lockdep: Implement find_usage_*wards by BFS
[linux-2.6.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_forwards(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_backwards(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
990 /*
991  * Print a dependency chain entry (this is only done when a deadlock
992  * has been detected):
993  */
994 static noinline int
995 print_circular_bug_entry(struct lock_list *target, unsigned int depth)
996 {
997         if (debug_locks_silent)
998                 return 0;
999         printk("\n-> #%u", depth);
1000         print_lock_name(target->class);
1001         printk(":\n");
1002         print_stack_trace(&target->trace, 6);
1003
1004         return 0;
1005 }
1006
1007 /*
1008  * When a circular dependency is detected, print the
1009  * header first:
1010  */
1011 static noinline int
1012 print_circular_bug_header(struct lock_list *entry, unsigned int depth,
1013                         struct held_lock *check_src,
1014                         struct held_lock *check_tgt)
1015 {
1016         struct task_struct *curr = current;
1017
1018         if (debug_locks_silent)
1019                 return 0;
1020
1021         printk("\n=======================================================\n");
1022         printk(  "[ INFO: possible circular locking dependency detected ]\n");
1023         print_kernel_version();
1024         printk(  "-------------------------------------------------------\n");
1025         printk("%s/%d is trying to acquire lock:\n",
1026                 curr->comm, task_pid_nr(curr));
1027         print_lock(check_src);
1028         printk("\nbut task is already holding lock:\n");
1029         print_lock(check_tgt);
1030         printk("\nwhich lock already depends on the new lock.\n\n");
1031         printk("\nthe existing dependency chain (in reverse order) is:\n");
1032
1033         print_circular_bug_entry(entry, depth);
1034
1035         return 0;
1036 }
1037
1038 static inline int class_equal(struct lock_list *entry, void *data)
1039 {
1040         return entry->class == data;
1041 }
1042
1043 static noinline int print_circular_bug(struct lock_list *this,
1044                                 struct lock_list *target,
1045                                 struct held_lock *check_src,
1046                                 struct held_lock *check_tgt)
1047 {
1048         struct task_struct *curr = current;
1049         struct lock_list *parent;
1050         unsigned long depth;
1051
1052         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1053                 return 0;
1054
1055         if (!save_trace(&this->trace))
1056                 return 0;
1057
1058         depth = get_lock_depth(target);
1059
1060         print_circular_bug_header(target, depth, check_src, check_tgt);
1061
1062         parent = get_lock_parent(target);
1063
1064         while (parent) {
1065                 print_circular_bug_entry(parent, --depth);
1066                 parent = get_lock_parent(parent);
1067         }
1068
1069         printk("\nother info that might help us debug this:\n\n");
1070         lockdep_print_held_locks(curr);
1071
1072         printk("\nstack backtrace:\n");
1073         dump_stack();
1074
1075         return 0;
1076 }
1077
1078 static noinline int print_bfs_bug(int ret)
1079 {
1080         if (!debug_locks_off_graph_unlock())
1081                 return 0;
1082
1083         WARN(1, "lockdep bfs error:%d\n", ret);
1084
1085         return 0;
1086 }
1087
1088 unsigned long __lockdep_count_forward_deps(struct lock_class *class,
1089                                            unsigned int depth)
1090 {
1091         struct lock_list *entry;
1092         unsigned long ret = 1;
1093
1094         if (lockdep_dependency_visit(class, depth))
1095                 return 0;
1096
1097         /*
1098          * Recurse this class's dependency list:
1099          */
1100         list_for_each_entry(entry, &class->locks_after, entry)
1101                 ret += __lockdep_count_forward_deps(entry->class, depth + 1);
1102
1103         return ret;
1104 }
1105
1106 unsigned long lockdep_count_forward_deps(struct lock_class *class)
1107 {
1108         unsigned long ret, flags;
1109
1110         local_irq_save(flags);
1111         __raw_spin_lock(&lockdep_lock);
1112         ret = __lockdep_count_forward_deps(class, 0);
1113         __raw_spin_unlock(&lockdep_lock);
1114         local_irq_restore(flags);
1115
1116         return ret;
1117 }
1118
1119 unsigned long __lockdep_count_backward_deps(struct lock_class *class,
1120                                             unsigned int depth)
1121 {
1122         struct lock_list *entry;
1123         unsigned long ret = 1;
1124
1125         if (lockdep_dependency_visit(class, depth))
1126                 return 0;
1127         /*
1128          * Recurse this class's dependency list:
1129          */
1130         list_for_each_entry(entry, &class->locks_before, entry)
1131                 ret += __lockdep_count_backward_deps(entry->class, depth + 1);
1132
1133         return ret;
1134 }
1135
1136 unsigned long lockdep_count_backward_deps(struct lock_class *class)
1137 {
1138         unsigned long ret, flags;
1139
1140         local_irq_save(flags);
1141         __raw_spin_lock(&lockdep_lock);
1142         ret = __lockdep_count_backward_deps(class, 0);
1143         __raw_spin_unlock(&lockdep_lock);
1144         local_irq_restore(flags);
1145
1146         return ret;
1147 }
1148
1149 /*
1150  * Prove that the dependency graph starting at <entry> can not
1151  * lead to <target>. Print an error and return 0 if it does.
1152  */
1153 static noinline int
1154 check_noncircular(struct lock_list *root, struct lock_class *target,
1155                 struct lock_list **target_entry)
1156 {
1157         int result;
1158
1159         debug_atomic_inc(&nr_cyclic_checks);
1160
1161         result = __bfs_forwards(root, target, class_equal, target_entry);
1162
1163         return result;
1164 }
1165
1166 #if defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING)
1167 /*
1168  * Forwards and backwards subgraph searching, for the purposes of
1169  * proving that two subgraphs can be connected by a new dependency
1170  * without creating any illegal irq-safe -> irq-unsafe lock dependency.
1171  */
1172 static struct lock_class *forwards_match, *backwards_match;
1173
1174
1175 #define   BFS_PROCESS_RET(ret)  do { \
1176                                         if (ret < 0) \
1177                                                 return print_bfs_bug(ret); \
1178                                         if (ret == 1) \
1179                                                 return 1; \
1180                                 } while (0)
1181
1182 static inline int usage_match(struct lock_list *entry, void *bit)
1183 {
1184         return entry->class->usage_mask & (1 << (enum lock_usage_bit)bit);
1185 }
1186
1187
1188
1189 /*
1190  * Find a node in the forwards-direction dependency sub-graph starting
1191  * at @root->class that matches @bit.
1192  *
1193  * Return 0 if such a node exists in the subgraph, and put that node
1194  * into *@target_entry.
1195  *
1196  * Return 1 otherwise and keep *@target_entry unchanged.
1197  * Return <0 on error.
1198  */
1199 static int
1200 find_usage_forwards(struct lock_list *root, enum lock_usage_bit bit,
1201                         struct lock_list **target_entry)
1202 {
1203         int result;
1204
1205         debug_atomic_inc(&nr_find_usage_forwards_checks);
1206
1207         result = __bfs_forwards(root, (void *)bit, usage_match, target_entry);
1208
1209         return result;
1210 }
1211
1212 /*
1213  * Find a node in the backwards-direction dependency sub-graph starting
1214  * at @root->class that matches @bit.
1215  *
1216  * Return 0 if such a node exists in the subgraph, and put that node
1217  * into *@target_entry.
1218  *
1219  * Return 1 otherwise and keep *@target_entry unchanged.
1220  * Return <0 on error.
1221  */
1222 static int
1223 find_usage_backwards(struct lock_list *root, enum lock_usage_bit bit,
1224                         struct lock_list **target_entry)
1225 {
1226         int result;
1227
1228         debug_atomic_inc(&nr_find_usage_backwards_checks);
1229
1230         result = __bfs_backwards(root, (void *)bit, usage_match, target_entry);
1231
1232         return result;
1233 }
1234
1235
1236 static int
1237 print_bad_irq_dependency(struct task_struct *curr,
1238                          struct held_lock *prev,
1239                          struct held_lock *next,
1240                          enum lock_usage_bit bit1,
1241                          enum lock_usage_bit bit2,
1242                          const char *irqclass)
1243 {
1244         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1245                 return 0;
1246
1247         printk("\n======================================================\n");
1248         printk(  "[ INFO: %s-safe -> %s-unsafe lock order detected ]\n",
1249                 irqclass, irqclass);
1250         print_kernel_version();
1251         printk(  "------------------------------------------------------\n");
1252         printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] is trying to acquire:\n",
1253                 curr->comm, task_pid_nr(curr),
1254                 curr->hardirq_context, hardirq_count() >> HARDIRQ_SHIFT,
1255                 curr->softirq_context, softirq_count() >> SOFTIRQ_SHIFT,
1256                 curr->hardirqs_enabled,
1257                 curr->softirqs_enabled);
1258         print_lock(next);
1259
1260         printk("\nand this task is already holding:\n");
1261         print_lock(prev);
1262         printk("which would create a new lock dependency:\n");
1263         print_lock_name(hlock_class(prev));
1264         printk(" ->");
1265         print_lock_name(hlock_class(next));
1266         printk("\n");
1267
1268         printk("\nbut this new dependency connects a %s-irq-safe lock:\n",
1269                 irqclass);
1270         print_lock_name(backwards_match);
1271         printk("\n... which became %s-irq-safe at:\n", irqclass);
1272
1273         print_stack_trace(backwards_match->usage_traces + bit1, 1);
1274
1275         printk("\nto a %s-irq-unsafe lock:\n", irqclass);
1276         print_lock_name(forwards_match);
1277         printk("\n... which became %s-irq-unsafe at:\n", irqclass);
1278         printk("...");
1279
1280         print_stack_trace(forwards_match->usage_traces + bit2, 1);
1281
1282         printk("\nother info that might help us debug this:\n\n");
1283         lockdep_print_held_locks(curr);
1284
1285         printk("\nthe %s-irq-safe lock's dependencies:\n", irqclass);
1286         print_lock_dependencies(backwards_match, 0);
1287
1288         printk("\nthe %s-irq-unsafe lock's dependencies:\n", irqclass);
1289         print_lock_dependencies(forwards_match, 0);
1290
1291         printk("\nstack backtrace:\n");
1292         dump_stack();
1293
1294         return 0;
1295 }
1296
1297 static int
1298 check_usage(struct task_struct *curr, struct held_lock *prev,
1299             struct held_lock *next, enum lock_usage_bit bit_backwards,
1300             enum lock_usage_bit bit_forwards, const char *irqclass)
1301 {
1302         int ret;
1303         struct lock_list this;
1304         struct lock_list *uninitialized_var(target_entry);
1305
1306         this.parent = NULL;
1307
1308         this.class = hlock_class(prev);
1309         ret = find_usage_backwards(&this, bit_backwards, &target_entry);
1310         BFS_PROCESS_RET(ret);
1311         backwards_match = target_entry->class;
1312
1313         this.class = hlock_class(next);
1314         ret = find_usage_forwards(&this, bit_forwards, &target_entry);
1315         BFS_PROCESS_RET(ret);
1316         forwards_match = target_entry->class;
1317
1318         return print_bad_irq_dependency(curr, prev, next,
1319                         bit_backwards, bit_forwards, irqclass);
1320 }
1321
1322 static const char *state_names[] = {
1323 #define LOCKDEP_STATE(__STATE) \
1324         __stringify(__STATE),
1325 #include "lockdep_states.h"
1326 #undef LOCKDEP_STATE
1327 };
1328
1329 static const char *state_rnames[] = {
1330 #define LOCKDEP_STATE(__STATE) \
1331         __stringify(__STATE)"-READ",
1332 #include "lockdep_states.h"
1333 #undef LOCKDEP_STATE
1334 };
1335
1336 static inline const char *state_name(enum lock_usage_bit bit)
1337 {
1338         return (bit & 1) ? state_rnames[bit >> 2] : state_names[bit >> 2];
1339 }
1340
1341 static int exclusive_bit(int new_bit)
1342 {
1343         /*
1344          * USED_IN
1345          * USED_IN_READ
1346          * ENABLED
1347          * ENABLED_READ
1348          *
1349          * bit 0 - write/read
1350          * bit 1 - used_in/enabled
1351          * bit 2+  state
1352          */
1353
1354         int state = new_bit & ~3;
1355         int dir = new_bit & 2;
1356
1357         /*
1358          * keep state, bit flip the direction and strip read.
1359          */
1360         return state | (dir ^ 2);
1361 }
1362
1363 static int check_irq_usage(struct task_struct *curr, struct held_lock *prev,
1364                            struct held_lock *next, enum lock_usage_bit bit)
1365 {
1366         /*
1367          * Prove that the new dependency does not connect a hardirq-safe
1368          * lock with a hardirq-unsafe lock - to achieve this we search
1369          * the backwards-subgraph starting at <prev>, and the
1370          * forwards-subgraph starting at <next>:
1371          */
1372         if (!check_usage(curr, prev, next, bit,
1373                            exclusive_bit(bit), state_name(bit)))
1374                 return 0;
1375
1376         bit++; /* _READ */
1377
1378         /*
1379          * Prove that the new dependency does not connect a hardirq-safe-read
1380          * lock with a hardirq-unsafe lock - to achieve this we search
1381          * the backwards-subgraph starting at <prev>, and the
1382          * forwards-subgraph starting at <next>:
1383          */
1384         if (!check_usage(curr, prev, next, bit,
1385                            exclusive_bit(bit), state_name(bit)))
1386                 return 0;
1387
1388         return 1;
1389 }
1390
1391 static int
1392 check_prev_add_irq(struct task_struct *curr, struct held_lock *prev,
1393                 struct held_lock *next)
1394 {
1395 #define LOCKDEP_STATE(__STATE)                                          \
1396         if (!check_irq_usage(curr, prev, next, LOCK_USED_IN_##__STATE)) \
1397                 return 0;
1398 #include "lockdep_states.h"
1399 #undef LOCKDEP_STATE
1400
1401         return 1;
1402 }
1403
1404 static void inc_chains(void)
1405 {
1406         if (current->hardirq_context)
1407                 nr_hardirq_chains++;
1408         else {
1409                 if (current->softirq_context)
1410                         nr_softirq_chains++;
1411                 else
1412                         nr_process_chains++;
1413         }
1414 }
1415
1416 #else
1417
1418 static inline int
1419 check_prev_add_irq(struct task_struct *curr, struct held_lock *prev,
1420                 struct held_lock *next)
1421 {
1422         return 1;
1423 }
1424
1425 static inline void inc_chains(void)
1426 {
1427         nr_process_chains++;
1428 }
1429
1430 #endif
1431
1432 static int
1433 print_deadlock_bug(struct task_struct *curr, struct held_lock *prev,
1434                    struct held_lock *next)
1435 {
1436         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1437                 return 0;
1438
1439         printk("\n=============================================\n");
1440         printk(  "[ INFO: possible recursive locking detected ]\n");
1441         print_kernel_version();
1442         printk(  "---------------------------------------------\n");
1443         printk("%s/%d is trying to acquire lock:\n",
1444                 curr->comm, task_pid_nr(curr));
1445         print_lock(next);
1446         printk("\nbut task is already holding lock:\n");
1447         print_lock(prev);
1448
1449         printk("\nother info that might help us debug this:\n");
1450         lockdep_print_held_locks(curr);
1451
1452         printk("\nstack backtrace:\n");
1453         dump_stack();
1454
1455         return 0;
1456 }
1457
1458 /*
1459  * Check whether we are holding such a class already.
1460  *
1461  * (Note that this has to be done separately, because the graph cannot
1462  * detect such classes of deadlocks.)
1463  *
1464  * Returns: 0 on deadlock detected, 1 on OK, 2 on recursive read
1465  */
1466 static int
1467 check_deadlock(struct task_struct *curr, struct held_lock *next,
1468                struct lockdep_map *next_instance, int read)
1469 {
1470         struct held_lock *prev;
1471         struct held_lock *nest = NULL;
1472         int i;
1473
1474         for (i = 0; i < curr->lockdep_depth; i++) {
1475                 prev = curr->held_locks + i;
1476
1477                 if (prev->instance == next->nest_lock)
1478                         nest = prev;
1479
1480                 if (hlock_class(prev) != hlock_class(next))
1481                         continue;
1482
1483                 /*
1484                  * Allow read-after-read recursion of the same
1485                  * lock class (i.e. read_lock(lock)+read_lock(lock)):
1486                  */
1487                 if ((read == 2) && prev->read)
1488                         return 2;
1489
1490                 /*
1491                  * We're holding the nest_lock, which serializes this lock's
1492                  * nesting behaviour.
1493                  */
1494                 if (nest)
1495                         return 2;
1496
1497                 return print_deadlock_bug(curr, prev, next);
1498         }
1499         return 1;
1500 }
1501
1502 /*
1503  * There was a chain-cache miss, and we are about to add a new dependency
1504  * to a previous lock. We recursively validate the following rules:
1505  *
1506  *  - would the adding of the <prev> -> <next> dependency create a
1507  *    circular dependency in the graph? [== circular deadlock]
1508  *
1509  *  - does the new prev->next dependency connect any hardirq-safe lock
1510  *    (in the full backwards-subgraph starting at <prev>) with any
1511  *    hardirq-unsafe lock (in the full forwards-subgraph starting at
1512  *    <next>)? [== illegal lock inversion with hardirq contexts]
1513  *
1514  *  - does the new prev->next dependency connect any softirq-safe lock
1515  *    (in the full backwards-subgraph starting at <prev>) with any
1516  *    softirq-unsafe lock (in the full forwards-subgraph starting at
1517  *    <next>)? [== illegal lock inversion with softirq contexts]
1518  *
1519  * any of these scenarios could lead to a deadlock.
1520  *
1521  * Then if all the validations pass, we add the forwards and backwards
1522  * dependency.
1523  */
1524 static int
1525 check_prev_add(struct task_struct *curr, struct held_lock *prev,
1526                struct held_lock *next, int distance)
1527 {
1528         struct lock_list *entry;
1529         int ret;
1530         struct lock_list this;
1531         struct lock_list *uninitialized_var(target_entry);
1532
1533         /*
1534          * Prove that the new <prev> -> <next> dependency would not
1535          * create a circular dependency in the graph. (We do this by
1536          * forward-recursing into the graph starting at <next>, and
1537          * checking whether we can reach <prev>.)
1538          *
1539          * We are using global variables to control the recursion, to
1540          * keep the stackframe size of the recursive functions low:
1541          */
1542         this.class = hlock_class(next);
1543         this.parent = NULL;
1544         ret = check_noncircular(&this, hlock_class(prev), &target_entry);
1545         if (unlikely(!ret))
1546                 return print_circular_bug(&this, target_entry, next, prev);
1547         else if (unlikely(ret < 0))
1548                 return print_bfs_bug(ret);
1549
1550         if (!check_prev_add_irq(curr, prev, next))
1551                 return 0;
1552
1553         /*
1554          * For recursive read-locks we do all the dependency checks,
1555          * but we dont store read-triggered dependencies (only
1556          * write-triggered dependencies). This ensures that only the
1557          * write-side dependencies matter, and that if for example a
1558          * write-lock never takes any other locks, then the reads are
1559          * equivalent to a NOP.
1560          */
1561         if (next->read == 2 || prev->read == 2)
1562                 return 1;
1563         /*
1564          * Is the <prev> -> <next> dependency already present?
1565          *
1566          * (this may occur even though this is a new chain: consider
1567          *  e.g. the L1 -> L2 -> L3 -> L4 and the L5 -> L1 -> L2 -> L3
1568          *  chains - the second one will be new, but L1 already has
1569          *  L2 added to its dependency list, due to the first chain.)
1570          */
1571         list_for_each_entry(entry, &hlock_class(prev)->locks_after, entry) {
1572                 if (entry->class == hlock_class(next)) {
1573                         if (distance == 1)
1574                                 entry->distance = 1;
1575                         return 2;
1576                 }
1577         }
1578
1579         /*
1580          * Ok, all validations passed, add the new lock
1581          * to the previous lock's dependency list:
1582          */
1583         ret = add_lock_to_list(hlock_class(prev), hlock_class(next),
1584                                &hlock_class(prev)->locks_after,
1585                                next->acquire_ip, distance);
1586
1587         if (!ret)
1588                 return 0;
1589
1590         ret = add_lock_to_list(hlock_class(next), hlock_class(prev),
1591                                &hlock_class(next)->locks_before,
1592                                next->acquire_ip, distance);
1593         if (!ret)
1594                 return 0;
1595
1596         /*
1597          * Debugging printouts:
1598          */
1599         if (verbose(hlock_class(prev)) || verbose(hlock_class(next))) {
1600                 graph_unlock();
1601                 printk("\n new dependency: ");
1602                 print_lock_name(hlock_class(prev));
1603                 printk(" => ");
1604                 print_lock_name(hlock_class(next));
1605                 printk("\n");
1606                 dump_stack();
1607                 return graph_lock();
1608         }
1609         return 1;
1610 }
1611
1612 /*
1613  * Add the dependency to all directly-previous locks that are 'relevant'.
1614  * The ones that are relevant are (in increasing distance from curr):
1615  * all consecutive trylock entries and the final non-trylock entry - or
1616  * the end of this context's lock-chain - whichever comes first.
1617  */
1618 static int
1619 check_prevs_add(struct task_struct *curr, struct held_lock *next)
1620 {
1621         int depth = curr->lockdep_depth;
1622         struct held_lock *hlock;
1623
1624         /*
1625          * Debugging checks.
1626          *
1627          * Depth must not be zero for a non-head lock:
1628          */
1629         if (!depth)
1630                 goto out_bug;
1631         /*
1632          * At least two relevant locks must exist for this
1633          * to be a head:
1634          */
1635         if (curr->held_locks[depth].irq_context !=
1636                         curr->held_locks[depth-1].irq_context)
1637                 goto out_bug;
1638
1639         for (;;) {
1640                 int distance = curr->lockdep_depth - depth + 1;
1641                 hlock = curr->held_locks + depth-1;
1642                 /*
1643                  * Only non-recursive-read entries get new dependencies
1644                  * added:
1645                  */
1646                 if (hlock->read != 2) {
1647                         if (!check_prev_add(curr, hlock, next, distance))
1648                                 return 0;
1649                         /*
1650                          * Stop after the first non-trylock entry,
1651                          * as non-trylock entries have added their
1652                          * own direct dependencies already, so this
1653                          * lock is connected to them indirectly:
1654                          */
1655                         if (!hlock->trylock)
1656                                 break;
1657                 }
1658                 depth--;
1659                 /*
1660                  * End of lock-stack?
1661                  */
1662                 if (!depth)
1663                         break;
1664                 /*
1665                  * Stop the search if we cross into another context:
1666                  */
1667                 if (curr->held_locks[depth].irq_context !=
1668                                 curr->held_locks[depth-1].irq_context)
1669                         break;
1670         }
1671         return 1;
1672 out_bug:
1673         if (!debug_locks_off_graph_unlock())
1674                 return 0;
1675
1676         WARN_ON(1);
1677
1678         return 0;
1679 }
1680
1681 unsigned long nr_lock_chains;
1682 struct lock_chain lock_chains[MAX_LOCKDEP_CHAINS];
1683 int nr_chain_hlocks;
1684 static u16 chain_hlocks[MAX_LOCKDEP_CHAIN_HLOCKS];
1685
1686 struct lock_class *lock_chain_get_class(struct lock_chain *chain, int i)
1687 {
1688         return lock_classes + chain_hlocks[chain->base + i];
1689 }
1690
1691 /*
1692  * Look up a dependency chain. If the key is not present yet then
1693  * add it and return 1 - in this case the new dependency chain is
1694  * validated. If the key is already hashed, return 0.
1695  * (On return with 1 graph_lock is held.)
1696  */
1697 static inline int lookup_chain_cache(struct task_struct *curr,
1698                                      struct held_lock *hlock,
1699                                      u64 chain_key)
1700 {
1701         struct lock_class *class = hlock_class(hlock);
1702         struct list_head *hash_head = chainhashentry(chain_key);
1703         struct lock_chain *chain;
1704         struct held_lock *hlock_curr, *hlock_next;
1705         int i, j, n, cn;
1706
1707         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
1708                 return 0;
1709         /*
1710          * We can walk it lock-free, because entries only get added
1711          * to the hash:
1712          */
1713         list_for_each_entry(chain, hash_head, entry) {
1714                 if (chain->chain_key == chain_key) {
1715 cache_hit:
1716                         debug_atomic_inc(&chain_lookup_hits);
1717                         if (very_verbose(class))
1718                                 printk("\nhash chain already cached, key: "
1719                                         "%016Lx tail class: [%p] %s\n",
1720                                         (unsigned long long)chain_key,
1721                                         class->key, class->name);
1722                         return 0;
1723                 }
1724         }
1725         if (very_verbose(class))
1726                 printk("\nnew hash chain, key: %016Lx tail class: [%p] %s\n",
1727                         (unsigned long long)chain_key, class->key, class->name);
1728         /*
1729          * Allocate a new chain entry from the static array, and add
1730          * it to the hash:
1731          */
1732         if (!graph_lock())
1733                 return 0;
1734         /*
1735          * We have to walk the chain again locked - to avoid duplicates:
1736          */
1737         list_for_each_entry(chain, hash_head, entry) {
1738                 if (chain->chain_key == chain_key) {
1739                         graph_unlock();
1740                         goto cache_hit;
1741                 }
1742         }
1743         if (unlikely(nr_lock_chains >= MAX_LOCKDEP_CHAINS)) {
1744                 if (!debug_locks_off_graph_unlock())
1745                         return 0;
1746
1747                 printk("BUG: MAX_LOCKDEP_CHAINS too low!\n");
1748                 printk("turning off the locking correctness validator.\n");
1749                 dump_stack();
1750                 return 0;
1751         }
1752         chain = lock_chains + nr_lock_chains++;
1753         chain->chain_key = chain_key;
1754         chain->irq_context = hlock->irq_context;
1755         /* Find the first held_lock of current chain */
1756         hlock_next = hlock;
1757         for (i = curr->lockdep_depth - 1; i >= 0; i--) {
1758                 hlock_curr = curr->held_locks + i;
1759                 if (hlock_curr->irq_context != hlock_next->irq_context)
1760                         break;
1761                 hlock_next = hlock;
1762         }
1763         i++;
1764         chain->depth = curr->lockdep_depth + 1 - i;
1765         cn = nr_chain_hlocks;
1766         while (cn + chain->depth <= MAX_LOCKDEP_CHAIN_HLOCKS) {
1767                 n = cmpxchg(&nr_chain_hlocks, cn, cn + chain->depth);
1768                 if (n == cn)
1769                         break;
1770                 cn = n;
1771         }
1772         if (likely(cn + chain->depth <= MAX_LOCKDEP_CHAIN_HLOCKS)) {
1773                 chain->base = cn;
1774                 for (j = 0; j < chain->depth - 1; j++, i++) {
1775                         int lock_id = curr->held_locks[i].class_idx - 1;
1776                         chain_hlocks[chain->base + j] = lock_id;
1777                 }
1778                 chain_hlocks[chain->base + j] = class - lock_classes;
1779         }
1780         list_add_tail_rcu(&chain->entry, hash_head);
1781         debug_atomic_inc(&chain_lookup_misses);
1782         inc_chains();
1783
1784         return 1;
1785 }
1786
1787 static int validate_chain(struct task_struct *curr, struct lockdep_map *lock,
1788                 struct held_lock *hlock, int chain_head, u64 chain_key)
1789 {
1790         /*
1791          * Trylock needs to maintain the stack of held locks, but it
1792          * does not add new dependencies, because trylock can be done
1793          * in any order.
1794          *
1795          * We look up the chain_key and do the O(N^2) check and update of
1796          * the dependencies only if this is a new dependency chain.
1797          * (If lookup_chain_cache() returns with 1 it acquires
1798          * graph_lock for us)
1799          */
1800         if (!hlock->trylock && (hlock->check == 2) &&
1801             lookup_chain_cache(curr, hlock, chain_key)) {
1802                 /*
1803                  * Check whether last held lock:
1804                  *
1805                  * - is irq-safe, if this lock is irq-unsafe
1806                  * - is softirq-safe, if this lock is hardirq-unsafe
1807                  *
1808                  * And check whether the new lock's dependency graph
1809                  * could lead back to the previous lock.
1810                  *
1811                  * any of these scenarios could lead to a deadlock. If
1812                  * All validations
1813                  */
1814                 int ret = check_deadlock(curr, hlock, lock, hlock->read);
1815
1816                 if (!ret)
1817                         return 0;
1818                 /*
1819                  * Mark recursive read, as we jump over it when
1820                  * building dependencies (just like we jump over
1821                  * trylock entries):
1822                  */
1823                 if (ret == 2)
1824                         hlock->read = 2;
1825                 /*
1826                  * Add dependency only if this lock is not the head
1827                  * of the chain, and if it's not a secondary read-lock:
1828                  */
1829                 if (!chain_head && ret != 2)
1830                         if (!check_prevs_add(curr, hlock))
1831                                 return 0;
1832                 graph_unlock();
1833         } else
1834                 /* after lookup_chain_cache(): */
1835                 if (unlikely(!debug_locks))
1836                         return 0;
1837
1838         return 1;
1839 }
1840 #else
1841 static inline int validate_chain(struct task_struct *curr,
1842                 struct lockdep_map *lock, struct held_lock *hlock,
1843                 int chain_head, u64 chain_key)
1844 {
1845         return 1;
1846 }
1847 #endif
1848
1849 /*
1850  * We are building curr_chain_key incrementally, so double-check
1851  * it from scratch, to make sure that it's done correctly:
1852  */
1853 static void check_chain_key(struct task_struct *curr)
1854 {
1855 #ifdef CONFIG_DEBUG_LOCKDEP
1856         struct held_lock *hlock, *prev_hlock = NULL;
1857         unsigned int i, id;
1858         u64 chain_key = 0;
1859
1860         for (i = 0; i < curr->lockdep_depth; i++) {
1861                 hlock = curr->held_locks + i;
1862                 if (chain_key != hlock->prev_chain_key) {
1863                         debug_locks_off();
1864                         WARN(1, "hm#1, depth: %u [%u], %016Lx != %016Lx\n",
1865                                 curr->lockdep_depth, i,
1866                                 (unsigned long long)chain_key,
1867                                 (unsigned long long)hlock->prev_chain_key);
1868                         return;
1869                 }
1870                 id = hlock->class_idx - 1;
1871                 if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
1872                         return;
1873
1874                 if (prev_hlock && (prev_hlock->irq_context !=
1875                                                         hlock->irq_context))
1876                         chain_key = 0;
1877                 chain_key = iterate_chain_key(chain_key, id);
1878                 prev_hlock = hlock;
1879         }
1880         if (chain_key != curr->curr_chain_key) {
1881                 debug_locks_off();
1882                 WARN(1, "hm#2, depth: %u [%u], %016Lx != %016Lx\n",
1883                         curr->lockdep_depth, i,
1884                         (unsigned long long)chain_key,
1885                         (unsigned long long)curr->curr_chain_key);
1886         }
1887 #endif
1888 }
1889
1890 static int
1891 print_usage_bug(struct task_struct *curr, struct held_lock *this,
1892                 enum lock_usage_bit prev_bit, enum lock_usage_bit new_bit)
1893 {
1894         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1895                 return 0;
1896
1897         printk("\n=================================\n");
1898         printk(  "[ INFO: inconsistent lock state ]\n");
1899         print_kernel_version();
1900         printk(  "---------------------------------\n");
1901
1902         printk("inconsistent {%s} -> {%s} usage.\n",
1903                 usage_str[prev_bit], usage_str[new_bit]);
1904
1905         printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] takes:\n",
1906                 curr->comm, task_pid_nr(curr),
1907                 trace_hardirq_context(curr), hardirq_count() >> HARDIRQ_SHIFT,
1908                 trace_softirq_context(curr), softirq_count() >> SOFTIRQ_SHIFT,
1909                 trace_hardirqs_enabled(curr),
1910                 trace_softirqs_enabled(curr));
1911         print_lock(this);
1912
1913         printk("{%s} state was registered at:\n", usage_str[prev_bit]);
1914         print_stack_trace(hlock_class(this)->usage_traces + prev_bit, 1);
1915
1916         print_irqtrace_events(curr);
1917         printk("\nother info that might help us debug this:\n");
1918         lockdep_print_held_locks(curr);
1919
1920         printk("\nstack backtrace:\n");
1921         dump_stack();
1922
1923         return 0;
1924 }
1925
1926 /*
1927  * Print out an error if an invalid bit is set:
1928  */
1929 static inline int
1930 valid_state(struct task_struct *curr, struct held_lock *this,
1931             enum lock_usage_bit new_bit, enum lock_usage_bit bad_bit)
1932 {
1933         if (unlikely(hlock_class(this)->usage_mask & (1 << bad_bit)))
1934                 return print_usage_bug(curr, this, bad_bit, new_bit);
1935         return 1;
1936 }
1937
1938 static int mark_lock(struct task_struct *curr, struct held_lock *this,
1939                      enum lock_usage_bit new_bit);
1940
1941 #if defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING)
1942
1943 /*
1944  * print irq inversion bug:
1945  */
1946 static int
1947 print_irq_inversion_bug(struct task_struct *curr, struct lock_class *other,
1948                         struct held_lock *this, int forwards,
1949                         const char *irqclass)
1950 {
1951         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1952                 return 0;
1953
1954         printk("\n=========================================================\n");
1955         printk(  "[ INFO: possible irq lock inversion dependency detected ]\n");
1956         print_kernel_version();
1957         printk(  "---------------------------------------------------------\n");
1958         printk("%s/%d just changed the state of lock:\n",
1959                 curr->comm, task_pid_nr(curr));
1960         print_lock(this);
1961         if (forwards)
1962                 printk("but this lock took another, %s-unsafe lock in the past:\n", irqclass);
1963         else
1964                 printk("but this lock was taken by another, %s-safe lock in the past:\n", irqclass);
1965         print_lock_name(other);
1966         printk("\n\nand interrupts could create inverse lock ordering between them.\n\n");
1967
1968         printk("\nother info that might help us debug this:\n");
1969         lockdep_print_held_locks(curr);
1970
1971         printk("\nthe first lock's dependencies:\n");
1972         print_lock_dependencies(hlock_class(this), 0);
1973
1974         printk("\nthe second lock's dependencies:\n");
1975         print_lock_dependencies(other, 0);
1976
1977         printk("\nstack backtrace:\n");
1978         dump_stack();
1979
1980         return 0;
1981 }
1982
1983 /*
1984  * Prove that in the forwards-direction subgraph starting at <this>
1985  * there is no lock matching <mask>:
1986  */
1987 static int
1988 check_usage_forwards(struct task_struct *curr, struct held_lock *this,
1989                      enum lock_usage_bit bit, const char *irqclass)
1990 {
1991         int ret;
1992         struct lock_list root;
1993         struct lock_list *uninitialized_var(target_entry);
1994
1995         root.parent = NULL;
1996         root.class = hlock_class(this);
1997         ret = find_usage_forwards(&root, bit, &target_entry);
1998         BFS_PROCESS_RET(ret);
1999
2000         return print_irq_inversion_bug(curr, target_entry->class,
2001                                         this, 1, irqclass);
2002 }
2003
2004 /*
2005  * Prove that in the backwards-direction subgraph starting at <this>
2006  * there is no lock matching <mask>:
2007  */
2008 static int
2009 check_usage_backwards(struct task_struct *curr, struct held_lock *this,
2010                       enum lock_usage_bit bit, const char *irqclass)
2011 {
2012         int ret;
2013         struct lock_list root;
2014         struct lock_list *uninitialized_var(target_entry);
2015
2016         root.parent = NULL;
2017         root.class = hlock_class(this);
2018         ret = find_usage_backwards(&root, bit, &target_entry);
2019         BFS_PROCESS_RET(ret);
2020
2021         return print_irq_inversion_bug(curr, target_entry->class,
2022                                         this, 1, irqclass);
2023 }
2024
2025 void print_irqtrace_events(struct task_struct *curr)
2026 {
2027         printk("irq event stamp: %u\n", curr->irq_events);
2028         printk("hardirqs last  enabled at (%u): ", curr->hardirq_enable_event);
2029         print_ip_sym(curr->hardirq_enable_ip);
2030         printk("hardirqs last disabled at (%u): ", curr->hardirq_disable_event);
2031         print_ip_sym(curr->hardirq_disable_ip);
2032         printk("softirqs last  enabled at (%u): ", curr->softirq_enable_event);
2033         print_ip_sym(curr->softirq_enable_ip);
2034         printk("softirqs last disabled at (%u): ", curr->softirq_disable_event);
2035         print_ip_sym(curr->softirq_disable_ip);
2036 }
2037
2038 static int HARDIRQ_verbose(struct lock_class *class)
2039 {
2040 #if HARDIRQ_VERBOSE
2041         return class_filter(class);
2042 #endif
2043         return 0;
2044 }
2045
2046 static int SOFTIRQ_verbose(struct lock_class *class)
2047 {
2048 #if SOFTIRQ_VERBOSE
2049         return class_filter(class);
2050 #endif
2051         return 0;
2052 }
2053
2054 static int RECLAIM_FS_verbose(struct lock_class *class)
2055 {
2056 #if RECLAIM_VERBOSE
2057         return class_filter(class);
2058 #endif
2059         return 0;
2060 }
2061
2062 #define STRICT_READ_CHECKS      1
2063
2064 static int (*state_verbose_f[])(struct lock_class *class) = {
2065 #define LOCKDEP_STATE(__STATE) \
2066         __STATE##_verbose,
2067 #include "lockdep_states.h"
2068 #undef LOCKDEP_STATE
2069 };
2070
2071 static inline int state_verbose(enum lock_usage_bit bit,
2072                                 struct lock_class *class)
2073 {
2074         return state_verbose_f[bit >> 2](class);
2075 }
2076
2077 typedef int (*check_usage_f)(struct task_struct *, struct held_lock *,
2078                              enum lock_usage_bit bit, const char *name);
2079
2080 static int
2081 mark_lock_irq(struct task_struct *curr, struct held_lock *this,
2082                 enum lock_usage_bit new_bit)
2083 {
2084         int excl_bit = exclusive_bit(new_bit);
2085         int read = new_bit & 1;
2086         int dir = new_bit & 2;
2087
2088         /*
2089          * mark USED_IN has to look forwards -- to ensure no dependency
2090          * has ENABLED state, which would allow recursion deadlocks.
2091          *
2092          * mark ENABLED has to look backwards -- to ensure no dependee
2093          * has USED_IN state, which, again, would allow  recursion deadlocks.
2094          */
2095         check_usage_f usage = dir ?
2096                 check_usage_backwards : check_usage_forwards;
2097
2098         /*
2099          * Validate that this particular lock does not have conflicting
2100          * usage states.
2101          */
2102         if (!valid_state(curr, this, new_bit, excl_bit))
2103                 return 0;
2104
2105         /*
2106          * Validate that the lock dependencies don't have conflicting usage
2107          * states.
2108          */
2109         if ((!read || !dir || STRICT_READ_CHECKS) &&
2110                         !usage(curr, this, excl_bit, state_name(new_bit & ~1)))
2111                 return 0;
2112
2113         /*
2114          * Check for read in write conflicts
2115          */
2116         if (!read) {
2117                 if (!valid_state(curr, this, new_bit, excl_bit + 1))
2118                         return 0;
2119
2120                 if (STRICT_READ_CHECKS &&
2121                         !usage(curr, this, excl_bit + 1,
2122                                 state_name(new_bit + 1)))
2123                         return 0;
2124         }
2125
2126         if (state_verbose(new_bit, hlock_class(this)))
2127                 return 2;
2128
2129         return 1;
2130 }
2131
2132 enum mark_type {
2133 #define LOCKDEP_STATE(__STATE)  __STATE,
2134 #include "lockdep_states.h"
2135 #undef LOCKDEP_STATE
2136 };
2137
2138 /*
2139  * Mark all held locks with a usage bit:
2140  */
2141 static int
2142 mark_held_locks(struct task_struct *curr, enum mark_type mark)
2143 {
2144         enum lock_usage_bit usage_bit;
2145         struct held_lock *hlock;
2146         int i;
2147
2148         for (i = 0; i < curr->lockdep_depth; i++) {
2149                 hlock = curr->held_locks + i;
2150
2151                 usage_bit = 2 + (mark << 2); /* ENABLED */
2152                 if (hlock->read)
2153                         usage_bit += 1; /* READ */
2154
2155                 BUG_ON(usage_bit >= LOCK_USAGE_STATES);
2156
2157                 if (!mark_lock(curr, hlock, usage_bit))
2158                         return 0;
2159         }
2160
2161         return 1;
2162 }
2163
2164 /*
2165  * Debugging helper: via this flag we know that we are in
2166  * 'early bootup code', and will warn about any invalid irqs-on event:
2167  */
2168 static int early_boot_irqs_enabled;
2169
2170 void early_boot_irqs_off(void)
2171 {
2172         early_boot_irqs_enabled = 0;
2173 }
2174
2175 void early_boot_irqs_on(void)
2176 {
2177         early_boot_irqs_enabled = 1;
2178 }
2179
2180 /*
2181  * Hardirqs will be enabled:
2182  */
2183 void trace_hardirqs_on_caller(unsigned long ip)
2184 {
2185         struct task_struct *curr = current;
2186
2187         time_hardirqs_on(CALLER_ADDR0, ip);
2188
2189         if (unlikely(!debug_locks || current->lockdep_recursion))
2190                 return;
2191
2192         if (DEBUG_LOCKS_WARN_ON(unlikely(!early_boot_irqs_enabled)))
2193                 return;
2194
2195         if (unlikely(curr->hardirqs_enabled)) {
2196                 debug_atomic_inc(&redundant_hardirqs_on);
2197                 return;
2198         }
2199         /* we'll do an OFF -> ON transition: */
2200         curr->hardirqs_enabled = 1;
2201
2202         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2203                 return;
2204         if (DEBUG_LOCKS_WARN_ON(current->hardirq_context))
2205                 return;
2206         /*
2207          * We are going to turn hardirqs on, so set the
2208          * usage bit for all held locks:
2209          */
2210         if (!mark_held_locks(curr, HARDIRQ))
2211                 return;
2212         /*
2213          * If we have softirqs enabled, then set the usage
2214          * bit for all held locks. (disabled hardirqs prevented
2215          * this bit from being set before)
2216          */
2217         if (curr->softirqs_enabled)
2218                 if (!mark_held_locks(curr, SOFTIRQ))
2219                         return;
2220
2221         curr->hardirq_enable_ip = ip;
2222         curr->hardirq_enable_event = ++curr->irq_events;
2223         debug_atomic_inc(&hardirqs_on_events);
2224 }
2225 EXPORT_SYMBOL(trace_hardirqs_on_caller);
2226
2227 void trace_hardirqs_on(void)
2228 {
2229         trace_hardirqs_on_caller(CALLER_ADDR0);
2230 }
2231 EXPORT_SYMBOL(trace_hardirqs_on);
2232
2233 /*
2234  * Hardirqs were disabled:
2235  */
2236 void trace_hardirqs_off_caller(unsigned long ip)
2237 {
2238         struct task_struct *curr = current;
2239
2240         time_hardirqs_off(CALLER_ADDR0, ip);
2241
2242         if (unlikely(!debug_locks || current->lockdep_recursion))
2243                 return;
2244
2245         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2246                 return;
2247
2248         if (curr->hardirqs_enabled) {
2249                 /*
2250                  * We have done an ON -> OFF transition:
2251                  */
2252                 curr->hardirqs_enabled = 0;
2253                 curr->hardirq_disable_ip = ip;
2254                 curr->hardirq_disable_event = ++curr->irq_events;
2255                 debug_atomic_inc(&hardirqs_off_events);
2256         } else
2257                 debug_atomic_inc(&redundant_hardirqs_off);
2258 }
2259 EXPORT_SYMBOL(trace_hardirqs_off_caller);
2260
2261 void trace_hardirqs_off(void)
2262 {
2263         trace_hardirqs_off_caller(CALLER_ADDR0);
2264 }
2265 EXPORT_SYMBOL(trace_hardirqs_off);
2266
2267 /*
2268  * Softirqs will be enabled:
2269  */
2270 void trace_softirqs_on(unsigned long ip)
2271 {
2272         struct task_struct *curr = current;
2273
2274         if (unlikely(!debug_locks))
2275                 return;
2276
2277         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2278                 return;
2279
2280         if (curr->softirqs_enabled) {
2281                 debug_atomic_inc(&redundant_softirqs_on);
2282                 return;
2283         }
2284
2285         /*
2286          * We'll do an OFF -> ON transition:
2287          */
2288         curr->softirqs_enabled = 1;
2289         curr->softirq_enable_ip = ip;
2290         curr->softirq_enable_event = ++curr->irq_events;
2291         debug_atomic_inc(&softirqs_on_events);
2292         /*
2293          * We are going to turn softirqs on, so set the
2294          * usage bit for all held locks, if hardirqs are
2295          * enabled too:
2296          */
2297         if (curr->hardirqs_enabled)
2298                 mark_held_locks(curr, SOFTIRQ);
2299 }
2300
2301 /*
2302  * Softirqs were disabled:
2303  */
2304 void trace_softirqs_off(unsigned long ip)
2305 {
2306         struct task_struct *curr = current;
2307
2308         if (unlikely(!debug_locks))
2309                 return;
2310
2311         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2312                 return;
2313
2314         if (curr->softirqs_enabled) {
2315                 /*
2316                  * We have done an ON -> OFF transition:
2317                  */
2318                 curr->softirqs_enabled = 0;
2319                 curr->softirq_disable_ip = ip;
2320                 curr->softirq_disable_event = ++curr->irq_events;
2321                 debug_atomic_inc(&softirqs_off_events);
2322                 DEBUG_LOCKS_WARN_ON(!softirq_count());
2323         } else
2324                 debug_atomic_inc(&redundant_softirqs_off);
2325 }
2326
2327 static void __lockdep_trace_alloc(gfp_t gfp_mask, unsigned long flags)
2328 {
2329         struct task_struct *curr = current;
2330
2331         if (unlikely(!debug_locks))
2332                 return;
2333
2334         /* no reclaim without waiting on it */
2335         if (!(gfp_mask & __GFP_WAIT))
2336                 return;
2337
2338         /* this guy won't enter reclaim */
2339         if ((curr->flags & PF_MEMALLOC) && !(gfp_mask & __GFP_NOMEMALLOC))
2340                 return;
2341
2342         /* We're only interested __GFP_FS allocations for now */
2343         if (!(gfp_mask & __GFP_FS))
2344                 return;
2345
2346         if (DEBUG_LOCKS_WARN_ON(irqs_disabled_flags(flags)))
2347                 return;
2348
2349         mark_held_locks(curr, RECLAIM_FS);
2350 }
2351
2352 static void check_flags(unsigned long flags);
2353
2354 void lockdep_trace_alloc(gfp_t gfp_mask)
2355 {
2356         unsigned long flags;
2357
2358         if (unlikely(current->lockdep_recursion))
2359                 return;
2360
2361         raw_local_irq_save(flags);
2362         check_flags(flags);
2363         current->lockdep_recursion = 1;
2364         __lockdep_trace_alloc(gfp_mask, flags);
2365         current->lockdep_recursion = 0;
2366         raw_local_irq_restore(flags);
2367 }
2368
2369 static int mark_irqflags(struct task_struct *curr, struct held_lock *hlock)
2370 {
2371         /*
2372          * If non-trylock use in a hardirq or softirq context, then
2373          * mark the lock as used in these contexts:
2374          */
2375         if (!hlock->trylock) {
2376                 if (hlock->read) {
2377                         if (curr->hardirq_context)
2378                                 if (!mark_lock(curr, hlock,
2379                                                 LOCK_USED_IN_HARDIRQ_READ))
2380                                         return 0;
2381                         if (curr->softirq_context)
2382                                 if (!mark_lock(curr, hlock,
2383                                                 LOCK_USED_IN_SOFTIRQ_READ))
2384                                         return 0;
2385                 } else {
2386                         if (curr->hardirq_context)
2387                                 if (!mark_lock(curr, hlock, LOCK_USED_IN_HARDIRQ))
2388                                         return 0;
2389                         if (curr->softirq_context)
2390                                 if (!mark_lock(curr, hlock, LOCK_USED_IN_SOFTIRQ))
2391                                         return 0;
2392                 }
2393         }
2394         if (!hlock->hardirqs_off) {
2395                 if (hlock->read) {
2396                         if (!mark_lock(curr, hlock,
2397                                         LOCK_ENABLED_HARDIRQ_READ))
2398                                 return 0;
2399                         if (curr->softirqs_enabled)
2400                                 if (!mark_lock(curr, hlock,
2401                                                 LOCK_ENABLED_SOFTIRQ_READ))
2402                                         return 0;
2403                 } else {
2404                         if (!mark_lock(curr, hlock,
2405                                         LOCK_ENABLED_HARDIRQ))
2406                                 return 0;
2407                         if (curr->softirqs_enabled)
2408                                 if (!mark_lock(curr, hlock,
2409                                                 LOCK_ENABLED_SOFTIRQ))
2410                                         return 0;
2411                 }
2412         }
2413
2414         /*
2415          * We reuse the irq context infrastructure more broadly as a general
2416          * context checking code. This tests GFP_FS recursion (a lock taken
2417          * during reclaim for a GFP_FS allocation is held over a GFP_FS
2418          * allocation).
2419          */
2420         if (!hlock->trylock && (curr->lockdep_reclaim_gfp & __GFP_FS)) {
2421                 if (hlock->read) {
2422                         if (!mark_lock(curr, hlock, LOCK_USED_IN_RECLAIM_FS_READ))
2423                                         return 0;
2424                 } else {
2425                         if (!mark_lock(curr, hlock, LOCK_USED_IN_RECLAIM_FS))
2426                                         return 0;
2427                 }
2428         }
2429
2430         return 1;
2431 }
2432
2433 static int separate_irq_context(struct task_struct *curr,
2434                 struct held_lock *hlock)
2435 {
2436         unsigned int depth = curr->lockdep_depth;
2437
2438         /*
2439          * Keep track of points where we cross into an interrupt context:
2440          */
2441         hlock->irq_context = 2*(curr->hardirq_context ? 1 : 0) +
2442                                 curr->softirq_context;
2443         if (depth) {
2444                 struct held_lock *prev_hlock;
2445
2446                 prev_hlock = curr->held_locks + depth-1;
2447                 /*
2448                  * If we cross into another context, reset the
2449                  * hash key (this also prevents the checking and the
2450                  * adding of the dependency to 'prev'):
2451                  */
2452                 if (prev_hlock->irq_context != hlock->irq_context)
2453                         return 1;
2454         }
2455         return 0;
2456 }
2457
2458 #else
2459
2460 static inline
2461 int mark_lock_irq(struct task_struct *curr, struct held_lock *this,
2462                 enum lock_usage_bit new_bit)
2463 {
2464         WARN_ON(1);
2465         return 1;
2466 }
2467
2468 static inline int mark_irqflags(struct task_struct *curr,
2469                 struct held_lock *hlock)
2470 {
2471         return 1;
2472 }
2473
2474 static inline int separate_irq_context(struct task_struct *curr,
2475                 struct held_lock *hlock)
2476 {
2477         return 0;
2478 }
2479
2480 void lockdep_trace_alloc(gfp_t gfp_mask)
2481 {
2482 }
2483
2484 #endif
2485
2486 /*
2487  * Mark a lock with a usage bit, and validate the state transition:
2488  */
2489 static int mark_lock(struct task_struct *curr, struct held_lock *this,
2490                              enum lock_usage_bit new_bit)
2491 {
2492         unsigned int new_mask = 1 << new_bit, ret = 1;
2493
2494         /*
2495          * If already set then do not dirty the cacheline,
2496          * nor do any checks:
2497          */
2498         if (likely(hlock_class(this)->usage_mask & new_mask))
2499                 return 1;
2500
2501         if (!graph_lock())
2502                 return 0;
2503         /*
2504          * Make sure we didnt race:
2505          */
2506         if (unlikely(hlock_class(this)->usage_mask & new_mask)) {
2507                 graph_unlock();
2508                 return 1;
2509         }
2510
2511         hlock_class(this)->usage_mask |= new_mask;
2512
2513         if (!save_trace(hlock_class(this)->usage_traces + new_bit))
2514                 return 0;
2515
2516         switch (new_bit) {
2517 #define LOCKDEP_STATE(__STATE)                  \
2518         case LOCK_USED_IN_##__STATE:            \
2519         case LOCK_USED_IN_##__STATE##_READ:     \
2520         case LOCK_ENABLED_##__STATE:            \
2521         case LOCK_ENABLED_##__STATE##_READ:
2522 #include "lockdep_states.h"
2523 #undef LOCKDEP_STATE
2524                 ret = mark_lock_irq(curr, this, new_bit);
2525                 if (!ret)
2526                         return 0;
2527                 break;
2528         case LOCK_USED:
2529                 debug_atomic_dec(&nr_unused_locks);
2530                 break;
2531         default:
2532                 if (!debug_locks_off_graph_unlock())
2533                         return 0;
2534                 WARN_ON(1);
2535                 return 0;
2536         }
2537
2538         graph_unlock();
2539
2540         /*
2541          * We must printk outside of the graph_lock:
2542          */
2543         if (ret == 2) {
2544                 printk("\nmarked lock as {%s}:\n", usage_str[new_bit]);
2545                 print_lock(this);
2546                 print_irqtrace_events(curr);
2547                 dump_stack();
2548         }
2549
2550         return ret;
2551 }
2552
2553 /*
2554  * Initialize a lock instance's lock-class mapping info:
2555  */
2556 void lockdep_init_map(struct lockdep_map *lock, const char *name,
2557                       struct lock_class_key *key, int subclass)
2558 {
2559         lock->class_cache = NULL;
2560 #ifdef CONFIG_LOCK_STAT
2561         lock->cpu = raw_smp_processor_id();
2562 #endif
2563
2564         if (DEBUG_LOCKS_WARN_ON(!name)) {
2565                 lock->name = "NULL";
2566                 return;
2567         }
2568
2569         lock->name = name;
2570
2571         if (DEBUG_LOCKS_WARN_ON(!key))
2572                 return;
2573         /*
2574          * Sanity check, the lock-class key must be persistent:
2575          */
2576         if (!static_obj(key)) {
2577                 printk("BUG: key %p not in .data!\n", key);
2578                 DEBUG_LOCKS_WARN_ON(1);
2579                 return;
2580         }
2581         lock->key = key;
2582
2583         if (unlikely(!debug_locks))
2584                 return;
2585
2586         if (subclass)
2587                 register_lock_class(lock, subclass, 1);
2588 }
2589 EXPORT_SYMBOL_GPL(lockdep_init_map);
2590
2591 /*
2592  * This gets called for every mutex_lock*()/spin_lock*() operation.
2593  * We maintain the dependency maps and validate the locking attempt:
2594  */
2595 static int __lock_acquire(struct lockdep_map *lock, unsigned int subclass,
2596                           int trylock, int read, int check, int hardirqs_off,
2597                           struct lockdep_map *nest_lock, unsigned long ip)
2598 {
2599         struct task_struct *curr = current;
2600         struct lock_class *class = NULL;
2601         struct held_lock *hlock;
2602         unsigned int depth, id;
2603         int chain_head = 0;
2604         u64 chain_key;
2605
2606         if (!prove_locking)
2607                 check = 1;
2608
2609         if (unlikely(!debug_locks))
2610                 return 0;
2611
2612         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2613                 return 0;
2614
2615         if (unlikely(subclass >= MAX_LOCKDEP_SUBCLASSES)) {
2616                 debug_locks_off();
2617                 printk("BUG: MAX_LOCKDEP_SUBCLASSES too low!\n");
2618                 printk("turning off the locking correctness validator.\n");
2619                 dump_stack();
2620                 return 0;
2621         }
2622
2623         if (!subclass)
2624                 class = lock->class_cache;
2625         /*
2626          * Not cached yet or subclass?
2627          */
2628         if (unlikely(!class)) {
2629                 class = register_lock_class(lock, subclass, 0);
2630                 if (!class)
2631                         return 0;
2632         }
2633         debug_atomic_inc((atomic_t *)&class->ops);
2634         if (very_verbose(class)) {
2635                 printk("\nacquire class [%p] %s", class->key, class->name);
2636                 if (class->name_version > 1)
2637                         printk("#%d", class->name_version);
2638                 printk("\n");
2639                 dump_stack();
2640         }
2641
2642         /*
2643          * Add the lock to the list of currently held locks.
2644          * (we dont increase the depth just yet, up until the
2645          * dependency checks are done)
2646          */
2647         depth = curr->lockdep_depth;
2648         if (DEBUG_LOCKS_WARN_ON(depth >= MAX_LOCK_DEPTH))
2649                 return 0;
2650
2651         hlock = curr->held_locks + depth;
2652         if (DEBUG_LOCKS_WARN_ON(!class))
2653                 return 0;
2654         hlock->class_idx = class - lock_classes + 1;
2655         hlock->acquire_ip = ip;
2656         hlock->instance = lock;
2657         hlock->nest_lock = nest_lock;
2658         hlock->trylock = trylock;
2659         hlock->read = read;
2660         hlock->check = check;
2661         hlock->hardirqs_off = !!hardirqs_off;
2662 #ifdef CONFIG_LOCK_STAT
2663         hlock->waittime_stamp = 0;
2664         hlock->holdtime_stamp = sched_clock();
2665 #endif
2666
2667         if (check == 2 && !mark_irqflags(curr, hlock))
2668                 return 0;
2669
2670         /* mark it as used: */
2671         if (!mark_lock(curr, hlock, LOCK_USED))
2672                 return 0;
2673
2674         /*
2675          * Calculate the chain hash: it's the combined hash of all the
2676          * lock keys along the dependency chain. We save the hash value
2677          * at every step so that we can get the current hash easily
2678          * after unlock. The chain hash is then used to cache dependency
2679          * results.
2680          *
2681          * The 'key ID' is what is the most compact key value to drive
2682          * the hash, not class->key.
2683          */
2684         id = class - lock_classes;
2685         if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
2686                 return 0;
2687
2688         chain_key = curr->curr_chain_key;
2689         if (!depth) {
2690                 if (DEBUG_LOCKS_WARN_ON(chain_key != 0))
2691                         return 0;
2692                 chain_head = 1;
2693         }
2694
2695         hlock->prev_chain_key = chain_key;
2696         if (separate_irq_context(curr, hlock)) {
2697                 chain_key = 0;
2698                 chain_head = 1;
2699         }
2700         chain_key = iterate_chain_key(chain_key, id);
2701
2702         if (!validate_chain(curr, lock, hlock, chain_head, chain_key))
2703                 return 0;
2704
2705         curr->curr_chain_key = chain_key;
2706         curr->lockdep_depth++;
2707         check_chain_key(curr);
2708 #ifdef CONFIG_DEBUG_LOCKDEP
2709         if (unlikely(!debug_locks))
2710                 return 0;
2711 #endif
2712         if (unlikely(curr->lockdep_depth >= MAX_LOCK_DEPTH)) {
2713                 debug_locks_off();
2714                 printk("BUG: MAX_LOCK_DEPTH too low!\n");
2715                 printk("turning off the locking correctness validator.\n");
2716                 dump_stack();
2717                 return 0;
2718         }
2719
2720         if (unlikely(curr->lockdep_depth > max_lockdep_depth))
2721                 max_lockdep_depth = curr->lockdep_depth;
2722
2723         return 1;
2724 }
2725
2726 static int
2727 print_unlock_inbalance_bug(struct task_struct *curr, struct lockdep_map *lock,
2728                            unsigned long ip)
2729 {
2730         if (!debug_locks_off())
2731                 return 0;
2732         if (debug_locks_silent)
2733                 return 0;
2734
2735         printk("\n=====================================\n");
2736         printk(  "[ BUG: bad unlock balance detected! ]\n");
2737         printk(  "-------------------------------------\n");
2738         printk("%s/%d is trying to release lock (",
2739                 curr->comm, task_pid_nr(curr));
2740         print_lockdep_cache(lock);
2741         printk(") at:\n");
2742         print_ip_sym(ip);
2743         printk("but there are no more locks to release!\n");
2744         printk("\nother info that might help us debug this:\n");
2745         lockdep_print_held_locks(curr);
2746
2747         printk("\nstack backtrace:\n");
2748         dump_stack();
2749
2750         return 0;
2751 }
2752
2753 /*
2754  * Common debugging checks for both nested and non-nested unlock:
2755  */
2756 static int check_unlock(struct task_struct *curr, struct lockdep_map *lock,
2757                         unsigned long ip)
2758 {
2759         if (unlikely(!debug_locks))
2760                 return 0;
2761         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2762                 return 0;
2763
2764         if (curr->lockdep_depth <= 0)
2765                 return print_unlock_inbalance_bug(curr, lock, ip);
2766
2767         return 1;
2768 }
2769
2770 static int
2771 __lock_set_class(struct lockdep_map *lock, const char *name,
2772                  struct lock_class_key *key, unsigned int subclass,
2773                  unsigned long ip)
2774 {
2775         struct task_struct *curr = current;
2776         struct held_lock *hlock, *prev_hlock;
2777         struct lock_class *class;
2778         unsigned int depth;
2779         int i;
2780
2781         depth = curr->lockdep_depth;
2782         if (DEBUG_LOCKS_WARN_ON(!depth))
2783                 return 0;
2784
2785         prev_hlock = NULL;
2786         for (i = depth-1; i >= 0; i--) {
2787                 hlock = curr->held_locks + i;
2788                 /*
2789                  * We must not cross into another context:
2790                  */
2791                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
2792                         break;
2793                 if (hlock->instance == lock)
2794                         goto found_it;
2795                 prev_hlock = hlock;
2796         }
2797         return print_unlock_inbalance_bug(curr, lock, ip);
2798
2799 found_it:
2800         lockdep_init_map(lock, name, key, 0);
2801         class = register_lock_class(lock, subclass, 0);
2802         hlock->class_idx = class - lock_classes + 1;
2803
2804         curr->lockdep_depth = i;
2805         curr->curr_chain_key = hlock->prev_chain_key;
2806
2807         for (; i < depth; i++) {
2808                 hlock = curr->held_locks + i;
2809                 if (!__lock_acquire(hlock->instance,
2810                         hlock_class(hlock)->subclass, hlock->trylock,
2811                                 hlock->read, hlock->check, hlock->hardirqs_off,
2812                                 hlock->nest_lock, hlock->acquire_ip))
2813                         return 0;
2814         }
2815
2816         if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth))
2817                 return 0;
2818         return 1;
2819 }
2820
2821 /*
2822  * Remove the lock to the list of currently held locks in a
2823  * potentially non-nested (out of order) manner. This is a
2824  * relatively rare operation, as all the unlock APIs default
2825  * to nested mode (which uses lock_release()):
2826  */
2827 static int
2828 lock_release_non_nested(struct task_struct *curr,
2829                         struct lockdep_map *lock, unsigned long ip)
2830 {
2831         struct held_lock *hlock, *prev_hlock;
2832         unsigned int depth;
2833         int i;
2834
2835         /*
2836          * Check whether the lock exists in the current stack
2837          * of held locks:
2838          */
2839         depth = curr->lockdep_depth;
2840         if (DEBUG_LOCKS_WARN_ON(!depth))
2841                 return 0;
2842
2843         prev_hlock = NULL;
2844         for (i = depth-1; i >= 0; i--) {
2845                 hlock = curr->held_locks + i;
2846                 /*
2847                  * We must not cross into another context:
2848                  */
2849                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
2850                         break;
2851                 if (hlock->instance == lock)
2852                         goto found_it;
2853                 prev_hlock = hlock;
2854         }
2855         return print_unlock_inbalance_bug(curr, lock, ip);
2856
2857 found_it:
2858         lock_release_holdtime(hlock);
2859
2860         /*
2861          * We have the right lock to unlock, 'hlock' points to it.
2862          * Now we remove it from the stack, and add back the other
2863          * entries (if any), recalculating the hash along the way:
2864          */
2865         curr->lockdep_depth = i;
2866         curr->curr_chain_key = hlock->prev_chain_key;
2867
2868         for (i++; i < depth; i++) {
2869                 hlock = curr->held_locks + i;
2870                 if (!__lock_acquire(hlock->instance,
2871                         hlock_class(hlock)->subclass, hlock->trylock,
2872                                 hlock->read, hlock->check, hlock->hardirqs_off,
2873                                 hlock->nest_lock, hlock->acquire_ip))
2874                         return 0;
2875         }
2876
2877         if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth - 1))
2878                 return 0;
2879         return 1;
2880 }
2881
2882 /*
2883  * Remove the lock to the list of currently held locks - this gets
2884  * called on mutex_unlock()/spin_unlock*() (or on a failed
2885  * mutex_lock_interruptible()). This is done for unlocks that nest
2886  * perfectly. (i.e. the current top of the lock-stack is unlocked)
2887  */
2888 static int lock_release_nested(struct task_struct *curr,
2889                                struct lockdep_map *lock, unsigned long ip)
2890 {
2891         struct held_lock *hlock;
2892         unsigned int depth;
2893
2894         /*
2895          * Pop off the top of the lock stack:
2896          */
2897         depth = curr->lockdep_depth - 1;
2898         hlock = curr->held_locks + depth;
2899
2900         /*
2901          * Is the unlock non-nested:
2902          */
2903         if (hlock->instance != lock)
2904                 return lock_release_non_nested(curr, lock, ip);
2905         curr->lockdep_depth--;
2906
2907         if (DEBUG_LOCKS_WARN_ON(!depth && (hlock->prev_chain_key != 0)))
2908                 return 0;
2909
2910         curr->curr_chain_key = hlock->prev_chain_key;
2911
2912         lock_release_holdtime(hlock);
2913
2914 #ifdef CONFIG_DEBUG_LOCKDEP
2915         hlock->prev_chain_key = 0;
2916         hlock->class_idx = 0;
2917         hlock->acquire_ip = 0;
2918         hlock->irq_context = 0;
2919 #endif
2920         return 1;
2921 }
2922
2923 /*
2924  * Remove the lock to the list of currently held locks - this gets
2925  * called on mutex_unlock()/spin_unlock*() (or on a failed
2926  * mutex_lock_interruptible()). This is done for unlocks that nest
2927  * perfectly. (i.e. the current top of the lock-stack is unlocked)
2928  */
2929 static void
2930 __lock_release(struct lockdep_map *lock, int nested, unsigned long ip)
2931 {
2932         struct task_struct *curr = current;
2933
2934         if (!check_unlock(curr, lock, ip))
2935                 return;
2936
2937         if (nested) {
2938                 if (!lock_release_nested(curr, lock, ip))
2939                         return;
2940         } else {
2941                 if (!lock_release_non_nested(curr, lock, ip))
2942                         return;
2943         }
2944
2945         check_chain_key(curr);
2946 }
2947
2948 /*
2949  * Check whether we follow the irq-flags state precisely:
2950  */
2951 static void check_flags(unsigned long flags)
2952 {
2953 #if defined(CONFIG_PROVE_LOCKING) && defined(CONFIG_DEBUG_LOCKDEP) && \
2954     defined(CONFIG_TRACE_IRQFLAGS)
2955         if (!debug_locks)
2956                 return;
2957
2958         if (irqs_disabled_flags(flags)) {
2959                 if (DEBUG_LOCKS_WARN_ON(current->hardirqs_enabled)) {
2960                         printk("possible reason: unannotated irqs-off.\n");
2961                 }
2962         } else {
2963                 if (DEBUG_LOCKS_WARN_ON(!current->hardirqs_enabled)) {
2964                         printk("possible reason: unannotated irqs-on.\n");
2965                 }
2966         }
2967
2968         /*
2969          * We dont accurately track softirq state in e.g.
2970          * hardirq contexts (such as on 4KSTACKS), so only
2971          * check if not in hardirq contexts:
2972          */
2973         if (!hardirq_count()) {
2974                 if (softirq_count())
2975                         DEBUG_LOCKS_WARN_ON(current->softirqs_enabled);
2976                 else
2977                         DEBUG_LOCKS_WARN_ON(!current->softirqs_enabled);
2978         }
2979
2980         if (!debug_locks)
2981                 print_irqtrace_events(current);
2982 #endif
2983 }
2984
2985 void lock_set_class(struct lockdep_map *lock, const char *name,
2986                     struct lock_class_key *key, unsigned int subclass,
2987                     unsigned long ip)
2988 {
2989         unsigned long flags;
2990
2991         if (unlikely(current->lockdep_recursion))
2992                 return;
2993
2994         raw_local_irq_save(flags);
2995         current->lockdep_recursion = 1;
2996         check_flags(flags);
2997         if (__lock_set_class(lock, name, key, subclass, ip))
2998                 check_chain_key(current);
2999         current->lockdep_recursion = 0;
3000         raw_local_irq_restore(flags);
3001 }
3002 EXPORT_SYMBOL_GPL(lock_set_class);
3003
3004 /*
3005  * We are not always called with irqs disabled - do that here,
3006  * and also avoid lockdep recursion:
3007  */
3008 void lock_acquire(struct lockdep_map *lock, unsigned int subclass,
3009                           int trylock, int read, int check,
3010                           struct lockdep_map *nest_lock, unsigned long ip)
3011 {
3012         unsigned long flags;
3013
3014         trace_lock_acquire(lock, subclass, trylock, read, check, nest_lock, ip);
3015
3016         if (unlikely(current->lockdep_recursion))
3017                 return;
3018
3019         raw_local_irq_save(flags);
3020         check_flags(flags);
3021
3022         current->lockdep_recursion = 1;
3023         __lock_acquire(lock, subclass, trylock, read, check,
3024                        irqs_disabled_flags(flags), nest_lock, ip);
3025         current->lockdep_recursion = 0;
3026         raw_local_irq_restore(flags);
3027 }
3028 EXPORT_SYMBOL_GPL(lock_acquire);
3029
3030 void lock_release(struct lockdep_map *lock, int nested,
3031                           unsigned long ip)
3032 {
3033         unsigned long flags;
3034
3035         trace_lock_release(lock, nested, ip);
3036
3037         if (unlikely(current->lockdep_recursion))
3038                 return;
3039
3040         raw_local_irq_save(flags);
3041         check_flags(flags);
3042         current->lockdep_recursion = 1;
3043         __lock_release(lock, nested, ip);
3044         current->lockdep_recursion = 0;
3045         raw_local_irq_restore(flags);
3046 }
3047 EXPORT_SYMBOL_GPL(lock_release);
3048
3049 void lockdep_set_current_reclaim_state(gfp_t gfp_mask)
3050 {
3051         current->lockdep_reclaim_gfp = gfp_mask;
3052 }
3053
3054 void lockdep_clear_current_reclaim_state(void)
3055 {
3056         current->lockdep_reclaim_gfp = 0;
3057 }
3058
3059 #ifdef CONFIG_LOCK_STAT
3060 static int
3061 print_lock_contention_bug(struct task_struct *curr, struct lockdep_map *lock,
3062                            unsigned long ip)
3063 {
3064         if (!debug_locks_off())
3065                 return 0;
3066         if (debug_locks_silent)
3067                 return 0;
3068
3069         printk("\n=================================\n");
3070         printk(  "[ BUG: bad contention detected! ]\n");
3071         printk(  "---------------------------------\n");
3072         printk("%s/%d is trying to contend lock (",
3073                 curr->comm, task_pid_nr(curr));
3074         print_lockdep_cache(lock);
3075         printk(") at:\n");
3076         print_ip_sym(ip);
3077         printk("but there are no locks held!\n");
3078         printk("\nother info that might help us debug this:\n");
3079         lockdep_print_held_locks(curr);
3080
3081         printk("\nstack backtrace:\n");
3082         dump_stack();
3083
3084         return 0;
3085 }
3086
3087 static void
3088 __lock_contended(struct lockdep_map *lock, unsigned long ip)
3089 {
3090         struct task_struct *curr = current;
3091         struct held_lock *hlock, *prev_hlock;
3092         struct lock_class_stats *stats;
3093         unsigned int depth;
3094         int i, contention_point, contending_point;
3095
3096         depth = curr->lockdep_depth;
3097         if (DEBUG_LOCKS_WARN_ON(!depth))
3098                 return;
3099
3100         prev_hlock = NULL;
3101         for (i = depth-1; i >= 0; i--) {
3102                 hlock = curr->held_locks + i;
3103                 /*
3104                  * We must not cross into another context:
3105                  */
3106                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3107                         break;
3108                 if (hlock->instance == lock)
3109                         goto found_it;
3110                 prev_hlock = hlock;
3111         }
3112         print_lock_contention_bug(curr, lock, ip);
3113         return;
3114
3115 found_it:
3116         hlock->waittime_stamp = sched_clock();
3117
3118         contention_point = lock_point(hlock_class(hlock)->contention_point, ip);
3119         contending_point = lock_point(hlock_class(hlock)->contending_point,
3120                                       lock->ip);
3121
3122         stats = get_lock_stats(hlock_class(hlock));
3123         if (contention_point < LOCKSTAT_POINTS)
3124                 stats->contention_point[contention_point]++;
3125         if (contending_point < LOCKSTAT_POINTS)
3126                 stats->contending_point[contending_point]++;
3127         if (lock->cpu != smp_processor_id())
3128                 stats->bounces[bounce_contended + !!hlock->read]++;
3129         put_lock_stats(stats);
3130 }
3131
3132 static void
3133 __lock_acquired(struct lockdep_map *lock, unsigned long ip)
3134 {
3135         struct task_struct *curr = current;
3136         struct held_lock *hlock, *prev_hlock;
3137         struct lock_class_stats *stats;
3138         unsigned int depth;
3139         u64 now;
3140         s64 waittime = 0;
3141         int i, cpu;
3142
3143         depth = curr->lockdep_depth;
3144         if (DEBUG_LOCKS_WARN_ON(!depth))
3145                 return;
3146
3147         prev_hlock = NULL;
3148         for (i = depth-1; i >= 0; i--) {
3149                 hlock = curr->held_locks + i;
3150                 /*
3151                  * We must not cross into another context:
3152                  */
3153                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3154                         break;
3155                 if (hlock->instance == lock)
3156                         goto found_it;
3157                 prev_hlock = hlock;
3158         }
3159         print_lock_contention_bug(curr, lock, _RET_IP_);
3160         return;
3161
3162 found_it:
3163         cpu = smp_processor_id();
3164         if (hlock->waittime_stamp) {
3165                 now = sched_clock();
3166                 waittime = now - hlock->waittime_stamp;
3167                 hlock->holdtime_stamp = now;
3168         }
3169
3170         trace_lock_acquired(lock, ip, waittime);
3171
3172         stats = get_lock_stats(hlock_class(hlock));
3173         if (waittime) {
3174                 if (hlock->read)
3175                         lock_time_inc(&stats->read_waittime, waittime);
3176                 else
3177                         lock_time_inc(&stats->write_waittime, waittime);
3178         }
3179         if (lock->cpu != cpu)
3180                 stats->bounces[bounce_acquired + !!hlock->read]++;
3181         put_lock_stats(stats);
3182
3183         lock->cpu = cpu;
3184         lock->ip = ip;
3185 }
3186
3187 void lock_contended(struct lockdep_map *lock, unsigned long ip)
3188 {
3189         unsigned long flags;
3190
3191         trace_lock_contended(lock, ip);
3192
3193         if (unlikely(!lock_stat))
3194                 return;
3195
3196         if (unlikely(current->lockdep_recursion))
3197                 return;
3198
3199         raw_local_irq_save(flags);
3200         check_flags(flags);
3201         current->lockdep_recursion = 1;
3202         __lock_contended(lock, ip);
3203         current->lockdep_recursion = 0;
3204         raw_local_irq_restore(flags);
3205 }
3206 EXPORT_SYMBOL_GPL(lock_contended);
3207
3208 void lock_acquired(struct lockdep_map *lock, unsigned long ip)
3209 {
3210         unsigned long flags;
3211
3212         if (unlikely(!lock_stat))
3213                 return;
3214
3215         if (unlikely(current->lockdep_recursion))
3216                 return;
3217
3218         raw_local_irq_save(flags);
3219         check_flags(flags);
3220         current->lockdep_recursion = 1;
3221         __lock_acquired(lock, ip);
3222         current->lockdep_recursion = 0;
3223         raw_local_irq_restore(flags);
3224 }
3225 EXPORT_SYMBOL_GPL(lock_acquired);
3226 #endif
3227
3228 /*
3229  * Used by the testsuite, sanitize the validator state
3230  * after a simulated failure:
3231  */
3232
3233 void lockdep_reset(void)
3234 {
3235         unsigned long flags;
3236         int i;
3237
3238         raw_local_irq_save(flags);
3239         current->curr_chain_key = 0;
3240         current->lockdep_depth = 0;
3241         current->lockdep_recursion = 0;
3242         memset(current->held_locks, 0, MAX_LOCK_DEPTH*sizeof(struct held_lock));
3243         nr_hardirq_chains = 0;
3244         nr_softirq_chains = 0;
3245         nr_process_chains = 0;
3246         debug_locks = 1;
3247         for (i = 0; i < CHAINHASH_SIZE; i++)
3248                 INIT_LIST_HEAD(chainhash_table + i);
3249         raw_local_irq_restore(flags);
3250 }
3251
3252 static void zap_class(struct lock_class *class)
3253 {
3254         int i;
3255
3256         /*
3257          * Remove all dependencies this lock is
3258          * involved in:
3259          */
3260         for (i = 0; i < nr_list_entries; i++) {
3261                 if (list_entries[i].class == class)
3262                         list_del_rcu(&list_entries[i].entry);
3263         }
3264         /*
3265          * Unhash the class and remove it from the all_lock_classes list:
3266          */
3267         list_del_rcu(&class->hash_entry);
3268         list_del_rcu(&class->lock_entry);
3269
3270         class->key = NULL;
3271 }
3272
3273 static inline int within(const void *addr, void *start, unsigned long size)
3274 {
3275         return addr >= start && addr < start + size;
3276 }
3277
3278 void lockdep_free_key_range(void *start, unsigned long size)
3279 {
3280         struct lock_class *class, *next;
3281         struct list_head *head;
3282         unsigned long flags;
3283         int i;
3284         int locked;
3285
3286         raw_local_irq_save(flags);
3287         locked = graph_lock();
3288
3289         /*
3290          * Unhash all classes that were created by this module:
3291          */
3292         for (i = 0; i < CLASSHASH_SIZE; i++) {
3293                 head = classhash_table + i;
3294                 if (list_empty(head))
3295                         continue;
3296                 list_for_each_entry_safe(class, next, head, hash_entry) {
3297                         if (within(class->key, start, size))
3298                                 zap_class(class);
3299                         else if (within(class->name, start, size))
3300                                 zap_class(class);
3301                 }
3302         }
3303
3304         if (locked)
3305                 graph_unlock();
3306         raw_local_irq_restore(flags);
3307 }
3308
3309 void lockdep_reset_lock(struct lockdep_map *lock)
3310 {
3311         struct lock_class *class, *next;
3312         struct list_head *head;
3313         unsigned long flags;
3314         int i, j;
3315         int locked;
3316
3317         raw_local_irq_save(flags);
3318
3319         /*
3320          * Remove all classes this lock might have:
3321          */
3322         for (j = 0; j < MAX_LOCKDEP_SUBCLASSES; j++) {
3323                 /*
3324                  * If the class exists we look it up and zap it:
3325                  */
3326                 class = look_up_lock_class(lock, j);
3327                 if (class)
3328                         zap_class(class);
3329         }
3330         /*
3331          * Debug check: in the end all mapped classes should
3332          * be gone.
3333          */
3334         locked = graph_lock();
3335         for (i = 0; i < CLASSHASH_SIZE; i++) {
3336                 head = classhash_table + i;
3337                 if (list_empty(head))
3338                         continue;
3339                 list_for_each_entry_safe(class, next, head, hash_entry) {
3340                         if (unlikely(class == lock->class_cache)) {
3341                                 if (debug_locks_off_graph_unlock())
3342                                         WARN_ON(1);
3343                                 goto out_restore;
3344                         }
3345                 }
3346         }
3347         if (locked)
3348                 graph_unlock();
3349
3350 out_restore:
3351         raw_local_irq_restore(flags);
3352 }
3353
3354 void lockdep_init(void)
3355 {
3356         int i;
3357
3358         /*
3359          * Some architectures have their own start_kernel()
3360          * code which calls lockdep_init(), while we also
3361          * call lockdep_init() from the start_kernel() itself,
3362          * and we want to initialize the hashes only once:
3363          */
3364         if (lockdep_initialized)
3365                 return;
3366
3367         for (i = 0; i < CLASSHASH_SIZE; i++)
3368                 INIT_LIST_HEAD(classhash_table + i);
3369
3370         for (i = 0; i < CHAINHASH_SIZE; i++)
3371                 INIT_LIST_HEAD(chainhash_table + i);
3372
3373         lockdep_initialized = 1;
3374 }
3375
3376 void __init lockdep_info(void)
3377 {
3378         printk("Lock dependency validator: Copyright (c) 2006 Red Hat, Inc., Ingo Molnar\n");
3379
3380         printk("... MAX_LOCKDEP_SUBCLASSES:  %lu\n", MAX_LOCKDEP_SUBCLASSES);
3381         printk("... MAX_LOCK_DEPTH:          %lu\n", MAX_LOCK_DEPTH);
3382         printk("... MAX_LOCKDEP_KEYS:        %lu\n", MAX_LOCKDEP_KEYS);
3383         printk("... CLASSHASH_SIZE:          %lu\n", CLASSHASH_SIZE);
3384         printk("... MAX_LOCKDEP_ENTRIES:     %lu\n", MAX_LOCKDEP_ENTRIES);
3385         printk("... MAX_LOCKDEP_CHAINS:      %lu\n", MAX_LOCKDEP_CHAINS);
3386         printk("... CHAINHASH_SIZE:          %lu\n", CHAINHASH_SIZE);
3387
3388         printk(" memory used by lock dependency info: %lu kB\n",
3389                 (sizeof(struct lock_class) * MAX_LOCKDEP_KEYS +
3390                 sizeof(struct list_head) * CLASSHASH_SIZE +
3391                 sizeof(struct lock_list) * MAX_LOCKDEP_ENTRIES +
3392                 sizeof(struct lock_chain) * MAX_LOCKDEP_CHAINS +
3393                 sizeof(struct list_head) * CHAINHASH_SIZE) / 1024);
3394
3395         printk(" per task-struct memory footprint: %lu bytes\n",
3396                 sizeof(struct held_lock) * MAX_LOCK_DEPTH);
3397
3398 #ifdef CONFIG_DEBUG_LOCKDEP
3399         if (lockdep_init_error) {
3400                 printk("WARNING: lockdep init error! Arch code didn't call lockdep_init() early enough?\n");
3401                 printk("Call stack leading to lockdep invocation was:\n");
3402                 print_stack_trace(&lockdep_init_trace, 0);
3403         }
3404 #endif
3405 }
3406
3407 static void
3408 print_freed_lock_bug(struct task_struct *curr, const void *mem_from,
3409                      const void *mem_to, struct held_lock *hlock)
3410 {
3411         if (!debug_locks_off())
3412                 return;
3413         if (debug_locks_silent)
3414                 return;
3415
3416         printk("\n=========================\n");
3417         printk(  "[ BUG: held lock freed! ]\n");
3418         printk(  "-------------------------\n");
3419         printk("%s/%d is freeing memory %p-%p, with a lock still held there!\n",
3420                 curr->comm, task_pid_nr(curr), mem_from, mem_to-1);
3421         print_lock(hlock);
3422         lockdep_print_held_locks(curr);
3423
3424         printk("\nstack backtrace:\n");
3425         dump_stack();
3426 }
3427
3428 static inline int not_in_range(const void* mem_from, unsigned long mem_len,
3429                                 const void* lock_from, unsigned long lock_len)
3430 {
3431         return lock_from + lock_len <= mem_from ||
3432                 mem_from + mem_len <= lock_from;
3433 }
3434
3435 /*
3436  * Called when kernel memory is freed (or unmapped), or if a lock
3437  * is destroyed or reinitialized - this code checks whether there is
3438  * any held lock in the memory range of <from> to <to>:
3439  */
3440 void debug_check_no_locks_freed(const void *mem_from, unsigned long mem_len)
3441 {
3442         struct task_struct *curr = current;
3443         struct held_lock *hlock;
3444         unsigned long flags;
3445         int i;
3446
3447         if (unlikely(!debug_locks))
3448                 return;
3449
3450         local_irq_save(flags);
3451         for (i = 0; i < curr->lockdep_depth; i++) {
3452                 hlock = curr->held_locks + i;
3453
3454                 if (not_in_range(mem_from, mem_len, hlock->instance,
3455                                         sizeof(*hlock->instance)))
3456                         continue;
3457
3458                 print_freed_lock_bug(curr, mem_from, mem_from + mem_len, hlock);
3459                 break;
3460         }
3461         local_irq_restore(flags);
3462 }
3463 EXPORT_SYMBOL_GPL(debug_check_no_locks_freed);
3464
3465 static void print_held_locks_bug(struct task_struct *curr)
3466 {
3467         if (!debug_locks_off())
3468                 return;
3469         if (debug_locks_silent)
3470                 return;
3471
3472         printk("\n=====================================\n");
3473         printk(  "[ BUG: lock held at task exit time! ]\n");
3474         printk(  "-------------------------------------\n");
3475         printk("%s/%d is exiting with locks still held!\n",
3476                 curr->comm, task_pid_nr(curr));
3477         lockdep_print_held_locks(curr);
3478
3479         printk("\nstack backtrace:\n");
3480         dump_stack();
3481 }
3482
3483 void debug_check_no_locks_held(struct task_struct *task)
3484 {
3485         if (unlikely(task->lockdep_depth > 0))
3486                 print_held_locks_bug(task);
3487 }
3488
3489 void debug_show_all_locks(void)
3490 {
3491         struct task_struct *g, *p;
3492         int count = 10;
3493         int unlock = 1;
3494
3495         if (unlikely(!debug_locks)) {
3496                 printk("INFO: lockdep is turned off.\n");
3497                 return;
3498         }
3499         printk("\nShowing all locks held in the system:\n");
3500
3501         /*
3502          * Here we try to get the tasklist_lock as hard as possible,
3503          * if not successful after 2 seconds we ignore it (but keep
3504          * trying). This is to enable a debug printout even if a
3505          * tasklist_lock-holding task deadlocks or crashes.
3506          */
3507 retry:
3508         if (!read_trylock(&tasklist_lock)) {
3509                 if (count == 10)
3510                         printk("hm, tasklist_lock locked, retrying... ");
3511                 if (count) {
3512                         count--;
3513                         printk(" #%d", 10-count);
3514                         mdelay(200);
3515                         goto retry;
3516                 }
3517                 printk(" ignoring it.\n");
3518                 unlock = 0;
3519         } else {
3520                 if (count != 10)
3521                         printk(KERN_CONT " locked it.\n");
3522         }
3523
3524         do_each_thread(g, p) {
3525                 /*
3526                  * It's not reliable to print a task's held locks
3527                  * if it's not sleeping (or if it's not the current
3528                  * task):
3529                  */
3530                 if (p->state == TASK_RUNNING && p != current)
3531                         continue;
3532                 if (p->lockdep_depth)
3533                         lockdep_print_held_locks(p);
3534                 if (!unlock)
3535                         if (read_trylock(&tasklist_lock))
3536                                 unlock = 1;
3537         } while_each_thread(g, p);
3538
3539         printk("\n");
3540         printk("=============================================\n\n");
3541
3542         if (unlock)
3543                 read_unlock(&tasklist_lock);
3544 }
3545 EXPORT_SYMBOL_GPL(debug_show_all_locks);
3546
3547 /*
3548  * Careful: only use this function if you are sure that
3549  * the task cannot run in parallel!
3550  */
3551 void __debug_show_held_locks(struct task_struct *task)
3552 {
3553         if (unlikely(!debug_locks)) {
3554                 printk("INFO: lockdep is turned off.\n");
3555                 return;
3556         }
3557         lockdep_print_held_locks(task);
3558 }
3559 EXPORT_SYMBOL_GPL(__debug_show_held_locks);
3560
3561 void debug_show_held_locks(struct task_struct *task)
3562 {
3563                 __debug_show_held_locks(task);
3564 }
3565 EXPORT_SYMBOL_GPL(debug_show_held_locks);
3566
3567 void lockdep_sys_exit(void)
3568 {
3569         struct task_struct *curr = current;
3570
3571         if (unlikely(curr->lockdep_depth)) {
3572                 if (!debug_locks_off())
3573                         return;
3574                 printk("\n================================================\n");
3575                 printk(  "[ BUG: lock held when returning to user space! ]\n");
3576                 printk(  "------------------------------------------------\n");
3577                 printk("%s/%d is leaving the kernel with locks still held!\n",
3578                                 curr->comm, curr->pid);
3579                 lockdep_print_held_locks(curr);
3580         }
3581 }