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