MIPS: SEAD3: Use symbolic addresses from sead-addr.h in LED driver.
[linux-drm-fsl-dcu.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 red black 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
54  *   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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
65
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/export.h>
73 #include <linux/kthread.h>
74 #include <linux/rbtree.h>
75 #include <linux/fs.h>
76 #include <linux/debugfs.h>
77 #include <linux/seq_file.h>
78 #include <linux/cpumask.h>
79 #include <linux/spinlock.h>
80 #include <linux/mutex.h>
81 #include <linux/rcupdate.h>
82 #include <linux/stacktrace.h>
83 #include <linux/cache.h>
84 #include <linux/percpu.h>
85 #include <linux/hardirq.h>
86 #include <linux/mmzone.h>
87 #include <linux/slab.h>
88 #include <linux/thread_info.h>
89 #include <linux/err.h>
90 #include <linux/uaccess.h>
91 #include <linux/string.h>
92 #include <linux/nodemask.h>
93 #include <linux/mm.h>
94 #include <linux/workqueue.h>
95 #include <linux/crc32.h>
96
97 #include <asm/sections.h>
98 #include <asm/processor.h>
99 #include <linux/atomic.h>
100
101 #include <linux/kasan.h>
102 #include <linux/kmemcheck.h>
103 #include <linux/kmemleak.h>
104 #include <linux/memory_hotplug.h>
105
106 /*
107  * Kmemleak configuration and common defines.
108  */
109 #define MAX_TRACE               16      /* stack trace length */
110 #define MSECS_MIN_AGE           5000    /* minimum object age for reporting */
111 #define SECS_FIRST_SCAN         60      /* delay before the first scan */
112 #define SECS_SCAN_WAIT          600     /* subsequent auto scanning delay */
113 #define MAX_SCAN_SIZE           4096    /* maximum size of a scanned block */
114
115 #define BYTES_PER_POINTER       sizeof(void *)
116
117 /* GFP bitmask for kmemleak internal allocations */
118 #define gfp_kmemleak_mask(gfp)  (((gfp) & (GFP_KERNEL | GFP_ATOMIC)) | \
119                                  __GFP_NORETRY | __GFP_NOMEMALLOC | \
120                                  __GFP_NOWARN)
121
122 /* scanning area inside a memory block */
123 struct kmemleak_scan_area {
124         struct hlist_node node;
125         unsigned long start;
126         size_t size;
127 };
128
129 #define KMEMLEAK_GREY   0
130 #define KMEMLEAK_BLACK  -1
131
132 /*
133  * Structure holding the metadata for each allocated memory block.
134  * Modifications to such objects should be made while holding the
135  * object->lock. Insertions or deletions from object_list, gray_list or
136  * rb_node are already protected by the corresponding locks or mutex (see
137  * the notes on locking above). These objects are reference-counted
138  * (use_count) and freed using the RCU mechanism.
139  */
140 struct kmemleak_object {
141         spinlock_t lock;
142         unsigned long flags;            /* object status flags */
143         struct list_head object_list;
144         struct list_head gray_list;
145         struct rb_node rb_node;
146         struct rcu_head rcu;            /* object_list lockless traversal */
147         /* object usage count; object freed when use_count == 0 */
148         atomic_t use_count;
149         unsigned long pointer;
150         size_t size;
151         /* minimum number of a pointers found before it is considered leak */
152         int min_count;
153         /* the total number of pointers found pointing to this object */
154         int count;
155         /* checksum for detecting modified objects */
156         u32 checksum;
157         /* memory ranges to be scanned inside an object (empty for all) */
158         struct hlist_head area_list;
159         unsigned long trace[MAX_TRACE];
160         unsigned int trace_len;
161         unsigned long jiffies;          /* creation timestamp */
162         pid_t pid;                      /* pid of the current task */
163         char comm[TASK_COMM_LEN];       /* executable name */
164 };
165
166 /* flag representing the memory block allocation status */
167 #define OBJECT_ALLOCATED        (1 << 0)
168 /* flag set after the first reporting of an unreference object */
169 #define OBJECT_REPORTED         (1 << 1)
170 /* flag set to not scan the object */
171 #define OBJECT_NO_SCAN          (1 << 2)
172
173 /* number of bytes to print per line; must be 16 or 32 */
174 #define HEX_ROW_SIZE            16
175 /* number of bytes to print at a time (1, 2, 4, 8) */
176 #define HEX_GROUP_SIZE          1
177 /* include ASCII after the hex output */
178 #define HEX_ASCII               1
179 /* max number of lines to be printed */
180 #define HEX_MAX_LINES           2
181
182 /* the list of all allocated objects */
183 static LIST_HEAD(object_list);
184 /* the list of gray-colored objects (see color_gray comment below) */
185 static LIST_HEAD(gray_list);
186 /* search tree for object boundaries */
187 static struct rb_root object_tree_root = RB_ROOT;
188 /* rw_lock protecting the access to object_list and object_tree_root */
189 static DEFINE_RWLOCK(kmemleak_lock);
190
191 /* allocation caches for kmemleak internal data */
192 static struct kmem_cache *object_cache;
193 static struct kmem_cache *scan_area_cache;
194
195 /* set if tracing memory operations is enabled */
196 static int kmemleak_enabled;
197 /* set in the late_initcall if there were no errors */
198 static int kmemleak_initialized;
199 /* enables or disables early logging of the memory operations */
200 static int kmemleak_early_log = 1;
201 /* set if a kmemleak warning was issued */
202 static int kmemleak_warning;
203 /* set if a fatal kmemleak error has occurred */
204 static int kmemleak_error;
205
206 /* minimum and maximum address that may be valid pointers */
207 static unsigned long min_addr = ULONG_MAX;
208 static unsigned long max_addr;
209
210 static struct task_struct *scan_thread;
211 /* used to avoid reporting of recently allocated objects */
212 static unsigned long jiffies_min_age;
213 static unsigned long jiffies_last_scan;
214 /* delay between automatic memory scannings */
215 static signed long jiffies_scan_wait;
216 /* enables or disables the task stacks scanning */
217 static int kmemleak_stack_scan = 1;
218 /* protects the memory scanning, parameters and debug/kmemleak file access */
219 static DEFINE_MUTEX(scan_mutex);
220 /* setting kmemleak=on, will set this var, skipping the disable */
221 static int kmemleak_skip_disable;
222 /* If there are leaks that can be reported */
223 static bool kmemleak_found_leaks;
224
225 /*
226  * Early object allocation/freeing logging. Kmemleak is initialized after the
227  * kernel allocator. However, both the kernel allocator and kmemleak may
228  * allocate memory blocks which need to be tracked. Kmemleak defines an
229  * arbitrary buffer to hold the allocation/freeing information before it is
230  * fully initialized.
231  */
232
233 /* kmemleak operation type for early logging */
234 enum {
235         KMEMLEAK_ALLOC,
236         KMEMLEAK_ALLOC_PERCPU,
237         KMEMLEAK_FREE,
238         KMEMLEAK_FREE_PART,
239         KMEMLEAK_FREE_PERCPU,
240         KMEMLEAK_NOT_LEAK,
241         KMEMLEAK_IGNORE,
242         KMEMLEAK_SCAN_AREA,
243         KMEMLEAK_NO_SCAN
244 };
245
246 /*
247  * Structure holding the information passed to kmemleak callbacks during the
248  * early logging.
249  */
250 struct early_log {
251         int op_type;                    /* kmemleak operation type */
252         const void *ptr;                /* allocated/freed memory block */
253         size_t size;                    /* memory block size */
254         int min_count;                  /* minimum reference count */
255         unsigned long trace[MAX_TRACE]; /* stack trace */
256         unsigned int trace_len;         /* stack trace length */
257 };
258
259 /* early logging buffer and current position */
260 static struct early_log
261         early_log[CONFIG_DEBUG_KMEMLEAK_EARLY_LOG_SIZE] __initdata;
262 static int crt_early_log __initdata;
263
264 static void kmemleak_disable(void);
265
266 /*
267  * Print a warning and dump the stack trace.
268  */
269 #define kmemleak_warn(x...)     do {            \
270         pr_warning(x);                          \
271         dump_stack();                           \
272         kmemleak_warning = 1;                   \
273 } while (0)
274
275 /*
276  * Macro invoked when a serious kmemleak condition occurred and cannot be
277  * recovered from. Kmemleak will be disabled and further allocation/freeing
278  * tracing no longer available.
279  */
280 #define kmemleak_stop(x...)     do {    \
281         kmemleak_warn(x);               \
282         kmemleak_disable();             \
283 } while (0)
284
285 /*
286  * Printing of the objects hex dump to the seq file. The number of lines to be
287  * printed is limited to HEX_MAX_LINES to prevent seq file spamming. The
288  * actual number of printed bytes depends on HEX_ROW_SIZE. It must be called
289  * with the object->lock held.
290  */
291 static void hex_dump_object(struct seq_file *seq,
292                             struct kmemleak_object *object)
293 {
294         const u8 *ptr = (const u8 *)object->pointer;
295         int i, len, remaining;
296         unsigned char linebuf[HEX_ROW_SIZE * 5];
297
298         /* limit the number of lines to HEX_MAX_LINES */
299         remaining = len =
300                 min(object->size, (size_t)(HEX_MAX_LINES * HEX_ROW_SIZE));
301
302         seq_printf(seq, "  hex dump (first %d bytes):\n", len);
303         for (i = 0; i < len; i += HEX_ROW_SIZE) {
304                 int linelen = min(remaining, HEX_ROW_SIZE);
305
306                 remaining -= HEX_ROW_SIZE;
307                 hex_dump_to_buffer(ptr + i, linelen, HEX_ROW_SIZE,
308                                    HEX_GROUP_SIZE, linebuf, sizeof(linebuf),
309                                    HEX_ASCII);
310                 seq_printf(seq, "    %s\n", linebuf);
311         }
312 }
313
314 /*
315  * Object colors, encoded with count and min_count:
316  * - white - orphan object, not enough references to it (count < min_count)
317  * - gray  - not orphan, not marked as false positive (min_count == 0) or
318  *              sufficient references to it (count >= min_count)
319  * - black - ignore, it doesn't contain references (e.g. text section)
320  *              (min_count == -1). No function defined for this color.
321  * Newly created objects don't have any color assigned (object->count == -1)
322  * before the next memory scan when they become white.
323  */
324 static bool color_white(const struct kmemleak_object *object)
325 {
326         return object->count != KMEMLEAK_BLACK &&
327                 object->count < object->min_count;
328 }
329
330 static bool color_gray(const struct kmemleak_object *object)
331 {
332         return object->min_count != KMEMLEAK_BLACK &&
333                 object->count >= object->min_count;
334 }
335
336 /*
337  * Objects are considered unreferenced only if their color is white, they have
338  * not be deleted and have a minimum age to avoid false positives caused by
339  * pointers temporarily stored in CPU registers.
340  */
341 static bool unreferenced_object(struct kmemleak_object *object)
342 {
343         return (color_white(object) && object->flags & OBJECT_ALLOCATED) &&
344                 time_before_eq(object->jiffies + jiffies_min_age,
345                                jiffies_last_scan);
346 }
347
348 /*
349  * Printing of the unreferenced objects information to the seq file. The
350  * print_unreferenced function must be called with the object->lock held.
351  */
352 static void print_unreferenced(struct seq_file *seq,
353                                struct kmemleak_object *object)
354 {
355         int i;
356         unsigned int msecs_age = jiffies_to_msecs(jiffies - object->jiffies);
357
358         seq_printf(seq, "unreferenced object 0x%08lx (size %zu):\n",
359                    object->pointer, object->size);
360         seq_printf(seq, "  comm \"%s\", pid %d, jiffies %lu (age %d.%03ds)\n",
361                    object->comm, object->pid, object->jiffies,
362                    msecs_age / 1000, msecs_age % 1000);
363         hex_dump_object(seq, object);
364         seq_printf(seq, "  backtrace:\n");
365
366         for (i = 0; i < object->trace_len; i++) {
367                 void *ptr = (void *)object->trace[i];
368                 seq_printf(seq, "    [<%p>] %pS\n", ptr, ptr);
369         }
370 }
371
372 /*
373  * Print the kmemleak_object information. This function is used mainly for
374  * debugging special cases when kmemleak operations. It must be called with
375  * the object->lock held.
376  */
377 static void dump_object_info(struct kmemleak_object *object)
378 {
379         struct stack_trace trace;
380
381         trace.nr_entries = object->trace_len;
382         trace.entries = object->trace;
383
384         pr_notice("Object 0x%08lx (size %zu):\n",
385                   object->pointer, object->size);
386         pr_notice("  comm \"%s\", pid %d, jiffies %lu\n",
387                   object->comm, object->pid, object->jiffies);
388         pr_notice("  min_count = %d\n", object->min_count);
389         pr_notice("  count = %d\n", object->count);
390         pr_notice("  flags = 0x%lx\n", object->flags);
391         pr_notice("  checksum = %u\n", object->checksum);
392         pr_notice("  backtrace:\n");
393         print_stack_trace(&trace, 4);
394 }
395
396 /*
397  * Look-up a memory block metadata (kmemleak_object) in the object search
398  * tree based on a pointer value. If alias is 0, only values pointing to the
399  * beginning of the memory block are allowed. The kmemleak_lock must be held
400  * when calling this function.
401  */
402 static struct kmemleak_object *lookup_object(unsigned long ptr, int alias)
403 {
404         struct rb_node *rb = object_tree_root.rb_node;
405
406         while (rb) {
407                 struct kmemleak_object *object =
408                         rb_entry(rb, struct kmemleak_object, rb_node);
409                 if (ptr < object->pointer)
410                         rb = object->rb_node.rb_left;
411                 else if (object->pointer + object->size <= ptr)
412                         rb = object->rb_node.rb_right;
413                 else if (object->pointer == ptr || alias)
414                         return object;
415                 else {
416                         kmemleak_warn("Found object by alias at 0x%08lx\n",
417                                       ptr);
418                         dump_object_info(object);
419                         break;
420                 }
421         }
422         return NULL;
423 }
424
425 /*
426  * Increment the object use_count. Return 1 if successful or 0 otherwise. Note
427  * that once an object's use_count reached 0, the RCU freeing was already
428  * registered and the object should no longer be used. This function must be
429  * called under the protection of rcu_read_lock().
430  */
431 static int get_object(struct kmemleak_object *object)
432 {
433         return atomic_inc_not_zero(&object->use_count);
434 }
435
436 /*
437  * RCU callback to free a kmemleak_object.
438  */
439 static void free_object_rcu(struct rcu_head *rcu)
440 {
441         struct hlist_node *tmp;
442         struct kmemleak_scan_area *area;
443         struct kmemleak_object *object =
444                 container_of(rcu, struct kmemleak_object, rcu);
445
446         /*
447          * Once use_count is 0 (guaranteed by put_object), there is no other
448          * code accessing this object, hence no need for locking.
449          */
450         hlist_for_each_entry_safe(area, tmp, &object->area_list, node) {
451                 hlist_del(&area->node);
452                 kmem_cache_free(scan_area_cache, area);
453         }
454         kmem_cache_free(object_cache, object);
455 }
456
457 /*
458  * Decrement the object use_count. Once the count is 0, free the object using
459  * an RCU callback. Since put_object() may be called via the kmemleak_free() ->
460  * delete_object() path, the delayed RCU freeing ensures that there is no
461  * recursive call to the kernel allocator. Lock-less RCU object_list traversal
462  * is also possible.
463  */
464 static void put_object(struct kmemleak_object *object)
465 {
466         if (!atomic_dec_and_test(&object->use_count))
467                 return;
468
469         /* should only get here after delete_object was called */
470         WARN_ON(object->flags & OBJECT_ALLOCATED);
471
472         call_rcu(&object->rcu, free_object_rcu);
473 }
474
475 /*
476  * Look up an object in the object search tree and increase its use_count.
477  */
478 static struct kmemleak_object *find_and_get_object(unsigned long ptr, int alias)
479 {
480         unsigned long flags;
481         struct kmemleak_object *object = NULL;
482
483         rcu_read_lock();
484         read_lock_irqsave(&kmemleak_lock, flags);
485         if (ptr >= min_addr && ptr < max_addr)
486                 object = lookup_object(ptr, alias);
487         read_unlock_irqrestore(&kmemleak_lock, flags);
488
489         /* check whether the object is still available */
490         if (object && !get_object(object))
491                 object = NULL;
492         rcu_read_unlock();
493
494         return object;
495 }
496
497 /*
498  * Save stack trace to the given array of MAX_TRACE size.
499  */
500 static int __save_stack_trace(unsigned long *trace)
501 {
502         struct stack_trace stack_trace;
503
504         stack_trace.max_entries = MAX_TRACE;
505         stack_trace.nr_entries = 0;
506         stack_trace.entries = trace;
507         stack_trace.skip = 2;
508         save_stack_trace(&stack_trace);
509
510         return stack_trace.nr_entries;
511 }
512
513 /*
514  * Create the metadata (struct kmemleak_object) corresponding to an allocated
515  * memory block and add it to the object_list and object_tree_root.
516  */
517 static struct kmemleak_object *create_object(unsigned long ptr, size_t size,
518                                              int min_count, gfp_t gfp)
519 {
520         unsigned long flags;
521         struct kmemleak_object *object, *parent;
522         struct rb_node **link, *rb_parent;
523
524         object = kmem_cache_alloc(object_cache, gfp_kmemleak_mask(gfp));
525         if (!object) {
526                 pr_warning("Cannot allocate a kmemleak_object structure\n");
527                 kmemleak_disable();
528                 return NULL;
529         }
530
531         INIT_LIST_HEAD(&object->object_list);
532         INIT_LIST_HEAD(&object->gray_list);
533         INIT_HLIST_HEAD(&object->area_list);
534         spin_lock_init(&object->lock);
535         atomic_set(&object->use_count, 1);
536         object->flags = OBJECT_ALLOCATED;
537         object->pointer = ptr;
538         object->size = size;
539         object->min_count = min_count;
540         object->count = 0;                      /* white color initially */
541         object->jiffies = jiffies;
542         object->checksum = 0;
543
544         /* task information */
545         if (in_irq()) {
546                 object->pid = 0;
547                 strncpy(object->comm, "hardirq", sizeof(object->comm));
548         } else if (in_softirq()) {
549                 object->pid = 0;
550                 strncpy(object->comm, "softirq", sizeof(object->comm));
551         } else {
552                 object->pid = current->pid;
553                 /*
554                  * There is a small chance of a race with set_task_comm(),
555                  * however using get_task_comm() here may cause locking
556                  * dependency issues with current->alloc_lock. In the worst
557                  * case, the command line is not correct.
558                  */
559                 strncpy(object->comm, current->comm, sizeof(object->comm));
560         }
561
562         /* kernel backtrace */
563         object->trace_len = __save_stack_trace(object->trace);
564
565         write_lock_irqsave(&kmemleak_lock, flags);
566
567         min_addr = min(min_addr, ptr);
568         max_addr = max(max_addr, ptr + size);
569         link = &object_tree_root.rb_node;
570         rb_parent = NULL;
571         while (*link) {
572                 rb_parent = *link;
573                 parent = rb_entry(rb_parent, struct kmemleak_object, rb_node);
574                 if (ptr + size <= parent->pointer)
575                         link = &parent->rb_node.rb_left;
576                 else if (parent->pointer + parent->size <= ptr)
577                         link = &parent->rb_node.rb_right;
578                 else {
579                         kmemleak_stop("Cannot insert 0x%lx into the object "
580                                       "search tree (overlaps existing)\n",
581                                       ptr);
582                         kmem_cache_free(object_cache, object);
583                         object = parent;
584                         spin_lock(&object->lock);
585                         dump_object_info(object);
586                         spin_unlock(&object->lock);
587                         goto out;
588                 }
589         }
590         rb_link_node(&object->rb_node, rb_parent, link);
591         rb_insert_color(&object->rb_node, &object_tree_root);
592
593         list_add_tail_rcu(&object->object_list, &object_list);
594 out:
595         write_unlock_irqrestore(&kmemleak_lock, flags);
596         return object;
597 }
598
599 /*
600  * Remove the metadata (struct kmemleak_object) for a memory block from the
601  * object_list and object_tree_root and decrement its use_count.
602  */
603 static void __delete_object(struct kmemleak_object *object)
604 {
605         unsigned long flags;
606
607         write_lock_irqsave(&kmemleak_lock, flags);
608         rb_erase(&object->rb_node, &object_tree_root);
609         list_del_rcu(&object->object_list);
610         write_unlock_irqrestore(&kmemleak_lock, flags);
611
612         WARN_ON(!(object->flags & OBJECT_ALLOCATED));
613         WARN_ON(atomic_read(&object->use_count) < 2);
614
615         /*
616          * Locking here also ensures that the corresponding memory block
617          * cannot be freed when it is being scanned.
618          */
619         spin_lock_irqsave(&object->lock, flags);
620         object->flags &= ~OBJECT_ALLOCATED;
621         spin_unlock_irqrestore(&object->lock, flags);
622         put_object(object);
623 }
624
625 /*
626  * Look up the metadata (struct kmemleak_object) corresponding to ptr and
627  * delete it.
628  */
629 static void delete_object_full(unsigned long ptr)
630 {
631         struct kmemleak_object *object;
632
633         object = find_and_get_object(ptr, 0);
634         if (!object) {
635 #ifdef DEBUG
636                 kmemleak_warn("Freeing unknown object at 0x%08lx\n",
637                               ptr);
638 #endif
639                 return;
640         }
641         __delete_object(object);
642         put_object(object);
643 }
644
645 /*
646  * Look up the metadata (struct kmemleak_object) corresponding to ptr and
647  * delete it. If the memory block is partially freed, the function may create
648  * additional metadata for the remaining parts of the block.
649  */
650 static void delete_object_part(unsigned long ptr, size_t size)
651 {
652         struct kmemleak_object *object;
653         unsigned long start, end;
654
655         object = find_and_get_object(ptr, 1);
656         if (!object) {
657 #ifdef DEBUG
658                 kmemleak_warn("Partially freeing unknown object at 0x%08lx "
659                               "(size %zu)\n", ptr, size);
660 #endif
661                 return;
662         }
663         __delete_object(object);
664
665         /*
666          * Create one or two objects that may result from the memory block
667          * split. Note that partial freeing is only done by free_bootmem() and
668          * this happens before kmemleak_init() is called. The path below is
669          * only executed during early log recording in kmemleak_init(), so
670          * GFP_KERNEL is enough.
671          */
672         start = object->pointer;
673         end = object->pointer + object->size;
674         if (ptr > start)
675                 create_object(start, ptr - start, object->min_count,
676                               GFP_KERNEL);
677         if (ptr + size < end)
678                 create_object(ptr + size, end - ptr - size, object->min_count,
679                               GFP_KERNEL);
680
681         put_object(object);
682 }
683
684 static void __paint_it(struct kmemleak_object *object, int color)
685 {
686         object->min_count = color;
687         if (color == KMEMLEAK_BLACK)
688                 object->flags |= OBJECT_NO_SCAN;
689 }
690
691 static void paint_it(struct kmemleak_object *object, int color)
692 {
693         unsigned long flags;
694
695         spin_lock_irqsave(&object->lock, flags);
696         __paint_it(object, color);
697         spin_unlock_irqrestore(&object->lock, flags);
698 }
699
700 static void paint_ptr(unsigned long ptr, int color)
701 {
702         struct kmemleak_object *object;
703
704         object = find_and_get_object(ptr, 0);
705         if (!object) {
706                 kmemleak_warn("Trying to color unknown object "
707                               "at 0x%08lx as %s\n", ptr,
708                               (color == KMEMLEAK_GREY) ? "Grey" :
709                               (color == KMEMLEAK_BLACK) ? "Black" : "Unknown");
710                 return;
711         }
712         paint_it(object, color);
713         put_object(object);
714 }
715
716 /*
717  * Mark an object permanently as gray-colored so that it can no longer be
718  * reported as a leak. This is used in general to mark a false positive.
719  */
720 static void make_gray_object(unsigned long ptr)
721 {
722         paint_ptr(ptr, KMEMLEAK_GREY);
723 }
724
725 /*
726  * Mark the object as black-colored so that it is ignored from scans and
727  * reporting.
728  */
729 static void make_black_object(unsigned long ptr)
730 {
731         paint_ptr(ptr, KMEMLEAK_BLACK);
732 }
733
734 /*
735  * Add a scanning area to the object. If at least one such area is added,
736  * kmemleak will only scan these ranges rather than the whole memory block.
737  */
738 static void add_scan_area(unsigned long ptr, size_t size, gfp_t gfp)
739 {
740         unsigned long flags;
741         struct kmemleak_object *object;
742         struct kmemleak_scan_area *area;
743
744         object = find_and_get_object(ptr, 1);
745         if (!object) {
746                 kmemleak_warn("Adding scan area to unknown object at 0x%08lx\n",
747                               ptr);
748                 return;
749         }
750
751         area = kmem_cache_alloc(scan_area_cache, gfp_kmemleak_mask(gfp));
752         if (!area) {
753                 pr_warning("Cannot allocate a scan area\n");
754                 goto out;
755         }
756
757         spin_lock_irqsave(&object->lock, flags);
758         if (size == SIZE_MAX) {
759                 size = object->pointer + object->size - ptr;
760         } else if (ptr + size > object->pointer + object->size) {
761                 kmemleak_warn("Scan area larger than object 0x%08lx\n", ptr);
762                 dump_object_info(object);
763                 kmem_cache_free(scan_area_cache, area);
764                 goto out_unlock;
765         }
766
767         INIT_HLIST_NODE(&area->node);
768         area->start = ptr;
769         area->size = size;
770
771         hlist_add_head(&area->node, &object->area_list);
772 out_unlock:
773         spin_unlock_irqrestore(&object->lock, flags);
774 out:
775         put_object(object);
776 }
777
778 /*
779  * Set the OBJECT_NO_SCAN flag for the object corresponding to the give
780  * pointer. Such object will not be scanned by kmemleak but references to it
781  * are searched.
782  */
783 static void object_no_scan(unsigned long ptr)
784 {
785         unsigned long flags;
786         struct kmemleak_object *object;
787
788         object = find_and_get_object(ptr, 0);
789         if (!object) {
790                 kmemleak_warn("Not scanning unknown object at 0x%08lx\n", ptr);
791                 return;
792         }
793
794         spin_lock_irqsave(&object->lock, flags);
795         object->flags |= OBJECT_NO_SCAN;
796         spin_unlock_irqrestore(&object->lock, flags);
797         put_object(object);
798 }
799
800 /*
801  * Log an early kmemleak_* call to the early_log buffer. These calls will be
802  * processed later once kmemleak is fully initialized.
803  */
804 static void __init log_early(int op_type, const void *ptr, size_t size,
805                              int min_count)
806 {
807         unsigned long flags;
808         struct early_log *log;
809
810         if (kmemleak_error) {
811                 /* kmemleak stopped recording, just count the requests */
812                 crt_early_log++;
813                 return;
814         }
815
816         if (crt_early_log >= ARRAY_SIZE(early_log)) {
817                 kmemleak_disable();
818                 return;
819         }
820
821         /*
822          * There is no need for locking since the kernel is still in UP mode
823          * at this stage. Disabling the IRQs is enough.
824          */
825         local_irq_save(flags);
826         log = &early_log[crt_early_log];
827         log->op_type = op_type;
828         log->ptr = ptr;
829         log->size = size;
830         log->min_count = min_count;
831         log->trace_len = __save_stack_trace(log->trace);
832         crt_early_log++;
833         local_irq_restore(flags);
834 }
835
836 /*
837  * Log an early allocated block and populate the stack trace.
838  */
839 static void early_alloc(struct early_log *log)
840 {
841         struct kmemleak_object *object;
842         unsigned long flags;
843         int i;
844
845         if (!kmemleak_enabled || !log->ptr || IS_ERR(log->ptr))
846                 return;
847
848         /*
849          * RCU locking needed to ensure object is not freed via put_object().
850          */
851         rcu_read_lock();
852         object = create_object((unsigned long)log->ptr, log->size,
853                                log->min_count, GFP_ATOMIC);
854         if (!object)
855                 goto out;
856         spin_lock_irqsave(&object->lock, flags);
857         for (i = 0; i < log->trace_len; i++)
858                 object->trace[i] = log->trace[i];
859         object->trace_len = log->trace_len;
860         spin_unlock_irqrestore(&object->lock, flags);
861 out:
862         rcu_read_unlock();
863 }
864
865 /*
866  * Log an early allocated block and populate the stack trace.
867  */
868 static void early_alloc_percpu(struct early_log *log)
869 {
870         unsigned int cpu;
871         const void __percpu *ptr = log->ptr;
872
873         for_each_possible_cpu(cpu) {
874                 log->ptr = per_cpu_ptr(ptr, cpu);
875                 early_alloc(log);
876         }
877 }
878
879 /**
880  * kmemleak_alloc - register a newly allocated object
881  * @ptr:        pointer to beginning of the object
882  * @size:       size of the object
883  * @min_count:  minimum number of references to this object. If during memory
884  *              scanning a number of references less than @min_count is found,
885  *              the object is reported as a memory leak. If @min_count is 0,
886  *              the object is never reported as a leak. If @min_count is -1,
887  *              the object is ignored (not scanned and not reported as a leak)
888  * @gfp:        kmalloc() flags used for kmemleak internal memory allocations
889  *
890  * This function is called from the kernel allocators when a new object
891  * (memory block) is allocated (kmem_cache_alloc, kmalloc, vmalloc etc.).
892  */
893 void __ref kmemleak_alloc(const void *ptr, size_t size, int min_count,
894                           gfp_t gfp)
895 {
896         pr_debug("%s(0x%p, %zu, %d)\n", __func__, ptr, size, min_count);
897
898         if (kmemleak_enabled && ptr && !IS_ERR(ptr))
899                 create_object((unsigned long)ptr, size, min_count, gfp);
900         else if (kmemleak_early_log)
901                 log_early(KMEMLEAK_ALLOC, ptr, size, min_count);
902 }
903 EXPORT_SYMBOL_GPL(kmemleak_alloc);
904
905 /**
906  * kmemleak_alloc_percpu - register a newly allocated __percpu object
907  * @ptr:        __percpu pointer to beginning of the object
908  * @size:       size of the object
909  *
910  * This function is called from the kernel percpu allocator when a new object
911  * (memory block) is allocated (alloc_percpu). It assumes GFP_KERNEL
912  * allocation.
913  */
914 void __ref kmemleak_alloc_percpu(const void __percpu *ptr, size_t size)
915 {
916         unsigned int cpu;
917
918         pr_debug("%s(0x%p, %zu)\n", __func__, ptr, size);
919
920         /*
921          * Percpu allocations are only scanned and not reported as leaks
922          * (min_count is set to 0).
923          */
924         if (kmemleak_enabled && ptr && !IS_ERR(ptr))
925                 for_each_possible_cpu(cpu)
926                         create_object((unsigned long)per_cpu_ptr(ptr, cpu),
927                                       size, 0, GFP_KERNEL);
928         else if (kmemleak_early_log)
929                 log_early(KMEMLEAK_ALLOC_PERCPU, ptr, size, 0);
930 }
931 EXPORT_SYMBOL_GPL(kmemleak_alloc_percpu);
932
933 /**
934  * kmemleak_free - unregister a previously registered object
935  * @ptr:        pointer to beginning of the object
936  *
937  * This function is called from the kernel allocators when an object (memory
938  * block) is freed (kmem_cache_free, kfree, vfree etc.).
939  */
940 void __ref kmemleak_free(const void *ptr)
941 {
942         pr_debug("%s(0x%p)\n", __func__, ptr);
943
944         if (kmemleak_enabled && ptr && !IS_ERR(ptr))
945                 delete_object_full((unsigned long)ptr);
946         else if (kmemleak_early_log)
947                 log_early(KMEMLEAK_FREE, ptr, 0, 0);
948 }
949 EXPORT_SYMBOL_GPL(kmemleak_free);
950
951 /**
952  * kmemleak_free_part - partially unregister a previously registered object
953  * @ptr:        pointer to the beginning or inside the object. This also
954  *              represents the start of the range to be freed
955  * @size:       size to be unregistered
956  *
957  * This function is called when only a part of a memory block is freed
958  * (usually from the bootmem allocator).
959  */
960 void __ref kmemleak_free_part(const void *ptr, size_t size)
961 {
962         pr_debug("%s(0x%p)\n", __func__, ptr);
963
964         if (kmemleak_enabled && ptr && !IS_ERR(ptr))
965                 delete_object_part((unsigned long)ptr, size);
966         else if (kmemleak_early_log)
967                 log_early(KMEMLEAK_FREE_PART, ptr, size, 0);
968 }
969 EXPORT_SYMBOL_GPL(kmemleak_free_part);
970
971 /**
972  * kmemleak_free_percpu - unregister a previously registered __percpu object
973  * @ptr:        __percpu pointer to beginning of the object
974  *
975  * This function is called from the kernel percpu allocator when an object
976  * (memory block) is freed (free_percpu).
977  */
978 void __ref kmemleak_free_percpu(const void __percpu *ptr)
979 {
980         unsigned int cpu;
981
982         pr_debug("%s(0x%p)\n", __func__, ptr);
983
984         if (kmemleak_enabled && ptr && !IS_ERR(ptr))
985                 for_each_possible_cpu(cpu)
986                         delete_object_full((unsigned long)per_cpu_ptr(ptr,
987                                                                       cpu));
988         else if (kmemleak_early_log)
989                 log_early(KMEMLEAK_FREE_PERCPU, ptr, 0, 0);
990 }
991 EXPORT_SYMBOL_GPL(kmemleak_free_percpu);
992
993 /**
994  * kmemleak_update_trace - update object allocation stack trace
995  * @ptr:        pointer to beginning of the object
996  *
997  * Override the object allocation stack trace for cases where the actual
998  * allocation place is not always useful.
999  */
1000 void __ref kmemleak_update_trace(const void *ptr)
1001 {
1002         struct kmemleak_object *object;
1003         unsigned long flags;
1004
1005         pr_debug("%s(0x%p)\n", __func__, ptr);
1006
1007         if (!kmemleak_enabled || IS_ERR_OR_NULL(ptr))
1008                 return;
1009
1010         object = find_and_get_object((unsigned long)ptr, 1);
1011         if (!object) {
1012 #ifdef DEBUG
1013                 kmemleak_warn("Updating stack trace for unknown object at %p\n",
1014                               ptr);
1015 #endif
1016                 return;
1017         }
1018
1019         spin_lock_irqsave(&object->lock, flags);
1020         object->trace_len = __save_stack_trace(object->trace);
1021         spin_unlock_irqrestore(&object->lock, flags);
1022
1023         put_object(object);
1024 }
1025 EXPORT_SYMBOL(kmemleak_update_trace);
1026
1027 /**
1028  * kmemleak_not_leak - mark an allocated object as false positive
1029  * @ptr:        pointer to beginning of the object
1030  *
1031  * Calling this function on an object will cause the memory block to no longer
1032  * be reported as leak and always be scanned.
1033  */
1034 void __ref kmemleak_not_leak(const void *ptr)
1035 {
1036         pr_debug("%s(0x%p)\n", __func__, ptr);
1037
1038         if (kmemleak_enabled && ptr && !IS_ERR(ptr))
1039                 make_gray_object((unsigned long)ptr);
1040         else if (kmemleak_early_log)
1041                 log_early(KMEMLEAK_NOT_LEAK, ptr, 0, 0);
1042 }
1043 EXPORT_SYMBOL(kmemleak_not_leak);
1044
1045 /**
1046  * kmemleak_ignore - ignore an allocated object
1047  * @ptr:        pointer to beginning of the object
1048  *
1049  * Calling this function on an object will cause the memory block to be
1050  * ignored (not scanned and not reported as a leak). This is usually done when
1051  * it is known that the corresponding block is not a leak and does not contain
1052  * any references to other allocated memory blocks.
1053  */
1054 void __ref kmemleak_ignore(const void *ptr)
1055 {
1056         pr_debug("%s(0x%p)\n", __func__, ptr);
1057
1058         if (kmemleak_enabled && ptr && !IS_ERR(ptr))
1059                 make_black_object((unsigned long)ptr);
1060         else if (kmemleak_early_log)
1061                 log_early(KMEMLEAK_IGNORE, ptr, 0, 0);
1062 }
1063 EXPORT_SYMBOL(kmemleak_ignore);
1064
1065 /**
1066  * kmemleak_scan_area - limit the range to be scanned in an allocated object
1067  * @ptr:        pointer to beginning or inside the object. This also
1068  *              represents the start of the scan area
1069  * @size:       size of the scan area
1070  * @gfp:        kmalloc() flags used for kmemleak internal memory allocations
1071  *
1072  * This function is used when it is known that only certain parts of an object
1073  * contain references to other objects. Kmemleak will only scan these areas
1074  * reducing the number false negatives.
1075  */
1076 void __ref kmemleak_scan_area(const void *ptr, size_t size, gfp_t gfp)
1077 {
1078         pr_debug("%s(0x%p)\n", __func__, ptr);
1079
1080         if (kmemleak_enabled && ptr && size && !IS_ERR(ptr))
1081                 add_scan_area((unsigned long)ptr, size, gfp);
1082         else if (kmemleak_early_log)
1083                 log_early(KMEMLEAK_SCAN_AREA, ptr, size, 0);
1084 }
1085 EXPORT_SYMBOL(kmemleak_scan_area);
1086
1087 /**
1088  * kmemleak_no_scan - do not scan an allocated object
1089  * @ptr:        pointer to beginning of the object
1090  *
1091  * This function notifies kmemleak not to scan the given memory block. Useful
1092  * in situations where it is known that the given object does not contain any
1093  * references to other objects. Kmemleak will not scan such objects reducing
1094  * the number of false negatives.
1095  */
1096 void __ref kmemleak_no_scan(const void *ptr)
1097 {
1098         pr_debug("%s(0x%p)\n", __func__, ptr);
1099
1100         if (kmemleak_enabled && ptr && !IS_ERR(ptr))
1101                 object_no_scan((unsigned long)ptr);
1102         else if (kmemleak_early_log)
1103                 log_early(KMEMLEAK_NO_SCAN, ptr, 0, 0);
1104 }
1105 EXPORT_SYMBOL(kmemleak_no_scan);
1106
1107 /*
1108  * Update an object's checksum and return true if it was modified.
1109  */
1110 static bool update_checksum(struct kmemleak_object *object)
1111 {
1112         u32 old_csum = object->checksum;
1113
1114         if (!kmemcheck_is_obj_initialized(object->pointer, object->size))
1115                 return false;
1116
1117         kasan_disable_current();
1118         object->checksum = crc32(0, (void *)object->pointer, object->size);
1119         kasan_enable_current();
1120
1121         return object->checksum != old_csum;
1122 }
1123
1124 /*
1125  * Memory scanning is a long process and it needs to be interruptable. This
1126  * function checks whether such interrupt condition occurred.
1127  */
1128 static int scan_should_stop(void)
1129 {
1130         if (!kmemleak_enabled)
1131                 return 1;
1132
1133         /*
1134          * This function may be called from either process or kthread context,
1135          * hence the need to check for both stop conditions.
1136          */
1137         if (current->mm)
1138                 return signal_pending(current);
1139         else
1140                 return kthread_should_stop();
1141
1142         return 0;
1143 }
1144
1145 /*
1146  * Scan a memory block (exclusive range) for valid pointers and add those
1147  * found to the gray list.
1148  */
1149 static void scan_block(void *_start, void *_end,
1150                        struct kmemleak_object *scanned, int allow_resched)
1151 {
1152         unsigned long *ptr;
1153         unsigned long *start = PTR_ALIGN(_start, BYTES_PER_POINTER);
1154         unsigned long *end = _end - (BYTES_PER_POINTER - 1);
1155
1156         for (ptr = start; ptr < end; ptr++) {
1157                 struct kmemleak_object *object;
1158                 unsigned long flags;
1159                 unsigned long pointer;
1160
1161                 if (allow_resched)
1162                         cond_resched();
1163                 if (scan_should_stop())
1164                         break;
1165
1166                 /* don't scan uninitialized memory */
1167                 if (!kmemcheck_is_obj_initialized((unsigned long)ptr,
1168                                                   BYTES_PER_POINTER))
1169                         continue;
1170
1171                 kasan_disable_current();
1172                 pointer = *ptr;
1173                 kasan_enable_current();
1174
1175                 object = find_and_get_object(pointer, 1);
1176                 if (!object)
1177                         continue;
1178                 if (object == scanned) {
1179                         /* self referenced, ignore */
1180                         put_object(object);
1181                         continue;
1182                 }
1183
1184                 /*
1185                  * Avoid the lockdep recursive warning on object->lock being
1186                  * previously acquired in scan_object(). These locks are
1187                  * enclosed by scan_mutex.
1188                  */
1189                 spin_lock_irqsave_nested(&object->lock, flags,
1190                                          SINGLE_DEPTH_NESTING);
1191                 if (!color_white(object)) {
1192                         /* non-orphan, ignored or new */
1193                         spin_unlock_irqrestore(&object->lock, flags);
1194                         put_object(object);
1195                         continue;
1196                 }
1197
1198                 /*
1199                  * Increase the object's reference count (number of pointers
1200                  * to the memory block). If this count reaches the required
1201                  * minimum, the object's color will become gray and it will be
1202                  * added to the gray_list.
1203                  */
1204                 object->count++;
1205                 if (color_gray(object)) {
1206                         list_add_tail(&object->gray_list, &gray_list);
1207                         spin_unlock_irqrestore(&object->lock, flags);
1208                         continue;
1209                 }
1210
1211                 spin_unlock_irqrestore(&object->lock, flags);
1212                 put_object(object);
1213         }
1214 }
1215
1216 /*
1217  * Scan a memory block corresponding to a kmemleak_object. A condition is
1218  * that object->use_count >= 1.
1219  */
1220 static void scan_object(struct kmemleak_object *object)
1221 {
1222         struct kmemleak_scan_area *area;
1223         unsigned long flags;
1224
1225         /*
1226          * Once the object->lock is acquired, the corresponding memory block
1227          * cannot be freed (the same lock is acquired in delete_object).
1228          */
1229         spin_lock_irqsave(&object->lock, flags);
1230         if (object->flags & OBJECT_NO_SCAN)
1231                 goto out;
1232         if (!(object->flags & OBJECT_ALLOCATED))
1233                 /* already freed object */
1234                 goto out;
1235         if (hlist_empty(&object->area_list)) {
1236                 void *start = (void *)object->pointer;
1237                 void *end = (void *)(object->pointer + object->size);
1238
1239                 while (start < end && (object->flags & OBJECT_ALLOCATED) &&
1240                        !(object->flags & OBJECT_NO_SCAN)) {
1241                         scan_block(start, min(start + MAX_SCAN_SIZE, end),
1242                                    object, 0);
1243                         start += MAX_SCAN_SIZE;
1244
1245                         spin_unlock_irqrestore(&object->lock, flags);
1246                         cond_resched();
1247                         spin_lock_irqsave(&object->lock, flags);
1248                 }
1249         } else
1250                 hlist_for_each_entry(area, &object->area_list, node)
1251                         scan_block((void *)area->start,
1252                                    (void *)(area->start + area->size),
1253                                    object, 0);
1254 out:
1255         spin_unlock_irqrestore(&object->lock, flags);
1256 }
1257
1258 /*
1259  * Scan the objects already referenced (gray objects). More objects will be
1260  * referenced and, if there are no memory leaks, all the objects are scanned.
1261  */
1262 static void scan_gray_list(void)
1263 {
1264         struct kmemleak_object *object, *tmp;
1265
1266         /*
1267          * The list traversal is safe for both tail additions and removals
1268          * from inside the loop. The kmemleak objects cannot be freed from
1269          * outside the loop because their use_count was incremented.
1270          */
1271         object = list_entry(gray_list.next, typeof(*object), gray_list);
1272         while (&object->gray_list != &gray_list) {
1273                 cond_resched();
1274
1275                 /* may add new objects to the list */
1276                 if (!scan_should_stop())
1277                         scan_object(object);
1278
1279                 tmp = list_entry(object->gray_list.next, typeof(*object),
1280                                  gray_list);
1281
1282                 /* remove the object from the list and release it */
1283                 list_del(&object->gray_list);
1284                 put_object(object);
1285
1286                 object = tmp;
1287         }
1288         WARN_ON(!list_empty(&gray_list));
1289 }
1290
1291 /*
1292  * Scan data sections and all the referenced memory blocks allocated via the
1293  * kernel's standard allocators. This function must be called with the
1294  * scan_mutex held.
1295  */
1296 static void kmemleak_scan(void)
1297 {
1298         unsigned long flags;
1299         struct kmemleak_object *object;
1300         int i;
1301         int new_leaks = 0;
1302
1303         jiffies_last_scan = jiffies;
1304
1305         /* prepare the kmemleak_object's */
1306         rcu_read_lock();
1307         list_for_each_entry_rcu(object, &object_list, object_list) {
1308                 spin_lock_irqsave(&object->lock, flags);
1309 #ifdef DEBUG
1310                 /*
1311                  * With a few exceptions there should be a maximum of
1312                  * 1 reference to any object at this point.
1313                  */
1314                 if (atomic_read(&object->use_count) > 1) {
1315                         pr_debug("object->use_count = %d\n",
1316                                  atomic_read(&object->use_count));
1317                         dump_object_info(object);
1318                 }
1319 #endif
1320                 /* reset the reference count (whiten the object) */
1321                 object->count = 0;
1322                 if (color_gray(object) && get_object(object))
1323                         list_add_tail(&object->gray_list, &gray_list);
1324
1325                 spin_unlock_irqrestore(&object->lock, flags);
1326         }
1327         rcu_read_unlock();
1328
1329         /* data/bss scanning */
1330         scan_block(_sdata, _edata, NULL, 1);
1331         scan_block(__bss_start, __bss_stop, NULL, 1);
1332
1333 #ifdef CONFIG_SMP
1334         /* per-cpu sections scanning */
1335         for_each_possible_cpu(i)
1336                 scan_block(__per_cpu_start + per_cpu_offset(i),
1337                            __per_cpu_end + per_cpu_offset(i), NULL, 1);
1338 #endif
1339
1340         /*
1341          * Struct page scanning for each node.
1342          */
1343         get_online_mems();
1344         for_each_online_node(i) {
1345                 unsigned long start_pfn = node_start_pfn(i);
1346                 unsigned long end_pfn = node_end_pfn(i);
1347                 unsigned long pfn;
1348
1349                 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1350                         struct page *page;
1351
1352                         if (!pfn_valid(pfn))
1353                                 continue;
1354                         page = pfn_to_page(pfn);
1355                         /* only scan if page is in use */
1356                         if (page_count(page) == 0)
1357                                 continue;
1358                         scan_block(page, page + 1, NULL, 1);
1359                 }
1360         }
1361         put_online_mems();
1362
1363         /*
1364          * Scanning the task stacks (may introduce false negatives).
1365          */
1366         if (kmemleak_stack_scan) {
1367                 struct task_struct *p, *g;
1368
1369                 read_lock(&tasklist_lock);
1370                 do_each_thread(g, p) {
1371                         scan_block(task_stack_page(p), task_stack_page(p) +
1372                                    THREAD_SIZE, NULL, 0);
1373                 } while_each_thread(g, p);
1374                 read_unlock(&tasklist_lock);
1375         }
1376
1377         /*
1378          * Scan the objects already referenced from the sections scanned
1379          * above.
1380          */
1381         scan_gray_list();
1382
1383         /*
1384          * Check for new or unreferenced objects modified since the previous
1385          * scan and color them gray until the next scan.
1386          */
1387         rcu_read_lock();
1388         list_for_each_entry_rcu(object, &object_list, object_list) {
1389                 spin_lock_irqsave(&object->lock, flags);
1390                 if (color_white(object) && (object->flags & OBJECT_ALLOCATED)
1391                     && update_checksum(object) && get_object(object)) {
1392                         /* color it gray temporarily */
1393                         object->count = object->min_count;
1394                         list_add_tail(&object->gray_list, &gray_list);
1395                 }
1396                 spin_unlock_irqrestore(&object->lock, flags);
1397         }
1398         rcu_read_unlock();
1399
1400         /*
1401          * Re-scan the gray list for modified unreferenced objects.
1402          */
1403         scan_gray_list();
1404
1405         /*
1406          * If scanning was stopped do not report any new unreferenced objects.
1407          */
1408         if (scan_should_stop())
1409                 return;
1410
1411         /*
1412          * Scanning result reporting.
1413          */
1414         rcu_read_lock();
1415         list_for_each_entry_rcu(object, &object_list, object_list) {
1416                 spin_lock_irqsave(&object->lock, flags);
1417                 if (unreferenced_object(object) &&
1418                     !(object->flags & OBJECT_REPORTED)) {
1419                         object->flags |= OBJECT_REPORTED;
1420                         new_leaks++;
1421                 }
1422                 spin_unlock_irqrestore(&object->lock, flags);
1423         }
1424         rcu_read_unlock();
1425
1426         if (new_leaks) {
1427                 kmemleak_found_leaks = true;
1428
1429                 pr_info("%d new suspected memory leaks (see "
1430                         "/sys/kernel/debug/kmemleak)\n", new_leaks);
1431         }
1432
1433 }
1434
1435 /*
1436  * Thread function performing automatic memory scanning. Unreferenced objects
1437  * at the end of a memory scan are reported but only the first time.
1438  */
1439 static int kmemleak_scan_thread(void *arg)
1440 {
1441         static int first_run = 1;
1442
1443         pr_info("Automatic memory scanning thread started\n");
1444         set_user_nice(current, 10);
1445
1446         /*
1447          * Wait before the first scan to allow the system to fully initialize.
1448          */
1449         if (first_run) {
1450                 first_run = 0;
1451                 ssleep(SECS_FIRST_SCAN);
1452         }
1453
1454         while (!kthread_should_stop()) {
1455                 signed long timeout = jiffies_scan_wait;
1456
1457                 mutex_lock(&scan_mutex);
1458                 kmemleak_scan();
1459                 mutex_unlock(&scan_mutex);
1460
1461                 /* wait before the next scan */
1462                 while (timeout && !kthread_should_stop())
1463                         timeout = schedule_timeout_interruptible(timeout);
1464         }
1465
1466         pr_info("Automatic memory scanning thread ended\n");
1467
1468         return 0;
1469 }
1470
1471 /*
1472  * Start the automatic memory scanning thread. This function must be called
1473  * with the scan_mutex held.
1474  */
1475 static void start_scan_thread(void)
1476 {
1477         if (scan_thread)
1478                 return;
1479         scan_thread = kthread_run(kmemleak_scan_thread, NULL, "kmemleak");
1480         if (IS_ERR(scan_thread)) {
1481                 pr_warning("Failed to create the scan thread\n");
1482                 scan_thread = NULL;
1483         }
1484 }
1485
1486 /*
1487  * Stop the automatic memory scanning thread. This function must be called
1488  * with the scan_mutex held.
1489  */
1490 static void stop_scan_thread(void)
1491 {
1492         if (scan_thread) {
1493                 kthread_stop(scan_thread);
1494                 scan_thread = NULL;
1495         }
1496 }
1497
1498 /*
1499  * Iterate over the object_list and return the first valid object at or after
1500  * the required position with its use_count incremented. The function triggers
1501  * a memory scanning when the pos argument points to the first position.
1502  */
1503 static void *kmemleak_seq_start(struct seq_file *seq, loff_t *pos)
1504 {
1505         struct kmemleak_object *object;
1506         loff_t n = *pos;
1507         int err;
1508
1509         err = mutex_lock_interruptible(&scan_mutex);
1510         if (err < 0)
1511                 return ERR_PTR(err);
1512
1513         rcu_read_lock();
1514         list_for_each_entry_rcu(object, &object_list, object_list) {
1515                 if (n-- > 0)
1516                         continue;
1517                 if (get_object(object))
1518                         goto out;
1519         }
1520         object = NULL;
1521 out:
1522         return object;
1523 }
1524
1525 /*
1526  * Return the next object in the object_list. The function decrements the
1527  * use_count of the previous object and increases that of the next one.
1528  */
1529 static void *kmemleak_seq_next(struct seq_file *seq, void *v, loff_t *pos)
1530 {
1531         struct kmemleak_object *prev_obj = v;
1532         struct kmemleak_object *next_obj = NULL;
1533         struct kmemleak_object *obj = prev_obj;
1534
1535         ++(*pos);
1536
1537         list_for_each_entry_continue_rcu(obj, &object_list, object_list) {
1538                 if (get_object(obj)) {
1539                         next_obj = obj;
1540                         break;
1541                 }
1542         }
1543
1544         put_object(prev_obj);
1545         return next_obj;
1546 }
1547
1548 /*
1549  * Decrement the use_count of the last object required, if any.
1550  */
1551 static void kmemleak_seq_stop(struct seq_file *seq, void *v)
1552 {
1553         if (!IS_ERR(v)) {
1554                 /*
1555                  * kmemleak_seq_start may return ERR_PTR if the scan_mutex
1556                  * waiting was interrupted, so only release it if !IS_ERR.
1557                  */
1558                 rcu_read_unlock();
1559                 mutex_unlock(&scan_mutex);
1560                 if (v)
1561                         put_object(v);
1562         }
1563 }
1564
1565 /*
1566  * Print the information for an unreferenced object to the seq file.
1567  */
1568 static int kmemleak_seq_show(struct seq_file *seq, void *v)
1569 {
1570         struct kmemleak_object *object = v;
1571         unsigned long flags;
1572
1573         spin_lock_irqsave(&object->lock, flags);
1574         if ((object->flags & OBJECT_REPORTED) && unreferenced_object(object))
1575                 print_unreferenced(seq, object);
1576         spin_unlock_irqrestore(&object->lock, flags);
1577         return 0;
1578 }
1579
1580 static const struct seq_operations kmemleak_seq_ops = {
1581         .start = kmemleak_seq_start,
1582         .next  = kmemleak_seq_next,
1583         .stop  = kmemleak_seq_stop,
1584         .show  = kmemleak_seq_show,
1585 };
1586
1587 static int kmemleak_open(struct inode *inode, struct file *file)
1588 {
1589         return seq_open(file, &kmemleak_seq_ops);
1590 }
1591
1592 static int dump_str_object_info(const char *str)
1593 {
1594         unsigned long flags;
1595         struct kmemleak_object *object;
1596         unsigned long addr;
1597
1598         if (kstrtoul(str, 0, &addr))
1599                 return -EINVAL;
1600         object = find_and_get_object(addr, 0);
1601         if (!object) {
1602                 pr_info("Unknown object at 0x%08lx\n", addr);
1603                 return -EINVAL;
1604         }
1605
1606         spin_lock_irqsave(&object->lock, flags);
1607         dump_object_info(object);
1608         spin_unlock_irqrestore(&object->lock, flags);
1609
1610         put_object(object);
1611         return 0;
1612 }
1613
1614 /*
1615  * We use grey instead of black to ensure we can do future scans on the same
1616  * objects. If we did not do future scans these black objects could
1617  * potentially contain references to newly allocated objects in the future and
1618  * we'd end up with false positives.
1619  */
1620 static void kmemleak_clear(void)
1621 {
1622         struct kmemleak_object *object;
1623         unsigned long flags;
1624
1625         rcu_read_lock();
1626         list_for_each_entry_rcu(object, &object_list, object_list) {
1627                 spin_lock_irqsave(&object->lock, flags);
1628                 if ((object->flags & OBJECT_REPORTED) &&
1629                     unreferenced_object(object))
1630                         __paint_it(object, KMEMLEAK_GREY);
1631                 spin_unlock_irqrestore(&object->lock, flags);
1632         }
1633         rcu_read_unlock();
1634
1635         kmemleak_found_leaks = false;
1636 }
1637
1638 static void __kmemleak_do_cleanup(void);
1639
1640 /*
1641  * File write operation to configure kmemleak at run-time. The following
1642  * commands can be written to the /sys/kernel/debug/kmemleak file:
1643  *   off        - disable kmemleak (irreversible)
1644  *   stack=on   - enable the task stacks scanning
1645  *   stack=off  - disable the tasks stacks scanning
1646  *   scan=on    - start the automatic memory scanning thread
1647  *   scan=off   - stop the automatic memory scanning thread
1648  *   scan=...   - set the automatic memory scanning period in seconds (0 to
1649  *                disable it)
1650  *   scan       - trigger a memory scan
1651  *   clear      - mark all current reported unreferenced kmemleak objects as
1652  *                grey to ignore printing them, or free all kmemleak objects
1653  *                if kmemleak has been disabled.
1654  *   dump=...   - dump information about the object found at the given address
1655  */
1656 static ssize_t kmemleak_write(struct file *file, const char __user *user_buf,
1657                               size_t size, loff_t *ppos)
1658 {
1659         char buf[64];
1660         int buf_size;
1661         int ret;
1662
1663         buf_size = min(size, (sizeof(buf) - 1));
1664         if (strncpy_from_user(buf, user_buf, buf_size) < 0)
1665                 return -EFAULT;
1666         buf[buf_size] = 0;
1667
1668         ret = mutex_lock_interruptible(&scan_mutex);
1669         if (ret < 0)
1670                 return ret;
1671
1672         if (strncmp(buf, "clear", 5) == 0) {
1673                 if (kmemleak_enabled)
1674                         kmemleak_clear();
1675                 else
1676                         __kmemleak_do_cleanup();
1677                 goto out;
1678         }
1679
1680         if (!kmemleak_enabled) {
1681                 ret = -EBUSY;
1682                 goto out;
1683         }
1684
1685         if (strncmp(buf, "off", 3) == 0)
1686                 kmemleak_disable();
1687         else if (strncmp(buf, "stack=on", 8) == 0)
1688                 kmemleak_stack_scan = 1;
1689         else if (strncmp(buf, "stack=off", 9) == 0)
1690                 kmemleak_stack_scan = 0;
1691         else if (strncmp(buf, "scan=on", 7) == 0)
1692                 start_scan_thread();
1693         else if (strncmp(buf, "scan=off", 8) == 0)
1694                 stop_scan_thread();
1695         else if (strncmp(buf, "scan=", 5) == 0) {
1696                 unsigned long secs;
1697
1698                 ret = kstrtoul(buf + 5, 0, &secs);
1699                 if (ret < 0)
1700                         goto out;
1701                 stop_scan_thread();
1702                 if (secs) {
1703                         jiffies_scan_wait = msecs_to_jiffies(secs * 1000);
1704                         start_scan_thread();
1705                 }
1706         } else if (strncmp(buf, "scan", 4) == 0)
1707                 kmemleak_scan();
1708         else if (strncmp(buf, "dump=", 5) == 0)
1709                 ret = dump_str_object_info(buf + 5);
1710         else
1711                 ret = -EINVAL;
1712
1713 out:
1714         mutex_unlock(&scan_mutex);
1715         if (ret < 0)
1716                 return ret;
1717
1718         /* ignore the rest of the buffer, only one command at a time */
1719         *ppos += size;
1720         return size;
1721 }
1722
1723 static const struct file_operations kmemleak_fops = {
1724         .owner          = THIS_MODULE,
1725         .open           = kmemleak_open,
1726         .read           = seq_read,
1727         .write          = kmemleak_write,
1728         .llseek         = seq_lseek,
1729         .release        = seq_release,
1730 };
1731
1732 static void __kmemleak_do_cleanup(void)
1733 {
1734         struct kmemleak_object *object;
1735
1736         rcu_read_lock();
1737         list_for_each_entry_rcu(object, &object_list, object_list)
1738                 delete_object_full(object->pointer);
1739         rcu_read_unlock();
1740 }
1741
1742 /*
1743  * Stop the memory scanning thread and free the kmemleak internal objects if
1744  * no previous scan thread (otherwise, kmemleak may still have some useful
1745  * information on memory leaks).
1746  */
1747 static void kmemleak_do_cleanup(struct work_struct *work)
1748 {
1749         mutex_lock(&scan_mutex);
1750         stop_scan_thread();
1751
1752         if (!kmemleak_found_leaks)
1753                 __kmemleak_do_cleanup();
1754         else
1755                 pr_info("Kmemleak disabled without freeing internal data. "
1756                         "Reclaim the memory with \"echo clear > /sys/kernel/debug/kmemleak\"\n");
1757         mutex_unlock(&scan_mutex);
1758 }
1759
1760 static DECLARE_WORK(cleanup_work, kmemleak_do_cleanup);
1761
1762 /*
1763  * Disable kmemleak. No memory allocation/freeing will be traced once this
1764  * function is called. Disabling kmemleak is an irreversible operation.
1765  */
1766 static void kmemleak_disable(void)
1767 {
1768         /* atomically check whether it was already invoked */
1769         if (cmpxchg(&kmemleak_error, 0, 1))
1770                 return;
1771
1772         /* stop any memory operation tracing */
1773         kmemleak_enabled = 0;
1774
1775         /* check whether it is too early for a kernel thread */
1776         if (kmemleak_initialized)
1777                 schedule_work(&cleanup_work);
1778
1779         pr_info("Kernel memory leak detector disabled\n");
1780 }
1781
1782 /*
1783  * Allow boot-time kmemleak disabling (enabled by default).
1784  */
1785 static int kmemleak_boot_config(char *str)
1786 {
1787         if (!str)
1788                 return -EINVAL;
1789         if (strcmp(str, "off") == 0)
1790                 kmemleak_disable();
1791         else if (strcmp(str, "on") == 0)
1792                 kmemleak_skip_disable = 1;
1793         else
1794                 return -EINVAL;
1795         return 0;
1796 }
1797 early_param("kmemleak", kmemleak_boot_config);
1798
1799 static void __init print_log_trace(struct early_log *log)
1800 {
1801         struct stack_trace trace;
1802
1803         trace.nr_entries = log->trace_len;
1804         trace.entries = log->trace;
1805
1806         pr_notice("Early log backtrace:\n");
1807         print_stack_trace(&trace, 2);
1808 }
1809
1810 /*
1811  * Kmemleak initialization.
1812  */
1813 void __init kmemleak_init(void)
1814 {
1815         int i;
1816         unsigned long flags;
1817
1818 #ifdef CONFIG_DEBUG_KMEMLEAK_DEFAULT_OFF
1819         if (!kmemleak_skip_disable) {
1820                 kmemleak_early_log = 0;
1821                 kmemleak_disable();
1822                 return;
1823         }
1824 #endif
1825
1826         jiffies_min_age = msecs_to_jiffies(MSECS_MIN_AGE);
1827         jiffies_scan_wait = msecs_to_jiffies(SECS_SCAN_WAIT * 1000);
1828
1829         object_cache = KMEM_CACHE(kmemleak_object, SLAB_NOLEAKTRACE);
1830         scan_area_cache = KMEM_CACHE(kmemleak_scan_area, SLAB_NOLEAKTRACE);
1831
1832         if (crt_early_log >= ARRAY_SIZE(early_log))
1833                 pr_warning("Early log buffer exceeded (%d), please increase "
1834                            "DEBUG_KMEMLEAK_EARLY_LOG_SIZE\n", crt_early_log);
1835
1836         /* the kernel is still in UP mode, so disabling the IRQs is enough */
1837         local_irq_save(flags);
1838         kmemleak_early_log = 0;
1839         if (kmemleak_error) {
1840                 local_irq_restore(flags);
1841                 return;
1842         } else
1843                 kmemleak_enabled = 1;
1844         local_irq_restore(flags);
1845
1846         /*
1847          * This is the point where tracking allocations is safe. Automatic
1848          * scanning is started during the late initcall. Add the early logged
1849          * callbacks to the kmemleak infrastructure.
1850          */
1851         for (i = 0; i < crt_early_log; i++) {
1852                 struct early_log *log = &early_log[i];
1853
1854                 switch (log->op_type) {
1855                 case KMEMLEAK_ALLOC:
1856                         early_alloc(log);
1857                         break;
1858                 case KMEMLEAK_ALLOC_PERCPU:
1859                         early_alloc_percpu(log);
1860                         break;
1861                 case KMEMLEAK_FREE:
1862                         kmemleak_free(log->ptr);
1863                         break;
1864                 case KMEMLEAK_FREE_PART:
1865                         kmemleak_free_part(log->ptr, log->size);
1866                         break;
1867                 case KMEMLEAK_FREE_PERCPU:
1868                         kmemleak_free_percpu(log->ptr);
1869                         break;
1870                 case KMEMLEAK_NOT_LEAK:
1871                         kmemleak_not_leak(log->ptr);
1872                         break;
1873                 case KMEMLEAK_IGNORE:
1874                         kmemleak_ignore(log->ptr);
1875                         break;
1876                 case KMEMLEAK_SCAN_AREA:
1877                         kmemleak_scan_area(log->ptr, log->size, GFP_KERNEL);
1878                         break;
1879                 case KMEMLEAK_NO_SCAN:
1880                         kmemleak_no_scan(log->ptr);
1881                         break;
1882                 default:
1883                         kmemleak_warn("Unknown early log operation: %d\n",
1884                                       log->op_type);
1885                 }
1886
1887                 if (kmemleak_warning) {
1888                         print_log_trace(log);
1889                         kmemleak_warning = 0;
1890                 }
1891         }
1892 }
1893
1894 /*
1895  * Late initialization function.
1896  */
1897 static int __init kmemleak_late_init(void)
1898 {
1899         struct dentry *dentry;
1900
1901         kmemleak_initialized = 1;
1902
1903         if (kmemleak_error) {
1904                 /*
1905                  * Some error occurred and kmemleak was disabled. There is a
1906                  * small chance that kmemleak_disable() was called immediately
1907                  * after setting kmemleak_initialized and we may end up with
1908                  * two clean-up threads but serialized by scan_mutex.
1909                  */
1910                 schedule_work(&cleanup_work);
1911                 return -ENOMEM;
1912         }
1913
1914         dentry = debugfs_create_file("kmemleak", S_IRUGO, NULL, NULL,
1915                                      &kmemleak_fops);
1916         if (!dentry)
1917                 pr_warning("Failed to create the debugfs kmemleak file\n");
1918         mutex_lock(&scan_mutex);
1919         start_scan_thread();
1920         mutex_unlock(&scan_mutex);
1921
1922         pr_info("Kernel memory leak detector initialized\n");
1923
1924         return 0;
1925 }
1926 late_initcall(kmemleak_late_init);