lockdep: Introduce match function to BFS
[pandora-kernel.git] / kernel / lockdep.c
1 /*
2  * kernel/lockdep.c
3  *
4  * Runtime locking correctness validator
5  *
6  * Started by Ingo Molnar:
7  *
8  *  Copyright (C) 2006,2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
9  *  Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
10  *
11  * this code maps all the lock dependencies as they occur in a live kernel
12  * and will warn about the following classes of locking bugs:
13  *
14  * - lock inversion scenarios
15  * - circular lock dependencies
16  * - hardirq/softirq safe/unsafe locking bugs
17  *
18  * Bugs are reported even if the current locking scenario does not cause
19  * any deadlock at this point.
20  *
21  * I.e. if anytime in the past two locks were taken in a different order,
22  * even if it happened for another task, even if those were different
23  * locks (but of the same class as this lock), this code will detect it.
24  *
25  * Thanks to Arjan van de Ven for coming up with the initial idea of
26  * mapping lock dependencies runtime.
27  */
28 #define DISABLE_BRANCH_PROFILING
29 #include <linux/mutex.h>
30 #include <linux/sched.h>
31 #include <linux/delay.h>
32 #include <linux/module.h>
33 #include <linux/proc_fs.h>
34 #include <linux/seq_file.h>
35 #include <linux/spinlock.h>
36 #include <linux/kallsyms.h>
37 #include <linux/interrupt.h>
38 #include <linux/stacktrace.h>
39 #include <linux/debug_locks.h>
40 #include <linux/irqflags.h>
41 #include <linux/utsname.h>
42 #include <linux/hash.h>
43 #include <linux/ftrace.h>
44 #include <linux/stringify.h>
45 #include <linux/bitops.h>
46 #include <asm/sections.h>
47
48 #include "lockdep_internals.h"
49
50 #define CREATE_TRACE_POINTS
51 #include <trace/events/lockdep.h>
52
53 #ifdef CONFIG_PROVE_LOCKING
54 int prove_locking = 1;
55 module_param(prove_locking, int, 0644);
56 #else
57 #define prove_locking 0
58 #endif
59
60 #ifdef CONFIG_LOCK_STAT
61 int lock_stat = 1;
62 module_param(lock_stat, int, 0644);
63 #else
64 #define lock_stat 0
65 #endif
66
67 /*
68  * lockdep_lock: protects the lockdep graph, the hashes and the
69  *               class/list/hash allocators.
70  *
71  * This is one of the rare exceptions where it's justified
72  * to use a raw spinlock - we really dont want the spinlock
73  * code to recurse back into the lockdep code...
74  */
75 static raw_spinlock_t lockdep_lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
76
77 static int graph_lock(void)
78 {
79         __raw_spin_lock(&lockdep_lock);
80         /*
81          * Make sure that if another CPU detected a bug while
82          * walking the graph we dont change it (while the other
83          * CPU is busy printing out stuff with the graph lock
84          * dropped already)
85          */
86         if (!debug_locks) {
87                 __raw_spin_unlock(&lockdep_lock);
88                 return 0;
89         }
90         /* prevent any recursions within lockdep from causing deadlocks */
91         current->lockdep_recursion++;
92         return 1;
93 }
94
95 static inline int graph_unlock(void)
96 {
97         if (debug_locks && !__raw_spin_is_locked(&lockdep_lock))
98                 return DEBUG_LOCKS_WARN_ON(1);
99
100         current->lockdep_recursion--;
101         __raw_spin_unlock(&lockdep_lock);
102         return 0;
103 }
104
105 /*
106  * Turn lock debugging off and return with 0 if it was off already,
107  * and also release the graph lock:
108  */
109 static inline int debug_locks_off_graph_unlock(void)
110 {
111         int ret = debug_locks_off();
112
113         __raw_spin_unlock(&lockdep_lock);
114
115         return ret;
116 }
117
118 static int lockdep_initialized;
119
120 unsigned long nr_list_entries;
121 struct lock_list list_entries[MAX_LOCKDEP_ENTRIES];
122
123 /*
124  * All data structures here are protected by the global debug_lock.
125  *
126  * Mutex key structs only get allocated, once during bootup, and never
127  * get freed - this significantly simplifies the debugging code.
128  */
129 unsigned long nr_lock_classes;
130 static struct lock_class lock_classes[MAX_LOCKDEP_KEYS];
131
132 static inline struct lock_class *hlock_class(struct held_lock *hlock)
133 {
134         if (!hlock->class_idx) {
135                 DEBUG_LOCKS_WARN_ON(1);
136                 return NULL;
137         }
138         return lock_classes + hlock->class_idx - 1;
139 }
140
141 #ifdef CONFIG_LOCK_STAT
142 static DEFINE_PER_CPU(struct lock_class_stats[MAX_LOCKDEP_KEYS], lock_stats);
143
144 static int lock_point(unsigned long points[], unsigned long ip)
145 {
146         int i;
147
148         for (i = 0; i < LOCKSTAT_POINTS; i++) {
149                 if (points[i] == 0) {
150                         points[i] = ip;
151                         break;
152                 }
153                 if (points[i] == ip)
154                         break;
155         }
156
157         return i;
158 }
159
160 static void lock_time_inc(struct lock_time *lt, s64 time)
161 {
162         if (time > lt->max)
163                 lt->max = time;
164
165         if (time < lt->min || !lt->min)
166                 lt->min = time;
167
168         lt->total += time;
169         lt->nr++;
170 }
171
172 static inline void lock_time_add(struct lock_time *src, struct lock_time *dst)
173 {
174         dst->min += src->min;
175         dst->max += src->max;
176         dst->total += src->total;
177         dst->nr += src->nr;
178 }
179
180 struct lock_class_stats lock_stats(struct lock_class *class)
181 {
182         struct lock_class_stats stats;
183         int cpu, i;
184
185         memset(&stats, 0, sizeof(struct lock_class_stats));
186         for_each_possible_cpu(cpu) {
187                 struct lock_class_stats *pcs =
188                         &per_cpu(lock_stats, cpu)[class - lock_classes];
189
190                 for (i = 0; i < ARRAY_SIZE(stats.contention_point); i++)
191                         stats.contention_point[i] += pcs->contention_point[i];
192
193                 for (i = 0; i < ARRAY_SIZE(stats.contending_point); i++)
194                         stats.contending_point[i] += pcs->contending_point[i];
195
196                 lock_time_add(&pcs->read_waittime, &stats.read_waittime);
197                 lock_time_add(&pcs->write_waittime, &stats.write_waittime);
198
199                 lock_time_add(&pcs->read_holdtime, &stats.read_holdtime);
200                 lock_time_add(&pcs->write_holdtime, &stats.write_holdtime);
201
202                 for (i = 0; i < ARRAY_SIZE(stats.bounces); i++)
203                         stats.bounces[i] += pcs->bounces[i];
204         }
205
206         return stats;
207 }
208
209 void clear_lock_stats(struct lock_class *class)
210 {
211         int cpu;
212
213         for_each_possible_cpu(cpu) {
214                 struct lock_class_stats *cpu_stats =
215                         &per_cpu(lock_stats, cpu)[class - lock_classes];
216
217                 memset(cpu_stats, 0, sizeof(struct lock_class_stats));
218         }
219         memset(class->contention_point, 0, sizeof(class->contention_point));
220         memset(class->contending_point, 0, sizeof(class->contending_point));
221 }
222
223 static struct lock_class_stats *get_lock_stats(struct lock_class *class)
224 {
225         return &get_cpu_var(lock_stats)[class - lock_classes];
226 }
227
228 static void put_lock_stats(struct lock_class_stats *stats)
229 {
230         put_cpu_var(lock_stats);
231 }
232
233 static void lock_release_holdtime(struct held_lock *hlock)
234 {
235         struct lock_class_stats *stats;
236         s64 holdtime;
237
238         if (!lock_stat)
239                 return;
240
241         holdtime = sched_clock() - hlock->holdtime_stamp;
242
243         stats = get_lock_stats(hlock_class(hlock));
244         if (hlock->read)
245                 lock_time_inc(&stats->read_holdtime, holdtime);
246         else
247                 lock_time_inc(&stats->write_holdtime, holdtime);
248         put_lock_stats(stats);
249 }
250 #else
251 static inline void lock_release_holdtime(struct held_lock *hlock)
252 {
253 }
254 #endif
255
256 /*
257  * We keep a global list of all lock classes. The list only grows,
258  * never shrinks. The list is only accessed with the lockdep
259  * spinlock lock held.
260  */
261 LIST_HEAD(all_lock_classes);
262
263 /*
264  * The lockdep classes are in a hash-table as well, for fast lookup:
265  */
266 #define CLASSHASH_BITS          (MAX_LOCKDEP_KEYS_BITS - 1)
267 #define CLASSHASH_SIZE          (1UL << CLASSHASH_BITS)
268 #define __classhashfn(key)      hash_long((unsigned long)key, CLASSHASH_BITS)
269 #define classhashentry(key)     (classhash_table + __classhashfn((key)))
270
271 static struct list_head classhash_table[CLASSHASH_SIZE];
272
273 /*
274  * We put the lock dependency chains into a hash-table as well, to cache
275  * their existence:
276  */
277 #define CHAINHASH_BITS          (MAX_LOCKDEP_CHAINS_BITS-1)
278 #define CHAINHASH_SIZE          (1UL << CHAINHASH_BITS)
279 #define __chainhashfn(chain)    hash_long(chain, CHAINHASH_BITS)
280 #define chainhashentry(chain)   (chainhash_table + __chainhashfn((chain)))
281
282 static struct list_head chainhash_table[CHAINHASH_SIZE];
283
284 /*
285  * The hash key of the lock dependency chains is a hash itself too:
286  * it's a hash of all locks taken up to that lock, including that lock.
287  * It's a 64-bit hash, because it's important for the keys to be
288  * unique.
289  */
290 #define iterate_chain_key(key1, key2) \
291         (((key1) << MAX_LOCKDEP_KEYS_BITS) ^ \
292         ((key1) >> (64-MAX_LOCKDEP_KEYS_BITS)) ^ \
293         (key2))
294
295 void lockdep_off(void)
296 {
297         current->lockdep_recursion++;
298 }
299 EXPORT_SYMBOL(lockdep_off);
300
301 void lockdep_on(void)
302 {
303         current->lockdep_recursion--;
304 }
305 EXPORT_SYMBOL(lockdep_on);
306
307 /*
308  * Debugging switches:
309  */
310
311 #define VERBOSE                 0
312 #define VERY_VERBOSE            0
313
314 #if VERBOSE
315 # define HARDIRQ_VERBOSE        1
316 # define SOFTIRQ_VERBOSE        1
317 # define RECLAIM_VERBOSE        1
318 #else
319 # define HARDIRQ_VERBOSE        0
320 # define SOFTIRQ_VERBOSE        0
321 # define RECLAIM_VERBOSE        0
322 #endif
323
324 #if VERBOSE || HARDIRQ_VERBOSE || SOFTIRQ_VERBOSE || RECLAIM_VERBOSE
325 /*
326  * Quick filtering for interesting events:
327  */
328 static int class_filter(struct lock_class *class)
329 {
330 #if 0
331         /* Example */
332         if (class->name_version == 1 &&
333                         !strcmp(class->name, "lockname"))
334                 return 1;
335         if (class->name_version == 1 &&
336                         !strcmp(class->name, "&struct->lockfield"))
337                 return 1;
338 #endif
339         /* Filter everything else. 1 would be to allow everything else */
340         return 0;
341 }
342 #endif
343
344 static int verbose(struct lock_class *class)
345 {
346 #if VERBOSE
347         return class_filter(class);
348 #endif
349         return 0;
350 }
351
352 /*
353  * Stack-trace: tightly packed array of stack backtrace
354  * addresses. Protected by the graph_lock.
355  */
356 unsigned long nr_stack_trace_entries;
357 static unsigned long stack_trace[MAX_STACK_TRACE_ENTRIES];
358
359 static int save_trace(struct stack_trace *trace)
360 {
361         trace->nr_entries = 0;
362         trace->max_entries = MAX_STACK_TRACE_ENTRIES - nr_stack_trace_entries;
363         trace->entries = stack_trace + nr_stack_trace_entries;
364
365         trace->skip = 3;
366
367         save_stack_trace(trace);
368
369         trace->max_entries = trace->nr_entries;
370
371         nr_stack_trace_entries += trace->nr_entries;
372
373         if (nr_stack_trace_entries == MAX_STACK_TRACE_ENTRIES) {
374                 if (!debug_locks_off_graph_unlock())
375                         return 0;
376
377                 printk("BUG: MAX_STACK_TRACE_ENTRIES too low!\n");
378                 printk("turning off the locking correctness validator.\n");
379                 dump_stack();
380
381                 return 0;
382         }
383
384         return 1;
385 }
386
387 unsigned int nr_hardirq_chains;
388 unsigned int nr_softirq_chains;
389 unsigned int nr_process_chains;
390 unsigned int max_lockdep_depth;
391 unsigned int max_recursion_depth;
392
393 static unsigned int lockdep_dependency_gen_id;
394
395 static bool lockdep_dependency_visit(struct lock_class *source,
396                                      unsigned int depth)
397 {
398         if (!depth)
399                 lockdep_dependency_gen_id++;
400         if (source->dep_gen_id == lockdep_dependency_gen_id)
401                 return true;
402         source->dep_gen_id = lockdep_dependency_gen_id;
403         return false;
404 }
405
406 #ifdef CONFIG_DEBUG_LOCKDEP
407 /*
408  * We cannot printk in early bootup code. Not even early_printk()
409  * might work. So we mark any initialization errors and printk
410  * about it later on, in lockdep_info().
411  */
412 static int lockdep_init_error;
413 static unsigned long lockdep_init_trace_data[20];
414 static struct stack_trace lockdep_init_trace = {
415         .max_entries = ARRAY_SIZE(lockdep_init_trace_data),
416         .entries = lockdep_init_trace_data,
417 };
418
419 /*
420  * Various lockdep statistics:
421  */
422 atomic_t chain_lookup_hits;
423 atomic_t chain_lookup_misses;
424 atomic_t hardirqs_on_events;
425 atomic_t hardirqs_off_events;
426 atomic_t redundant_hardirqs_on;
427 atomic_t redundant_hardirqs_off;
428 atomic_t softirqs_on_events;
429 atomic_t softirqs_off_events;
430 atomic_t redundant_softirqs_on;
431 atomic_t redundant_softirqs_off;
432 atomic_t nr_unused_locks;
433 atomic_t nr_cyclic_checks;
434 atomic_t nr_cyclic_check_recursions;
435 atomic_t nr_find_usage_forwards_checks;
436 atomic_t nr_find_usage_forwards_recursions;
437 atomic_t nr_find_usage_backwards_checks;
438 atomic_t nr_find_usage_backwards_recursions;
439 #endif
440
441 /*
442  * Locking printouts:
443  */
444
445 #define __USAGE(__STATE)                                                \
446         [LOCK_USED_IN_##__STATE] = "IN-"__stringify(__STATE)"-W",       \
447         [LOCK_ENABLED_##__STATE] = __stringify(__STATE)"-ON-W",         \
448         [LOCK_USED_IN_##__STATE##_READ] = "IN-"__stringify(__STATE)"-R",\
449         [LOCK_ENABLED_##__STATE##_READ] = __stringify(__STATE)"-ON-R",
450
451 static const char *usage_str[] =
452 {
453 #define LOCKDEP_STATE(__STATE) __USAGE(__STATE)
454 #include "lockdep_states.h"
455 #undef LOCKDEP_STATE
456         [LOCK_USED] = "INITIAL USE",
457 };
458
459 const char * __get_key_name(struct lockdep_subclass_key *key, char *str)
460 {
461         return kallsyms_lookup((unsigned long)key, NULL, NULL, NULL, str);
462 }
463
464 static inline unsigned long lock_flag(enum lock_usage_bit bit)
465 {
466         return 1UL << bit;
467 }
468
469 static char get_usage_char(struct lock_class *class, enum lock_usage_bit bit)
470 {
471         char c = '.';
472
473         if (class->usage_mask & lock_flag(bit + 2))
474                 c = '+';
475         if (class->usage_mask & lock_flag(bit)) {
476                 c = '-';
477                 if (class->usage_mask & lock_flag(bit + 2))
478                         c = '?';
479         }
480
481         return c;
482 }
483
484 void get_usage_chars(struct lock_class *class, char usage[LOCK_USAGE_CHARS])
485 {
486         int i = 0;
487
488 #define LOCKDEP_STATE(__STATE)                                          \
489         usage[i++] = get_usage_char(class, LOCK_USED_IN_##__STATE);     \
490         usage[i++] = get_usage_char(class, LOCK_USED_IN_##__STATE##_READ);
491 #include "lockdep_states.h"
492 #undef LOCKDEP_STATE
493
494         usage[i] = '\0';
495 }
496
497 static void print_lock_name(struct lock_class *class)
498 {
499         char str[KSYM_NAME_LEN], usage[LOCK_USAGE_CHARS];
500         const char *name;
501
502         get_usage_chars(class, usage);
503
504         name = class->name;
505         if (!name) {
506                 name = __get_key_name(class->key, str);
507                 printk(" (%s", name);
508         } else {
509                 printk(" (%s", name);
510                 if (class->name_version > 1)
511                         printk("#%d", class->name_version);
512                 if (class->subclass)
513                         printk("/%d", class->subclass);
514         }
515         printk("){%s}", usage);
516 }
517
518 static void print_lockdep_cache(struct lockdep_map *lock)
519 {
520         const char *name;
521         char str[KSYM_NAME_LEN];
522
523         name = lock->name;
524         if (!name)
525                 name = __get_key_name(lock->key->subkeys, str);
526
527         printk("%s", name);
528 }
529
530 static void print_lock(struct held_lock *hlock)
531 {
532         print_lock_name(hlock_class(hlock));
533         printk(", at: ");
534         print_ip_sym(hlock->acquire_ip);
535 }
536
537 static void lockdep_print_held_locks(struct task_struct *curr)
538 {
539         int i, depth = curr->lockdep_depth;
540
541         if (!depth) {
542                 printk("no locks held by %s/%d.\n", curr->comm, task_pid_nr(curr));
543                 return;
544         }
545         printk("%d lock%s held by %s/%d:\n",
546                 depth, depth > 1 ? "s" : "", curr->comm, task_pid_nr(curr));
547
548         for (i = 0; i < depth; i++) {
549                 printk(" #%d: ", i);
550                 print_lock(curr->held_locks + i);
551         }
552 }
553
554 static void print_lock_class_header(struct lock_class *class, int depth)
555 {
556         int bit;
557
558         printk("%*s->", depth, "");
559         print_lock_name(class);
560         printk(" ops: %lu", class->ops);
561         printk(" {\n");
562
563         for (bit = 0; bit < LOCK_USAGE_STATES; bit++) {
564                 if (class->usage_mask & (1 << bit)) {
565                         int len = depth;
566
567                         len += printk("%*s   %s", depth, "", usage_str[bit]);
568                         len += printk(" at:\n");
569                         print_stack_trace(class->usage_traces + bit, len);
570                 }
571         }
572         printk("%*s }\n", depth, "");
573
574         printk("%*s ... key      at: ",depth,"");
575         print_ip_sym((unsigned long)class->key);
576 }
577
578 /*
579  * printk all lock dependencies starting at <entry>:
580  */
581 static void __used
582 print_lock_dependencies(struct lock_class *class, int depth)
583 {
584         struct lock_list *entry;
585
586         if (lockdep_dependency_visit(class, depth))
587                 return;
588
589         if (DEBUG_LOCKS_WARN_ON(depth >= 20))
590                 return;
591
592         print_lock_class_header(class, depth);
593
594         list_for_each_entry(entry, &class->locks_after, entry) {
595                 if (DEBUG_LOCKS_WARN_ON(!entry->class))
596                         return;
597
598                 print_lock_dependencies(entry->class, depth + 1);
599
600                 printk("%*s ... acquired at:\n",depth,"");
601                 print_stack_trace(&entry->trace, 2);
602                 printk("\n");
603         }
604 }
605
606 static void print_kernel_version(void)
607 {
608         printk("%s %.*s\n", init_utsname()->release,
609                 (int)strcspn(init_utsname()->version, " "),
610                 init_utsname()->version);
611 }
612
613 static int very_verbose(struct lock_class *class)
614 {
615 #if VERY_VERBOSE
616         return class_filter(class);
617 #endif
618         return 0;
619 }
620
621 /*
622  * Is this the address of a static object:
623  */
624 static int static_obj(void *obj)
625 {
626         unsigned long start = (unsigned long) &_stext,
627                       end   = (unsigned long) &_end,
628                       addr  = (unsigned long) obj;
629 #ifdef CONFIG_SMP
630         int i;
631 #endif
632
633         /*
634          * static variable?
635          */
636         if ((addr >= start) && (addr < end))
637                 return 1;
638
639 #ifdef CONFIG_SMP
640         /*
641          * percpu var?
642          */
643         for_each_possible_cpu(i) {
644                 start = (unsigned long) &__per_cpu_start + per_cpu_offset(i);
645                 end   = (unsigned long) &__per_cpu_start + PERCPU_ENOUGH_ROOM
646                                         + per_cpu_offset(i);
647
648                 if ((addr >= start) && (addr < end))
649                         return 1;
650         }
651 #endif
652
653         /*
654          * module var?
655          */
656         return is_module_address(addr);
657 }
658
659 /*
660  * To make lock name printouts unique, we calculate a unique
661  * class->name_version generation counter:
662  */
663 static int count_matching_names(struct lock_class *new_class)
664 {
665         struct lock_class *class;
666         int count = 0;
667
668         if (!new_class->name)
669                 return 0;
670
671         list_for_each_entry(class, &all_lock_classes, lock_entry) {
672                 if (new_class->key - new_class->subclass == class->key)
673                         return class->name_version;
674                 if (class->name && !strcmp(class->name, new_class->name))
675                         count = max(count, class->name_version);
676         }
677
678         return count + 1;
679 }
680
681 /*
682  * Register a lock's class in the hash-table, if the class is not present
683  * yet. Otherwise we look it up. We cache the result in the lock object
684  * itself, so actual lookup of the hash should be once per lock object.
685  */
686 static inline struct lock_class *
687 look_up_lock_class(struct lockdep_map *lock, unsigned int subclass)
688 {
689         struct lockdep_subclass_key *key;
690         struct list_head *hash_head;
691         struct lock_class *class;
692
693 #ifdef CONFIG_DEBUG_LOCKDEP
694         /*
695          * If the architecture calls into lockdep before initializing
696          * the hashes then we'll warn about it later. (we cannot printk
697          * right now)
698          */
699         if (unlikely(!lockdep_initialized)) {
700                 lockdep_init();
701                 lockdep_init_error = 1;
702                 save_stack_trace(&lockdep_init_trace);
703         }
704 #endif
705
706         /*
707          * Static locks do not have their class-keys yet - for them the key
708          * is the lock object itself:
709          */
710         if (unlikely(!lock->key))
711                 lock->key = (void *)lock;
712
713         /*
714          * NOTE: the class-key must be unique. For dynamic locks, a static
715          * lock_class_key variable is passed in through the mutex_init()
716          * (or spin_lock_init()) call - which acts as the key. For static
717          * locks we use the lock object itself as the key.
718          */
719         BUILD_BUG_ON(sizeof(struct lock_class_key) >
720                         sizeof(struct lockdep_map));
721
722         key = lock->key->subkeys + subclass;
723
724         hash_head = classhashentry(key);
725
726         /*
727          * We can walk the hash lockfree, because the hash only
728          * grows, and we are careful when adding entries to the end:
729          */
730         list_for_each_entry(class, hash_head, hash_entry) {
731                 if (class->key == key) {
732                         WARN_ON_ONCE(class->name != lock->name);
733                         return class;
734                 }
735         }
736
737         return NULL;
738 }
739
740 /*
741  * Register a lock's class in the hash-table, if the class is not present
742  * yet. Otherwise we look it up. We cache the result in the lock object
743  * itself, so actual lookup of the hash should be once per lock object.
744  */
745 static inline struct lock_class *
746 register_lock_class(struct lockdep_map *lock, unsigned int subclass, int force)
747 {
748         struct lockdep_subclass_key *key;
749         struct list_head *hash_head;
750         struct lock_class *class;
751         unsigned long flags;
752
753         class = look_up_lock_class(lock, subclass);
754         if (likely(class))
755                 return class;
756
757         /*
758          * Debug-check: all keys must be persistent!
759          */
760         if (!static_obj(lock->key)) {
761                 debug_locks_off();
762                 printk("INFO: trying to register non-static key.\n");
763                 printk("the code is fine but needs lockdep annotation.\n");
764                 printk("turning off the locking correctness validator.\n");
765                 dump_stack();
766
767                 return NULL;
768         }
769
770         key = lock->key->subkeys + subclass;
771         hash_head = classhashentry(key);
772
773         raw_local_irq_save(flags);
774         if (!graph_lock()) {
775                 raw_local_irq_restore(flags);
776                 return NULL;
777         }
778         /*
779          * We have to do the hash-walk again, to avoid races
780          * with another CPU:
781          */
782         list_for_each_entry(class, hash_head, hash_entry)
783                 if (class->key == key)
784                         goto out_unlock_set;
785         /*
786          * Allocate a new key from the static array, and add it to
787          * the hash:
788          */
789         if (nr_lock_classes >= MAX_LOCKDEP_KEYS) {
790                 if (!debug_locks_off_graph_unlock()) {
791                         raw_local_irq_restore(flags);
792                         return NULL;
793                 }
794                 raw_local_irq_restore(flags);
795
796                 printk("BUG: MAX_LOCKDEP_KEYS too low!\n");
797                 printk("turning off the locking correctness validator.\n");
798                 dump_stack();
799                 return NULL;
800         }
801         class = lock_classes + nr_lock_classes++;
802         debug_atomic_inc(&nr_unused_locks);
803         class->key = key;
804         class->name = lock->name;
805         class->subclass = subclass;
806         INIT_LIST_HEAD(&class->lock_entry);
807         INIT_LIST_HEAD(&class->locks_before);
808         INIT_LIST_HEAD(&class->locks_after);
809         class->name_version = count_matching_names(class);
810         /*
811          * We use RCU's safe list-add method to make
812          * parallel walking of the hash-list safe:
813          */
814         list_add_tail_rcu(&class->hash_entry, hash_head);
815         /*
816          * Add it to the global list of classes:
817          */
818         list_add_tail_rcu(&class->lock_entry, &all_lock_classes);
819
820         if (verbose(class)) {
821                 graph_unlock();
822                 raw_local_irq_restore(flags);
823
824                 printk("\nnew class %p: %s", class->key, class->name);
825                 if (class->name_version > 1)
826                         printk("#%d", class->name_version);
827                 printk("\n");
828                 dump_stack();
829
830                 raw_local_irq_save(flags);
831                 if (!graph_lock()) {
832                         raw_local_irq_restore(flags);
833                         return NULL;
834                 }
835         }
836 out_unlock_set:
837         graph_unlock();
838         raw_local_irq_restore(flags);
839
840         if (!subclass || force)
841                 lock->class_cache = class;
842
843         if (DEBUG_LOCKS_WARN_ON(class->subclass != subclass))
844                 return NULL;
845
846         return class;
847 }
848
849 #ifdef CONFIG_PROVE_LOCKING
850 /*
851  * Allocate a lockdep entry. (assumes the graph_lock held, returns
852  * with NULL on failure)
853  */
854 static struct lock_list *alloc_list_entry(void)
855 {
856         if (nr_list_entries >= MAX_LOCKDEP_ENTRIES) {
857                 if (!debug_locks_off_graph_unlock())
858                         return NULL;
859
860                 printk("BUG: MAX_LOCKDEP_ENTRIES too low!\n");
861                 printk("turning off the locking correctness validator.\n");
862                 dump_stack();
863                 return NULL;
864         }
865         return list_entries + nr_list_entries++;
866 }
867
868 /*
869  * Add a new dependency to the head of the list:
870  */
871 static int add_lock_to_list(struct lock_class *class, struct lock_class *this,
872                             struct list_head *head, unsigned long ip, int distance)
873 {
874         struct lock_list *entry;
875         /*
876          * Lock not present yet - get a new dependency struct and
877          * add it to the list:
878          */
879         entry = alloc_list_entry();
880         if (!entry)
881                 return 0;
882
883         if (!save_trace(&entry->trace))
884                 return 0;
885
886         entry->class = this;
887         entry->distance = distance;
888         /*
889          * Since we never remove from the dependency list, the list can
890          * be walked lockless by other CPUs, it's only allocation
891          * that must be protected by the spinlock. But this also means
892          * we must make new entries visible only once writes to the
893          * entry become visible - hence the RCU op:
894          */
895         list_add_tail_rcu(&entry->entry, head);
896
897         return 1;
898 }
899
900 unsigned long bfs_accessed[BITS_TO_LONGS(MAX_LOCKDEP_ENTRIES)];
901 static struct circular_queue  lock_cq;
902
903 static int __bfs(struct lock_list *source_entry,
904                         void *data,
905                         int (*match)(struct lock_list *entry, void *data),
906                         struct lock_list **target_entry,
907                         int forward)
908 {
909         struct lock_list *entry;
910         struct list_head *head;
911         struct circular_queue *cq = &lock_cq;
912         int ret = 1;
913
914         if (match(source_entry, data)) {
915                 *target_entry = source_entry;
916                 ret = 0;
917                 goto exit;
918         }
919
920         if (forward)
921                 head = &source_entry->class->locks_after;
922         else
923                 head = &source_entry->class->locks_before;
924
925         if (list_empty(head))
926                 goto exit;
927
928         __cq_init(cq);
929         __cq_enqueue(cq, (unsigned long)source_entry);
930
931         while (!__cq_empty(cq)) {
932                 struct lock_list *lock;
933
934                 __cq_dequeue(cq, (unsigned long *)&lock);
935
936                 if (!lock->class) {
937                         ret = -2;
938                         goto exit;
939                 }
940
941                 if (forward)
942                         head = &lock->class->locks_after;
943                 else
944                         head = &lock->class->locks_before;
945
946                 list_for_each_entry(entry, head, entry) {
947                         if (!lock_accessed(entry)) {
948                                 mark_lock_accessed(entry, lock);
949                                 if (match(entry, data)) {
950                                         *target_entry = entry;
951                                         ret = 0;
952                                         goto exit;
953                                 }
954
955                                 if (__cq_enqueue(cq, (unsigned long)entry)) {
956                                         ret = -1;
957                                         goto exit;
958                                 }
959                         }
960                 }
961         }
962 exit:
963         return ret;
964 }
965
966 static inline int __bfs_forward(struct lock_list *src_entry,
967                         void *data,
968                         int (*match)(struct lock_list *entry, void *data),
969                         struct lock_list **target_entry)
970 {
971         return __bfs(src_entry, data, match, target_entry, 1);
972
973 }
974
975 static inline int __bfs_backward(struct lock_list *src_entry,
976                         void *data,
977                         int (*match)(struct lock_list *entry, void *data),
978                         struct lock_list **target_entry)
979 {
980         return __bfs(src_entry, data, match, target_entry, 0);
981
982 }
983
984 /*
985  * Recursive, forwards-direction lock-dependency checking, used for
986  * both noncyclic checking and for hardirq-unsafe/softirq-unsafe
987  * checking.
988  *
989  * (to keep the stackframe of the recursive functions small we
990  *  use these global variables, and we also mark various helper
991  *  functions as noinline.)
992  */
993 static struct held_lock *check_source, *check_target;
994
995 /*
996  * Print a dependency chain entry (this is only done when a deadlock
997  * has been detected):
998  */
999 static noinline int
1000 print_circular_bug_entry(struct lock_list *target, unsigned int depth)
1001 {
1002         if (debug_locks_silent)
1003                 return 0;
1004         printk("\n-> #%u", depth);
1005         print_lock_name(target->class);
1006         printk(":\n");
1007         print_stack_trace(&target->trace, 6);
1008
1009         return 0;
1010 }
1011
1012 /*
1013  * When a circular dependency is detected, print the
1014  * header first:
1015  */
1016 static noinline int
1017 print_circular_bug_header(struct lock_list *entry, unsigned int depth)
1018 {
1019         struct task_struct *curr = current;
1020
1021         if (debug_locks_silent)
1022                 return 0;
1023
1024         printk("\n=======================================================\n");
1025         printk(  "[ INFO: possible circular locking dependency detected ]\n");
1026         print_kernel_version();
1027         printk(  "-------------------------------------------------------\n");
1028         printk("%s/%d is trying to acquire lock:\n",
1029                 curr->comm, task_pid_nr(curr));
1030         print_lock(check_source);
1031         printk("\nbut task is already holding lock:\n");
1032         print_lock(check_target);
1033         printk("\nwhich lock already depends on the new lock.\n\n");
1034         printk("\nthe existing dependency chain (in reverse order) is:\n");
1035
1036         print_circular_bug_entry(entry, depth);
1037
1038         return 0;
1039 }
1040
1041 static inline int class_equal(struct lock_list *entry, void *data)
1042 {
1043         return entry->class == data;
1044 }
1045
1046 static noinline int print_circular_bug(void)
1047 {
1048         struct task_struct *curr = current;
1049         struct lock_list this;
1050         struct lock_list *target;
1051         struct lock_list *parent;
1052         int result;
1053         unsigned long depth;
1054
1055         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1056                 return 0;
1057
1058         this.class = hlock_class(check_source);
1059         this.parent = NULL;
1060         if (!save_trace(&this.trace))
1061                 return 0;
1062
1063         result = __bfs_forward(&this,
1064                         hlock_class(check_target),
1065                         class_equal,
1066                         &target);
1067         if (result) {
1068                 printk("\n%s:search shortest path failed:%d\n", __func__,
1069                         result);
1070                 return 0;
1071         }
1072
1073         depth = get_lock_depth(target);
1074
1075         print_circular_bug_header(target, depth);
1076
1077         parent = get_lock_parent(target);
1078
1079         while (parent) {
1080                 print_circular_bug_entry(parent, --depth);
1081                 parent = get_lock_parent(parent);
1082         }
1083
1084         printk("\nother info that might help us debug this:\n\n");
1085         lockdep_print_held_locks(curr);
1086
1087         printk("\nstack backtrace:\n");
1088         dump_stack();
1089
1090         return 0;
1091 }
1092
1093 #define RECURSION_LIMIT 40
1094
1095 static int noinline print_infinite_recursion_bug(void)
1096 {
1097         if (!debug_locks_off_graph_unlock())
1098                 return 0;
1099
1100         WARN_ON(1);
1101
1102         return 0;
1103 }
1104
1105 unsigned long __lockdep_count_forward_deps(struct lock_class *class,
1106                                            unsigned int depth)
1107 {
1108         struct lock_list *entry;
1109         unsigned long ret = 1;
1110
1111         if (lockdep_dependency_visit(class, depth))
1112                 return 0;
1113
1114         /*
1115          * Recurse this class's dependency list:
1116          */
1117         list_for_each_entry(entry, &class->locks_after, entry)
1118                 ret += __lockdep_count_forward_deps(entry->class, depth + 1);
1119
1120         return ret;
1121 }
1122
1123 unsigned long lockdep_count_forward_deps(struct lock_class *class)
1124 {
1125         unsigned long ret, flags;
1126
1127         local_irq_save(flags);
1128         __raw_spin_lock(&lockdep_lock);
1129         ret = __lockdep_count_forward_deps(class, 0);
1130         __raw_spin_unlock(&lockdep_lock);
1131         local_irq_restore(flags);
1132
1133         return ret;
1134 }
1135
1136 unsigned long __lockdep_count_backward_deps(struct lock_class *class,
1137                                             unsigned int depth)
1138 {
1139         struct lock_list *entry;
1140         unsigned long ret = 1;
1141
1142         if (lockdep_dependency_visit(class, depth))
1143                 return 0;
1144         /*
1145          * Recurse this class's dependency list:
1146          */
1147         list_for_each_entry(entry, &class->locks_before, entry)
1148                 ret += __lockdep_count_backward_deps(entry->class, depth + 1);
1149
1150         return ret;
1151 }
1152
1153 unsigned long lockdep_count_backward_deps(struct lock_class *class)
1154 {
1155         unsigned long ret, flags;
1156
1157         local_irq_save(flags);
1158         __raw_spin_lock(&lockdep_lock);
1159         ret = __lockdep_count_backward_deps(class, 0);
1160         __raw_spin_unlock(&lockdep_lock);
1161         local_irq_restore(flags);
1162
1163         return ret;
1164 }
1165
1166 /*
1167  * Prove that the dependency graph starting at <entry> can not
1168  * lead to <target>. Print an error and return 0 if it does.
1169  */
1170 static noinline int
1171 check_noncircular(struct lock_class *source, unsigned int depth)
1172 {
1173         struct lock_list *entry;
1174
1175         if (lockdep_dependency_visit(source, depth))
1176                 return 1;
1177
1178         debug_atomic_inc(&nr_cyclic_check_recursions);
1179         if (depth > max_recursion_depth)
1180                 max_recursion_depth = depth;
1181         if (depth >= RECURSION_LIMIT)
1182                 return print_infinite_recursion_bug();
1183         /*
1184          * Check this lock's dependency list:
1185          */
1186         list_for_each_entry(entry, &source->locks_after, entry) {
1187                 if (entry->class == hlock_class(check_target))
1188                         return 2;
1189                 debug_atomic_inc(&nr_cyclic_checks);
1190                 if (check_noncircular(entry->class, depth+1) == 2)
1191                         return 2;
1192         }
1193         return 1;
1194 }
1195
1196
1197 #if defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING)
1198 /*
1199  * Forwards and backwards subgraph searching, for the purposes of
1200  * proving that two subgraphs can be connected by a new dependency
1201  * without creating any illegal irq-safe -> irq-unsafe lock dependency.
1202  */
1203 static enum lock_usage_bit find_usage_bit;
1204 static struct lock_class *forwards_match, *backwards_match;
1205
1206 /*
1207  * Find a node in the forwards-direction dependency sub-graph starting
1208  * at <source> that matches <find_usage_bit>.
1209  *
1210  * Return 2 if such a node exists in the subgraph, and put that node
1211  * into <forwards_match>.
1212  *
1213  * Return 1 otherwise and keep <forwards_match> unchanged.
1214  * Return 0 on error.
1215  */
1216 static noinline int
1217 find_usage_forwards(struct lock_class *source, unsigned int depth)
1218 {
1219         struct lock_list *entry;
1220         int ret;
1221
1222         if (lockdep_dependency_visit(source, depth))
1223                 return 1;
1224
1225         if (depth > max_recursion_depth)
1226                 max_recursion_depth = depth;
1227         if (depth >= RECURSION_LIMIT)
1228                 return print_infinite_recursion_bug();
1229
1230         debug_atomic_inc(&nr_find_usage_forwards_checks);
1231         if (source->usage_mask & (1 << find_usage_bit)) {
1232                 forwards_match = source;
1233                 return 2;
1234         }
1235
1236         /*
1237          * Check this lock's dependency list:
1238          */
1239         list_for_each_entry(entry, &source->locks_after, entry) {
1240                 debug_atomic_inc(&nr_find_usage_forwards_recursions);
1241                 ret = find_usage_forwards(entry->class, depth+1);
1242                 if (ret == 2 || ret == 0)
1243                         return ret;
1244         }
1245         return 1;
1246 }
1247
1248 /*
1249  * Find a node in the backwards-direction dependency sub-graph starting
1250  * at <source> that matches <find_usage_bit>.
1251  *
1252  * Return 2 if such a node exists in the subgraph, and put that node
1253  * into <backwards_match>.
1254  *
1255  * Return 1 otherwise and keep <backwards_match> unchanged.
1256  * Return 0 on error.
1257  */
1258 static noinline int
1259 find_usage_backwards(struct lock_class *source, unsigned int depth)
1260 {
1261         struct lock_list *entry;
1262         int ret;
1263
1264         if (lockdep_dependency_visit(source, depth))
1265                 return 1;
1266
1267         if (!__raw_spin_is_locked(&lockdep_lock))
1268                 return DEBUG_LOCKS_WARN_ON(1);
1269
1270         if (depth > max_recursion_depth)
1271                 max_recursion_depth = depth;
1272         if (depth >= RECURSION_LIMIT)
1273                 return print_infinite_recursion_bug();
1274
1275         debug_atomic_inc(&nr_find_usage_backwards_checks);
1276         if (source->usage_mask & (1 << find_usage_bit)) {
1277                 backwards_match = source;
1278                 return 2;
1279         }
1280
1281         if (!source && debug_locks_off_graph_unlock()) {
1282                 WARN_ON(1);
1283                 return 0;
1284         }
1285
1286         /*
1287          * Check this lock's dependency list:
1288          */
1289         list_for_each_entry(entry, &source->locks_before, entry) {
1290                 debug_atomic_inc(&nr_find_usage_backwards_recursions);
1291                 ret = find_usage_backwards(entry->class, depth+1);
1292                 if (ret == 2 || ret == 0)
1293                         return ret;
1294         }
1295         return 1;
1296 }
1297
1298 static int
1299 print_bad_irq_dependency(struct task_struct *curr,
1300                          struct held_lock *prev,
1301                          struct held_lock *next,
1302                          enum lock_usage_bit bit1,
1303                          enum lock_usage_bit bit2,
1304                          const char *irqclass)
1305 {
1306         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1307                 return 0;
1308
1309         printk("\n======================================================\n");
1310         printk(  "[ INFO: %s-safe -> %s-unsafe lock order detected ]\n",
1311                 irqclass, irqclass);
1312         print_kernel_version();
1313         printk(  "------------------------------------------------------\n");
1314         printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] is trying to acquire:\n",
1315                 curr->comm, task_pid_nr(curr),
1316                 curr->hardirq_context, hardirq_count() >> HARDIRQ_SHIFT,
1317                 curr->softirq_context, softirq_count() >> SOFTIRQ_SHIFT,
1318                 curr->hardirqs_enabled,
1319                 curr->softirqs_enabled);
1320         print_lock(next);
1321
1322         printk("\nand this task is already holding:\n");
1323         print_lock(prev);
1324         printk("which would create a new lock dependency:\n");
1325         print_lock_name(hlock_class(prev));
1326         printk(" ->");
1327         print_lock_name(hlock_class(next));
1328         printk("\n");
1329
1330         printk("\nbut this new dependency connects a %s-irq-safe lock:\n",
1331                 irqclass);
1332         print_lock_name(backwards_match);
1333         printk("\n... which became %s-irq-safe at:\n", irqclass);
1334
1335         print_stack_trace(backwards_match->usage_traces + bit1, 1);
1336
1337         printk("\nto a %s-irq-unsafe lock:\n", irqclass);
1338         print_lock_name(forwards_match);
1339         printk("\n... which became %s-irq-unsafe at:\n", irqclass);
1340         printk("...");
1341
1342         print_stack_trace(forwards_match->usage_traces + bit2, 1);
1343
1344         printk("\nother info that might help us debug this:\n\n");
1345         lockdep_print_held_locks(curr);
1346
1347         printk("\nthe %s-irq-safe lock's dependencies:\n", irqclass);
1348         print_lock_dependencies(backwards_match, 0);
1349
1350         printk("\nthe %s-irq-unsafe lock's dependencies:\n", irqclass);
1351         print_lock_dependencies(forwards_match, 0);
1352
1353         printk("\nstack backtrace:\n");
1354         dump_stack();
1355
1356         return 0;
1357 }
1358
1359 static int
1360 check_usage(struct task_struct *curr, struct held_lock *prev,
1361             struct held_lock *next, enum lock_usage_bit bit_backwards,
1362             enum lock_usage_bit bit_forwards, const char *irqclass)
1363 {
1364         int ret;
1365
1366         find_usage_bit = bit_backwards;
1367         /* fills in <backwards_match> */
1368         ret = find_usage_backwards(hlock_class(prev), 0);
1369         if (!ret || ret == 1)
1370                 return ret;
1371
1372         find_usage_bit = bit_forwards;
1373         ret = find_usage_forwards(hlock_class(next), 0);
1374         if (!ret || ret == 1)
1375                 return ret;
1376         /* ret == 2 */
1377         return print_bad_irq_dependency(curr, prev, next,
1378                         bit_backwards, bit_forwards, irqclass);
1379 }
1380
1381 static const char *state_names[] = {
1382 #define LOCKDEP_STATE(__STATE) \
1383         __stringify(__STATE),
1384 #include "lockdep_states.h"
1385 #undef LOCKDEP_STATE
1386 };
1387
1388 static const char *state_rnames[] = {
1389 #define LOCKDEP_STATE(__STATE) \
1390         __stringify(__STATE)"-READ",
1391 #include "lockdep_states.h"
1392 #undef LOCKDEP_STATE
1393 };
1394
1395 static inline const char *state_name(enum lock_usage_bit bit)
1396 {
1397         return (bit & 1) ? state_rnames[bit >> 2] : state_names[bit >> 2];
1398 }
1399
1400 static int exclusive_bit(int new_bit)
1401 {
1402         /*
1403          * USED_IN
1404          * USED_IN_READ
1405          * ENABLED
1406          * ENABLED_READ
1407          *
1408          * bit 0 - write/read
1409          * bit 1 - used_in/enabled
1410          * bit 2+  state
1411          */
1412
1413         int state = new_bit & ~3;
1414         int dir = new_bit & 2;
1415
1416         /*
1417          * keep state, bit flip the direction and strip read.
1418          */
1419         return state | (dir ^ 2);
1420 }
1421
1422 static int check_irq_usage(struct task_struct *curr, struct held_lock *prev,
1423                            struct held_lock *next, enum lock_usage_bit bit)
1424 {
1425         /*
1426          * Prove that the new dependency does not connect a hardirq-safe
1427          * lock with a hardirq-unsafe lock - to achieve this we search
1428          * the backwards-subgraph starting at <prev>, and the
1429          * forwards-subgraph starting at <next>:
1430          */
1431         if (!check_usage(curr, prev, next, bit,
1432                            exclusive_bit(bit), state_name(bit)))
1433                 return 0;
1434
1435         bit++; /* _READ */
1436
1437         /*
1438          * Prove that the new dependency does not connect a hardirq-safe-read
1439          * lock with a hardirq-unsafe lock - to achieve this we search
1440          * the backwards-subgraph starting at <prev>, and the
1441          * forwards-subgraph starting at <next>:
1442          */
1443         if (!check_usage(curr, prev, next, bit,
1444                            exclusive_bit(bit), state_name(bit)))
1445                 return 0;
1446
1447         return 1;
1448 }
1449
1450 static int
1451 check_prev_add_irq(struct task_struct *curr, struct held_lock *prev,
1452                 struct held_lock *next)
1453 {
1454 #define LOCKDEP_STATE(__STATE)                                          \
1455         if (!check_irq_usage(curr, prev, next, LOCK_USED_IN_##__STATE)) \
1456                 return 0;
1457 #include "lockdep_states.h"
1458 #undef LOCKDEP_STATE
1459
1460         return 1;
1461 }
1462
1463 static void inc_chains(void)
1464 {
1465         if (current->hardirq_context)
1466                 nr_hardirq_chains++;
1467         else {
1468                 if (current->softirq_context)
1469                         nr_softirq_chains++;
1470                 else
1471                         nr_process_chains++;
1472         }
1473 }
1474
1475 #else
1476
1477 static inline int
1478 check_prev_add_irq(struct task_struct *curr, struct held_lock *prev,
1479                 struct held_lock *next)
1480 {
1481         return 1;
1482 }
1483
1484 static inline void inc_chains(void)
1485 {
1486         nr_process_chains++;
1487 }
1488
1489 #endif
1490
1491 static int
1492 print_deadlock_bug(struct task_struct *curr, struct held_lock *prev,
1493                    struct held_lock *next)
1494 {
1495         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1496                 return 0;
1497
1498         printk("\n=============================================\n");
1499         printk(  "[ INFO: possible recursive locking detected ]\n");
1500         print_kernel_version();
1501         printk(  "---------------------------------------------\n");
1502         printk("%s/%d is trying to acquire lock:\n",
1503                 curr->comm, task_pid_nr(curr));
1504         print_lock(next);
1505         printk("\nbut task is already holding lock:\n");
1506         print_lock(prev);
1507
1508         printk("\nother info that might help us debug this:\n");
1509         lockdep_print_held_locks(curr);
1510
1511         printk("\nstack backtrace:\n");
1512         dump_stack();
1513
1514         return 0;
1515 }
1516
1517 /*
1518  * Check whether we are holding such a class already.
1519  *
1520  * (Note that this has to be done separately, because the graph cannot
1521  * detect such classes of deadlocks.)
1522  *
1523  * Returns: 0 on deadlock detected, 1 on OK, 2 on recursive read
1524  */
1525 static int
1526 check_deadlock(struct task_struct *curr, struct held_lock *next,
1527                struct lockdep_map *next_instance, int read)
1528 {
1529         struct held_lock *prev;
1530         struct held_lock *nest = NULL;
1531         int i;
1532
1533         for (i = 0; i < curr->lockdep_depth; i++) {
1534                 prev = curr->held_locks + i;
1535
1536                 if (prev->instance == next->nest_lock)
1537                         nest = prev;
1538
1539                 if (hlock_class(prev) != hlock_class(next))
1540                         continue;
1541
1542                 /*
1543                  * Allow read-after-read recursion of the same
1544                  * lock class (i.e. read_lock(lock)+read_lock(lock)):
1545                  */
1546                 if ((read == 2) && prev->read)
1547                         return 2;
1548
1549                 /*
1550                  * We're holding the nest_lock, which serializes this lock's
1551                  * nesting behaviour.
1552                  */
1553                 if (nest)
1554                         return 2;
1555
1556                 return print_deadlock_bug(curr, prev, next);
1557         }
1558         return 1;
1559 }
1560
1561 /*
1562  * There was a chain-cache miss, and we are about to add a new dependency
1563  * to a previous lock. We recursively validate the following rules:
1564  *
1565  *  - would the adding of the <prev> -> <next> dependency create a
1566  *    circular dependency in the graph? [== circular deadlock]
1567  *
1568  *  - does the new prev->next dependency connect any hardirq-safe lock
1569  *    (in the full backwards-subgraph starting at <prev>) with any
1570  *    hardirq-unsafe lock (in the full forwards-subgraph starting at
1571  *    <next>)? [== illegal lock inversion with hardirq contexts]
1572  *
1573  *  - does the new prev->next dependency connect any softirq-safe lock
1574  *    (in the full backwards-subgraph starting at <prev>) with any
1575  *    softirq-unsafe lock (in the full forwards-subgraph starting at
1576  *    <next>)? [== illegal lock inversion with softirq contexts]
1577  *
1578  * any of these scenarios could lead to a deadlock.
1579  *
1580  * Then if all the validations pass, we add the forwards and backwards
1581  * dependency.
1582  */
1583 static int
1584 check_prev_add(struct task_struct *curr, struct held_lock *prev,
1585                struct held_lock *next, int distance)
1586 {
1587         struct lock_list *entry;
1588         int ret;
1589
1590         /*
1591          * Prove that the new <prev> -> <next> dependency would not
1592          * create a circular dependency in the graph. (We do this by
1593          * forward-recursing into the graph starting at <next>, and
1594          * checking whether we can reach <prev>.)
1595          *
1596          * We are using global variables to control the recursion, to
1597          * keep the stackframe size of the recursive functions low:
1598          */
1599         check_source = next;
1600         check_target = prev;
1601
1602         if (check_noncircular(hlock_class(next), 0) == 2)
1603                 return print_circular_bug();
1604
1605         if (!check_prev_add_irq(curr, prev, next))
1606                 return 0;
1607
1608         /*
1609          * For recursive read-locks we do all the dependency checks,
1610          * but we dont store read-triggered dependencies (only
1611          * write-triggered dependencies). This ensures that only the
1612          * write-side dependencies matter, and that if for example a
1613          * write-lock never takes any other locks, then the reads are
1614          * equivalent to a NOP.
1615          */
1616         if (next->read == 2 || prev->read == 2)
1617                 return 1;
1618         /*
1619          * Is the <prev> -> <next> dependency already present?
1620          *
1621          * (this may occur even though this is a new chain: consider
1622          *  e.g. the L1 -> L2 -> L3 -> L4 and the L5 -> L1 -> L2 -> L3
1623          *  chains - the second one will be new, but L1 already has
1624          *  L2 added to its dependency list, due to the first chain.)
1625          */
1626         list_for_each_entry(entry, &hlock_class(prev)->locks_after, entry) {
1627                 if (entry->class == hlock_class(next)) {
1628                         if (distance == 1)
1629                                 entry->distance = 1;
1630                         return 2;
1631                 }
1632         }
1633
1634         /*
1635          * Ok, all validations passed, add the new lock
1636          * to the previous lock's dependency list:
1637          */
1638         ret = add_lock_to_list(hlock_class(prev), hlock_class(next),
1639                                &hlock_class(prev)->locks_after,
1640                                next->acquire_ip, distance);
1641
1642         if (!ret)
1643                 return 0;
1644
1645         ret = add_lock_to_list(hlock_class(next), hlock_class(prev),
1646                                &hlock_class(next)->locks_before,
1647                                next->acquire_ip, distance);
1648         if (!ret)
1649                 return 0;
1650
1651         /*
1652          * Debugging printouts:
1653          */
1654         if (verbose(hlock_class(prev)) || verbose(hlock_class(next))) {
1655                 graph_unlock();
1656                 printk("\n new dependency: ");
1657                 print_lock_name(hlock_class(prev));
1658                 printk(" => ");
1659                 print_lock_name(hlock_class(next));
1660                 printk("\n");
1661                 dump_stack();
1662                 return graph_lock();
1663         }
1664         return 1;
1665 }
1666
1667 /*
1668  * Add the dependency to all directly-previous locks that are 'relevant'.
1669  * The ones that are relevant are (in increasing distance from curr):
1670  * all consecutive trylock entries and the final non-trylock entry - or
1671  * the end of this context's lock-chain - whichever comes first.
1672  */
1673 static int
1674 check_prevs_add(struct task_struct *curr, struct held_lock *next)
1675 {
1676         int depth = curr->lockdep_depth;
1677         struct held_lock *hlock;
1678
1679         /*
1680          * Debugging checks.
1681          *
1682          * Depth must not be zero for a non-head lock:
1683          */
1684         if (!depth)
1685                 goto out_bug;
1686         /*
1687          * At least two relevant locks must exist for this
1688          * to be a head:
1689          */
1690         if (curr->held_locks[depth].irq_context !=
1691                         curr->held_locks[depth-1].irq_context)
1692                 goto out_bug;
1693
1694         for (;;) {
1695                 int distance = curr->lockdep_depth - depth + 1;
1696                 hlock = curr->held_locks + depth-1;
1697                 /*
1698                  * Only non-recursive-read entries get new dependencies
1699                  * added:
1700                  */
1701                 if (hlock->read != 2) {
1702                         if (!check_prev_add(curr, hlock, next, distance))
1703                                 return 0;
1704                         /*
1705                          * Stop after the first non-trylock entry,
1706                          * as non-trylock entries have added their
1707                          * own direct dependencies already, so this
1708                          * lock is connected to them indirectly:
1709                          */
1710                         if (!hlock->trylock)
1711                                 break;
1712                 }
1713                 depth--;
1714                 /*
1715                  * End of lock-stack?
1716                  */
1717                 if (!depth)
1718                         break;
1719                 /*
1720                  * Stop the search if we cross into another context:
1721                  */
1722                 if (curr->held_locks[depth].irq_context !=
1723                                 curr->held_locks[depth-1].irq_context)
1724                         break;
1725         }
1726         return 1;
1727 out_bug:
1728         if (!debug_locks_off_graph_unlock())
1729                 return 0;
1730
1731         WARN_ON(1);
1732
1733         return 0;
1734 }
1735
1736 unsigned long nr_lock_chains;
1737 struct lock_chain lock_chains[MAX_LOCKDEP_CHAINS];
1738 int nr_chain_hlocks;
1739 static u16 chain_hlocks[MAX_LOCKDEP_CHAIN_HLOCKS];
1740
1741 struct lock_class *lock_chain_get_class(struct lock_chain *chain, int i)
1742 {
1743         return lock_classes + chain_hlocks[chain->base + i];
1744 }
1745
1746 /*
1747  * Look up a dependency chain. If the key is not present yet then
1748  * add it and return 1 - in this case the new dependency chain is
1749  * validated. If the key is already hashed, return 0.
1750  * (On return with 1 graph_lock is held.)
1751  */
1752 static inline int lookup_chain_cache(struct task_struct *curr,
1753                                      struct held_lock *hlock,
1754                                      u64 chain_key)
1755 {
1756         struct lock_class *class = hlock_class(hlock);
1757         struct list_head *hash_head = chainhashentry(chain_key);
1758         struct lock_chain *chain;
1759         struct held_lock *hlock_curr, *hlock_next;
1760         int i, j, n, cn;
1761
1762         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
1763                 return 0;
1764         /*
1765          * We can walk it lock-free, because entries only get added
1766          * to the hash:
1767          */
1768         list_for_each_entry(chain, hash_head, entry) {
1769                 if (chain->chain_key == chain_key) {
1770 cache_hit:
1771                         debug_atomic_inc(&chain_lookup_hits);
1772                         if (very_verbose(class))
1773                                 printk("\nhash chain already cached, key: "
1774                                         "%016Lx tail class: [%p] %s\n",
1775                                         (unsigned long long)chain_key,
1776                                         class->key, class->name);
1777                         return 0;
1778                 }
1779         }
1780         if (very_verbose(class))
1781                 printk("\nnew hash chain, key: %016Lx tail class: [%p] %s\n",
1782                         (unsigned long long)chain_key, class->key, class->name);
1783         /*
1784          * Allocate a new chain entry from the static array, and add
1785          * it to the hash:
1786          */
1787         if (!graph_lock())
1788                 return 0;
1789         /*
1790          * We have to walk the chain again locked - to avoid duplicates:
1791          */
1792         list_for_each_entry(chain, hash_head, entry) {
1793                 if (chain->chain_key == chain_key) {
1794                         graph_unlock();
1795                         goto cache_hit;
1796                 }
1797         }
1798         if (unlikely(nr_lock_chains >= MAX_LOCKDEP_CHAINS)) {
1799                 if (!debug_locks_off_graph_unlock())
1800                         return 0;
1801
1802                 printk("BUG: MAX_LOCKDEP_CHAINS too low!\n");
1803                 printk("turning off the locking correctness validator.\n");
1804                 dump_stack();
1805                 return 0;
1806         }
1807         chain = lock_chains + nr_lock_chains++;
1808         chain->chain_key = chain_key;
1809         chain->irq_context = hlock->irq_context;
1810         /* Find the first held_lock of current chain */
1811         hlock_next = hlock;
1812         for (i = curr->lockdep_depth - 1; i >= 0; i--) {
1813                 hlock_curr = curr->held_locks + i;
1814                 if (hlock_curr->irq_context != hlock_next->irq_context)
1815                         break;
1816                 hlock_next = hlock;
1817         }
1818         i++;
1819         chain->depth = curr->lockdep_depth + 1 - i;
1820         cn = nr_chain_hlocks;
1821         while (cn + chain->depth <= MAX_LOCKDEP_CHAIN_HLOCKS) {
1822                 n = cmpxchg(&nr_chain_hlocks, cn, cn + chain->depth);
1823                 if (n == cn)
1824                         break;
1825                 cn = n;
1826         }
1827         if (likely(cn + chain->depth <= MAX_LOCKDEP_CHAIN_HLOCKS)) {
1828                 chain->base = cn;
1829                 for (j = 0; j < chain->depth - 1; j++, i++) {
1830                         int lock_id = curr->held_locks[i].class_idx - 1;
1831                         chain_hlocks[chain->base + j] = lock_id;
1832                 }
1833                 chain_hlocks[chain->base + j] = class - lock_classes;
1834         }
1835         list_add_tail_rcu(&chain->entry, hash_head);
1836         debug_atomic_inc(&chain_lookup_misses);
1837         inc_chains();
1838
1839         return 1;
1840 }
1841
1842 static int validate_chain(struct task_struct *curr, struct lockdep_map *lock,
1843                 struct held_lock *hlock, int chain_head, u64 chain_key)
1844 {
1845         /*
1846          * Trylock needs to maintain the stack of held locks, but it
1847          * does not add new dependencies, because trylock can be done
1848          * in any order.
1849          *
1850          * We look up the chain_key and do the O(N^2) check and update of
1851          * the dependencies only if this is a new dependency chain.
1852          * (If lookup_chain_cache() returns with 1 it acquires
1853          * graph_lock for us)
1854          */
1855         if (!hlock->trylock && (hlock->check == 2) &&
1856             lookup_chain_cache(curr, hlock, chain_key)) {
1857                 /*
1858                  * Check whether last held lock:
1859                  *
1860                  * - is irq-safe, if this lock is irq-unsafe
1861                  * - is softirq-safe, if this lock is hardirq-unsafe
1862                  *
1863                  * And check whether the new lock's dependency graph
1864                  * could lead back to the previous lock.
1865                  *
1866                  * any of these scenarios could lead to a deadlock. If
1867                  * All validations
1868                  */
1869                 int ret = check_deadlock(curr, hlock, lock, hlock->read);
1870
1871                 if (!ret)
1872                         return 0;
1873                 /*
1874                  * Mark recursive read, as we jump over it when
1875                  * building dependencies (just like we jump over
1876                  * trylock entries):
1877                  */
1878                 if (ret == 2)
1879                         hlock->read = 2;
1880                 /*
1881                  * Add dependency only if this lock is not the head
1882                  * of the chain, and if it's not a secondary read-lock:
1883                  */
1884                 if (!chain_head && ret != 2)
1885                         if (!check_prevs_add(curr, hlock))
1886                                 return 0;
1887                 graph_unlock();
1888         } else
1889                 /* after lookup_chain_cache(): */
1890                 if (unlikely(!debug_locks))
1891                         return 0;
1892
1893         return 1;
1894 }
1895 #else
1896 static inline int validate_chain(struct task_struct *curr,
1897                 struct lockdep_map *lock, struct held_lock *hlock,
1898                 int chain_head, u64 chain_key)
1899 {
1900         return 1;
1901 }
1902 #endif
1903
1904 /*
1905  * We are building curr_chain_key incrementally, so double-check
1906  * it from scratch, to make sure that it's done correctly:
1907  */
1908 static void check_chain_key(struct task_struct *curr)
1909 {
1910 #ifdef CONFIG_DEBUG_LOCKDEP
1911         struct held_lock *hlock, *prev_hlock = NULL;
1912         unsigned int i, id;
1913         u64 chain_key = 0;
1914
1915         for (i = 0; i < curr->lockdep_depth; i++) {
1916                 hlock = curr->held_locks + i;
1917                 if (chain_key != hlock->prev_chain_key) {
1918                         debug_locks_off();
1919                         WARN(1, "hm#1, depth: %u [%u], %016Lx != %016Lx\n",
1920                                 curr->lockdep_depth, i,
1921                                 (unsigned long long)chain_key,
1922                                 (unsigned long long)hlock->prev_chain_key);
1923                         return;
1924                 }
1925                 id = hlock->class_idx - 1;
1926                 if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
1927                         return;
1928
1929                 if (prev_hlock && (prev_hlock->irq_context !=
1930                                                         hlock->irq_context))
1931                         chain_key = 0;
1932                 chain_key = iterate_chain_key(chain_key, id);
1933                 prev_hlock = hlock;
1934         }
1935         if (chain_key != curr->curr_chain_key) {
1936                 debug_locks_off();
1937                 WARN(1, "hm#2, depth: %u [%u], %016Lx != %016Lx\n",
1938                         curr->lockdep_depth, i,
1939                         (unsigned long long)chain_key,
1940                         (unsigned long long)curr->curr_chain_key);
1941         }
1942 #endif
1943 }
1944
1945 static int
1946 print_usage_bug(struct task_struct *curr, struct held_lock *this,
1947                 enum lock_usage_bit prev_bit, enum lock_usage_bit new_bit)
1948 {
1949         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1950                 return 0;
1951
1952         printk("\n=================================\n");
1953         printk(  "[ INFO: inconsistent lock state ]\n");
1954         print_kernel_version();
1955         printk(  "---------------------------------\n");
1956
1957         printk("inconsistent {%s} -> {%s} usage.\n",
1958                 usage_str[prev_bit], usage_str[new_bit]);
1959
1960         printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] takes:\n",
1961                 curr->comm, task_pid_nr(curr),
1962                 trace_hardirq_context(curr), hardirq_count() >> HARDIRQ_SHIFT,
1963                 trace_softirq_context(curr), softirq_count() >> SOFTIRQ_SHIFT,
1964                 trace_hardirqs_enabled(curr),
1965                 trace_softirqs_enabled(curr));
1966         print_lock(this);
1967
1968         printk("{%s} state was registered at:\n", usage_str[prev_bit]);
1969         print_stack_trace(hlock_class(this)->usage_traces + prev_bit, 1);
1970
1971         print_irqtrace_events(curr);
1972         printk("\nother info that might help us debug this:\n");
1973         lockdep_print_held_locks(curr);
1974
1975         printk("\nstack backtrace:\n");
1976         dump_stack();
1977
1978         return 0;
1979 }
1980
1981 /*
1982  * Print out an error if an invalid bit is set:
1983  */
1984 static inline int
1985 valid_state(struct task_struct *curr, struct held_lock *this,
1986             enum lock_usage_bit new_bit, enum lock_usage_bit bad_bit)
1987 {
1988         if (unlikely(hlock_class(this)->usage_mask & (1 << bad_bit)))
1989                 return print_usage_bug(curr, this, bad_bit, new_bit);
1990         return 1;
1991 }
1992
1993 static int mark_lock(struct task_struct *curr, struct held_lock *this,
1994                      enum lock_usage_bit new_bit);
1995
1996 #if defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING)
1997
1998 /*
1999  * print irq inversion bug:
2000  */
2001 static int
2002 print_irq_inversion_bug(struct task_struct *curr, struct lock_class *other,
2003                         struct held_lock *this, int forwards,
2004                         const char *irqclass)
2005 {
2006         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
2007                 return 0;
2008
2009         printk("\n=========================================================\n");
2010         printk(  "[ INFO: possible irq lock inversion dependency detected ]\n");
2011         print_kernel_version();
2012         printk(  "---------------------------------------------------------\n");
2013         printk("%s/%d just changed the state of lock:\n",
2014                 curr->comm, task_pid_nr(curr));
2015         print_lock(this);
2016         if (forwards)
2017                 printk("but this lock took another, %s-unsafe lock in the past:\n", irqclass);
2018         else
2019                 printk("but this lock was taken by another, %s-safe lock in the past:\n", irqclass);
2020         print_lock_name(other);
2021         printk("\n\nand interrupts could create inverse lock ordering between them.\n\n");
2022
2023         printk("\nother info that might help us debug this:\n");
2024         lockdep_print_held_locks(curr);
2025
2026         printk("\nthe first lock's dependencies:\n");
2027         print_lock_dependencies(hlock_class(this), 0);
2028
2029         printk("\nthe second lock's dependencies:\n");
2030         print_lock_dependencies(other, 0);
2031
2032         printk("\nstack backtrace:\n");
2033         dump_stack();
2034
2035         return 0;
2036 }
2037
2038 /*
2039  * Prove that in the forwards-direction subgraph starting at <this>
2040  * there is no lock matching <mask>:
2041  */
2042 static int
2043 check_usage_forwards(struct task_struct *curr, struct held_lock *this,
2044                      enum lock_usage_bit bit, const char *irqclass)
2045 {
2046         int ret;
2047
2048         find_usage_bit = bit;
2049         /* fills in <forwards_match> */
2050         ret = find_usage_forwards(hlock_class(this), 0);
2051         if (!ret || ret == 1)
2052                 return ret;
2053
2054         return print_irq_inversion_bug(curr, forwards_match, this, 1, irqclass);
2055 }
2056
2057 /*
2058  * Prove that in the backwards-direction subgraph starting at <this>
2059  * there is no lock matching <mask>:
2060  */
2061 static int
2062 check_usage_backwards(struct task_struct *curr, struct held_lock *this,
2063                       enum lock_usage_bit bit, const char *irqclass)
2064 {
2065         int ret;
2066
2067         find_usage_bit = bit;
2068         /* fills in <backwards_match> */
2069         ret = find_usage_backwards(hlock_class(this), 0);
2070         if (!ret || ret == 1)
2071                 return ret;
2072
2073         return print_irq_inversion_bug(curr, backwards_match, this, 0, irqclass);
2074 }
2075
2076 void print_irqtrace_events(struct task_struct *curr)
2077 {
2078         printk("irq event stamp: %u\n", curr->irq_events);
2079         printk("hardirqs last  enabled at (%u): ", curr->hardirq_enable_event);
2080         print_ip_sym(curr->hardirq_enable_ip);
2081         printk("hardirqs last disabled at (%u): ", curr->hardirq_disable_event);
2082         print_ip_sym(curr->hardirq_disable_ip);
2083         printk("softirqs last  enabled at (%u): ", curr->softirq_enable_event);
2084         print_ip_sym(curr->softirq_enable_ip);
2085         printk("softirqs last disabled at (%u): ", curr->softirq_disable_event);
2086         print_ip_sym(curr->softirq_disable_ip);
2087 }
2088
2089 static int HARDIRQ_verbose(struct lock_class *class)
2090 {
2091 #if HARDIRQ_VERBOSE
2092         return class_filter(class);
2093 #endif
2094         return 0;
2095 }
2096
2097 static int SOFTIRQ_verbose(struct lock_class *class)
2098 {
2099 #if SOFTIRQ_VERBOSE
2100         return class_filter(class);
2101 #endif
2102         return 0;
2103 }
2104
2105 static int RECLAIM_FS_verbose(struct lock_class *class)
2106 {
2107 #if RECLAIM_VERBOSE
2108         return class_filter(class);
2109 #endif
2110         return 0;
2111 }
2112
2113 #define STRICT_READ_CHECKS      1
2114
2115 static int (*state_verbose_f[])(struct lock_class *class) = {
2116 #define LOCKDEP_STATE(__STATE) \
2117         __STATE##_verbose,
2118 #include "lockdep_states.h"
2119 #undef LOCKDEP_STATE
2120 };
2121
2122 static inline int state_verbose(enum lock_usage_bit bit,
2123                                 struct lock_class *class)
2124 {
2125         return state_verbose_f[bit >> 2](class);
2126 }
2127
2128 typedef int (*check_usage_f)(struct task_struct *, struct held_lock *,
2129                              enum lock_usage_bit bit, const char *name);
2130
2131 static int
2132 mark_lock_irq(struct task_struct *curr, struct held_lock *this,
2133                 enum lock_usage_bit new_bit)
2134 {
2135         int excl_bit = exclusive_bit(new_bit);
2136         int read = new_bit & 1;
2137         int dir = new_bit & 2;
2138
2139         /*
2140          * mark USED_IN has to look forwards -- to ensure no dependency
2141          * has ENABLED state, which would allow recursion deadlocks.
2142          *
2143          * mark ENABLED has to look backwards -- to ensure no dependee
2144          * has USED_IN state, which, again, would allow  recursion deadlocks.
2145          */
2146         check_usage_f usage = dir ?
2147                 check_usage_backwards : check_usage_forwards;
2148
2149         /*
2150          * Validate that this particular lock does not have conflicting
2151          * usage states.
2152          */
2153         if (!valid_state(curr, this, new_bit, excl_bit))
2154                 return 0;
2155
2156         /*
2157          * Validate that the lock dependencies don't have conflicting usage
2158          * states.
2159          */
2160         if ((!read || !dir || STRICT_READ_CHECKS) &&
2161                         !usage(curr, this, excl_bit, state_name(new_bit & ~1)))
2162                 return 0;
2163
2164         /*
2165          * Check for read in write conflicts
2166          */
2167         if (!read) {
2168                 if (!valid_state(curr, this, new_bit, excl_bit + 1))
2169                         return 0;
2170
2171                 if (STRICT_READ_CHECKS &&
2172                         !usage(curr, this, excl_bit + 1,
2173                                 state_name(new_bit + 1)))
2174                         return 0;
2175         }
2176
2177         if (state_verbose(new_bit, hlock_class(this)))
2178                 return 2;
2179
2180         return 1;
2181 }
2182
2183 enum mark_type {
2184 #define LOCKDEP_STATE(__STATE)  __STATE,
2185 #include "lockdep_states.h"
2186 #undef LOCKDEP_STATE
2187 };
2188
2189 /*
2190  * Mark all held locks with a usage bit:
2191  */
2192 static int
2193 mark_held_locks(struct task_struct *curr, enum mark_type mark)
2194 {
2195         enum lock_usage_bit usage_bit;
2196         struct held_lock *hlock;
2197         int i;
2198
2199         for (i = 0; i < curr->lockdep_depth; i++) {
2200                 hlock = curr->held_locks + i;
2201
2202                 usage_bit = 2 + (mark << 2); /* ENABLED */
2203                 if (hlock->read)
2204                         usage_bit += 1; /* READ */
2205
2206                 BUG_ON(usage_bit >= LOCK_USAGE_STATES);
2207
2208                 if (!mark_lock(curr, hlock, usage_bit))
2209                         return 0;
2210         }
2211
2212         return 1;
2213 }
2214
2215 /*
2216  * Debugging helper: via this flag we know that we are in
2217  * 'early bootup code', and will warn about any invalid irqs-on event:
2218  */
2219 static int early_boot_irqs_enabled;
2220
2221 void early_boot_irqs_off(void)
2222 {
2223         early_boot_irqs_enabled = 0;
2224 }
2225
2226 void early_boot_irqs_on(void)
2227 {
2228         early_boot_irqs_enabled = 1;
2229 }
2230
2231 /*
2232  * Hardirqs will be enabled:
2233  */
2234 void trace_hardirqs_on_caller(unsigned long ip)
2235 {
2236         struct task_struct *curr = current;
2237
2238         time_hardirqs_on(CALLER_ADDR0, ip);
2239
2240         if (unlikely(!debug_locks || current->lockdep_recursion))
2241                 return;
2242
2243         if (DEBUG_LOCKS_WARN_ON(unlikely(!early_boot_irqs_enabled)))
2244                 return;
2245
2246         if (unlikely(curr->hardirqs_enabled)) {
2247                 debug_atomic_inc(&redundant_hardirqs_on);
2248                 return;
2249         }
2250         /* we'll do an OFF -> ON transition: */
2251         curr->hardirqs_enabled = 1;
2252
2253         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2254                 return;
2255         if (DEBUG_LOCKS_WARN_ON(current->hardirq_context))
2256                 return;
2257         /*
2258          * We are going to turn hardirqs on, so set the
2259          * usage bit for all held locks:
2260          */
2261         if (!mark_held_locks(curr, HARDIRQ))
2262                 return;
2263         /*
2264          * If we have softirqs enabled, then set the usage
2265          * bit for all held locks. (disabled hardirqs prevented
2266          * this bit from being set before)
2267          */
2268         if (curr->softirqs_enabled)
2269                 if (!mark_held_locks(curr, SOFTIRQ))
2270                         return;
2271
2272         curr->hardirq_enable_ip = ip;
2273         curr->hardirq_enable_event = ++curr->irq_events;
2274         debug_atomic_inc(&hardirqs_on_events);
2275 }
2276 EXPORT_SYMBOL(trace_hardirqs_on_caller);
2277
2278 void trace_hardirqs_on(void)
2279 {
2280         trace_hardirqs_on_caller(CALLER_ADDR0);
2281 }
2282 EXPORT_SYMBOL(trace_hardirqs_on);
2283
2284 /*
2285  * Hardirqs were disabled:
2286  */
2287 void trace_hardirqs_off_caller(unsigned long ip)
2288 {
2289         struct task_struct *curr = current;
2290
2291         time_hardirqs_off(CALLER_ADDR0, ip);
2292
2293         if (unlikely(!debug_locks || current->lockdep_recursion))
2294                 return;
2295
2296         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2297                 return;
2298
2299         if (curr->hardirqs_enabled) {
2300                 /*
2301                  * We have done an ON -> OFF transition:
2302                  */
2303                 curr->hardirqs_enabled = 0;
2304                 curr->hardirq_disable_ip = ip;
2305                 curr->hardirq_disable_event = ++curr->irq_events;
2306                 debug_atomic_inc(&hardirqs_off_events);
2307         } else
2308                 debug_atomic_inc(&redundant_hardirqs_off);
2309 }
2310 EXPORT_SYMBOL(trace_hardirqs_off_caller);
2311
2312 void trace_hardirqs_off(void)
2313 {
2314         trace_hardirqs_off_caller(CALLER_ADDR0);
2315 }
2316 EXPORT_SYMBOL(trace_hardirqs_off);
2317
2318 /*
2319  * Softirqs will be enabled:
2320  */
2321 void trace_softirqs_on(unsigned long ip)
2322 {
2323         struct task_struct *curr = current;
2324
2325         if (unlikely(!debug_locks))
2326                 return;
2327
2328         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2329                 return;
2330
2331         if (curr->softirqs_enabled) {
2332                 debug_atomic_inc(&redundant_softirqs_on);
2333                 return;
2334         }
2335
2336         /*
2337          * We'll do an OFF -> ON transition:
2338          */
2339         curr->softirqs_enabled = 1;
2340         curr->softirq_enable_ip = ip;
2341         curr->softirq_enable_event = ++curr->irq_events;
2342         debug_atomic_inc(&softirqs_on_events);
2343         /*
2344          * We are going to turn softirqs on, so set the
2345          * usage bit for all held locks, if hardirqs are
2346          * enabled too:
2347          */
2348         if (curr->hardirqs_enabled)
2349                 mark_held_locks(curr, SOFTIRQ);
2350 }
2351
2352 /*
2353  * Softirqs were disabled:
2354  */
2355 void trace_softirqs_off(unsigned long ip)
2356 {
2357         struct task_struct *curr = current;
2358
2359         if (unlikely(!debug_locks))
2360                 return;
2361
2362         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2363                 return;
2364
2365         if (curr->softirqs_enabled) {
2366                 /*
2367                  * We have done an ON -> OFF transition:
2368                  */
2369                 curr->softirqs_enabled = 0;
2370                 curr->softirq_disable_ip = ip;
2371                 curr->softirq_disable_event = ++curr->irq_events;
2372                 debug_atomic_inc(&softirqs_off_events);
2373                 DEBUG_LOCKS_WARN_ON(!softirq_count());
2374         } else
2375                 debug_atomic_inc(&redundant_softirqs_off);
2376 }
2377
2378 static void __lockdep_trace_alloc(gfp_t gfp_mask, unsigned long flags)
2379 {
2380         struct task_struct *curr = current;
2381
2382         if (unlikely(!debug_locks))
2383                 return;
2384
2385         /* no reclaim without waiting on it */
2386         if (!(gfp_mask & __GFP_WAIT))
2387                 return;
2388
2389         /* this guy won't enter reclaim */
2390         if ((curr->flags & PF_MEMALLOC) && !(gfp_mask & __GFP_NOMEMALLOC))
2391                 return;
2392
2393         /* We're only interested __GFP_FS allocations for now */
2394         if (!(gfp_mask & __GFP_FS))
2395                 return;
2396
2397         if (DEBUG_LOCKS_WARN_ON(irqs_disabled_flags(flags)))
2398                 return;
2399
2400         mark_held_locks(curr, RECLAIM_FS);
2401 }
2402
2403 static void check_flags(unsigned long flags);
2404
2405 void lockdep_trace_alloc(gfp_t gfp_mask)
2406 {
2407         unsigned long flags;
2408
2409         if (unlikely(current->lockdep_recursion))
2410                 return;
2411
2412         raw_local_irq_save(flags);
2413         check_flags(flags);
2414         current->lockdep_recursion = 1;
2415         __lockdep_trace_alloc(gfp_mask, flags);
2416         current->lockdep_recursion = 0;
2417         raw_local_irq_restore(flags);
2418 }
2419
2420 static int mark_irqflags(struct task_struct *curr, struct held_lock *hlock)
2421 {
2422         /*
2423          * If non-trylock use in a hardirq or softirq context, then
2424          * mark the lock as used in these contexts:
2425          */
2426         if (!hlock->trylock) {
2427                 if (hlock->read) {
2428                         if (curr->hardirq_context)
2429                                 if (!mark_lock(curr, hlock,
2430                                                 LOCK_USED_IN_HARDIRQ_READ))
2431                                         return 0;
2432                         if (curr->softirq_context)
2433                                 if (!mark_lock(curr, hlock,
2434                                                 LOCK_USED_IN_SOFTIRQ_READ))
2435                                         return 0;
2436                 } else {
2437                         if (curr->hardirq_context)
2438                                 if (!mark_lock(curr, hlock, LOCK_USED_IN_HARDIRQ))
2439                                         return 0;
2440                         if (curr->softirq_context)
2441                                 if (!mark_lock(curr, hlock, LOCK_USED_IN_SOFTIRQ))
2442                                         return 0;
2443                 }
2444         }
2445         if (!hlock->hardirqs_off) {
2446                 if (hlock->read) {
2447                         if (!mark_lock(curr, hlock,
2448                                         LOCK_ENABLED_HARDIRQ_READ))
2449                                 return 0;
2450                         if (curr->softirqs_enabled)
2451                                 if (!mark_lock(curr, hlock,
2452                                                 LOCK_ENABLED_SOFTIRQ_READ))
2453                                         return 0;
2454                 } else {
2455                         if (!mark_lock(curr, hlock,
2456                                         LOCK_ENABLED_HARDIRQ))
2457                                 return 0;
2458                         if (curr->softirqs_enabled)
2459                                 if (!mark_lock(curr, hlock,
2460                                                 LOCK_ENABLED_SOFTIRQ))
2461                                         return 0;
2462                 }
2463         }
2464
2465         /*
2466          * We reuse the irq context infrastructure more broadly as a general
2467          * context checking code. This tests GFP_FS recursion (a lock taken
2468          * during reclaim for a GFP_FS allocation is held over a GFP_FS
2469          * allocation).
2470          */
2471         if (!hlock->trylock && (curr->lockdep_reclaim_gfp & __GFP_FS)) {
2472                 if (hlock->read) {
2473                         if (!mark_lock(curr, hlock, LOCK_USED_IN_RECLAIM_FS_READ))
2474                                         return 0;
2475                 } else {
2476                         if (!mark_lock(curr, hlock, LOCK_USED_IN_RECLAIM_FS))
2477                                         return 0;
2478                 }
2479         }
2480
2481         return 1;
2482 }
2483
2484 static int separate_irq_context(struct task_struct *curr,
2485                 struct held_lock *hlock)
2486 {
2487         unsigned int depth = curr->lockdep_depth;
2488
2489         /*
2490          * Keep track of points where we cross into an interrupt context:
2491          */
2492         hlock->irq_context = 2*(curr->hardirq_context ? 1 : 0) +
2493                                 curr->softirq_context;
2494         if (depth) {
2495                 struct held_lock *prev_hlock;
2496
2497                 prev_hlock = curr->held_locks + depth-1;
2498                 /*
2499                  * If we cross into another context, reset the
2500                  * hash key (this also prevents the checking and the
2501                  * adding of the dependency to 'prev'):
2502                  */
2503                 if (prev_hlock->irq_context != hlock->irq_context)
2504                         return 1;
2505         }
2506         return 0;
2507 }
2508
2509 #else
2510
2511 static inline
2512 int mark_lock_irq(struct task_struct *curr, struct held_lock *this,
2513                 enum lock_usage_bit new_bit)
2514 {
2515         WARN_ON(1);
2516         return 1;
2517 }
2518
2519 static inline int mark_irqflags(struct task_struct *curr,
2520                 struct held_lock *hlock)
2521 {
2522         return 1;
2523 }
2524
2525 static inline int separate_irq_context(struct task_struct *curr,
2526                 struct held_lock *hlock)
2527 {
2528         return 0;
2529 }
2530
2531 void lockdep_trace_alloc(gfp_t gfp_mask)
2532 {
2533 }
2534
2535 #endif
2536
2537 /*
2538  * Mark a lock with a usage bit, and validate the state transition:
2539  */
2540 static int mark_lock(struct task_struct *curr, struct held_lock *this,
2541                              enum lock_usage_bit new_bit)
2542 {
2543         unsigned int new_mask = 1 << new_bit, ret = 1;
2544
2545         /*
2546          * If already set then do not dirty the cacheline,
2547          * nor do any checks:
2548          */
2549         if (likely(hlock_class(this)->usage_mask & new_mask))
2550                 return 1;
2551
2552         if (!graph_lock())
2553                 return 0;
2554         /*
2555          * Make sure we didnt race:
2556          */
2557         if (unlikely(hlock_class(this)->usage_mask & new_mask)) {
2558                 graph_unlock();
2559                 return 1;
2560         }
2561
2562         hlock_class(this)->usage_mask |= new_mask;
2563
2564         if (!save_trace(hlock_class(this)->usage_traces + new_bit))
2565                 return 0;
2566
2567         switch (new_bit) {
2568 #define LOCKDEP_STATE(__STATE)                  \
2569         case LOCK_USED_IN_##__STATE:            \
2570         case LOCK_USED_IN_##__STATE##_READ:     \
2571         case LOCK_ENABLED_##__STATE:            \
2572         case LOCK_ENABLED_##__STATE##_READ:
2573 #include "lockdep_states.h"
2574 #undef LOCKDEP_STATE
2575                 ret = mark_lock_irq(curr, this, new_bit);
2576                 if (!ret)
2577                         return 0;
2578                 break;
2579         case LOCK_USED:
2580                 debug_atomic_dec(&nr_unused_locks);
2581                 break;
2582         default:
2583                 if (!debug_locks_off_graph_unlock())
2584                         return 0;
2585                 WARN_ON(1);
2586                 return 0;
2587         }
2588
2589         graph_unlock();
2590
2591         /*
2592          * We must printk outside of the graph_lock:
2593          */
2594         if (ret == 2) {
2595                 printk("\nmarked lock as {%s}:\n", usage_str[new_bit]);
2596                 print_lock(this);
2597                 print_irqtrace_events(curr);
2598                 dump_stack();
2599         }
2600
2601         return ret;
2602 }
2603
2604 /*
2605  * Initialize a lock instance's lock-class mapping info:
2606  */
2607 void lockdep_init_map(struct lockdep_map *lock, const char *name,
2608                       struct lock_class_key *key, int subclass)
2609 {
2610         lock->class_cache = NULL;
2611 #ifdef CONFIG_LOCK_STAT
2612         lock->cpu = raw_smp_processor_id();
2613 #endif
2614
2615         if (DEBUG_LOCKS_WARN_ON(!name)) {
2616                 lock->name = "NULL";
2617                 return;
2618         }
2619
2620         lock->name = name;
2621
2622         if (DEBUG_LOCKS_WARN_ON(!key))
2623                 return;
2624         /*
2625          * Sanity check, the lock-class key must be persistent:
2626          */
2627         if (!static_obj(key)) {
2628                 printk("BUG: key %p not in .data!\n", key);
2629                 DEBUG_LOCKS_WARN_ON(1);
2630                 return;
2631         }
2632         lock->key = key;
2633
2634         if (unlikely(!debug_locks))
2635                 return;
2636
2637         if (subclass)
2638                 register_lock_class(lock, subclass, 1);
2639 }
2640 EXPORT_SYMBOL_GPL(lockdep_init_map);
2641
2642 /*
2643  * This gets called for every mutex_lock*()/spin_lock*() operation.
2644  * We maintain the dependency maps and validate the locking attempt:
2645  */
2646 static int __lock_acquire(struct lockdep_map *lock, unsigned int subclass,
2647                           int trylock, int read, int check, int hardirqs_off,
2648                           struct lockdep_map *nest_lock, unsigned long ip)
2649 {
2650         struct task_struct *curr = current;
2651         struct lock_class *class = NULL;
2652         struct held_lock *hlock;
2653         unsigned int depth, id;
2654         int chain_head = 0;
2655         u64 chain_key;
2656
2657         if (!prove_locking)
2658                 check = 1;
2659
2660         if (unlikely(!debug_locks))
2661                 return 0;
2662
2663         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2664                 return 0;
2665
2666         if (unlikely(subclass >= MAX_LOCKDEP_SUBCLASSES)) {
2667                 debug_locks_off();
2668                 printk("BUG: MAX_LOCKDEP_SUBCLASSES too low!\n");
2669                 printk("turning off the locking correctness validator.\n");
2670                 dump_stack();
2671                 return 0;
2672         }
2673
2674         if (!subclass)
2675                 class = lock->class_cache;
2676         /*
2677          * Not cached yet or subclass?
2678          */
2679         if (unlikely(!class)) {
2680                 class = register_lock_class(lock, subclass, 0);
2681                 if (!class)
2682                         return 0;
2683         }
2684         debug_atomic_inc((atomic_t *)&class->ops);
2685         if (very_verbose(class)) {
2686                 printk("\nacquire class [%p] %s", class->key, class->name);
2687                 if (class->name_version > 1)
2688                         printk("#%d", class->name_version);
2689                 printk("\n");
2690                 dump_stack();
2691         }
2692
2693         /*
2694          * Add the lock to the list of currently held locks.
2695          * (we dont increase the depth just yet, up until the
2696          * dependency checks are done)
2697          */
2698         depth = curr->lockdep_depth;
2699         if (DEBUG_LOCKS_WARN_ON(depth >= MAX_LOCK_DEPTH))
2700                 return 0;
2701
2702         hlock = curr->held_locks + depth;
2703         if (DEBUG_LOCKS_WARN_ON(!class))
2704                 return 0;
2705         hlock->class_idx = class - lock_classes + 1;
2706         hlock->acquire_ip = ip;
2707         hlock->instance = lock;
2708         hlock->nest_lock = nest_lock;
2709         hlock->trylock = trylock;
2710         hlock->read = read;
2711         hlock->check = check;
2712         hlock->hardirqs_off = !!hardirqs_off;
2713 #ifdef CONFIG_LOCK_STAT
2714         hlock->waittime_stamp = 0;
2715         hlock->holdtime_stamp = sched_clock();
2716 #endif
2717
2718         if (check == 2 && !mark_irqflags(curr, hlock))
2719                 return 0;
2720
2721         /* mark it as used: */
2722         if (!mark_lock(curr, hlock, LOCK_USED))
2723                 return 0;
2724
2725         /*
2726          * Calculate the chain hash: it's the combined hash of all the
2727          * lock keys along the dependency chain. We save the hash value
2728          * at every step so that we can get the current hash easily
2729          * after unlock. The chain hash is then used to cache dependency
2730          * results.
2731          *
2732          * The 'key ID' is what is the most compact key value to drive
2733          * the hash, not class->key.
2734          */
2735         id = class - lock_classes;
2736         if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
2737                 return 0;
2738
2739         chain_key = curr->curr_chain_key;
2740         if (!depth) {
2741                 if (DEBUG_LOCKS_WARN_ON(chain_key != 0))
2742                         return 0;
2743                 chain_head = 1;
2744         }
2745
2746         hlock->prev_chain_key = chain_key;
2747         if (separate_irq_context(curr, hlock)) {
2748                 chain_key = 0;
2749                 chain_head = 1;
2750         }
2751         chain_key = iterate_chain_key(chain_key, id);
2752
2753         if (!validate_chain(curr, lock, hlock, chain_head, chain_key))
2754                 return 0;
2755
2756         curr->curr_chain_key = chain_key;
2757         curr->lockdep_depth++;
2758         check_chain_key(curr);
2759 #ifdef CONFIG_DEBUG_LOCKDEP
2760         if (unlikely(!debug_locks))
2761                 return 0;
2762 #endif
2763         if (unlikely(curr->lockdep_depth >= MAX_LOCK_DEPTH)) {
2764                 debug_locks_off();
2765                 printk("BUG: MAX_LOCK_DEPTH too low!\n");
2766                 printk("turning off the locking correctness validator.\n");
2767                 dump_stack();
2768                 return 0;
2769         }
2770
2771         if (unlikely(curr->lockdep_depth > max_lockdep_depth))
2772                 max_lockdep_depth = curr->lockdep_depth;
2773
2774         return 1;
2775 }
2776
2777 static int
2778 print_unlock_inbalance_bug(struct task_struct *curr, struct lockdep_map *lock,
2779                            unsigned long ip)
2780 {
2781         if (!debug_locks_off())
2782                 return 0;
2783         if (debug_locks_silent)
2784                 return 0;
2785
2786         printk("\n=====================================\n");
2787         printk(  "[ BUG: bad unlock balance detected! ]\n");
2788         printk(  "-------------------------------------\n");
2789         printk("%s/%d is trying to release lock (",
2790                 curr->comm, task_pid_nr(curr));
2791         print_lockdep_cache(lock);
2792         printk(") at:\n");
2793         print_ip_sym(ip);
2794         printk("but there are no more locks to release!\n");
2795         printk("\nother info that might help us debug this:\n");
2796         lockdep_print_held_locks(curr);
2797
2798         printk("\nstack backtrace:\n");
2799         dump_stack();
2800
2801         return 0;
2802 }
2803
2804 /*
2805  * Common debugging checks for both nested and non-nested unlock:
2806  */
2807 static int check_unlock(struct task_struct *curr, struct lockdep_map *lock,
2808                         unsigned long ip)
2809 {
2810         if (unlikely(!debug_locks))
2811                 return 0;
2812         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2813                 return 0;
2814
2815         if (curr->lockdep_depth <= 0)
2816                 return print_unlock_inbalance_bug(curr, lock, ip);
2817
2818         return 1;
2819 }
2820
2821 static int
2822 __lock_set_class(struct lockdep_map *lock, const char *name,
2823                  struct lock_class_key *key, unsigned int subclass,
2824                  unsigned long ip)
2825 {
2826         struct task_struct *curr = current;
2827         struct held_lock *hlock, *prev_hlock;
2828         struct lock_class *class;
2829         unsigned int depth;
2830         int i;
2831
2832         depth = curr->lockdep_depth;
2833         if (DEBUG_LOCKS_WARN_ON(!depth))
2834                 return 0;
2835
2836         prev_hlock = NULL;
2837         for (i = depth-1; i >= 0; i--) {
2838                 hlock = curr->held_locks + i;
2839                 /*
2840                  * We must not cross into another context:
2841                  */
2842                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
2843                         break;
2844                 if (hlock->instance == lock)
2845                         goto found_it;
2846                 prev_hlock = hlock;
2847         }
2848         return print_unlock_inbalance_bug(curr, lock, ip);
2849
2850 found_it:
2851         lockdep_init_map(lock, name, key, 0);
2852         class = register_lock_class(lock, subclass, 0);
2853         hlock->class_idx = class - lock_classes + 1;
2854
2855         curr->lockdep_depth = i;
2856         curr->curr_chain_key = hlock->prev_chain_key;
2857
2858         for (; i < depth; i++) {
2859                 hlock = curr->held_locks + i;
2860                 if (!__lock_acquire(hlock->instance,
2861                         hlock_class(hlock)->subclass, hlock->trylock,
2862                                 hlock->read, hlock->check, hlock->hardirqs_off,
2863                                 hlock->nest_lock, hlock->acquire_ip))
2864                         return 0;
2865         }
2866
2867         if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth))
2868                 return 0;
2869         return 1;
2870 }
2871
2872 /*
2873  * Remove the lock to the list of currently held locks in a
2874  * potentially non-nested (out of order) manner. This is a
2875  * relatively rare operation, as all the unlock APIs default
2876  * to nested mode (which uses lock_release()):
2877  */
2878 static int
2879 lock_release_non_nested(struct task_struct *curr,
2880                         struct lockdep_map *lock, unsigned long ip)
2881 {
2882         struct held_lock *hlock, *prev_hlock;
2883         unsigned int depth;
2884         int i;
2885
2886         /*
2887          * Check whether the lock exists in the current stack
2888          * of held locks:
2889          */
2890         depth = curr->lockdep_depth;
2891         if (DEBUG_LOCKS_WARN_ON(!depth))
2892                 return 0;
2893
2894         prev_hlock = NULL;
2895         for (i = depth-1; i >= 0; i--) {
2896                 hlock = curr->held_locks + i;
2897                 /*
2898                  * We must not cross into another context:
2899                  */
2900                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
2901                         break;
2902                 if (hlock->instance == lock)
2903                         goto found_it;
2904                 prev_hlock = hlock;
2905         }
2906         return print_unlock_inbalance_bug(curr, lock, ip);
2907
2908 found_it:
2909         lock_release_holdtime(hlock);
2910
2911         /*
2912          * We have the right lock to unlock, 'hlock' points to it.
2913          * Now we remove it from the stack, and add back the other
2914          * entries (if any), recalculating the hash along the way:
2915          */
2916         curr->lockdep_depth = i;
2917         curr->curr_chain_key = hlock->prev_chain_key;
2918
2919         for (i++; i < depth; i++) {
2920                 hlock = curr->held_locks + i;
2921                 if (!__lock_acquire(hlock->instance,
2922                         hlock_class(hlock)->subclass, hlock->trylock,
2923                                 hlock->read, hlock->check, hlock->hardirqs_off,
2924                                 hlock->nest_lock, hlock->acquire_ip))
2925                         return 0;
2926         }
2927
2928         if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth - 1))
2929                 return 0;
2930         return 1;
2931 }
2932
2933 /*
2934  * Remove the lock to the list of currently held locks - this gets
2935  * called on mutex_unlock()/spin_unlock*() (or on a failed
2936  * mutex_lock_interruptible()). This is done for unlocks that nest
2937  * perfectly. (i.e. the current top of the lock-stack is unlocked)
2938  */
2939 static int lock_release_nested(struct task_struct *curr,
2940                                struct lockdep_map *lock, unsigned long ip)
2941 {
2942         struct held_lock *hlock;
2943         unsigned int depth;
2944
2945         /*
2946          * Pop off the top of the lock stack:
2947          */
2948         depth = curr->lockdep_depth - 1;
2949         hlock = curr->held_locks + depth;
2950
2951         /*
2952          * Is the unlock non-nested:
2953          */
2954         if (hlock->instance != lock)
2955                 return lock_release_non_nested(curr, lock, ip);
2956         curr->lockdep_depth--;
2957
2958         if (DEBUG_LOCKS_WARN_ON(!depth && (hlock->prev_chain_key != 0)))
2959                 return 0;
2960
2961         curr->curr_chain_key = hlock->prev_chain_key;
2962
2963         lock_release_holdtime(hlock);
2964
2965 #ifdef CONFIG_DEBUG_LOCKDEP
2966         hlock->prev_chain_key = 0;
2967         hlock->class_idx = 0;
2968         hlock->acquire_ip = 0;
2969         hlock->irq_context = 0;
2970 #endif
2971         return 1;
2972 }
2973
2974 /*
2975  * Remove the lock to the list of currently held locks - this gets
2976  * called on mutex_unlock()/spin_unlock*() (or on a failed
2977  * mutex_lock_interruptible()). This is done for unlocks that nest
2978  * perfectly. (i.e. the current top of the lock-stack is unlocked)
2979  */
2980 static void
2981 __lock_release(struct lockdep_map *lock, int nested, unsigned long ip)
2982 {
2983         struct task_struct *curr = current;
2984
2985         if (!check_unlock(curr, lock, ip))
2986                 return;
2987
2988         if (nested) {
2989                 if (!lock_release_nested(curr, lock, ip))
2990                         return;
2991         } else {
2992                 if (!lock_release_non_nested(curr, lock, ip))
2993                         return;
2994         }
2995
2996         check_chain_key(curr);
2997 }
2998
2999 /*
3000  * Check whether we follow the irq-flags state precisely:
3001  */
3002 static void check_flags(unsigned long flags)
3003 {
3004 #if defined(CONFIG_PROVE_LOCKING) && defined(CONFIG_DEBUG_LOCKDEP) && \
3005     defined(CONFIG_TRACE_IRQFLAGS)
3006         if (!debug_locks)
3007                 return;
3008
3009         if (irqs_disabled_flags(flags)) {
3010                 if (DEBUG_LOCKS_WARN_ON(current->hardirqs_enabled)) {
3011                         printk("possible reason: unannotated irqs-off.\n");
3012                 }
3013         } else {
3014                 if (DEBUG_LOCKS_WARN_ON(!current->hardirqs_enabled)) {
3015                         printk("possible reason: unannotated irqs-on.\n");
3016                 }
3017         }
3018
3019         /*
3020          * We dont accurately track softirq state in e.g.
3021          * hardirq contexts (such as on 4KSTACKS), so only
3022          * check if not in hardirq contexts:
3023          */
3024         if (!hardirq_count()) {
3025                 if (softirq_count())
3026                         DEBUG_LOCKS_WARN_ON(current->softirqs_enabled);
3027                 else
3028                         DEBUG_LOCKS_WARN_ON(!current->softirqs_enabled);
3029         }
3030
3031         if (!debug_locks)
3032                 print_irqtrace_events(current);
3033 #endif
3034 }
3035
3036 void lock_set_class(struct lockdep_map *lock, const char *name,
3037                     struct lock_class_key *key, unsigned int subclass,
3038                     unsigned long ip)
3039 {
3040         unsigned long flags;
3041
3042         if (unlikely(current->lockdep_recursion))
3043                 return;
3044
3045         raw_local_irq_save(flags);
3046         current->lockdep_recursion = 1;
3047         check_flags(flags);
3048         if (__lock_set_class(lock, name, key, subclass, ip))
3049                 check_chain_key(current);
3050         current->lockdep_recursion = 0;
3051         raw_local_irq_restore(flags);
3052 }
3053 EXPORT_SYMBOL_GPL(lock_set_class);
3054
3055 /*
3056  * We are not always called with irqs disabled - do that here,
3057  * and also avoid lockdep recursion:
3058  */
3059 void lock_acquire(struct lockdep_map *lock, unsigned int subclass,
3060                           int trylock, int read, int check,
3061                           struct lockdep_map *nest_lock, unsigned long ip)
3062 {
3063         unsigned long flags;
3064
3065         trace_lock_acquire(lock, subclass, trylock, read, check, nest_lock, ip);
3066
3067         if (unlikely(current->lockdep_recursion))
3068                 return;
3069
3070         raw_local_irq_save(flags);
3071         check_flags(flags);
3072
3073         current->lockdep_recursion = 1;
3074         __lock_acquire(lock, subclass, trylock, read, check,
3075                        irqs_disabled_flags(flags), nest_lock, ip);
3076         current->lockdep_recursion = 0;
3077         raw_local_irq_restore(flags);
3078 }
3079 EXPORT_SYMBOL_GPL(lock_acquire);
3080
3081 void lock_release(struct lockdep_map *lock, int nested,
3082                           unsigned long ip)
3083 {
3084         unsigned long flags;
3085
3086         trace_lock_release(lock, nested, ip);
3087
3088         if (unlikely(current->lockdep_recursion))
3089                 return;
3090
3091         raw_local_irq_save(flags);
3092         check_flags(flags);
3093         current->lockdep_recursion = 1;
3094         __lock_release(lock, nested, ip);
3095         current->lockdep_recursion = 0;
3096         raw_local_irq_restore(flags);
3097 }
3098 EXPORT_SYMBOL_GPL(lock_release);
3099
3100 void lockdep_set_current_reclaim_state(gfp_t gfp_mask)
3101 {
3102         current->lockdep_reclaim_gfp = gfp_mask;
3103 }
3104
3105 void lockdep_clear_current_reclaim_state(void)
3106 {
3107         current->lockdep_reclaim_gfp = 0;
3108 }
3109
3110 #ifdef CONFIG_LOCK_STAT
3111 static int
3112 print_lock_contention_bug(struct task_struct *curr, struct lockdep_map *lock,
3113                            unsigned long ip)
3114 {
3115         if (!debug_locks_off())
3116                 return 0;
3117         if (debug_locks_silent)
3118                 return 0;
3119
3120         printk("\n=================================\n");
3121         printk(  "[ BUG: bad contention detected! ]\n");
3122         printk(  "---------------------------------\n");
3123         printk("%s/%d is trying to contend lock (",
3124                 curr->comm, task_pid_nr(curr));
3125         print_lockdep_cache(lock);
3126         printk(") at:\n");
3127         print_ip_sym(ip);
3128         printk("but there are no locks held!\n");
3129         printk("\nother info that might help us debug this:\n");
3130         lockdep_print_held_locks(curr);
3131
3132         printk("\nstack backtrace:\n");
3133         dump_stack();
3134
3135         return 0;
3136 }
3137
3138 static void
3139 __lock_contended(struct lockdep_map *lock, unsigned long ip)
3140 {
3141         struct task_struct *curr = current;
3142         struct held_lock *hlock, *prev_hlock;
3143         struct lock_class_stats *stats;
3144         unsigned int depth;
3145         int i, contention_point, contending_point;
3146
3147         depth = curr->lockdep_depth;
3148         if (DEBUG_LOCKS_WARN_ON(!depth))
3149                 return;
3150
3151         prev_hlock = NULL;
3152         for (i = depth-1; i >= 0; i--) {
3153                 hlock = curr->held_locks + i;
3154                 /*
3155                  * We must not cross into another context:
3156                  */
3157                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3158                         break;
3159                 if (hlock->instance == lock)
3160                         goto found_it;
3161                 prev_hlock = hlock;
3162         }
3163         print_lock_contention_bug(curr, lock, ip);
3164         return;
3165
3166 found_it:
3167         hlock->waittime_stamp = sched_clock();
3168
3169         contention_point = lock_point(hlock_class(hlock)->contention_point, ip);
3170         contending_point = lock_point(hlock_class(hlock)->contending_point,
3171                                       lock->ip);
3172
3173         stats = get_lock_stats(hlock_class(hlock));
3174         if (contention_point < LOCKSTAT_POINTS)
3175                 stats->contention_point[contention_point]++;
3176         if (contending_point < LOCKSTAT_POINTS)
3177                 stats->contending_point[contending_point]++;
3178         if (lock->cpu != smp_processor_id())
3179                 stats->bounces[bounce_contended + !!hlock->read]++;
3180         put_lock_stats(stats);
3181 }
3182
3183 static void
3184 __lock_acquired(struct lockdep_map *lock, unsigned long ip)
3185 {
3186         struct task_struct *curr = current;
3187         struct held_lock *hlock, *prev_hlock;
3188         struct lock_class_stats *stats;
3189         unsigned int depth;
3190         u64 now;
3191         s64 waittime = 0;
3192         int i, cpu;
3193
3194         depth = curr->lockdep_depth;
3195         if (DEBUG_LOCKS_WARN_ON(!depth))
3196                 return;
3197
3198         prev_hlock = NULL;
3199         for (i = depth-1; i >= 0; i--) {
3200                 hlock = curr->held_locks + i;
3201                 /*
3202                  * We must not cross into another context:
3203                  */
3204                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3205                         break;
3206                 if (hlock->instance == lock)
3207                         goto found_it;
3208                 prev_hlock = hlock;
3209         }
3210         print_lock_contention_bug(curr, lock, _RET_IP_);
3211         return;
3212
3213 found_it:
3214         cpu = smp_processor_id();
3215         if (hlock->waittime_stamp) {
3216                 now = sched_clock();
3217                 waittime = now - hlock->waittime_stamp;
3218                 hlock->holdtime_stamp = now;
3219         }
3220
3221         trace_lock_acquired(lock, ip, waittime);
3222
3223         stats = get_lock_stats(hlock_class(hlock));
3224         if (waittime) {
3225                 if (hlock->read)
3226                         lock_time_inc(&stats->read_waittime, waittime);
3227                 else
3228                         lock_time_inc(&stats->write_waittime, waittime);
3229         }
3230         if (lock->cpu != cpu)
3231                 stats->bounces[bounce_acquired + !!hlock->read]++;
3232         put_lock_stats(stats);
3233
3234         lock->cpu = cpu;
3235         lock->ip = ip;
3236 }
3237
3238 void lock_contended(struct lockdep_map *lock, unsigned long ip)
3239 {
3240         unsigned long flags;
3241
3242         trace_lock_contended(lock, ip);
3243
3244         if (unlikely(!lock_stat))
3245                 return;
3246
3247         if (unlikely(current->lockdep_recursion))
3248                 return;
3249
3250         raw_local_irq_save(flags);
3251         check_flags(flags);
3252         current->lockdep_recursion = 1;
3253         __lock_contended(lock, ip);
3254         current->lockdep_recursion = 0;
3255         raw_local_irq_restore(flags);
3256 }
3257 EXPORT_SYMBOL_GPL(lock_contended);
3258
3259 void lock_acquired(struct lockdep_map *lock, unsigned long ip)
3260 {
3261         unsigned long flags;
3262
3263         if (unlikely(!lock_stat))
3264                 return;
3265
3266         if (unlikely(current->lockdep_recursion))
3267                 return;
3268
3269         raw_local_irq_save(flags);
3270         check_flags(flags);
3271         current->lockdep_recursion = 1;
3272         __lock_acquired(lock, ip);
3273         current->lockdep_recursion = 0;
3274         raw_local_irq_restore(flags);
3275 }
3276 EXPORT_SYMBOL_GPL(lock_acquired);
3277 #endif
3278
3279 /*
3280  * Used by the testsuite, sanitize the validator state
3281  * after a simulated failure:
3282  */
3283
3284 void lockdep_reset(void)
3285 {
3286         unsigned long flags;
3287         int i;
3288
3289         raw_local_irq_save(flags);
3290         current->curr_chain_key = 0;
3291         current->lockdep_depth = 0;
3292         current->lockdep_recursion = 0;
3293         memset(current->held_locks, 0, MAX_LOCK_DEPTH*sizeof(struct held_lock));
3294         nr_hardirq_chains = 0;
3295         nr_softirq_chains = 0;
3296         nr_process_chains = 0;
3297         debug_locks = 1;
3298         for (i = 0; i < CHAINHASH_SIZE; i++)
3299                 INIT_LIST_HEAD(chainhash_table + i);
3300         raw_local_irq_restore(flags);
3301 }
3302
3303 static void zap_class(struct lock_class *class)
3304 {
3305         int i;
3306
3307         /*
3308          * Remove all dependencies this lock is
3309          * involved in:
3310          */
3311         for (i = 0; i < nr_list_entries; i++) {
3312                 if (list_entries[i].class == class)
3313                         list_del_rcu(&list_entries[i].entry);
3314         }
3315         /*
3316          * Unhash the class and remove it from the all_lock_classes list:
3317          */
3318         list_del_rcu(&class->hash_entry);
3319         list_del_rcu(&class->lock_entry);
3320
3321         class->key = NULL;
3322 }
3323
3324 static inline int within(const void *addr, void *start, unsigned long size)
3325 {
3326         return addr >= start && addr < start + size;
3327 }
3328
3329 void lockdep_free_key_range(void *start, unsigned long size)
3330 {
3331         struct lock_class *class, *next;
3332         struct list_head *head;
3333         unsigned long flags;
3334         int i;
3335         int locked;
3336
3337         raw_local_irq_save(flags);
3338         locked = graph_lock();
3339
3340         /*
3341          * Unhash all classes that were created by this module:
3342          */
3343         for (i = 0; i < CLASSHASH_SIZE; i++) {
3344                 head = classhash_table + i;
3345                 if (list_empty(head))
3346                         continue;
3347                 list_for_each_entry_safe(class, next, head, hash_entry) {
3348                         if (within(class->key, start, size))
3349                                 zap_class(class);
3350                         else if (within(class->name, start, size))
3351                                 zap_class(class);
3352                 }
3353         }
3354
3355         if (locked)
3356                 graph_unlock();
3357         raw_local_irq_restore(flags);
3358 }
3359
3360 void lockdep_reset_lock(struct lockdep_map *lock)
3361 {
3362         struct lock_class *class, *next;
3363         struct list_head *head;
3364         unsigned long flags;
3365         int i, j;
3366         int locked;
3367
3368         raw_local_irq_save(flags);
3369
3370         /*
3371          * Remove all classes this lock might have:
3372          */
3373         for (j = 0; j < MAX_LOCKDEP_SUBCLASSES; j++) {
3374                 /*
3375                  * If the class exists we look it up and zap it:
3376                  */
3377                 class = look_up_lock_class(lock, j);
3378                 if (class)
3379                         zap_class(class);
3380         }
3381         /*
3382          * Debug check: in the end all mapped classes should
3383          * be gone.
3384          */
3385         locked = graph_lock();
3386         for (i = 0; i < CLASSHASH_SIZE; i++) {
3387                 head = classhash_table + i;
3388                 if (list_empty(head))
3389                         continue;
3390                 list_for_each_entry_safe(class, next, head, hash_entry) {
3391                         if (unlikely(class == lock->class_cache)) {
3392                                 if (debug_locks_off_graph_unlock())
3393                                         WARN_ON(1);
3394                                 goto out_restore;
3395                         }
3396                 }
3397         }
3398         if (locked)
3399                 graph_unlock();
3400
3401 out_restore:
3402         raw_local_irq_restore(flags);
3403 }
3404
3405 void lockdep_init(void)
3406 {
3407         int i;
3408
3409         /*
3410          * Some architectures have their own start_kernel()
3411          * code which calls lockdep_init(), while we also
3412          * call lockdep_init() from the start_kernel() itself,
3413          * and we want to initialize the hashes only once:
3414          */
3415         if (lockdep_initialized)
3416                 return;
3417
3418         for (i = 0; i < CLASSHASH_SIZE; i++)
3419                 INIT_LIST_HEAD(classhash_table + i);
3420
3421         for (i = 0; i < CHAINHASH_SIZE; i++)
3422                 INIT_LIST_HEAD(chainhash_table + i);
3423
3424         lockdep_initialized = 1;
3425 }
3426
3427 void __init lockdep_info(void)
3428 {
3429         printk("Lock dependency validator: Copyright (c) 2006 Red Hat, Inc., Ingo Molnar\n");
3430
3431         printk("... MAX_LOCKDEP_SUBCLASSES:  %lu\n", MAX_LOCKDEP_SUBCLASSES);
3432         printk("... MAX_LOCK_DEPTH:          %lu\n", MAX_LOCK_DEPTH);
3433         printk("... MAX_LOCKDEP_KEYS:        %lu\n", MAX_LOCKDEP_KEYS);
3434         printk("... CLASSHASH_SIZE:          %lu\n", CLASSHASH_SIZE);
3435         printk("... MAX_LOCKDEP_ENTRIES:     %lu\n", MAX_LOCKDEP_ENTRIES);
3436         printk("... MAX_LOCKDEP_CHAINS:      %lu\n", MAX_LOCKDEP_CHAINS);
3437         printk("... CHAINHASH_SIZE:          %lu\n", CHAINHASH_SIZE);
3438
3439         printk(" memory used by lock dependency info: %lu kB\n",
3440                 (sizeof(struct lock_class) * MAX_LOCKDEP_KEYS +
3441                 sizeof(struct list_head) * CLASSHASH_SIZE +
3442                 sizeof(struct lock_list) * MAX_LOCKDEP_ENTRIES +
3443                 sizeof(struct lock_chain) * MAX_LOCKDEP_CHAINS +
3444                 sizeof(struct list_head) * CHAINHASH_SIZE) / 1024);
3445
3446         printk(" per task-struct memory footprint: %lu bytes\n",
3447                 sizeof(struct held_lock) * MAX_LOCK_DEPTH);
3448
3449 #ifdef CONFIG_DEBUG_LOCKDEP
3450         if (lockdep_init_error) {
3451                 printk("WARNING: lockdep init error! Arch code didn't call lockdep_init() early enough?\n");
3452                 printk("Call stack leading to lockdep invocation was:\n");
3453                 print_stack_trace(&lockdep_init_trace, 0);
3454         }
3455 #endif
3456 }
3457
3458 static void
3459 print_freed_lock_bug(struct task_struct *curr, const void *mem_from,
3460                      const void *mem_to, struct held_lock *hlock)
3461 {
3462         if (!debug_locks_off())
3463                 return;
3464         if (debug_locks_silent)
3465                 return;
3466
3467         printk("\n=========================\n");
3468         printk(  "[ BUG: held lock freed! ]\n");
3469         printk(  "-------------------------\n");
3470         printk("%s/%d is freeing memory %p-%p, with a lock still held there!\n",
3471                 curr->comm, task_pid_nr(curr), mem_from, mem_to-1);
3472         print_lock(hlock);
3473         lockdep_print_held_locks(curr);
3474
3475         printk("\nstack backtrace:\n");
3476         dump_stack();
3477 }
3478
3479 static inline int not_in_range(const void* mem_from, unsigned long mem_len,
3480                                 const void* lock_from, unsigned long lock_len)
3481 {
3482         return lock_from + lock_len <= mem_from ||
3483                 mem_from + mem_len <= lock_from;
3484 }
3485
3486 /*
3487  * Called when kernel memory is freed (or unmapped), or if a lock
3488  * is destroyed or reinitialized - this code checks whether there is
3489  * any held lock in the memory range of <from> to <to>:
3490  */
3491 void debug_check_no_locks_freed(const void *mem_from, unsigned long mem_len)
3492 {
3493         struct task_struct *curr = current;
3494         struct held_lock *hlock;
3495         unsigned long flags;
3496         int i;
3497
3498         if (unlikely(!debug_locks))
3499                 return;
3500
3501         local_irq_save(flags);
3502         for (i = 0; i < curr->lockdep_depth; i++) {
3503                 hlock = curr->held_locks + i;
3504
3505                 if (not_in_range(mem_from, mem_len, hlock->instance,
3506                                         sizeof(*hlock->instance)))
3507                         continue;
3508
3509                 print_freed_lock_bug(curr, mem_from, mem_from + mem_len, hlock);
3510                 break;
3511         }
3512         local_irq_restore(flags);
3513 }
3514 EXPORT_SYMBOL_GPL(debug_check_no_locks_freed);
3515
3516 static void print_held_locks_bug(struct task_struct *curr)
3517 {
3518         if (!debug_locks_off())
3519                 return;
3520         if (debug_locks_silent)
3521                 return;
3522
3523         printk("\n=====================================\n");
3524         printk(  "[ BUG: lock held at task exit time! ]\n");
3525         printk(  "-------------------------------------\n");
3526         printk("%s/%d is exiting with locks still held!\n",
3527                 curr->comm, task_pid_nr(curr));
3528         lockdep_print_held_locks(curr);
3529
3530         printk("\nstack backtrace:\n");
3531         dump_stack();
3532 }
3533
3534 void debug_check_no_locks_held(struct task_struct *task)
3535 {
3536         if (unlikely(task->lockdep_depth > 0))
3537                 print_held_locks_bug(task);
3538 }
3539
3540 void debug_show_all_locks(void)
3541 {
3542         struct task_struct *g, *p;
3543         int count = 10;
3544         int unlock = 1;
3545
3546         if (unlikely(!debug_locks)) {
3547                 printk("INFO: lockdep is turned off.\n");
3548                 return;
3549         }
3550         printk("\nShowing all locks held in the system:\n");
3551
3552         /*
3553          * Here we try to get the tasklist_lock as hard as possible,
3554          * if not successful after 2 seconds we ignore it (but keep
3555          * trying). This is to enable a debug printout even if a
3556          * tasklist_lock-holding task deadlocks or crashes.
3557          */
3558 retry:
3559         if (!read_trylock(&tasklist_lock)) {
3560                 if (count == 10)
3561                         printk("hm, tasklist_lock locked, retrying... ");
3562                 if (count) {
3563                         count--;
3564                         printk(" #%d", 10-count);
3565                         mdelay(200);
3566                         goto retry;
3567                 }
3568                 printk(" ignoring it.\n");
3569                 unlock = 0;
3570         } else {
3571                 if (count != 10)
3572                         printk(KERN_CONT " locked it.\n");
3573         }
3574
3575         do_each_thread(g, p) {
3576                 /*
3577                  * It's not reliable to print a task's held locks
3578                  * if it's not sleeping (or if it's not the current
3579                  * task):
3580                  */
3581                 if (p->state == TASK_RUNNING && p != current)
3582                         continue;
3583                 if (p->lockdep_depth)
3584                         lockdep_print_held_locks(p);
3585                 if (!unlock)
3586                         if (read_trylock(&tasklist_lock))
3587                                 unlock = 1;
3588         } while_each_thread(g, p);
3589
3590         printk("\n");
3591         printk("=============================================\n\n");
3592
3593         if (unlock)
3594                 read_unlock(&tasklist_lock);
3595 }
3596 EXPORT_SYMBOL_GPL(debug_show_all_locks);
3597
3598 /*
3599  * Careful: only use this function if you are sure that
3600  * the task cannot run in parallel!
3601  */
3602 void __debug_show_held_locks(struct task_struct *task)
3603 {
3604         if (unlikely(!debug_locks)) {
3605                 printk("INFO: lockdep is turned off.\n");
3606                 return;
3607         }
3608         lockdep_print_held_locks(task);
3609 }
3610 EXPORT_SYMBOL_GPL(__debug_show_held_locks);
3611
3612 void debug_show_held_locks(struct task_struct *task)
3613 {
3614                 __debug_show_held_locks(task);
3615 }
3616 EXPORT_SYMBOL_GPL(debug_show_held_locks);
3617
3618 void lockdep_sys_exit(void)
3619 {
3620         struct task_struct *curr = current;
3621
3622         if (unlikely(curr->lockdep_depth)) {
3623                 if (!debug_locks_off())
3624                         return;
3625                 printk("\n================================================\n");
3626                 printk(  "[ BUG: lock held when returning to user space! ]\n");
3627                 printk(  "------------------------------------------------\n");
3628                 printk("%s/%d is leaving the kernel with locks still held!\n",
3629                                 curr->comm, curr->pid);
3630                 lockdep_print_held_locks(curr);
3631         }
3632 }