (use_count) and freed using the RCU mechanism.
126  */
127 struct kmemleak_object {
128         spinlock_t lock;
129         unsigned long flags;            /* object status flags */
130         struct list_head object_list;
131         struct list_head gray_list;
132         struct prio_tree_node tree_node;
133         struct rcu_head rcu;            /* object_list lockless traversal */
134         /* object usage count; object freed when use_count == 0 */
135         atomic_t use_count;
136         unsigned long pointer;
137         size_t size;
138         /* minimum number of a pointers found before it is considered leak */
139         int min_count;
140         /* the total number of pointers found pointing to this object */
141         int count;
142         /* memory ranges to be scanned inside an object (empty for all) */
143         struct hlist_head area_list;
144         unsigned long trace[MAX_TRACE];
145         unsigned int trace_len;
146         unsigned long jiffies;          /* creation timestamp */
147         pid_t pid;                      /* pid of the current task */
148         char comm[TASK_COMM_LEN];       /* executable name */
149 };
150
151 /* flag representing the memory block allocation status */
152 #define OBJECT_ALLOCATED        (1 << 0)
153 /* flag set after the first reporting of an unreference object */
154 #define OBJECT_REPORTED         (1 << 1)
155 /* flag set to not scan the object */
156 #define OBJECT_NO_SCAN          (1 << 2)
157
158 /* the list of all allocated objects */
159 static LIST_HEAD(object_list);
160 /* the list of gray-colored objects (see color_gray comment below) */
161 static LIST_HEAD(gray_list);
162 /* prio search tree for object boundaries */
163 static struct prio_tree_root object_tree_root;
164 /* rw_lock protecting the access to object_list and prio_tree_root */
165 static DEFINE_RWLOCK(kmemleak_lock);
166
167 /* allocation caches for kmemleak internal data */
168 static struct kmem_cache *object_cache;
169 static struct kmem_cache *scan_area_cache;
170
171 /* set if tracing memory operations is enabled */
172 static atomic_t kmemleak_enabled = ATOMIC_INIT(0);
173 /* set in the late_initcall if there were no errors */
174 static atomic_t kmemleak_initialized = ATOMIC_INIT(0);
175 /* enables or disables early logging of the memory operations */
176 static atomic_t kmemleak_early_log = ATOMIC_INIT(1);
177 /* set if a fata kmemleak error has occurred */
178 static atomic_t kmemleak_error = ATOMIC_INIT(0);
179
180 /* minimum and maximum address that may be valid pointers */
181 static unsigned long min_addr = ULONG_MAX;
182 static unsigned long max_addr;
183
184 /* used for yielding the CPU to other tasks during scanning */
185 static unsigned long next_scan_yield;
186 static struct task_struct *scan_thread;
187 static unsigned long jiffies_scan_yield;
188 static unsigned long jiffies_min_age;
189 /* delay between automatic memory scannings */
190 static signed long jiffies_scan_wait;
191 /* enables or disables the task stacks scanning */
192 static int kmemleak_stack_scan;
193 /* mutex protecting the memory scanning */
194 static DEFINE_MUTEX(scan_mutex);
195 /* mutex protecting the access to the /sys/kernel/debug/kmemleak file */
196 static DEFINE_MUTEX(kmemleak_mutex);
197
198 /* number of leaks reported (for limitation purposes) */
199 static int reported_leaks;
200
201 /*
202  * Early object allocation/freeing logging. Kkmemleak is initialized after the
203  * kernel allocator. However, both the kernel allocator and kmemleak may
204  * allocate memory blocks which need to be tracked. Kkmemleak defines an
205  * arbitrary buffer to hold the allocation/freeing information before it is
206  * fully initialized.
207  */
208
209 /* kmemleak operation type for early logging */
210 enum {
211         KMEMLEAK_ALLOC,
212         KMEMLEAK_FREE,
213         KMEMLEAK_NOT_LEAK,
214         KMEMLEAK_IGNORE,
215         KMEMLEAK_SCAN_AREA,
216         KMEMLEAK_NO_SCAN
217 };
218
219 /*
220  * Structure holding the information passed to kmemleak callbacks during the
221  * early logging.
222  */
223 struct early_log {
224         int op_type;                    /* kmemleak operation type */
225         const void *ptr;                /* allocated/freed memory block */
226         size_t size;                    /* memory block size */
227         int min_count;                  /* minimum reference count */
228         unsigned long offset;           /* scan area offset */
229         size_t length;                  /* scan area length */
230 };
231
232 /* early logging buffer and current position */
233 static struct early_log early_log[200];
234 static int crt_early_log;
235
236 static void kmemleak_disable(void);
237
238 /*
239  * Print a warning and dump the stack trace.
240  */
241 #define kmemleak_warn(x...)     do {    \
242         pr_warning(x);                  \
243         dump_stack();                   \
244 } while (0)
245
246 /*
247  * Macro invoked when a serious kmemleak condition occured and cannot be
248  * recovered from. Kkmemleak will be disabled and further allocation/freeing
249  * tracing no longer available.
250  */
251 #define kmemleak_panic(x...)    do {    \
252         kmemleak_warn(x);               \
253         kmemleak_disable();             \
254 } while (0)
255
256 /*
257  * Object colors, encoded with count and min_count:
258  * - white - orphan object, not enough references to it (count < min_count)
259  * - gray  - not orphan, not marked as false positive (min_count == 0) or
260  *              sufficient references to it (count >= min_count)
261  * - black - ignore, it doesn't contain references (e.g. text section)
262  *              (min_count == -1). No function defined for this color.
263  * Newly created objects don't have any color assigned (object->count == -1)
264  * before the next memory scan when they become white.
265  */
266 static int color_white(const struct kmemleak_object *object)
267 {
268         return object->count != -1 && object->count < object->min_count;
269 }
270
271 static int color_gray(const struct kmemleak_object *object)
272 {
273         return object->min_count != -1 && object->count >= object->min_count;
274 }
275
276 /*
277  * Objects are considered referenced if their color is gray and they have not
278  * been deleted.
279  */
280 static int referenced_object(struct kmemleak_object *object)
281 {
282         return (object->flags & OBJECT_ALLOCATED) && color_gray(object);
283 }
284
285 /*
286  * Objects are considered unreferenced only if their color is white, they have
287  * not be deleted and have a minimum age to avoid false positives caused by
288  * pointers temporarily stored in CPU registers.
289  */
290 static int unreferenced_object(struct kmemleak_object *object)
291 {
292         return (object->flags & OBJECT_ALLOCATED) && color_white(object) &&
293                 time_is_before_eq_jiffies(object->jiffies + jiffies_min_age);
294 }
295
296 /*
297  * Printing of the (un)referenced objects information, either to the seq file
298  * or to the kernel log. The print_referenced/print_unreferenced functions
299  * must be called with the object->lock held.
300  */
301 #define print_helper(seq, x...) do {    \
302         struct seq_file *s = (seq);     \
303         if (s)                          \
304                 seq_printf(s, x);       \
305         else                            \
306                 pr_info(x);             \
307 } while (0)
308
309 static void print_referenced(struct kmemleak_object *object)
310 {
311         pr_info("kmemleak: referenced object 0x%08lx (size %zu)\n",
312                 object->pointer, object->size);
313 }
314
315 static void print_unreferenced(struct seq_file *seq,
316                                struct kmemleak_object *object)
317 {
318         int i;
319
320         print_helper(seq, "kmemleak: unreferenced object 0x%08lx (size %zu):\n",
321                      object->pointer, object->size);
322         print_helper(seq, "  comm \"%s\", pid %d, jiffies %lu\n",
323                      object->comm, object->pid, object->jiffies);
324         print_helper(seq, "  backtrace:\n");
325
326         for (i = 0; i < object->trace_len; i++) {
327                 void *ptr = (void *)object->trace[i];
328                 print_helper(seq, "    [<%p>] %pS\n", ptr, ptr);
329         }
330 }
331
332 /*
333  * Print the kmemleak_object information. This function is used mainly for
334  * debugging special cases when kmemleak operations. It must be called with
335  * the object->lock held.
336  */
337 static void dump_object_info(struct kmemleak_object *object)
338 {
339         struct stack_trace trace;
340
341         trace.nr_entries = object->trace_len;
342         trace.entries = object->trace;
343
344         pr_notice("kmemleak: Object 0x%08lx (size %zu):\n",
345                   object->tree_node.start, object->size);
346         pr_notice("  comm \"%s\", pid %d, jiffies %lu\n",
347                   object->comm, object->pid, object->jiffies);
348         pr_notice("  min_count = %d\n", object->min_count);
349         pr_notice("  count = %d\n", object->count);
350         pr_notice("  backtrace:\n");
351         print_stack_trace(&trace, 4);
352 }
353
354 /*
355  * Look-up a memory block metadata (kmemleak_object) in the priority search
356  * tree based on a pointer value. If alias is 0, only values pointing to the
357  * beginning of the memory block are allowed. The kmemleak_lock must be held
358  * when calling this function.
359  */
360 static struct kmemleak_object *lookup_object(unsigned long ptr, int alias)
361 {
362         struct prio_tree_node *node;
363         struct prio_tree_iter iter;
364         struct kmemleak_object *object;
365
366         prio_tree_iter_init(&iter, &object_tree_root, ptr, ptr);
367         node = prio_tree_next(&iter);
368         if (node) {
369                 object = prio_tree_entry(node, struct kmemleak_object,
370                                          tree_node);
371                 if (!alias && object->pointer != ptr) {
372                         kmemleak_warn("kmemleak: Found object by alias");
373                         object = NULL;
374                 }
375         } else
376                 object = NULL;
377
378         return object;
379 }
380
381 /*
382  * Increment the object use_count. Return 1 if successful or 0 otherwise. Note
383  * that once an object's use_count reached 0, the RCU freeing was already
384  * registered and the object should no longer be used. This function must be
385  * called under the protection of rcu_read_lock().
386  */
387 static int get_object(struct kmemleak_object *object)
388 {
389         return atomic_inc_not_zero(&object->use_count);
390 }
391
392 /*
393  * RCU callback to free a kmemleak_object.
394  */
395 static void free_object_rcu(struct rcu_head *rcu)
396 {
397         struct hlist_node *elem, *tmp;
398         struct kmemleak_scan_area *area;
399         struct kmemleak_object *object =
400                 container_of(rcu, struct kmemleak_object, rcu);
401
402         /*
403          * Once use_count is 0 (guaranteed by put_object), there is no other
404          * code accessing this object, hence no need for locking.
405          */
406         hlist_for_each_entry_safe(area, elem, tmp, &object->area_list, node) {
407                 hlist_del(elem);
408                 kmem_cache_free(scan_area_cache, area);
409         }
410         kmem_cache_free(object_cache, object);
411 }
412
413 /*
414  * Decrement the object use_count. Once the count is 0, free the object using
415  * an RCU callback. Since put_object() may be called via the kmemleak_free() ->
416  * delete_object() path, the delayed RCU freeing ensures that there is no
417  * recursive call to the kernel allocator. Lock-less RCU object_list traversal
418  * is also possible.
419  */
420 static void put_object(struct kmemleak_object *object)
421 {
422         if (!atomic_dec_and_test(&object->use_count))
423                 return;
424
425         /* should only get here after delete_object was called */
426         WARN_ON(object->flags & OBJECT_ALLOCATED);
427
428         call_rcu(&object->rcu, free_object_rcu);
429 }
430
431 /*
432  * Look up an object in the prio search tree and increase its use_count.
433  */
434 static struct kmemleak_object *find_and_get_object(unsigned long ptr, int alias)
435 {
436         unsigned long flags;
437         struct kmemleak_object *object = NULL;
438
439         rcu_read_lock();
440         read_lock_irqsave(&kmemleak_lock, flags);
441         if (ptr >= min_addr && ptr < max_addr)
442                 object = lookup_object(ptr, alias);
443         read_unlock_irqrestore(&kmemleak_lock, flags);
444
445         /* check whether the object is still available */
446         if (object && !get_object(object))
447                 object = NULL;
448         rcu_read_unlock();
449
450         return object;
451 }
452
453 /*
454  * Create the metadata (struct kmemleak_object) corresponding to an allocated
455  * memory block and add it to the object_list and object_tree_root.
456  */
457 static void create_object(unsigned long ptr, size_t size, int min_count,
458                           gfp_t gfp)
459 {
460         unsigned long flags;
461         struct kmemleak_object *object;
462         struct prio_tree_node *node;
463         struct stack_trace trace;
464
465         object = kmem_cache_alloc(object_cache, gfp & ~GFP_SLAB_BUG_MASK);
466         if (!object) {
467                 kmemleak_panic("kmemleak: Cannot allocate a kmemleak_object "
468                                "structure\n");
469                 return;
470         }
471
472         INIT_LIST_HEAD(&object->object_list);
473         INIT_LIST_HEAD(&object->gray_list);
474         INIT_HLIST_HEAD(&object->area_list);
475         spin_lock_init(&object->lock);
476         atomic_set(&object->use_count, 1);
477         object->flags = OBJECT_ALLOCATED;
478         object->pointer = ptr;
479         object->size = size;
480         object->min_count = min_count;
481         object->count = -1;                     /* no color initially */
482         object->jiffies = jiffies;
483
484         /* task information */
485         if (in_irq()) {
486                 object->pid = 0;
487                 strncpy(object->comm, "hardirq", sizeof(object->comm));
488         } else if (in_softirq()) {
489                 object->pid = 0;
490                 strncpy(object->comm, "softirq", sizeof(object->comm));
491         } else {
492                 object->pid = current->pid;
493                 /*
494                  * There is a small chance of a race with set_task_comm(),
495                  * however using get_task_comm() here may cause locking
496                  * dependency issues with current->alloc_lock. In the worst
497                  * case, the command line is not correct.
498                  */
499                 strncpy(object->comm, current->comm, sizeof(object->comm));
500         }
501
502         /* kernel backtrace */
503         trace.max_entries = MAX_TRACE;
504         trace.nr_entries = 0;
505         trace.entries = object->trace;
506         trace.skip = 1;
507         save_stack_trace(&trace);
508         object->trace_len = trace.nr_entries;
509
510         INIT_PRIO_TREE_NODE(&object->tree_node);
511         object->tree_node.start = ptr;
512         object->tree_node.last = ptr + size - 1;
513
514         write_lock_irqsave(&kmemleak_lock, flags);
515         min_addr = min(min_addr, ptr);
516         max_addr = max(max_addr, ptr + size);
517         node = prio_tree_insert(&object_tree_root, &object->tree_node);
518         /*
519          * The code calling the kernel does not yet have the pointer to the
520          * memory block to be able to free it.  However, we still hold the
521          * kmemleak_lock here in case parts of the kernel started freeing
522          * random memory blocks.
523          */
524         if (node != &object->tree_node) {
525                 unsigned long flags;
526
527                 kmemleak_panic("kmemleak: Cannot insert 0x%lx into the object "
528                                "search tree (already existing)\n", ptr);
529                 object = lookup_object(ptr, 1);
530                 spin_lock_irqsave(&object->lock, flags);
531                 dump_object_info(object);
532                 spin_unlock_irqrestore(&object->lock, flags);
533
534                 goto out;
535         }
536         list_add_tail_rcu(&object->object_list, &object_list);
537 out:
538         write_unlock_irqrestore(&kmemleak_lock, flags);
539 }
540
541 /*
542  * Remove the metadata (struct kmemleak_object) for a memory block from the
543  * object_list and object_tree_root and decrement its use_count.
544  */
545 static void delete_object(unsigned long ptr)
546 {
547         unsigned long flags;
548         struct kmemleak_object *object;
549
550         write_lock_irqsave(&kmemleak_lock, flags);
551         object = lookup_object(ptr, 0);
552         if (!object) {
553                 kmemleak_warn("kmemleak: Freeing unknown object at 0x%08lx\n",
554                               ptr);
555                 write_unlock_irqrestore(&kmemleak_lock, flags);
556                 return;
557         }
558         prio_tree_remove(&object_tree_root, &object->tree_node);
559         list_del_rcu(&object->object_list);
560         write_unlock_irqrestore(&kmemleak_lock, flags);
561
562         WARN_ON(!(object->flags & OBJECT_ALLOCATED));
563         WARN_ON(atomic_read(&object->use_count) < 1);
564
565         /*
566          * Locking here also ensures that the corresponding memory block
567          * cannot be freed when it is being scanned.
568          */
569         spin_lock_irqsave(&object->lock, flags);
570         if (object->flags & OBJECT_REPORTED)
571                 print_referenced(object);
572         object->flags &= ~OBJECT_ALLOCATED;
573         spin_unlock_irqrestore(&object->lock, flags);
574         put_object(object);
575 }
576
577 /*
578  * Make a object permanently as gray-colored so that it can no longer be
579  * reported as a leak. This is used in general to mark a false positive.
580  */
581 static void make_gray_object(unsigned long ptr)
582 {
583         unsigned long flags;
584         struct kmemleak_object *object;
585
586         object = find_and_get_object(ptr, 0);
587         if (!object) {
588                 kmemleak_warn("kmemleak: Graying unknown object at 0x%08lx\n",
589                               ptr);
590                 return;
591         }
592
593         spin_lock_irqsave(&object->lock, flags);
594         object->min_count = 0;
595         spin_unlock_irqrestore(&object->lock, flags);
596         put_object(object);
597 }
598
599 /*
600  * Mark the object as black-colored so that it is ignored from scans and
601  * reporting.
602  */
603 static void make_black_object(unsigned long ptr)
604 {
605         unsigned long flags;
606         struct kmemleak_object *object;
607
608         object = find_and_get_object(ptr, 0);
609         if (!object) {
610                 kmemleak_warn("kmemleak: Blacking unknown object at 0x%08lx\n",
611                               ptr);
612                 return;
613         }
614
615         spin_lock_irqsave(&object->lock, flags);
616         object->min_count = -1;
617         spin_unlock_irqrestore(&object->lock, flags);
618         put_object(object);
619 }
620
621 /*
622  * Add a scanning area to the object. If at least one such area is added,
623  * kmemleak will only scan these ranges rather than the whole memory block.
624  */
625 static void add_scan_area(unsigned long ptr, unsigned long offset,
626                           size_t length, gfp_t gfp)
627 {
628         unsigned long flags;
629         struct kmemleak_object *object;
630         struct kmemleak_scan_area *area;
631
632         object = find_and_get_object(ptr, 0);
633         if (!object) {
634                 kmemleak_warn("kmemleak: Adding scan area to unknown "
635                               "object at 0x%08lx\n", ptr);
636                 return;
637         }
638
639         area = kmem_cache_alloc(scan_area_cache, gfp & ~GFP_SLAB_BUG_MASK);
640         if (!area) {
641                 kmemleak_warn("kmemleak: Cannot allocate a scan area\n");
642                 goto out;
643         }
644
645         spin_lock_irqsave(&object->lock, flags);
646         if (offset + length > object->size) {
647                 kmemleak_warn("kmemleak: Scan area larger than object "
648                               "0x%08lx\n", ptr);
649                 dump_object_info(object);
650                 kmem_cache_free(scan_area_cache, area);
651                 goto out_unlock;
652         }
653
654         INIT_HLIST_NODE(&area->node);
655         area->offset = offset;
656         area->length = length;
657
658         hlist_add_head(&area->node, &object->area_list);
659 out_unlock:
660         spin_unlock_irqrestore(&object->lock, flags);
661 out:
662         put_object(object);
663 }
664
665 /*
666  * Set the OBJECT_NO_SCAN flag for the object corresponding to the give
667  * pointer. Such object will not be scanned by kmemleak but references to it
668  * are searched.
669  */
670 static void object_no_scan(unsigned long ptr)
671 {
672         unsigned long flags;
673         struct kmemleak_object *object;
674
675         object = find_and_get_object(ptr, 0);
676         if (!object) {
677                 kmemleak_warn("kmemleak: Not scanning unknown object at "
678                               "0x%08lx\n", ptr);
679                 return;
680         }
681
682         spin_lock_irqsave(&object->lock, flags);
683         object->flags |= OBJECT_NO_SCAN;
684         spin_unlock_irqrestore(&object->lock, flags);
685         put_object(object);
686 }
687
688 /*
689  * Log an early kmemleak_* call to the early_log buffer. These calls will be
690  * processed later once kmemleak is fully initialized.
691  */
692 static void log_early(int op_type, const void *ptr, size_t size,
693                       int min_count, unsigned long offset, size_t length)
694 {
695         unsigned long flags;
696         struct early_log *log;
697
698         if (crt_early_log >= ARRAY_SIZE(early_log)) {
699                 kmemleak_panic("kmemleak: Early log buffer exceeded\n");
700                 return;
701         }
702
703         /*
704          * There is no need for locking since the kernel is still in UP mode
705          * at this stage. Disabling the IRQs is enough.
706          */
707         local_irq_save(flags);
708         log = &early_log[crt_early_log];
709         log->op_type = op_type;
710         log->ptr = ptr;
711         log->size = size;
712         log->min_count = min_count;
713         log->offset = offset;
714         log->length = length;
715         crt_early_log++;
716         local_irq_restore(flags);
717 }
718
719 /*
720  * Memory allocation function callback. This function is called from the
721  * kernel allocators when a new block is allocated (kmem_cache_alloc, kmalloc,
722  * vmalloc etc.).
723  */
724 void kmemleak_alloc(const void *ptr, size_t size, int min_count, gfp_t gfp)
725 {
726         pr_debug("%s(0x%p, %zu, %d)\n", __func__, ptr, size, min_count);
727
728         if (atomic_read(&kmemleak_enabled) && ptr && !IS_ERR(ptr))
729                 create_object((unsigned long)ptr, size, min_count, gfp);
730         else if (atomic_read(&kmemleak_early_log))
731                 log_early(KMEMLEAK_ALLOC, ptr, size, min_count, 0, 0);
732 }
733 EXPORT_SYMBOL_GPL(kmemleak_alloc);
734
735 /*
736  * Memory freeing function callback. This function is called from the kernel
737  * allocators when a block is freed (kmem_cache_free, kfree, vfree etc.).
738  */
739 void kmemleak_free(const void *ptr)
740 {
741         pr_debug("%s(0x%p)\n", __func__, ptr);
742
743         if (atomic_read(&kmemleak_enabled) && ptr && !IS_ERR(ptr))
744                 delete_object((unsigned long)ptr);
745         else if (atomic_read(&kmemleak_early_log))
746                 log_early(KMEMLEAK_FREE, ptr, 0, 0, 0, 0);
747 }
748 EXPORT_SYMBOL_GPL(kmemleak_free);
749
750 /*
751  * Mark an already allocated memory block as a false positive. This will cause
752  * the block to no longer be reported as leak and always be scanned.
753  */
754 void kmemleak_not_leak(const void *ptr)
755 {
756         pr_debug("%s(0x%p)\n", __func__, ptr);
757
758         if (atomic_read(&kmemleak_enabled) && ptr && !IS_ERR(ptr))
759                 make_gray_object((unsigned long)ptr);
760         else if (atomic_read(&kmemleak_early_log))
761                 log_early(KMEMLEAK_NOT_LEAK, ptr, 0, 0, 0, 0);
762 }
763 EXPORT_SYMBOL(kmemleak_not_leak);
764
765 /*
766  * Ignore a memory block. This is usually done when it is known that the
767  * corresponding block is not a leak and does not contain any references to
768  * other allocated memory blocks.
769  */
770 void kmemleak_ignore(const void *ptr)
771 {
772         pr_debug("%s(0x%p)\n", __func__, ptr);
773
774         if (atomic_read(&kmemleak_enabled) && ptr && !IS_ERR(ptr))
775                 make_black_object((unsigned long)ptr);
776         else if (atomic_read(&kmemleak_early_log))
777                 log_early(KMEMLEAK_IGNORE, ptr, 0, 0, 0, 0);
778 }
779 EXPORT_SYMBOL(kmemleak_ignore);
780
781 /*
782  * Limit the range to be scanned in an allocated memory block.
783  */
784 void kmemleak_scan_area(const void *ptr, unsigned long offset, size_t length,
785                         gfp_t gfp)
786 {
787         pr_debug("%s(0x%p)\n", __func__, ptr);
788
789         if (atomic_read(&kmemleak_enabled) && ptr && !IS_ERR(ptr))
790                 add_scan_area((unsigned long)ptr, offset, length, gfp);
791         else if (atomic_read(&kmemleak_early_log))
792                 log_early(KMEMLEAK_SCAN_AREA, ptr, 0, 0, offset, length);
793 }
794 EXPORT_SYMBOL(kmemleak_scan_area);
795
796 /*
797  * Inform kmemleak not to scan the given memory block.
798  */
799 void kmemleak_no_scan(const void *ptr)
800 {
801         pr_debug("%s(0x%p)\n", __func__, ptr);
802
803         if (atomic_read(&kmemleak_enabled) && ptr && !IS_ERR(ptr))
804                 object_no_scan((unsigned long)ptr);
805         else if (atomic_read(&kmemleak_early_log))
806                 log_early(KMEMLEAK_NO_SCAN, ptr, 0, 0, 0, 0);
807 }
808 EXPORT_SYMBOL(kmemleak_no_scan);
809
810 /*
811  * Yield the CPU so that other tasks get a chance to run.  The yielding is
812  * rate-limited to avoid excessive number of calls to the schedule() function
813  * during memory scanning.
814  */
815 static void scan_yield(void)
816 {
817         might_sleep();
818
819         if (time_is_before_eq_jiffies(next_scan_yield)) {
820                 schedule();
821                 next_scan_yield = jiffies + jiffies_scan_yield;
822         }
823 }
824
825 /*
826  * Memory scanning is a long process and it needs to be interruptable. This
827  * function checks whether such interrupt condition occured.
828  */
829 static int scan_should_stop(void)
830 {
831         if (!atomic_read(&kmemleak_enabled))
832                 return 1;
833
834         /*
835          * This function may be called from either process or kthread context,
836          * hence the need to check for both stop conditions.
837          */
838         if (current->mm)
839                 return signal_pending(current);
840         else
841                 return kthread_should_stop();
842
843         return 0;
844 }
845
846 /*
847  * Scan a memory block (exclusive range) for valid pointers and add those
848  * found to the gray list.
849  */
850 static void scan_block(void *_start, void *_end,
851                        struct kmemleak_object *scanned)
852 {
853         unsigned long *ptr;
854         unsigned long *start = PTR_ALIGN(_start, BYTES_PER_POINTER);
855         unsigned long *end = _end - (BYTES_PER_POINTER - 1);
856
857         for (ptr = start; ptr < end; ptr++) {
858                 unsigned long flags;
859                 unsigned long pointer = *ptr;
860                 struct kmemleak_object *object;
861
862                 if (scan_should_stop())
863                         break;
864
865                 /*
866                  * When scanning a memory block with a corresponding
867                  * kmemleak_object, the CPU yielding is handled in the calling
868                  * code since it holds the object->lock to avoid the block
869                  * freeing.
870                  */
871                 if (!scanned)
872                         scan_yield();
873
874                 object = find_and_get_object(pointer, 1);
875                 if (!object)
876                         continue;
877                 if (object == scanned) {
878                         /* self referenced, ignore */
879                         put_object(object);
880                         continue;
881                 }
882
883                 /*
884                  * Avoid the lockdep recursive warning on object->lock being
885                  * previously acquired in scan_object(). These locks are
886                  * enclosed by scan_mutex.
887                  */
888                 spin_lock_irqsave_nested(&object->lock, flags,
889                                          SINGLE_DEPTH_NESTING);
890                 if (!color_white(object)) {
891                         /* non-orphan, ignored or new */
892                         spin_unlock_irqrestore(&object->lock, flags);
893                         put_object(object);
894                         continue;
895                 }
896
897                 /*
898                  * Increase the object's reference count (number of pointers
899                  * to the memory block). If this count reaches the required
900                  * minimum, the object's color will become gray and it will be
901                  * added to the gray_list.
902                  */
903                 object->count++;
904                 if (color_gray(object))
905                         list_add_tail(&object->gray_list, &gray_list);
906                 else
907                         put_object(object);
908                 spin_unlock_irqrestore(&object->lock, flags);
909         }
910 }
911
912 /*
913  * Scan a memory block corresponding to a kmemleak_object. A condition is
914  * that object->use_count >= 1.
915  */
916 static void scan_object(struct kmemleak_object *object)
917 {
918         struct kmemleak_scan_area *area;
919         struct hlist_node *elem;
920         unsigned long flags;
921
922         /*
923          * Once the object->lock is aquired, the corresponding memory block
924          * cannot be freed (the same lock is aquired in delete_object).
925          */
926         spin_lock_irqsave(&object->lock, flags);
927         if (object->flags & OBJECT_NO_SCAN)
928                 goto out;
929         if (!(object->flags & OBJECT_ALLOCATED))
930                 /* already freed object */
931                 goto out;
932         if (hlist_empty(&object->area_list))
933                 scan_block((void *)object->pointer,
934                            (void *)(object->pointer + object->size), object);
935         else
936                 hlist_for_each_entry(area, elem, &object->area_list, node)
937                         scan_block((void *)(object->pointer + area->offset),
938                                    (void *)(object->pointer + area->offset
939                                             + area->length), object);
940 out:
941         spin_unlock_irqrestore(&object->lock, flags);
942 }
943
944 /*
945  * Scan data sections and all the referenced memory blocks allocated via the
946  * kernel's standard allocators. This function must be called with the
947  * scan_mutex held.
948  */
949 static void kmemleak_scan(void)
950 {
951         unsigned long flags;
952         struct kmemleak_object *object, *tmp;
953         struct task_struct *task;
954         int i;
955
956         /* prepare the kmemleak_object's */
957         rcu_read_lock();
958         list_for_each_entry_rcu(object, &object_list, object_list) {
959                 spin_lock_irqsave(&object->lock, flags);
960 #ifdef DEBUG
961                 /*
962                  * With a few exceptions there should be a maximum of
963                  * 1 reference to any object at this point.
964                  */
965                 if (atomic_read(&object->use_count) > 1) {
966                         pr_debug("kmemleak: object->use_count = %d\n",
967                                  atomic_read(&object->use_count));
968                         dump_object_info(object);
969                 }
970 #endif
971                 /* reset the reference count (whiten the object) */
972                 object->count = 0;
973                 if (color_gray(object) && get_object(object))
974                         list_add_tail(&object->gray_list, &gray_list);
975
976                 spin_unlock_irqrestore(&object->lock, flags);
977         }
978         rcu_read_unlock();
979
980         /* data/bss scanning */
981         scan_block(_sdata, _edata, NULL);
982         scan_block(__bss_start, __bss_stop, NULL);
983
984 #ifdef CONFIG_SMP
985         /* per-cpu sections scanning */
986         for_each_possible_cpu(i)
987                 scan_block(__per_cpu_start + per_cpu_offset(i),
988                            __per_cpu_end + per_cpu_offset(i), NULL);
989 #endif
990
991         /*
992          * Struct page scanning for each node. The code below is not yet safe
993          * with MEMORY_HOTPLUG.
994          */
995         for_each_online_node(i) {
996                 pg_data_t *pgdat = NODE_DATA(i);
997                 unsigned long start_pfn = pgdat->node_start_pfn;
998                 unsigned long end_pfn = start_pfn + pgdat->node_spanned_pages;
999                 unsigned long pfn;
1000
1001                 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1002                         struct page *page;
1003
1004                         if (!pfn_valid(pfn))
1005                                 continue;
1006                         page = pfn_to_page(pfn);
1007                         /* only scan if page is in use */
1008                         if (page_count(page) == 0)
1009                                 continue;
1010                         scan_block(page, page + 1, NULL);
1011                 }
1012         }
1013
1014         /*
1015          * Scanning the task stacks may introduce false negatives and it is
1016          * not enabled by default.
1017          */
1018         if (kmemleak_stack_scan) {
1019                 read_lock(&tasklist_lock);
1020                 for_each_process(task)
1021                         scan_block(task_stack_page(task),
1022                                    task_stack_page(task) + THREAD_SIZE, NULL);
1023                 read_unlock(&tasklist_lock);
1024         }
1025
1026         /*
1027          * Scan the objects already referenced from the sections scanned
1028          * above. More objects will be referenced and, if there are no memory
1029          * leaks, all the objects will be scanned. The list traversal is safe
1030          * for both tail additions and removals from inside the loop. The
1031          * kmemleak objects cannot be freed from outside the loop because their
1032          * use_count was increased.
1033          */
1034         object = list_entry(gray_list.next, typeof(*object), gray_list);
1035         while (&object->gray_list != &gray_list) {
1036                 scan_yield();
1037
1038                 /* may add new objects to the list */
1039                 if (!scan_should_stop())
1040                         scan_object(object);
1041
1042                 tmp = list_entry(object->gray_list.next, typeof(*object),
1043                                  gray_list);
1044
1045                 /* remove the object from the list and release it */
1046                 list_del(&object->gray_list);
1047                 put_object(object);
1048
1049                 object = tmp;
1050         }
1051         WARN_ON(!list_empty(&gray_list));
1052 }
1053
1054 /*
1055  * Thread function performing automatic memory scanning. Unreferenced objects
1056  * at the end of a memory scan are reported but only the first time.
1057  */
1058 static int kmemleak_scan_thread(void *arg)
1059 {
1060         static int first_run = 1;
1061
1062         pr_info("kmemleak: Automatic memory scanning thread started\n");
1063
1064         /*
1065          * Wait before the first scan to allow the system to fully initialize.
1066          */
1067         if (first_run) {
1068                 first_run = 0;
1069                 ssleep(SECS_FIRST_SCAN);
1070         }
1071
1072         while (!kthread_should_stop()) {
1073                 struct kmemleak_object *object;
1074                 signed long timeout = jiffies_scan_wait;
1075
1076                 mutex_lock(&scan_mutex);
1077
1078                 kmemleak_scan();
1079                 reported_leaks = 0;
1080
1081                 rcu_read_lock();
1082                 list_for_each_entry_rcu(object, &object_list, object_list) {
1083                         unsigned long flags;
1084
1085                         if (reported_leaks >= REPORTS_NR)
1086                                 break;
1087                         spin_lock_irqsave(&object->lock, flags);
1088                         if (!(object->flags & OBJECT_REPORTED) &&
1089                             unreferenced_object(object)) {
1090                                 print_unreferenced(NULL, object);
1091                                 object->flags |= OBJECT_REPORTED;
1092                                 reported_leaks++;
1093                         } else if ((object->flags & OBJECT_REPORTED) &&
1094                                    referenced_object(object)) {
1095                                 print_referenced(object);
1096                                 object->flags &= ~OBJECT_REPORTED;
1097                         }
1098                         spin_unlock_irqrestore(&object->lock, flags);
1099                 }
1100                 rcu_read_unlock();
1101
1102                 mutex_unlock(&scan_mutex);
1103                 /* wait before the next scan */
1104                 while (timeout && !kthread_should_stop())
1105                         timeout = schedule_timeout_interruptible(timeout);
1106         }
1107
1108         pr_info("kmemleak: Automatic memory scanning thread ended\n");
1109
1110         return 0;
1111 }
1112
1113 /*
1114  * Start the automatic memory scanning thread. This function must be called
1115  * with the kmemleak_mutex held.
1116  */
1117 void start_scan_thread(void)
1118 {
1119         if (scan_thread)
1120                 return;
1121         scan_thread = kthread_run(kmemleak_scan_thread, NULL, "kmemleak");
1122         if (IS_ERR(scan_thread)) {
1123                 pr_warning("kmemleak: Failed to create the scan thread\n");
1124                 scan_thread = NULL;
1125         }
1126 }
1127
1128 /*
1129  * Stop the automatic memory scanning thread. This function must be called
1130  * with the kmemleak_mutex held.
1131  */
1132 void stop_scan_thread(void)
1133 {
1134         if (scan_thread) {
1135                 kthread_stop(scan_thread);
1136                 scan_thread = NULL;
1137         }
1138 }
1139
1140 /*
1141  * Iterate over the object_list and return the first valid object at or after
1142  * the required position with its use_count incremented. The function triggers
1143  * a memory scanning when the pos argument points to the first position.
1144  */
1145 static void *kmemleak_seq_start(struct seq_file *seq, loff_t *pos)
1146 {
1147         struct kmemleak_object *object;
1148         loff_t n = *pos;
1149
1150         if (!n) {
1151                 kmemleak_scan();
1152                 reported_leaks = 0;
1153         }
1154         if (reported_leaks >= REPORTS_NR)
1155                 return NULL;
1156
1157         rcu_read_lock();
1158         list_for_each_entry_rcu(object, &object_list, object_list) {
1159                 if (n-- > 0)
1160                         continue;
1161                 if (get_object(object))
1162                         goto out;
1163         }
1164         object = NULL;
1165 out:
1166         rcu_read_unlock();
1167         return object;
1168 }
1169
1170 /*
1171  * Return the next object in the object_list. The function decrements the
1172  * use_count of the previous object and increases that of the next one.
1173  */
1174 static void *kmemleak_seq_next(struct seq_file *seq, void *v, loff_t *pos)
1175 {
1176         struct kmemleak_object *prev_obj = v;
1177         struct kmemleak_object *next_obj = NULL;
1178         struct list_head *n = &prev_obj->object_list;
1179
1180         ++(*pos);
1181         if (reported_leaks >= REPORTS_NR)
1182                 goto out;
1183
1184         rcu_read_lock();
1185         list_for_each_continue_rcu(n, &object_list) {
1186                 next_obj = list_entry(n, struct kmemleak_object, object_list);
1187                 if (get_object(next_obj))
1188                         break;
1189         }
1190         rcu_read_unlock();
1191 out:
1192         put_object(prev_obj);
1193         return next_obj;
1194 }
1195
1196 /*
1197  * Decrement the use_count of the last object required, if any.
1198  */
1199 static void kmemleak_seq_stop(struct seq_file *seq, void *v)
1200 {
1201         if (v)
1202                 put_object(v);
1203 }
1204
1205 /*
1206  * Print the information for an unreferenced object to the seq file.
1207  */
1208 static int kmemleak_seq_show(struct seq_file *seq, void *v)
1209 {
1210         struct kmemleak_object *object = v;
1211         unsigned long flags;
1212
1213         spin_lock_irqsave(&object->lock, flags);
1214         if (!unreferenced_object(object))
1215                 goto out;
1216         print_unreferenced(seq, object);
1217         reported_leaks++;
1218 out:
1219         spin_unlock_irqrestore(&object->lock, flags);
1220         return 0;
1221 }
1222
1223 static const struct seq_operations kmemleak_seq_ops = {
1224         .start = kmemleak_seq_start,
1225         .next  = kmemleak_seq_next,
1226         .stop  = kmemleak_seq_stop,
1227         .show  = kmemleak_seq_show,
1228 };
1229
1230 static int kmemleak_open(struct inode *inode, struct file *file)
1231 {
1232         int ret = 0;
1233
1234         if (!atomic_read(&kmemleak_enabled))
1235                 return -EBUSY;
1236
1237         ret = mutex_lock_interruptible(&kmemleak_mutex);
1238         if (ret < 0)
1239                 goto out;
1240         if (file->f_mode & FMODE_READ) {
1241                 ret = mutex_lock_interruptible(&scan_mutex);
1242                 if (ret < 0)
1243                         goto kmemleak_unlock;
1244                 ret = seq_open(file, &kmemleak_seq_ops);
1245                 if (ret < 0)
1246                         goto scan_unlock;
1247         }
1248         return ret;
1249
1250 scan_unlock:
1251         mutex_unlock(&scan_mutex);
1252 kmemleak_unlock:
1253         mutex_unlock(&kmemleak_mutex);
1254 out:
1255         return ret;
1256 }
1257
1258 static int kmemleak_release(struct inode *inode, struct file *file)
1259 {
1260         int ret = 0;
1261
1262         if (file->f_mode & FMODE_READ) {
1263                 seq_release(inode, file);
1264                 mutex_unlock(&scan_mutex);
1265         }
1266         mutex_unlock(&kmemleak_mutex);
1267
1268         return ret;
1269 }
1270
1271 /*
1272  * File write operation to configure kmemleak at run-time. The following
1273  * commands can be written to the /sys/kernel/debug/kmemleak file:
1274  *   off        - disable kmemleak (irreversible)
1275  *   stack=on   - enable the task stacks scanning
1276  *   stack=off  - disable the tasks stacks scanning
1277  *   scan=on    - start the automatic memory scanning thread
1278  *   scan=off   - stop the automatic memory scanning thread
1279  *   scan=...   - set the automatic memory scanning period in seconds (0 to
1280  *                disable it)
1281  */
1282 static ssize_t kmemleak_write(struct file *file, const char __user *user_buf,
1283                               size_t size, loff_t *ppos)
1284 {
1285         char buf[64];
1286         int buf_size;
1287
1288         if (!atomic_read(&kmemleak_enabled))
1289                 return -EBUSY;
1290
1291         buf_size = min(size, (sizeof(buf) - 1));
1292         if (strncpy_from_user(buf, user_buf, buf_size) < 0)
1293                 return -EFAULT;
1294         buf[buf_size] = 0;
1295
1296         if (strncmp(buf, "off", 3) == 0)
1297                 kmemleak_disable();
1298         else if (strncmp(buf, "stack=on", 8) == 0)
1299                 kmemleak_stack_scan = 1;
1300         else if (strncmp(buf, "stack=off", 9) == 0)
1301                 kmemleak_stack_scan = 0;
1302         else if (strncmp(buf, "scan=on", 7) == 0)
1303                 start_scan_thread();
1304         else if (strncmp(buf, "scan=off", 8) == 0)
1305                 stop_scan_thread();
1306         else if (strncmp(buf, "scan=", 5) == 0) {
1307                 unsigned long secs;
1308                 int err;
1309
1310                 err = strict_strtoul(buf + 5, 0, &secs);
1311                 if (err < 0)
1312                         return err;
1313                 stop_scan_thread();
1314                 if (secs) {
1315                         jiffies_scan_wait = msecs_to_jiffies(secs * 1000);
1316                         start_scan_thread();
1317                 }
1318         } else
1319                 return -EINVAL;
1320
1321         /* ignore the rest of the buffer, only one command at a time */
1322         *ppos += size;
1323         return size;
1324 }
1325
1326 static const struct file_operations kmemleak_fops = {
1327         .owner          = THIS_MODULE,
1328         .open           = kmemleak_open,
1329         .read           = seq_read,
1330         .write          = kmemleak_write,
1331         .llseek         = seq_lseek,
1332         .release        = kmemleak_release,
1333 };
1334
1335 /*
1336  * Perform the freeing of the kmemleak internal objects after waiting for any
1337  * current memory scan to complete.
1338  */
1339 static int kmemleak_cleanup_thread(void *arg)
1340 {
1341         struct kmemleak_object *object;
1342
1343         mutex_lock(&kmemleak_mutex);
1344         stop_scan_thread();
1345         mutex_unlock(&kmemleak_mutex);
1346
1347         mutex_lock(&scan_mutex);
1348         rcu_read_lock();
1349         list_for_each_entry_rcu(object, &object_list, object_list)
1350                 delete_object(object->pointer);
1351         rcu_read_unlock();
1352         mutex_unlock(&scan_mutex);
1353
1354         return 0;
1355 }
1356
1357 /*
1358  * Start the clean-up thread.
1359  */
1360 static void kmemleak_cleanup(void)
1361 {
1362         struct task_struct *cleanup_thread;
1363
1364         cleanup_thread = kthread_run(kmemleak_cleanup_thread, NULL,
1365                                      "kmemleak-clean");
1366         if (IS_ERR(cleanup_thread))
1367                 pr_warning("kmemleak: Failed to create the clean-up thread\n");
1368 }
1369
1370 /*
1371  * Disable kmemleak. No memory allocation/freeing will be traced once this
1372  * function is called. Disabling kmemleak is an irreversible operation.
1373  */
1374 static void kmemleak_disable(void)
1375 {
1376         /* atomically check whether it was already invoked */
1377         if (atomic_cmpxchg(&kmemleak_error, 0, 1))
1378                 return;
1379
1380         /* stop any memory operation tracing */
1381         atomic_set(&kmemleak_early_log, 0);
1382         atomic_set(&kmemleak_enabled, 0);
1383
1384         /* check whether it is too early for a kernel thread */
1385         if (atomic_read(&kmemleak_initialized))
1386                 kmemleak_cleanup();
1387
1388         pr_info("Kernel memory leak detector disabled\n");
1389 }
1390
1391 /*
1392  * Allow boot-time kmemleak disabling (enabled by default).
1393  */
1394 static int kmemleak_boot_config(char *str)
1395 {
1396         if (!str)
1397                 return -EINVAL;
1398         if (strcmp(str, "off") == 0)
1399                 kmemleak_disable();
1400         else if (strcmp(str, "on") != 0)
1401                 return -EINVAL;
1402         return 0;
1403 }
1404 early_param("kmemleak", kmemleak_boot_config);
1405
1406 /*
1407  * Kkmemleak initialization.
1408  */
1409 void __init kmemleak_init(void)
1410 {
1411         int i;
1412         unsigned long flags;
1413
1414         jiffies_scan_yield = msecs_to_jiffies(MSECS_SCAN_YIELD);
1415         jiffies_min_age = msecs_to_jiffies(MSECS_MIN_AGE);
1416         jiffies_scan_wait = msecs_to_jiffies(SECS_SCAN_WAIT * 1000);
1417
1418         object_cache = KMEM_CACHE(kmemleak_object, SLAB_NOLEAKTRACE);
1419         scan_area_cache = KMEM_CACHE(kmemleak_scan_area, SLAB_NOLEAKTRACE);
1420         INIT_PRIO_TREE_ROOT(&object_tree_root);
1421
1422         /* the kernel is still in UP mode, so disabling the IRQs is enough */
1423         local_irq_save(flags);
1424         if (!atomic_read(&kmemleak_error)) {
1425                 atomic_set(&kmemleak_enabled, 1);
1426                 atomic_set(&kmemleak_early_log, 0);
1427         }
1428         local_irq_restore(flags);
1429
1430         /*
1431          * This is the point where tracking allocations is safe. Automatic
1432          * scanning is started during the late initcall. Add the early logged
1433          * callbacks to the kmemleak infrastructure.
1434          */
1435         for (i = 0; i < crt_early_log; i++) {
1436                 struct early_log *log = &early_log[i];
1437
1438                 switch (log->op_type) {
1439                 case KMEMLEAK_ALLOC:
1440                         kmemleak_alloc(log->ptr, log->size, log->min_count,
1441                                        GFP_KERNEL);
1442                         break;
1443                 case KMEMLEAK_FREE:
1444                         kmemleak_free(log->ptr);
1445                         break;
1446                 case KMEMLEAK_NOT_LEAK:
1447                         kmemleak_not_leak(log->ptr);
1448                         break;
1449                 case KMEMLEAK_IGNORE:
1450                         kmemleak_ignore(log->ptr);
1451                         break;
1452                 case KMEMLEAK_SCAN_AREA:
1453                         kmemleak_scan_area(log->ptr, log->offset, log->length,
1454                                            GFP_KERNEL);
1455                         break;
1456                 case KMEMLEAK_NO_SCAN:
1457                         kmemleak_no_scan(log->ptr);
1458                         break;
1459                 default:
1460                         WARN_ON(1);
1461                 }
1462         }
1463 }
1464
1465 /*
1466  * Late initialization function.
1467  */
1468 static int __init kmemleak_late_init(void)
1469 {
1470         struct dentry *dentry;
1471
1472         atomic_set(&kmemleak_initialized, 1);
1473
1474         if (atomic_read(&kmemleak_error)) {
1475                 /*
1476                  * Some error occured and kmemleak was disabled. There is a
1477                  * small chance that kmemleak_disable() was called immediately
1478                  * after setting kmemleak_initialized and we may end up with
1479                  * two clean-up threads but serialized by scan_mutex.
1480                  */
1481                 kmemleak_cleanup();
1482                 return -ENOMEM;
1483         }
1484
1485         dentry = debugfs_create_file("kmemleak", S_IRUGO, NULL, NULL,
1486                                      &kmemleak_fops);
1487         if (!dentry)
1488                 pr_warning("kmemleak: Failed to create the debugfs kmemleak "
1489                            "file\n");
1490         mutex_lock(&kmemleak_mutex);
1491         start_scan_thread();
1492         mutex_unlock(&kmemleak_mutex);
1493
1494         pr_info("Kernel memory leak detector initialized\n");
1495
1496         return 0;
1497 }
1498 late_initcall(kmemleak_late_init);