47b201ecc6df5cd8ae30842bea386a6adad8f0bd
[linux-2.6.git] / kernel / lockdep.c
1 /*
2  * kernel/lockdep.c
3  *
4  * Runtime locking correctness validator
5  *
6  * Started by Ingo Molnar:
7  *
8  *  Copyright (C) 2006,2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
9  *  Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
10  *
11  * this code maps all the lock dependencies as they occur in a live kernel
12  * and will warn about the following classes of locking bugs:
13  *
14  * - lock inversion scenarios
15  * - circular lock dependencies
16  * - hardirq/softirq safe/unsafe locking bugs
17  *
18  * Bugs are reported even if the current locking scenario does not cause
19  * any deadlock at this point.
20  *
21  * I.e. if anytime in the past two locks were taken in a different order,
22  * even if it happened for another task, even if those were different
23  * locks (but of the same class as this lock), this code will detect it.
24  *
25  * Thanks to Arjan van de Ven for coming up with the initial idea of
26  * mapping lock dependencies runtime.
27  */
28 #define DISABLE_BRANCH_PROFILING
29 #include <linux/mutex.h>
30 #include <linux/sched.h>
31 #include <linux/delay.h>
32 #include <linux/module.h>
33 #include <linux/proc_fs.h>
34 #include <linux/seq_file.h>
35 #include <linux/spinlock.h>
36 #include <linux/kallsyms.h>
37 #include <linux/interrupt.h>
38 #include <linux/stacktrace.h>
39 #include <linux/debug_locks.h>
40 #include <linux/irqflags.h>
41 #include <linux/utsname.h>
42 #include <linux/hash.h>
43 #include <linux/ftrace.h>
44 #include <linux/stringify.h>
45
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 static 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 /*
901  * Recursive, forwards-direction lock-dependency checking, used for
902  * both noncyclic checking and for hardirq-unsafe/softirq-unsafe
903  * checking.
904  *
905  * (to keep the stackframe of the recursive functions small we
906  *  use these global variables, and we also mark various helper
907  *  functions as noinline.)
908  */
909 static struct held_lock *check_source, *check_target;
910
911 /*
912  * Print a dependency chain entry (this is only done when a deadlock
913  * has been detected):
914  */
915 static noinline int
916 print_circular_bug_entry(struct lock_list *target, unsigned int depth)
917 {
918         if (debug_locks_silent)
919                 return 0;
920         printk("\n-> #%u", depth);
921         print_lock_name(target->class);
922         printk(":\n");
923         print_stack_trace(&target->trace, 6);
924
925         return 0;
926 }
927
928 /*
929  * When a circular dependency is detected, print the
930  * header first:
931  */
932 static noinline int
933 print_circular_bug_header(struct lock_list *entry, unsigned int depth)
934 {
935         struct task_struct *curr = current;
936
937         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
938                 return 0;
939
940         printk("\n=======================================================\n");
941         printk(  "[ INFO: possible circular locking dependency detected ]\n");
942         print_kernel_version();
943         printk(  "-------------------------------------------------------\n");
944         printk("%s/%d is trying to acquire lock:\n",
945                 curr->comm, task_pid_nr(curr));
946         print_lock(check_source);
947         printk("\nbut task is already holding lock:\n");
948         print_lock(check_target);
949         printk("\nwhich lock already depends on the new lock.\n\n");
950         printk("\nthe existing dependency chain (in reverse order) is:\n");
951
952         print_circular_bug_entry(entry, depth);
953
954         return 0;
955 }
956
957 static noinline int print_circular_bug_tail(void)
958 {
959         struct task_struct *curr = current;
960         struct lock_list this;
961
962         if (debug_locks_silent)
963                 return 0;
964
965         this.class = hlock_class(check_source);
966         if (!save_trace(&this.trace))
967                 return 0;
968
969         print_circular_bug_entry(&this, 0);
970
971         printk("\nother info that might help us debug this:\n\n");
972         lockdep_print_held_locks(curr);
973
974         printk("\nstack backtrace:\n");
975         dump_stack();
976
977         return 0;
978 }
979
980 #define RECURSION_LIMIT 40
981
982 static int noinline print_infinite_recursion_bug(void)
983 {
984         if (!debug_locks_off_graph_unlock())
985                 return 0;
986
987         WARN_ON(1);
988
989         return 0;
990 }
991
992 unsigned long __lockdep_count_forward_deps(struct lock_class *class,
993                                            unsigned int depth)
994 {
995         struct lock_list *entry;
996         unsigned long ret = 1;
997
998         if (lockdep_dependency_visit(class, depth))
999                 return 0;
1000
1001         /*
1002          * Recurse this class's dependency list:
1003          */
1004         list_for_each_entry(entry, &class->locks_after, entry)
1005                 ret += __lockdep_count_forward_deps(entry->class, depth + 1);
1006
1007         return ret;
1008 }
1009
1010 unsigned long lockdep_count_forward_deps(struct lock_class *class)
1011 {
1012         unsigned long ret, flags;
1013
1014         local_irq_save(flags);
1015         __raw_spin_lock(&lockdep_lock);
1016         ret = __lockdep_count_forward_deps(class, 0);
1017         __raw_spin_unlock(&lockdep_lock);
1018         local_irq_restore(flags);
1019
1020         return ret;
1021 }
1022
1023 unsigned long __lockdep_count_backward_deps(struct lock_class *class,
1024                                             unsigned int depth)
1025 {
1026         struct lock_list *entry;
1027         unsigned long ret = 1;
1028
1029         if (lockdep_dependency_visit(class, depth))
1030                 return 0;
1031         /*
1032          * Recurse this class's dependency list:
1033          */
1034         list_for_each_entry(entry, &class->locks_before, entry)
1035                 ret += __lockdep_count_backward_deps(entry->class, depth + 1);
1036
1037         return ret;
1038 }
1039
1040 unsigned long lockdep_count_backward_deps(struct lock_class *class)
1041 {
1042         unsigned long ret, flags;
1043
1044         local_irq_save(flags);
1045         __raw_spin_lock(&lockdep_lock);
1046         ret = __lockdep_count_backward_deps(class, 0);
1047         __raw_spin_unlock(&lockdep_lock);
1048         local_irq_restore(flags);
1049
1050         return ret;
1051 }
1052
1053 /*
1054  * Prove that the dependency graph starting at <entry> can not
1055  * lead to <target>. Print an error and return 0 if it does.
1056  */
1057 static noinline int
1058 check_noncircular(struct lock_class *source, unsigned int depth)
1059 {
1060         struct lock_list *entry;
1061
1062         if (lockdep_dependency_visit(source, depth))
1063                 return 1;
1064
1065         debug_atomic_inc(&nr_cyclic_check_recursions);
1066         if (depth > max_recursion_depth)
1067                 max_recursion_depth = depth;
1068         if (depth >= RECURSION_LIMIT)
1069                 return print_infinite_recursion_bug();
1070         /*
1071          * Check this lock's dependency list:
1072          */
1073         list_for_each_entry(entry, &source->locks_after, entry) {
1074                 if (entry->class == hlock_class(check_target))
1075                         return print_circular_bug_header(entry, depth+1);
1076                 debug_atomic_inc(&nr_cyclic_checks);
1077                 if (!check_noncircular(entry->class, depth+1))
1078                         return print_circular_bug_entry(entry, depth+1);
1079         }
1080         return 1;
1081 }
1082
1083 #if defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING)
1084 /*
1085  * Forwards and backwards subgraph searching, for the purposes of
1086  * proving that two subgraphs can be connected by a new dependency
1087  * without creating any illegal irq-safe -> irq-unsafe lock dependency.
1088  */
1089 static enum lock_usage_bit find_usage_bit;
1090 static struct lock_class *forwards_match, *backwards_match;
1091
1092 /*
1093  * Find a node in the forwards-direction dependency sub-graph starting
1094  * at <source> that matches <find_usage_bit>.
1095  *
1096  * Return 2 if such a node exists in the subgraph, and put that node
1097  * into <forwards_match>.
1098  *
1099  * Return 1 otherwise and keep <forwards_match> unchanged.
1100  * Return 0 on error.
1101  */
1102 static noinline int
1103 find_usage_forwards(struct lock_class *source, unsigned int depth)
1104 {
1105         struct lock_list *entry;
1106         int ret;
1107
1108         if (lockdep_dependency_visit(source, depth))
1109                 return 1;
1110
1111         if (depth > max_recursion_depth)
1112                 max_recursion_depth = depth;
1113         if (depth >= RECURSION_LIMIT)
1114                 return print_infinite_recursion_bug();
1115
1116         debug_atomic_inc(&nr_find_usage_forwards_checks);
1117         if (source->usage_mask & (1 << find_usage_bit)) {
1118                 forwards_match = source;
1119                 return 2;
1120         }
1121
1122         /*
1123          * Check this lock's dependency list:
1124          */
1125         list_for_each_entry(entry, &source->locks_after, entry) {
1126                 debug_atomic_inc(&nr_find_usage_forwards_recursions);
1127                 ret = find_usage_forwards(entry->class, depth+1);
1128                 if (ret == 2 || ret == 0)
1129                         return ret;
1130         }
1131         return 1;
1132 }
1133
1134 /*
1135  * Find a node in the backwards-direction dependency sub-graph starting
1136  * at <source> that matches <find_usage_bit>.
1137  *
1138  * Return 2 if such a node exists in the subgraph, and put that node
1139  * into <backwards_match>.
1140  *
1141  * Return 1 otherwise and keep <backwards_match> unchanged.
1142  * Return 0 on error.
1143  */
1144 static noinline int
1145 find_usage_backwards(struct lock_class *source, unsigned int depth)
1146 {
1147         struct lock_list *entry;
1148         int ret;
1149
1150         if (lockdep_dependency_visit(source, depth))
1151                 return 1;
1152
1153         if (!__raw_spin_is_locked(&lockdep_lock))
1154                 return DEBUG_LOCKS_WARN_ON(1);
1155
1156         if (depth > max_recursion_depth)
1157                 max_recursion_depth = depth;
1158         if (depth >= RECURSION_LIMIT)
1159                 return print_infinite_recursion_bug();
1160
1161         debug_atomic_inc(&nr_find_usage_backwards_checks);
1162         if (source->usage_mask & (1 << find_usage_bit)) {
1163                 backwards_match = source;
1164                 return 2;
1165         }
1166
1167         if (!source && debug_locks_off_graph_unlock()) {
1168                 WARN_ON(1);
1169                 return 0;
1170         }
1171
1172         /*
1173          * Check this lock's dependency list:
1174          */
1175         list_for_each_entry(entry, &source->locks_before, entry) {
1176                 debug_atomic_inc(&nr_find_usage_backwards_recursions);
1177                 ret = find_usage_backwards(entry->class, depth+1);
1178                 if (ret == 2 || ret == 0)
1179                         return ret;
1180         }
1181         return 1;
1182 }
1183
1184 static int
1185 print_bad_irq_dependency(struct task_struct *curr,
1186                          struct held_lock *prev,
1187                          struct held_lock *next,
1188                          enum lock_usage_bit bit1,
1189                          enum lock_usage_bit bit2,
1190                          const char *irqclass)
1191 {
1192         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1193                 return 0;
1194
1195         printk("\n======================================================\n");
1196         printk(  "[ INFO: %s-safe -> %s-unsafe lock order detected ]\n",
1197                 irqclass, irqclass);
1198         print_kernel_version();
1199         printk(  "------------------------------------------------------\n");
1200         printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] is trying to acquire:\n",
1201                 curr->comm, task_pid_nr(curr),
1202                 curr->hardirq_context, hardirq_count() >> HARDIRQ_SHIFT,
1203                 curr->softirq_context, softirq_count() >> SOFTIRQ_SHIFT,
1204                 curr->hardirqs_enabled,
1205                 curr->softirqs_enabled);
1206         print_lock(next);
1207
1208         printk("\nand this task is already holding:\n");
1209         print_lock(prev);
1210         printk("which would create a new lock dependency:\n");
1211         print_lock_name(hlock_class(prev));
1212         printk(" ->");
1213         print_lock_name(hlock_class(next));
1214         printk("\n");
1215
1216         printk("\nbut this new dependency connects a %s-irq-safe lock:\n",
1217                 irqclass);
1218         print_lock_name(backwards_match);
1219         printk("\n... which became %s-irq-safe at:\n", irqclass);
1220
1221         print_stack_trace(backwards_match->usage_traces + bit1, 1);
1222
1223         printk("\nto a %s-irq-unsafe lock:\n", irqclass);
1224         print_lock_name(forwards_match);
1225         printk("\n... which became %s-irq-unsafe at:\n", irqclass);
1226         printk("...");
1227
1228         print_stack_trace(forwards_match->usage_traces + bit2, 1);
1229
1230         printk("\nother info that might help us debug this:\n\n");
1231         lockdep_print_held_locks(curr);
1232
1233         printk("\nthe %s-irq-safe lock's dependencies:\n", irqclass);
1234         print_lock_dependencies(backwards_match, 0);
1235
1236         printk("\nthe %s-irq-unsafe lock's dependencies:\n", irqclass);
1237         print_lock_dependencies(forwards_match, 0);
1238
1239         printk("\nstack backtrace:\n");
1240         dump_stack();
1241
1242         return 0;
1243 }
1244
1245 static int
1246 check_usage(struct task_struct *curr, struct held_lock *prev,
1247             struct held_lock *next, enum lock_usage_bit bit_backwards,
1248             enum lock_usage_bit bit_forwards, const char *irqclass)
1249 {
1250         int ret;
1251
1252         find_usage_bit = bit_backwards;
1253         /* fills in <backwards_match> */
1254         ret = find_usage_backwards(hlock_class(prev), 0);
1255         if (!ret || ret == 1)
1256                 return ret;
1257
1258         find_usage_bit = bit_forwards;
1259         ret = find_usage_forwards(hlock_class(next), 0);
1260         if (!ret || ret == 1)
1261                 return ret;
1262         /* ret == 2 */
1263         return print_bad_irq_dependency(curr, prev, next,
1264                         bit_backwards, bit_forwards, irqclass);
1265 }
1266
1267 static const char *state_names[] = {
1268 #define LOCKDEP_STATE(__STATE) \
1269         __stringify(__STATE),
1270 #include "lockdep_states.h"
1271 #undef LOCKDEP_STATE
1272 };
1273
1274 static const char *state_rnames[] = {
1275 #define LOCKDEP_STATE(__STATE) \
1276         __stringify(__STATE)"-READ",
1277 #include "lockdep_states.h"
1278 #undef LOCKDEP_STATE
1279 };
1280
1281 static inline const char *state_name(enum lock_usage_bit bit)
1282 {
1283         return (bit & 1) ? state_rnames[bit >> 2] : state_names[bit >> 2];
1284 }
1285
1286 static int exclusive_bit(int new_bit)
1287 {
1288         /*
1289          * USED_IN
1290          * USED_IN_READ
1291          * ENABLED
1292          * ENABLED_READ
1293          *
1294          * bit 0 - write/read
1295          * bit 1 - used_in/enabled
1296          * bit 2+  state
1297          */
1298
1299         int state = new_bit & ~3;
1300         int dir = new_bit & 2;
1301
1302         /*
1303          * keep state, bit flip the direction and strip read.
1304          */
1305         return state | (dir ^ 2);
1306 }
1307
1308 static int check_irq_usage(struct task_struct *curr, struct held_lock *prev,
1309                            struct held_lock *next, enum lock_usage_bit bit)
1310 {
1311         /*
1312          * Prove that the new dependency does not connect a hardirq-safe
1313          * lock with a hardirq-unsafe lock - to achieve this we search
1314          * the backwards-subgraph starting at <prev>, and the
1315          * forwards-subgraph starting at <next>:
1316          */
1317         if (!check_usage(curr, prev, next, bit,
1318                            exclusive_bit(bit), state_name(bit)))
1319                 return 0;
1320
1321         bit++; /* _READ */
1322
1323         /*
1324          * Prove that the new dependency does not connect a hardirq-safe-read
1325          * lock with a hardirq-unsafe lock - to achieve this we search
1326          * the backwards-subgraph starting at <prev>, and the
1327          * forwards-subgraph starting at <next>:
1328          */
1329         if (!check_usage(curr, prev, next, bit,
1330                            exclusive_bit(bit), state_name(bit)))
1331                 return 0;
1332
1333         return 1;
1334 }
1335
1336 static int
1337 check_prev_add_irq(struct task_struct *curr, struct held_lock *prev,
1338                 struct held_lock *next)
1339 {
1340 #define LOCKDEP_STATE(__STATE)                                          \
1341         if (!check_irq_usage(curr, prev, next, LOCK_USED_IN_##__STATE)) \
1342                 return 0;
1343 #include "lockdep_states.h"
1344 #undef LOCKDEP_STATE
1345
1346         return 1;
1347 }
1348
1349 static void inc_chains(void)
1350 {
1351         if (current->hardirq_context)
1352                 nr_hardirq_chains++;
1353         else {
1354                 if (current->softirq_context)
1355                         nr_softirq_chains++;
1356                 else
1357                         nr_process_chains++;
1358         }
1359 }
1360
1361 #else
1362
1363 static inline int
1364 check_prev_add_irq(struct task_struct *curr, struct held_lock *prev,
1365                 struct held_lock *next)
1366 {
1367         return 1;
1368 }
1369
1370 static inline void inc_chains(void)
1371 {
1372         nr_process_chains++;
1373 }
1374
1375 #endif
1376
1377 static int
1378 print_deadlock_bug(struct task_struct *curr, struct held_lock *prev,
1379                    struct held_lock *next)
1380 {
1381         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1382                 return 0;
1383
1384         printk("\n=============================================\n");
1385         printk(  "[ INFO: possible recursive locking detected ]\n");
1386         print_kernel_version();
1387         printk(  "---------------------------------------------\n");
1388         printk("%s/%d is trying to acquire lock:\n",
1389                 curr->comm, task_pid_nr(curr));
1390         print_lock(next);
1391         printk("\nbut task is already holding lock:\n");
1392         print_lock(prev);
1393
1394         printk("\nother info that might help us debug this:\n");
1395         lockdep_print_held_locks(curr);
1396
1397         printk("\nstack backtrace:\n");
1398         dump_stack();
1399
1400         return 0;
1401 }
1402
1403 /*
1404  * Check whether we are holding such a class already.
1405  *
1406  * (Note that this has to be done separately, because the graph cannot
1407  * detect such classes of deadlocks.)
1408  *
1409  * Returns: 0 on deadlock detected, 1 on OK, 2 on recursive read
1410  */
1411 static int
1412 check_deadlock(struct task_struct *curr, struct held_lock *next,
1413                struct lockdep_map *next_instance, int read)
1414 {
1415         struct held_lock *prev;
1416         struct held_lock *nest = NULL;
1417         int i;
1418
1419         for (i = 0; i < curr->lockdep_depth; i++) {
1420                 prev = curr->held_locks + i;
1421
1422                 if (prev->instance == next->nest_lock)
1423                         nest = prev;
1424
1425                 if (hlock_class(prev) != hlock_class(next))
1426                         continue;
1427
1428                 /*
1429                  * Allow read-after-read recursion of the same
1430                  * lock class (i.e. read_lock(lock)+read_lock(lock)):
1431                  */
1432                 if ((read == 2) && prev->read)
1433                         return 2;
1434
1435                 /*
1436                  * We're holding the nest_lock, which serializes this lock's
1437                  * nesting behaviour.
1438                  */
1439                 if (nest)
1440                         return 2;
1441
1442                 return print_deadlock_bug(curr, prev, next);
1443         }
1444         return 1;
1445 }
1446
1447 /*
1448  * There was a chain-cache miss, and we are about to add a new dependency
1449  * to a previous lock. We recursively validate the following rules:
1450  *
1451  *  - would the adding of the <prev> -> <next> dependency create a
1452  *    circular dependency in the graph? [== circular deadlock]
1453  *
1454  *  - does the new prev->next dependency connect any hardirq-safe lock
1455  *    (in the full backwards-subgraph starting at <prev>) with any
1456  *    hardirq-unsafe lock (in the full forwards-subgraph starting at
1457  *    <next>)? [== illegal lock inversion with hardirq contexts]
1458  *
1459  *  - does the new prev->next dependency connect any softirq-safe lock
1460  *    (in the full backwards-subgraph starting at <prev>) with any
1461  *    softirq-unsafe lock (in the full forwards-subgraph starting at
1462  *    <next>)? [== illegal lock inversion with softirq contexts]
1463  *
1464  * any of these scenarios could lead to a deadlock.
1465  *
1466  * Then if all the validations pass, we add the forwards and backwards
1467  * dependency.
1468  */
1469 static int
1470 check_prev_add(struct task_struct *curr, struct held_lock *prev,
1471                struct held_lock *next, int distance)
1472 {
1473         struct lock_list *entry;
1474         int ret;
1475
1476         /*
1477          * Prove that the new <prev> -> <next> dependency would not
1478          * create a circular dependency in the graph. (We do this by
1479          * forward-recursing into the graph starting at <next>, and
1480          * checking whether we can reach <prev>.)
1481          *
1482          * We are using global variables to control the recursion, to
1483          * keep the stackframe size of the recursive functions low:
1484          */
1485         check_source = next;
1486         check_target = prev;
1487         if (!(check_noncircular(hlock_class(next), 0)))
1488                 return print_circular_bug_tail();
1489
1490         if (!check_prev_add_irq(curr, prev, next))
1491                 return 0;
1492
1493         /*
1494          * For recursive read-locks we do all the dependency checks,
1495          * but we dont store read-triggered dependencies (only
1496          * write-triggered dependencies). This ensures that only the
1497          * write-side dependencies matter, and that if for example a
1498          * write-lock never takes any other locks, then the reads are
1499          * equivalent to a NOP.
1500          */
1501         if (next->read == 2 || prev->read == 2)
1502                 return 1;
1503         /*
1504          * Is the <prev> -> <next> dependency already present?
1505          *
1506          * (this may occur even though this is a new chain: consider
1507          *  e.g. the L1 -> L2 -> L3 -> L4 and the L5 -> L1 -> L2 -> L3
1508          *  chains - the second one will be new, but L1 already has
1509          *  L2 added to its dependency list, due to the first chain.)
1510          */
1511         list_for_each_entry(entry, &hlock_class(prev)->locks_after, entry) {
1512                 if (entry->class == hlock_class(next)) {
1513                         if (distance == 1)
1514                                 entry->distance = 1;
1515                         return 2;
1516                 }
1517         }
1518
1519         /*
1520          * Ok, all validations passed, add the new lock
1521          * to the previous lock's dependency list:
1522          */
1523         ret = add_lock_to_list(hlock_class(prev), hlock_class(next),
1524                                &hlock_class(prev)->locks_after,
1525                                next->acquire_ip, distance);
1526
1527         if (!ret)
1528                 return 0;
1529
1530         ret = add_lock_to_list(hlock_class(next), hlock_class(prev),
1531                                &hlock_class(next)->locks_before,
1532                                next->acquire_ip, distance);
1533         if (!ret)
1534                 return 0;
1535
1536         /*
1537          * Debugging printouts:
1538          */
1539         if (verbose(hlock_class(prev)) || verbose(hlock_class(next))) {
1540                 graph_unlock();
1541                 printk("\n new dependency: ");
1542                 print_lock_name(hlock_class(prev));
1543                 printk(" => ");
1544                 print_lock_name(hlock_class(next));
1545                 printk("\n");
1546                 dump_stack();
1547                 return graph_lock();
1548         }
1549         return 1;
1550 }
1551
1552 /*
1553  * Add the dependency to all directly-previous locks that are 'relevant'.
1554  * The ones that are relevant are (in increasing distance from curr):
1555  * all consecutive trylock entries and the final non-trylock entry - or
1556  * the end of this context's lock-chain - whichever comes first.
1557  */
1558 static int
1559 check_prevs_add(struct task_struct *curr, struct held_lock *next)
1560 {
1561         int depth = curr->lockdep_depth;
1562         struct held_lock *hlock;
1563
1564         /*
1565          * Debugging checks.
1566          *
1567          * Depth must not be zero for a non-head lock:
1568          */
1569         if (!depth)
1570                 goto out_bug;
1571         /*
1572          * At least two relevant locks must exist for this
1573          * to be a head:
1574          */
1575         if (curr->held_locks[depth].irq_context !=
1576                         curr->held_locks[depth-1].irq_context)
1577                 goto out_bug;
1578
1579         for (;;) {
1580                 int distance = curr->lockdep_depth - depth + 1;
1581                 hlock = curr->held_locks + depth-1;
1582                 /*
1583                  * Only non-recursive-read entries get new dependencies
1584                  * added:
1585                  */
1586                 if (hlock->read != 2) {
1587                         if (!check_prev_add(curr, hlock, next, distance))
1588                                 return 0;
1589                         /*
1590                          * Stop after the first non-trylock entry,
1591                          * as non-trylock entries have added their
1592                          * own direct dependencies already, so this
1593                          * lock is connected to them indirectly:
1594                          */
1595                         if (!hlock->trylock)
1596                                 break;
1597                 }
1598                 depth--;
1599                 /*
1600                  * End of lock-stack?
1601                  */
1602                 if (!depth)
1603                         break;
1604                 /*
1605                  * Stop the search if we cross into another context:
1606                  */
1607                 if (curr->held_locks[depth].irq_context !=
1608                                 curr->held_locks[depth-1].irq_context)
1609                         break;
1610         }
1611         return 1;
1612 out_bug:
1613         if (!debug_locks_off_graph_unlock())
1614                 return 0;
1615
1616         WARN_ON(1);
1617
1618         return 0;
1619 }
1620
1621 unsigned long nr_lock_chains;
1622 struct lock_chain lock_chains[MAX_LOCKDEP_CHAINS];
1623 int nr_chain_hlocks;
1624 static u16 chain_hlocks[MAX_LOCKDEP_CHAIN_HLOCKS];
1625
1626 struct lock_class *lock_chain_get_class(struct lock_chain *chain, int i)
1627 {
1628         return lock_classes + chain_hlocks[chain->base + i];
1629 }
1630
1631 /*
1632  * Look up a dependency chain. If the key is not present yet then
1633  * add it and return 1 - in this case the new dependency chain is
1634  * validated. If the key is already hashed, return 0.
1635  * (On return with 1 graph_lock is held.)
1636  */
1637 static inline int lookup_chain_cache(struct task_struct *curr,
1638                                      struct held_lock *hlock,
1639                                      u64 chain_key)
1640 {
1641         struct lock_class *class = hlock_class(hlock);
1642         struct list_head *hash_head = chainhashentry(chain_key);
1643         struct lock_chain *chain;
1644         struct held_lock *hlock_curr, *hlock_next;
1645         int i, j, n, cn;
1646
1647         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
1648                 return 0;
1649         /*
1650          * We can walk it lock-free, because entries only get added
1651          * to the hash:
1652          */
1653         list_for_each_entry(chain, hash_head, entry) {
1654                 if (chain->chain_key == chain_key) {
1655 cache_hit:
1656                         debug_atomic_inc(&chain_lookup_hits);
1657                         if (very_verbose(class))
1658                                 printk("\nhash chain already cached, key: "
1659                                         "%016Lx tail class: [%p] %s\n",
1660                                         (unsigned long long)chain_key,
1661                                         class->key, class->name);
1662                         return 0;
1663                 }
1664         }
1665         if (very_verbose(class))
1666                 printk("\nnew hash chain, key: %016Lx tail class: [%p] %s\n",
1667                         (unsigned long long)chain_key, class->key, class->name);
1668         /*
1669          * Allocate a new chain entry from the static array, and add
1670          * it to the hash:
1671          */
1672         if (!graph_lock())
1673                 return 0;
1674         /*
1675          * We have to walk the chain again locked - to avoid duplicates:
1676          */
1677         list_for_each_entry(chain, hash_head, entry) {
1678                 if (chain->chain_key == chain_key) {
1679                         graph_unlock();
1680                         goto cache_hit;
1681                 }
1682         }
1683         if (unlikely(nr_lock_chains >= MAX_LOCKDEP_CHAINS)) {
1684                 if (!debug_locks_off_graph_unlock())
1685                         return 0;
1686
1687                 printk("BUG: MAX_LOCKDEP_CHAINS too low!\n");
1688                 printk("turning off the locking correctness validator.\n");
1689                 dump_stack();
1690                 return 0;
1691         }
1692         chain = lock_chains + nr_lock_chains++;
1693         chain->chain_key = chain_key;
1694         chain->irq_context = hlock->irq_context;
1695         /* Find the first held_lock of current chain */
1696         hlock_next = hlock;
1697         for (i = curr->lockdep_depth - 1; i >= 0; i--) {
1698                 hlock_curr = curr->held_locks + i;
1699                 if (hlock_curr->irq_context != hlock_next->irq_context)
1700                         break;
1701                 hlock_next = hlock;
1702         }
1703         i++;
1704         chain->depth = curr->lockdep_depth + 1 - i;
1705         cn = nr_chain_hlocks;
1706         while (cn + chain->depth <= MAX_LOCKDEP_CHAIN_HLOCKS) {
1707                 n = cmpxchg(&nr_chain_hlocks, cn, cn + chain->depth);
1708                 if (n == cn)
1709                         break;
1710                 cn = n;
1711         }
1712         if (likely(cn + chain->depth <= MAX_LOCKDEP_CHAIN_HLOCKS)) {
1713                 chain->base = cn;
1714                 for (j = 0; j < chain->depth - 1; j++, i++) {
1715                         int lock_id = curr->held_locks[i].class_idx - 1;
1716                         chain_hlocks[chain->base + j] = lock_id;
1717                 }
1718                 chain_hlocks[chain->base + j] = class - lock_classes;
1719         }
1720         list_add_tail_rcu(&chain->entry, hash_head);
1721         debug_atomic_inc(&chain_lookup_misses);
1722         inc_chains();
1723
1724         return 1;
1725 }
1726
1727 static int validate_chain(struct task_struct *curr, struct lockdep_map *lock,
1728                 struct held_lock *hlock, int chain_head, u64 chain_key)
1729 {
1730         /*
1731          * Trylock needs to maintain the stack of held locks, but it
1732          * does not add new dependencies, because trylock can be done
1733          * in any order.
1734          *
1735          * We look up the chain_key and do the O(N^2) check and update of
1736          * the dependencies only if this is a new dependency chain.
1737          * (If lookup_chain_cache() returns with 1 it acquires
1738          * graph_lock for us)
1739          */
1740         if (!hlock->trylock && (hlock->check == 2) &&
1741             lookup_chain_cache(curr, hlock, chain_key)) {
1742                 /*
1743                  * Check whether last held lock:
1744                  *
1745                  * - is irq-safe, if this lock is irq-unsafe
1746                  * - is softirq-safe, if this lock is hardirq-unsafe
1747                  *
1748                  * And check whether the new lock's dependency graph
1749                  * could lead back to the previous lock.
1750                  *
1751                  * any of these scenarios could lead to a deadlock. If
1752                  * All validations
1753                  */
1754                 int ret = check_deadlock(curr, hlock, lock, hlock->read);
1755
1756                 if (!ret)
1757                         return 0;
1758                 /*
1759                  * Mark recursive read, as we jump over it when
1760                  * building dependencies (just like we jump over
1761                  * trylock entries):
1762                  */
1763                 if (ret == 2)
1764                         hlock->read = 2;
1765                 /*
1766                  * Add dependency only if this lock is not the head
1767                  * of the chain, and if it's not a secondary read-lock:
1768                  */
1769                 if (!chain_head && ret != 2)
1770                         if (!check_prevs_add(curr, hlock))
1771                                 return 0;
1772                 graph_unlock();
1773         } else
1774                 /* after lookup_chain_cache(): */
1775                 if (unlikely(!debug_locks))
1776                         return 0;
1777
1778         return 1;
1779 }
1780 #else
1781 static inline int validate_chain(struct task_struct *curr,
1782                 struct lockdep_map *lock, struct held_lock *hlock,
1783                 int chain_head, u64 chain_key)
1784 {
1785         return 1;
1786 }
1787 #endif
1788
1789 /*
1790  * We are building curr_chain_key incrementally, so double-check
1791  * it from scratch, to make sure that it's done correctly:
1792  */
1793 static void check_chain_key(struct task_struct *curr)
1794 {
1795 #ifdef CONFIG_DEBUG_LOCKDEP
1796         struct held_lock *hlock, *prev_hlock = NULL;
1797         unsigned int i, id;
1798         u64 chain_key = 0;
1799
1800         for (i = 0; i < curr->lockdep_depth; i++) {
1801                 hlock = curr->held_locks + i;
1802                 if (chain_key != hlock->prev_chain_key) {
1803                         debug_locks_off();
1804                         WARN(1, "hm#1, depth: %u [%u], %016Lx != %016Lx\n",
1805                                 curr->lockdep_depth, i,
1806                                 (unsigned long long)chain_key,
1807                                 (unsigned long long)hlock->prev_chain_key);
1808                         return;
1809                 }
1810                 id = hlock->class_idx - 1;
1811                 if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
1812                         return;
1813
1814                 if (prev_hlock && (prev_hlock->irq_context !=
1815                                                         hlock->irq_context))
1816                         chain_key = 0;
1817                 chain_key = iterate_chain_key(chain_key, id);
1818                 prev_hlock = hlock;
1819         }
1820         if (chain_key != curr->curr_chain_key) {
1821                 debug_locks_off();
1822                 WARN(1, "hm#2, depth: %u [%u], %016Lx != %016Lx\n",
1823                         curr->lockdep_depth, i,
1824                         (unsigned long long)chain_key,
1825                         (unsigned long long)curr->curr_chain_key);
1826         }
1827 #endif
1828 }
1829
1830 static int
1831 print_usage_bug(struct task_struct *curr, struct held_lock *this,
1832                 enum lock_usage_bit prev_bit, enum lock_usage_bit new_bit)
1833 {
1834         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1835                 return 0;
1836
1837         printk("\n=================================\n");
1838         printk(  "[ INFO: inconsistent lock state ]\n");
1839         print_kernel_version();
1840         printk(  "---------------------------------\n");
1841
1842         printk("inconsistent {%s} -> {%s} usage.\n",
1843                 usage_str[prev_bit], usage_str[new_bit]);
1844
1845         printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] takes:\n",
1846                 curr->comm, task_pid_nr(curr),
1847                 trace_hardirq_context(curr), hardirq_count() >> HARDIRQ_SHIFT,
1848                 trace_softirq_context(curr), softirq_count() >> SOFTIRQ_SHIFT,
1849                 trace_hardirqs_enabled(curr),
1850                 trace_softirqs_enabled(curr));
1851         print_lock(this);
1852
1853         printk("{%s} state was registered at:\n", usage_str[prev_bit]);
1854         print_stack_trace(hlock_class(this)->usage_traces + prev_bit, 1);
1855
1856         print_irqtrace_events(curr);
1857         printk("\nother info that might help us debug this:\n");
1858         lockdep_print_held_locks(curr);
1859
1860         printk("\nstack backtrace:\n");
1861         dump_stack();
1862
1863         return 0;
1864 }
1865
1866 /*
1867  * Print out an error if an invalid bit is set:
1868  */
1869 static inline int
1870 valid_state(struct task_struct *curr, struct held_lock *this,
1871             enum lock_usage_bit new_bit, enum lock_usage_bit bad_bit)
1872 {
1873         if (unlikely(hlock_class(this)->usage_mask & (1 << bad_bit)))
1874                 return print_usage_bug(curr, this, bad_bit, new_bit);
1875         return 1;
1876 }
1877
1878 static int mark_lock(struct task_struct *curr, struct held_lock *this,
1879                      enum lock_usage_bit new_bit);
1880
1881 #if defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING)
1882
1883 /*
1884  * print irq inversion bug:
1885  */
1886 static int
1887 print_irq_inversion_bug(struct task_struct *curr, struct lock_class *other,
1888                         struct held_lock *this, int forwards,
1889                         const char *irqclass)
1890 {
1891         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1892                 return 0;
1893
1894         printk("\n=========================================================\n");
1895         printk(  "[ INFO: possible irq lock inversion dependency detected ]\n");
1896         print_kernel_version();
1897         printk(  "---------------------------------------------------------\n");
1898         printk("%s/%d just changed the state of lock:\n",
1899                 curr->comm, task_pid_nr(curr));
1900         print_lock(this);
1901         if (forwards)
1902                 printk("but this lock took another, %s-unsafe lock in the past:\n", irqclass);
1903         else
1904                 printk("but this lock was taken by another, %s-safe lock in the past:\n", irqclass);
1905         print_lock_name(other);
1906         printk("\n\nand interrupts could create inverse lock ordering between them.\n\n");
1907
1908         printk("\nother info that might help us debug this:\n");
1909         lockdep_print_held_locks(curr);
1910
1911         printk("\nthe first lock's dependencies:\n");
1912         print_lock_dependencies(hlock_class(this), 0);
1913
1914         printk("\nthe second lock's dependencies:\n");
1915         print_lock_dependencies(other, 0);
1916
1917         printk("\nstack backtrace:\n");
1918         dump_stack();
1919
1920         return 0;
1921 }
1922
1923 /*
1924  * Prove that in the forwards-direction subgraph starting at <this>
1925  * there is no lock matching <mask>:
1926  */
1927 static int
1928 check_usage_forwards(struct task_struct *curr, struct held_lock *this,
1929                      enum lock_usage_bit bit, const char *irqclass)
1930 {
1931         int ret;
1932
1933         find_usage_bit = bit;
1934         /* fills in <forwards_match> */
1935         ret = find_usage_forwards(hlock_class(this), 0);
1936         if (!ret || ret == 1)
1937                 return ret;
1938
1939         return print_irq_inversion_bug(curr, forwards_match, this, 1, irqclass);
1940 }
1941
1942 /*
1943  * Prove that in the backwards-direction subgraph starting at <this>
1944  * there is no lock matching <mask>:
1945  */
1946 static int
1947 check_usage_backwards(struct task_struct *curr, struct held_lock *this,
1948                       enum lock_usage_bit bit, const char *irqclass)
1949 {
1950         int ret;
1951
1952         find_usage_bit = bit;
1953         /* fills in <backwards_match> */
1954         ret = find_usage_backwards(hlock_class(this), 0);
1955         if (!ret || ret == 1)
1956                 return ret;
1957
1958         return print_irq_inversion_bug(curr, backwards_match, this, 0, irqclass);
1959 }
1960
1961 void print_irqtrace_events(struct task_struct *curr)
1962 {
1963         printk("irq event stamp: %u\n", curr->irq_events);
1964         printk("hardirqs last  enabled at (%u): ", curr->hardirq_enable_event);
1965         print_ip_sym(curr->hardirq_enable_ip);
1966         printk("hardirqs last disabled at (%u): ", curr->hardirq_disable_event);
1967         print_ip_sym(curr->hardirq_disable_ip);
1968         printk("softirqs last  enabled at (%u): ", curr->softirq_enable_event);
1969         print_ip_sym(curr->softirq_enable_ip);
1970         printk("softirqs last disabled at (%u): ", curr->softirq_disable_event);
1971         print_ip_sym(curr->softirq_disable_ip);
1972 }
1973
1974 static int HARDIRQ_verbose(struct lock_class *class)
1975 {
1976 #if HARDIRQ_VERBOSE
1977         return class_filter(class);
1978 #endif
1979         return 0;
1980 }
1981
1982 static int SOFTIRQ_verbose(struct lock_class *class)
1983 {
1984 #if SOFTIRQ_VERBOSE
1985         return class_filter(class);
1986 #endif
1987         return 0;
1988 }
1989
1990 static int RECLAIM_FS_verbose(struct lock_class *class)
1991 {
1992 #if RECLAIM_VERBOSE
1993         return class_filter(class);
1994 #endif
1995         return 0;
1996 }
1997
1998 #define STRICT_READ_CHECKS      1
1999
2000 static int (*state_verbose_f[])(struct lock_class *class) = {
2001 #define LOCKDEP_STATE(__STATE) \
2002         __STATE##_verbose,
2003 #include "lockdep_states.h"
2004 #undef LOCKDEP_STATE
2005 };
2006
2007 static inline int state_verbose(enum lock_usage_bit bit,
2008                                 struct lock_class *class)
2009 {
2010         return state_verbose_f[bit >> 2](class);
2011 }
2012
2013 typedef int (*check_usage_f)(struct task_struct *, struct held_lock *,
2014                              enum lock_usage_bit bit, const char *name);
2015
2016 static int
2017 mark_lock_irq(struct task_struct *curr, struct held_lock *this,
2018                 enum lock_usage_bit new_bit)
2019 {
2020         int excl_bit = exclusive_bit(new_bit);
2021         int read = new_bit & 1;
2022         int dir = new_bit & 2;
2023
2024         /*
2025          * mark USED_IN has to look forwards -- to ensure no dependency
2026          * has ENABLED state, which would allow recursion deadlocks.
2027          *
2028          * mark ENABLED has to look backwards -- to ensure no dependee
2029          * has USED_IN state, which, again, would allow  recursion deadlocks.
2030          */
2031         check_usage_f usage = dir ?
2032                 check_usage_backwards : check_usage_forwards;
2033
2034         /*
2035          * Validate that this particular lock does not have conflicting
2036          * usage states.
2037          */
2038         if (!valid_state(curr, this, new_bit, excl_bit))
2039                 return 0;
2040
2041         /*
2042          * Validate that the lock dependencies don't have conflicting usage
2043          * states.
2044          */
2045         if ((!read || !dir || STRICT_READ_CHECKS) &&
2046                         !usage(curr, this, excl_bit, state_name(new_bit & ~1)))
2047                 return 0;
2048
2049         /*
2050          * Check for read in write conflicts
2051          */
2052         if (!read) {
2053                 if (!valid_state(curr, this, new_bit, excl_bit + 1))
2054                         return 0;
2055
2056                 if (STRICT_READ_CHECKS &&
2057                         !usage(curr, this, excl_bit + 1,
2058                                 state_name(new_bit + 1)))
2059                         return 0;
2060         }
2061
2062         if (state_verbose(new_bit, hlock_class(this)))
2063                 return 2;
2064
2065         return 1;
2066 }
2067
2068 enum mark_type {
2069 #define LOCKDEP_STATE(__STATE)  __STATE,
2070 #include "lockdep_states.h"
2071 #undef LOCKDEP_STATE
2072 };
2073
2074 /*
2075  * Mark all held locks with a usage bit:
2076  */
2077 static int
2078 mark_held_locks(struct task_struct *curr, enum mark_type mark)
2079 {
2080         enum lock_usage_bit usage_bit;
2081         struct held_lock *hlock;
2082         int i;
2083
2084         for (i = 0; i < curr->lockdep_depth; i++) {
2085                 hlock = curr->held_locks + i;
2086
2087                 usage_bit = 2 + (mark << 2); /* ENABLED */
2088                 if (hlock->read)
2089                         usage_bit += 1; /* READ */
2090
2091                 BUG_ON(usage_bit >= LOCK_USAGE_STATES);
2092
2093                 if (!mark_lock(curr, hlock, usage_bit))
2094                         return 0;
2095         }
2096
2097         return 1;
2098 }
2099
2100 /*
2101  * Debugging helper: via this flag we know that we are in
2102  * 'early bootup code', and will warn about any invalid irqs-on event:
2103  */
2104 static int early_boot_irqs_enabled;
2105
2106 void early_boot_irqs_off(void)
2107 {
2108         early_boot_irqs_enabled = 0;
2109 }
2110
2111 void early_boot_irqs_on(void)
2112 {
2113         early_boot_irqs_enabled = 1;
2114 }
2115
2116 /*
2117  * Hardirqs will be enabled:
2118  */
2119 void trace_hardirqs_on_caller(unsigned long ip)
2120 {
2121         struct task_struct *curr = current;
2122
2123         time_hardirqs_on(CALLER_ADDR0, ip);
2124
2125         if (unlikely(!debug_locks || current->lockdep_recursion))
2126                 return;
2127
2128         if (DEBUG_LOCKS_WARN_ON(unlikely(!early_boot_irqs_enabled)))
2129                 return;
2130
2131         if (unlikely(curr->hardirqs_enabled)) {
2132                 debug_atomic_inc(&redundant_hardirqs_on);
2133                 return;
2134         }
2135         /* we'll do an OFF -> ON transition: */
2136         curr->hardirqs_enabled = 1;
2137
2138         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2139                 return;
2140         if (DEBUG_LOCKS_WARN_ON(current->hardirq_context))
2141                 return;
2142         /*
2143          * We are going to turn hardirqs on, so set the
2144          * usage bit for all held locks:
2145          */
2146         if (!mark_held_locks(curr, HARDIRQ))
2147                 return;
2148         /*
2149          * If we have softirqs enabled, then set the usage
2150          * bit for all held locks. (disabled hardirqs prevented
2151          * this bit from being set before)
2152          */
2153         if (curr->softirqs_enabled)
2154                 if (!mark_held_locks(curr, SOFTIRQ))
2155                         return;
2156
2157         curr->hardirq_enable_ip = ip;
2158         curr->hardirq_enable_event = ++curr->irq_events;
2159         debug_atomic_inc(&hardirqs_on_events);
2160 }
2161 EXPORT_SYMBOL(trace_hardirqs_on_caller);
2162
2163 void trace_hardirqs_on(void)
2164 {
2165         trace_hardirqs_on_caller(CALLER_ADDR0);
2166 }
2167 EXPORT_SYMBOL(trace_hardirqs_on);
2168
2169 /*
2170  * Hardirqs were disabled:
2171  */
2172 void trace_hardirqs_off_caller(unsigned long ip)
2173 {
2174         struct task_struct *curr = current;
2175
2176         time_hardirqs_off(CALLER_ADDR0, ip);
2177
2178         if (unlikely(!debug_locks || current->lockdep_recursion))
2179                 return;
2180
2181         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2182                 return;
2183
2184         if (curr->hardirqs_enabled) {
2185                 /*
2186                  * We have done an ON -> OFF transition:
2187                  */
2188                 curr->hardirqs_enabled = 0;
2189                 curr->hardirq_disable_ip = ip;
2190                 curr->hardirq_disable_event = ++curr->irq_events;
2191                 debug_atomic_inc(&hardirqs_off_events);
2192         } else
2193                 debug_atomic_inc(&redundant_hardirqs_off);
2194 }
2195 EXPORT_SYMBOL(trace_hardirqs_off_caller);
2196
2197 void trace_hardirqs_off(void)
2198 {
2199         trace_hardirqs_off_caller(CALLER_ADDR0);
2200 }
2201 EXPORT_SYMBOL(trace_hardirqs_off);
2202
2203 /*
2204  * Softirqs will be enabled:
2205  */
2206 void trace_softirqs_on(unsigned long ip)
2207 {
2208         struct task_struct *curr = current;
2209
2210         if (unlikely(!debug_locks))
2211                 return;
2212
2213         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2214                 return;
2215
2216         if (curr->softirqs_enabled) {
2217                 debug_atomic_inc(&redundant_softirqs_on);
2218                 return;
2219         }
2220
2221         /*
2222          * We'll do an OFF -> ON transition:
2223          */
2224         curr->softirqs_enabled = 1;
2225         curr->softirq_enable_ip = ip;
2226         curr->softirq_enable_event = ++curr->irq_events;
2227         debug_atomic_inc(&softirqs_on_events);
2228         /*
2229          * We are going to turn softirqs on, so set the
2230          * usage bit for all held locks, if hardirqs are
2231          * enabled too:
2232          */
2233         if (curr->hardirqs_enabled)
2234                 mark_held_locks(curr, SOFTIRQ);
2235 }
2236
2237 /*
2238  * Softirqs were disabled:
2239  */
2240 void trace_softirqs_off(unsigned long ip)
2241 {
2242         struct task_struct *curr = current;
2243
2244         if (unlikely(!debug_locks))
2245                 return;
2246
2247         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2248                 return;
2249
2250         if (curr->softirqs_enabled) {
2251                 /*
2252                  * We have done an ON -> OFF transition:
2253                  */
2254                 curr->softirqs_enabled = 0;
2255                 curr->softirq_disable_ip = ip;
2256                 curr->softirq_disable_event = ++curr->irq_events;
2257                 debug_atomic_inc(&softirqs_off_events);
2258                 DEBUG_LOCKS_WARN_ON(!softirq_count());
2259         } else
2260                 debug_atomic_inc(&redundant_softirqs_off);
2261 }
2262
2263 static void __lockdep_trace_alloc(gfp_t gfp_mask, unsigned long flags)
2264 {
2265         struct task_struct *curr = current;
2266
2267         if (unlikely(!debug_locks))
2268                 return;
2269
2270         /* no reclaim without waiting on it */
2271         if (!(gfp_mask & __GFP_WAIT))
2272                 return;
2273
2274         /* this guy won't enter reclaim */
2275         if ((curr->flags & PF_MEMALLOC) && !(gfp_mask & __GFP_NOMEMALLOC))
2276                 return;
2277
2278         /* We're only interested __GFP_FS allocations for now */
2279         if (!(gfp_mask & __GFP_FS))
2280                 return;
2281
2282         if (DEBUG_LOCKS_WARN_ON(irqs_disabled_flags(flags)))
2283                 return;
2284
2285         mark_held_locks(curr, RECLAIM_FS);
2286 }
2287
2288 static void check_flags(unsigned long flags);
2289
2290 void lockdep_trace_alloc(gfp_t gfp_mask)
2291 {
2292         unsigned long flags;
2293
2294         if (unlikely(current->lockdep_recursion))
2295                 return;
2296
2297         raw_local_irq_save(flags);
2298         check_flags(flags);
2299         current->lockdep_recursion = 1;
2300         __lockdep_trace_alloc(gfp_mask, flags);
2301         current->lockdep_recursion = 0;
2302         raw_local_irq_restore(flags);
2303 }
2304
2305 static int mark_irqflags(struct task_struct *curr, struct held_lock *hlock)
2306 {
2307         /*
2308          * If non-trylock use in a hardirq or softirq context, then
2309          * mark the lock as used in these contexts:
2310          */
2311         if (!hlock->trylock) {
2312                 if (hlock->read) {
2313                         if (curr->hardirq_context)
2314                                 if (!mark_lock(curr, hlock,
2315                                                 LOCK_USED_IN_HARDIRQ_READ))
2316                                         return 0;
2317                         if (curr->softirq_context)
2318                                 if (!mark_lock(curr, hlock,
2319                                                 LOCK_USED_IN_SOFTIRQ_READ))
2320                                         return 0;
2321                 } else {
2322                         if (curr->hardirq_context)
2323                                 if (!mark_lock(curr, hlock, LOCK_USED_IN_HARDIRQ))
2324                                         return 0;
2325                         if (curr->softirq_context)
2326                                 if (!mark_lock(curr, hlock, LOCK_USED_IN_SOFTIRQ))
2327                                         return 0;
2328                 }
2329         }
2330         if (!hlock->hardirqs_off) {
2331                 if (hlock->read) {
2332                         if (!mark_lock(curr, hlock,
2333                                         LOCK_ENABLED_HARDIRQ_READ))
2334                                 return 0;
2335                         if (curr->softirqs_enabled)
2336                                 if (!mark_lock(curr, hlock,
2337                                                 LOCK_ENABLED_SOFTIRQ_READ))
2338                                         return 0;
2339                 } else {
2340                         if (!mark_lock(curr, hlock,
2341                                         LOCK_ENABLED_HARDIRQ))
2342                                 return 0;
2343                         if (curr->softirqs_enabled)
2344                                 if (!mark_lock(curr, hlock,
2345                                                 LOCK_ENABLED_SOFTIRQ))
2346                                         return 0;
2347                 }
2348         }
2349
2350         /*
2351          * We reuse the irq context infrastructure more broadly as a general
2352          * context checking code. This tests GFP_FS recursion (a lock taken
2353          * during reclaim for a GFP_FS allocation is held over a GFP_FS
2354          * allocation).
2355          */
2356         if (!hlock->trylock && (curr->lockdep_reclaim_gfp & __GFP_FS)) {
2357                 if (hlock->read) {
2358                         if (!mark_lock(curr, hlock, LOCK_USED_IN_RECLAIM_FS_READ))
2359                                         return 0;
2360                 } else {
2361                         if (!mark_lock(curr, hlock, LOCK_USED_IN_RECLAIM_FS))
2362                                         return 0;
2363                 }
2364         }
2365
2366         return 1;
2367 }
2368
2369 static int separate_irq_context(struct task_struct *curr,
2370                 struct held_lock *hlock)
2371 {
2372         unsigned int depth = curr->lockdep_depth;
2373
2374         /*
2375          * Keep track of points where we cross into an interrupt context:
2376          */
2377         hlock->irq_context = 2*(curr->hardirq_context ? 1 : 0) +
2378                                 curr->softirq_context;
2379         if (depth) {
2380                 struct held_lock *prev_hlock;
2381
2382                 prev_hlock = curr->held_locks + depth-1;
2383                 /*
2384                  * If we cross into another context, reset the
2385                  * hash key (this also prevents the checking and the
2386                  * adding of the dependency to 'prev'):
2387                  */
2388                 if (prev_hlock->irq_context != hlock->irq_context)
2389                         return 1;
2390         }
2391         return 0;
2392 }
2393
2394 #else
2395
2396 static inline
2397 int mark_lock_irq(struct task_struct *curr, struct held_lock *this,
2398                 enum lock_usage_bit new_bit)
2399 {
2400         WARN_ON(1);
2401         return 1;
2402 }
2403
2404 static inline int mark_irqflags(struct task_struct *curr,
2405                 struct held_lock *hlock)
2406 {
2407         return 1;
2408 }
2409
2410 static inline int separate_irq_context(struct task_struct *curr,
2411                 struct held_lock *hlock)
2412 {
2413         return 0;
2414 }
2415
2416 void lockdep_trace_alloc(gfp_t gfp_mask)
2417 {
2418 }
2419
2420 #endif
2421
2422 /*
2423  * Mark a lock with a usage bit, and validate the state transition:
2424  */
2425 static int mark_lock(struct task_struct *curr, struct held_lock *this,
2426                              enum lock_usage_bit new_bit)
2427 {
2428         unsigned int new_mask = 1 << new_bit, ret = 1;
2429
2430         /*
2431          * If already set then do not dirty the cacheline,
2432          * nor do any checks:
2433          */
2434         if (likely(hlock_class(this)->usage_mask & new_mask))
2435                 return 1;
2436
2437         if (!graph_lock())
2438                 return 0;
2439         /*
2440          * Make sure we didnt race:
2441          */
2442         if (unlikely(hlock_class(this)->usage_mask & new_mask)) {
2443                 graph_unlock();
2444                 return 1;
2445         }
2446
2447         hlock_class(this)->usage_mask |= new_mask;
2448
2449         if (!save_trace(hlock_class(this)->usage_traces + new_bit))
2450                 return 0;
2451
2452         switch (new_bit) {
2453 #define LOCKDEP_STATE(__STATE)                  \
2454         case LOCK_USED_IN_##__STATE:            \
2455         case LOCK_USED_IN_##__STATE##_READ:     \
2456         case LOCK_ENABLED_##__STATE:            \
2457         case LOCK_ENABLED_##__STATE##_READ:
2458 #include "lockdep_states.h"
2459 #undef LOCKDEP_STATE
2460                 ret = mark_lock_irq(curr, this, new_bit);
2461                 if (!ret)
2462                         return 0;
2463                 break;
2464         case LOCK_USED:
2465                 debug_atomic_dec(&nr_unused_locks);
2466                 break;
2467         default:
2468                 if (!debug_locks_off_graph_unlock())
2469                         return 0;
2470                 WARN_ON(1);
2471                 return 0;
2472         }
2473
2474         graph_unlock();
2475
2476         /*
2477          * We must printk outside of the graph_lock:
2478          */
2479         if (ret == 2) {
2480                 printk("\nmarked lock as {%s}:\n", usage_str[new_bit]);
2481                 print_lock(this);
2482                 print_irqtrace_events(curr);
2483                 dump_stack();
2484         }
2485
2486         return ret;
2487 }
2488
2489 /*
2490  * Initialize a lock instance's lock-class mapping info:
2491  */
2492 void lockdep_init_map(struct lockdep_map *lock, const char *name,
2493                       struct lock_class_key *key, int subclass)
2494 {
2495         if (unlikely(!debug_locks))
2496                 return;
2497
2498         if (DEBUG_LOCKS_WARN_ON(!key))
2499                 return;
2500         if (DEBUG_LOCKS_WARN_ON(!name))
2501                 return;
2502         /*
2503          * Sanity check, the lock-class key must be persistent:
2504          */
2505         if (!static_obj(key)) {
2506                 printk("BUG: key %p not in .data!\n", key);
2507                 DEBUG_LOCKS_WARN_ON(1);
2508                 return;
2509         }
2510         lock->name = name;
2511         lock->key = key;
2512         lock->class_cache = NULL;
2513 #ifdef CONFIG_LOCK_STAT
2514         lock->cpu = raw_smp_processor_id();
2515 #endif
2516         if (subclass)
2517                 register_lock_class(lock, subclass, 1);
2518 }
2519 EXPORT_SYMBOL_GPL(lockdep_init_map);
2520
2521 /*
2522  * This gets called for every mutex_lock*()/spin_lock*() operation.
2523  * We maintain the dependency maps and validate the locking attempt:
2524  */
2525 static int __lock_acquire(struct lockdep_map *lock, unsigned int subclass,
2526                           int trylock, int read, int check, int hardirqs_off,
2527                           struct lockdep_map *nest_lock, unsigned long ip)
2528 {
2529         struct task_struct *curr = current;
2530         struct lock_class *class = NULL;
2531         struct held_lock *hlock;
2532         unsigned int depth, id;
2533         int chain_head = 0;
2534         u64 chain_key;
2535
2536         if (!prove_locking)
2537                 check = 1;
2538
2539         if (unlikely(!debug_locks))
2540                 return 0;
2541
2542         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2543                 return 0;
2544
2545         if (unlikely(subclass >= MAX_LOCKDEP_SUBCLASSES)) {
2546                 debug_locks_off();
2547                 printk("BUG: MAX_LOCKDEP_SUBCLASSES too low!\n");
2548                 printk("turning off the locking correctness validator.\n");
2549                 dump_stack();
2550                 return 0;
2551         }
2552
2553         if (!subclass)
2554                 class = lock->class_cache;
2555         /*
2556          * Not cached yet or subclass?
2557          */
2558         if (unlikely(!class)) {
2559                 class = register_lock_class(lock, subclass, 0);
2560                 if (!class)
2561                         return 0;
2562         }
2563         debug_atomic_inc((atomic_t *)&class->ops);
2564         if (very_verbose(class)) {
2565                 printk("\nacquire class [%p] %s", class->key, class->name);
2566                 if (class->name_version > 1)
2567                         printk("#%d", class->name_version);
2568                 printk("\n");
2569                 dump_stack();
2570         }
2571
2572         /*
2573          * Add the lock to the list of currently held locks.
2574          * (we dont increase the depth just yet, up until the
2575          * dependency checks are done)
2576          */
2577         depth = curr->lockdep_depth;
2578         if (DEBUG_LOCKS_WARN_ON(depth >= MAX_LOCK_DEPTH))
2579                 return 0;
2580
2581         hlock = curr->held_locks + depth;
2582         if (DEBUG_LOCKS_WARN_ON(!class))
2583                 return 0;
2584         hlock->class_idx = class - lock_classes + 1;
2585         hlock->acquire_ip = ip;
2586         hlock->instance = lock;
2587         hlock->nest_lock = nest_lock;
2588         hlock->trylock = trylock;
2589         hlock->read = read;
2590         hlock->check = check;
2591         hlock->hardirqs_off = !!hardirqs_off;
2592 #ifdef CONFIG_LOCK_STAT
2593         hlock->waittime_stamp = 0;
2594         hlock->holdtime_stamp = sched_clock();
2595 #endif
2596
2597         if (check == 2 && !mark_irqflags(curr, hlock))
2598                 return 0;
2599
2600         /* mark it as used: */
2601         if (!mark_lock(curr, hlock, LOCK_USED))
2602                 return 0;
2603
2604         /*
2605          * Calculate the chain hash: it's the combined hash of all the
2606          * lock keys along the dependency chain. We save the hash value
2607          * at every step so that we can get the current hash easily
2608          * after unlock. The chain hash is then used to cache dependency
2609          * results.
2610          *
2611          * The 'key ID' is what is the most compact key value to drive
2612          * the hash, not class->key.
2613          */
2614         id = class - lock_classes;
2615         if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
2616                 return 0;
2617
2618         chain_key = curr->curr_chain_key;
2619         if (!depth) {
2620                 if (DEBUG_LOCKS_WARN_ON(chain_key != 0))
2621                         return 0;
2622                 chain_head = 1;
2623         }
2624
2625         hlock->prev_chain_key = chain_key;
2626         if (separate_irq_context(curr, hlock)) {
2627                 chain_key = 0;
2628                 chain_head = 1;
2629         }
2630         chain_key = iterate_chain_key(chain_key, id);
2631
2632         if (!validate_chain(curr, lock, hlock, chain_head, chain_key))
2633                 return 0;
2634
2635         curr->curr_chain_key = chain_key;
2636         curr->lockdep_depth++;
2637         check_chain_key(curr);
2638 #ifdef CONFIG_DEBUG_LOCKDEP
2639         if (unlikely(!debug_locks))
2640                 return 0;
2641 #endif
2642         if (unlikely(curr->lockdep_depth >= MAX_LOCK_DEPTH)) {
2643                 debug_locks_off();
2644                 printk("BUG: MAX_LOCK_DEPTH too low!\n");
2645                 printk("turning off the locking correctness validator.\n");
2646                 dump_stack();
2647                 return 0;
2648         }
2649
2650         if (unlikely(curr->lockdep_depth > max_lockdep_depth))
2651                 max_lockdep_depth = curr->lockdep_depth;
2652
2653         return 1;
2654 }
2655
2656 static int
2657 print_unlock_inbalance_bug(struct task_struct *curr, struct lockdep_map *lock,
2658                            unsigned long ip)
2659 {
2660         if (!debug_locks_off())
2661                 return 0;
2662         if (debug_locks_silent)
2663                 return 0;
2664
2665         printk("\n=====================================\n");
2666         printk(  "[ BUG: bad unlock balance detected! ]\n");
2667         printk(  "-------------------------------------\n");
2668         printk("%s/%d is trying to release lock (",
2669                 curr->comm, task_pid_nr(curr));
2670         print_lockdep_cache(lock);
2671         printk(") at:\n");
2672         print_ip_sym(ip);
2673         printk("but there are no more locks to release!\n");
2674         printk("\nother info that might help us debug this:\n");
2675         lockdep_print_held_locks(curr);
2676
2677         printk("\nstack backtrace:\n");
2678         dump_stack();
2679
2680         return 0;
2681 }
2682
2683 /*
2684  * Common debugging checks for both nested and non-nested unlock:
2685  */
2686 static int check_unlock(struct task_struct *curr, struct lockdep_map *lock,
2687                         unsigned long ip)
2688 {
2689         if (unlikely(!debug_locks))
2690                 return 0;
2691         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2692                 return 0;
2693
2694         if (curr->lockdep_depth <= 0)
2695                 return print_unlock_inbalance_bug(curr, lock, ip);
2696
2697         return 1;
2698 }
2699
2700 static int
2701 __lock_set_class(struct lockdep_map *lock, const char *name,
2702                  struct lock_class_key *key, unsigned int subclass,
2703                  unsigned long ip)
2704 {
2705         struct task_struct *curr = current;
2706         struct held_lock *hlock, *prev_hlock;
2707         struct lock_class *class;
2708         unsigned int depth;
2709         int i;
2710
2711         depth = curr->lockdep_depth;
2712         if (DEBUG_LOCKS_WARN_ON(!depth))
2713                 return 0;
2714
2715         prev_hlock = NULL;
2716         for (i = depth-1; i >= 0; i--) {
2717                 hlock = curr->held_locks + i;
2718                 /*
2719                  * We must not cross into another context:
2720                  */
2721                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
2722                         break;
2723                 if (hlock->instance == lock)
2724                         goto found_it;
2725                 prev_hlock = hlock;
2726         }
2727         return print_unlock_inbalance_bug(curr, lock, ip);
2728
2729 found_it:
2730         lockdep_init_map(lock, name, key, 0);
2731         class = register_lock_class(lock, subclass, 0);
2732         hlock->class_idx = class - lock_classes + 1;
2733
2734         curr->lockdep_depth = i;
2735         curr->curr_chain_key = hlock->prev_chain_key;
2736
2737         for (; i < depth; i++) {
2738                 hlock = curr->held_locks + i;
2739                 if (!__lock_acquire(hlock->instance,
2740                         hlock_class(hlock)->subclass, hlock->trylock,
2741                                 hlock->read, hlock->check, hlock->hardirqs_off,
2742                                 hlock->nest_lock, hlock->acquire_ip))
2743                         return 0;
2744         }
2745
2746         if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth))
2747                 return 0;
2748         return 1;
2749 }
2750
2751 /*
2752  * Remove the lock to the list of currently held locks in a
2753  * potentially non-nested (out of order) manner. This is a
2754  * relatively rare operation, as all the unlock APIs default
2755  * to nested mode (which uses lock_release()):
2756  */
2757 static int
2758 lock_release_non_nested(struct task_struct *curr,
2759                         struct lockdep_map *lock, unsigned long ip)
2760 {
2761         struct held_lock *hlock, *prev_hlock;
2762         unsigned int depth;
2763         int i;
2764
2765         /*
2766          * Check whether the lock exists in the current stack
2767          * of held locks:
2768          */
2769         depth = curr->lockdep_depth;
2770         if (DEBUG_LOCKS_WARN_ON(!depth))
2771                 return 0;
2772
2773         prev_hlock = NULL;
2774         for (i = depth-1; i >= 0; i--) {
2775                 hlock = curr->held_locks + i;
2776                 /*
2777                  * We must not cross into another context:
2778                  */
2779                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
2780                         break;
2781                 if (hlock->instance == lock)
2782                         goto found_it;
2783                 prev_hlock = hlock;
2784         }
2785         return print_unlock_inbalance_bug(curr, lock, ip);
2786
2787 found_it:
2788         lock_release_holdtime(hlock);
2789
2790         /*
2791          * We have the right lock to unlock, 'hlock' points to it.
2792          * Now we remove it from the stack, and add back the other
2793          * entries (if any), recalculating the hash along the way:
2794          */
2795         curr->lockdep_depth = i;
2796         curr->curr_chain_key = hlock->prev_chain_key;
2797
2798         for (i++; i < depth; i++) {
2799                 hlock = curr->held_locks + i;
2800                 if (!__lock_acquire(hlock->instance,
2801                         hlock_class(hlock)->subclass, hlock->trylock,
2802                                 hlock->read, hlock->check, hlock->hardirqs_off,
2803                                 hlock->nest_lock, hlock->acquire_ip))
2804                         return 0;
2805         }
2806
2807         if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth - 1))
2808                 return 0;
2809         return 1;
2810 }
2811
2812 /*
2813  * Remove the lock to the list of currently held locks - this gets
2814  * called on mutex_unlock()/spin_unlock*() (or on a failed
2815  * mutex_lock_interruptible()). This is done for unlocks that nest
2816  * perfectly. (i.e. the current top of the lock-stack is unlocked)
2817  */
2818 static int lock_release_nested(struct task_struct *curr,
2819                                struct lockdep_map *lock, unsigned long ip)
2820 {
2821         struct held_lock *hlock;
2822         unsigned int depth;
2823
2824         /*
2825          * Pop off the top of the lock stack:
2826          */
2827         depth = curr->lockdep_depth - 1;
2828         hlock = curr->held_locks + depth;
2829
2830         /*
2831          * Is the unlock non-nested:
2832          */
2833         if (hlock->instance != lock)
2834                 return lock_release_non_nested(curr, lock, ip);
2835         curr->lockdep_depth--;
2836
2837         if (DEBUG_LOCKS_WARN_ON(!depth && (hlock->prev_chain_key != 0)))
2838                 return 0;
2839
2840         curr->curr_chain_key = hlock->prev_chain_key;
2841
2842         lock_release_holdtime(hlock);
2843
2844 #ifdef CONFIG_DEBUG_LOCKDEP
2845         hlock->prev_chain_key = 0;
2846         hlock->class_idx = 0;
2847         hlock->acquire_ip = 0;
2848         hlock->irq_context = 0;
2849 #endif
2850         return 1;
2851 }
2852
2853 /*
2854  * Remove the lock to the list of currently held locks - this gets
2855  * called on mutex_unlock()/spin_unlock*() (or on a failed
2856  * mutex_lock_interruptible()). This is done for unlocks that nest
2857  * perfectly. (i.e. the current top of the lock-stack is unlocked)
2858  */
2859 static void
2860 __lock_release(struct lockdep_map *lock, int nested, unsigned long ip)
2861 {
2862         struct task_struct *curr = current;
2863
2864         if (!check_unlock(curr, lock, ip))
2865                 return;
2866
2867         if (nested) {
2868                 if (!lock_release_nested(curr, lock, ip))
2869                         return;
2870         } else {
2871                 if (!lock_release_non_nested(curr, lock, ip))
2872                         return;
2873         }
2874
2875         check_chain_key(curr);
2876 }
2877
2878 /*
2879  * Check whether we follow the irq-flags state precisely:
2880  */
2881 static void check_flags(unsigned long flags)
2882 {
2883 #if defined(CONFIG_PROVE_LOCKING) && defined(CONFIG_DEBUG_LOCKDEP) && \
2884     defined(CONFIG_TRACE_IRQFLAGS)
2885         if (!debug_locks)
2886                 return;
2887
2888         if (irqs_disabled_flags(flags)) {
2889                 if (DEBUG_LOCKS_WARN_ON(current->hardirqs_enabled)) {
2890                         printk("possible reason: unannotated irqs-off.\n");
2891                 }
2892         } else {
2893                 if (DEBUG_LOCKS_WARN_ON(!current->hardirqs_enabled)) {
2894                         printk("possible reason: unannotated irqs-on.\n");
2895                 }
2896         }
2897
2898         /*
2899          * We dont accurately track softirq state in e.g.
2900          * hardirq contexts (such as on 4KSTACKS), so only
2901          * check if not in hardirq contexts:
2902          */
2903         if (!hardirq_count()) {
2904                 if (softirq_count())
2905                         DEBUG_LOCKS_WARN_ON(current->softirqs_enabled);
2906                 else
2907                         DEBUG_LOCKS_WARN_ON(!current->softirqs_enabled);
2908         }
2909
2910         if (!debug_locks)
2911                 print_irqtrace_events(current);
2912 #endif
2913 }
2914
2915 void lock_set_class(struct lockdep_map *lock, const char *name,
2916                     struct lock_class_key *key, unsigned int subclass,
2917                     unsigned long ip)
2918 {
2919         unsigned long flags;
2920
2921         if (unlikely(current->lockdep_recursion))
2922                 return;
2923
2924         raw_local_irq_save(flags);
2925         current->lockdep_recursion = 1;
2926         check_flags(flags);
2927         if (__lock_set_class(lock, name, key, subclass, ip))
2928                 check_chain_key(current);
2929         current->lockdep_recursion = 0;
2930         raw_local_irq_restore(flags);
2931 }
2932 EXPORT_SYMBOL_GPL(lock_set_class);
2933
2934 /*
2935  * We are not always called with irqs disabled - do that here,
2936  * and also avoid lockdep recursion:
2937  */
2938 void lock_acquire(struct lockdep_map *lock, unsigned int subclass,
2939                           int trylock, int read, int check,
2940                           struct lockdep_map *nest_lock, unsigned long ip)
2941 {
2942         unsigned long flags;
2943
2944         trace_lock_acquire(lock, subclass, trylock, read, check, nest_lock, ip);
2945
2946         if (unlikely(current->lockdep_recursion))
2947                 return;
2948
2949         raw_local_irq_save(flags);
2950         check_flags(flags);
2951
2952         current->lockdep_recursion = 1;
2953         __lock_acquire(lock, subclass, trylock, read, check,
2954                        irqs_disabled_flags(flags), nest_lock, ip);
2955         current->lockdep_recursion = 0;
2956         raw_local_irq_restore(flags);
2957 }
2958 EXPORT_SYMBOL_GPL(lock_acquire);
2959
2960 void lock_release(struct lockdep_map *lock, int nested,
2961                           unsigned long ip)
2962 {
2963         unsigned long flags;
2964
2965         trace_lock_release(lock, nested, ip);
2966
2967         if (unlikely(current->lockdep_recursion))
2968                 return;
2969
2970         raw_local_irq_save(flags);
2971         check_flags(flags);
2972         current->lockdep_recursion = 1;
2973         __lock_release(lock, nested, ip);
2974         current->lockdep_recursion = 0;
2975         raw_local_irq_restore(flags);
2976 }
2977 EXPORT_SYMBOL_GPL(lock_release);
2978
2979 void lockdep_set_current_reclaim_state(gfp_t gfp_mask)
2980 {
2981         current->lockdep_reclaim_gfp = gfp_mask;
2982 }
2983
2984 void lockdep_clear_current_reclaim_state(void)
2985 {
2986         current->lockdep_reclaim_gfp = 0;
2987 }
2988
2989 #ifdef CONFIG_LOCK_STAT
2990 static int
2991 print_lock_contention_bug(struct task_struct *curr, struct lockdep_map *lock,
2992                            unsigned long ip)
2993 {
2994         if (!debug_locks_off())
2995                 return 0;
2996         if (debug_locks_silent)
2997                 return 0;
2998
2999         printk("\n=================================\n");
3000         printk(  "[ BUG: bad contention detected! ]\n");
3001         printk(  "---------------------------------\n");
3002         printk("%s/%d is trying to contend lock (",
3003                 curr->comm, task_pid_nr(curr));
3004         print_lockdep_cache(lock);
3005         printk(") at:\n");
3006         print_ip_sym(ip);
3007         printk("but there are no locks held!\n");
3008         printk("\nother info that might help us debug this:\n");
3009         lockdep_print_held_locks(curr);
3010
3011         printk("\nstack backtrace:\n");
3012         dump_stack();
3013
3014         return 0;
3015 }
3016
3017 static void
3018 __lock_contended(struct lockdep_map *lock, unsigned long ip)
3019 {
3020         struct task_struct *curr = current;
3021         struct held_lock *hlock, *prev_hlock;
3022         struct lock_class_stats *stats;
3023         unsigned int depth;
3024         int i, contention_point, contending_point;
3025
3026         depth = curr->lockdep_depth;
3027         if (DEBUG_LOCKS_WARN_ON(!depth))
3028                 return;
3029
3030         prev_hlock = NULL;
3031         for (i = depth-1; i >= 0; i--) {
3032                 hlock = curr->held_locks + i;
3033                 /*
3034                  * We must not cross into another context:
3035                  */
3036                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3037                         break;
3038                 if (hlock->instance == lock)
3039                         goto found_it;
3040                 prev_hlock = hlock;
3041         }
3042         print_lock_contention_bug(curr, lock, ip);
3043         return;
3044
3045 found_it:
3046         hlock->waittime_stamp = sched_clock();
3047
3048         contention_point = lock_point(hlock_class(hlock)->contention_point, ip);
3049         contending_point = lock_point(hlock_class(hlock)->contending_point,
3050                                       lock->ip);
3051
3052         stats = get_lock_stats(hlock_class(hlock));
3053         if (contention_point < LOCKSTAT_POINTS)
3054                 stats->contention_point[contention_point]++;
3055         if (contending_point < LOCKSTAT_POINTS)
3056                 stats->contending_point[contending_point]++;
3057         if (lock->cpu != smp_processor_id())
3058                 stats->bounces[bounce_contended + !!hlock->read]++;
3059         put_lock_stats(stats);
3060 }
3061
3062 static void
3063 __lock_acquired(struct lockdep_map *lock, unsigned long ip)
3064 {
3065         struct task_struct *curr = current;
3066         struct held_lock *hlock, *prev_hlock;
3067         struct lock_class_stats *stats;
3068         unsigned int depth;
3069         u64 now;
3070         s64 waittime = 0;
3071         int i, cpu;
3072
3073         depth = curr->lockdep_depth;
3074         if (DEBUG_LOCKS_WARN_ON(!depth))
3075                 return;
3076
3077         prev_hlock = NULL;
3078         for (i = depth-1; i >= 0; i--) {
3079                 hlock = curr->held_locks + i;
3080                 /*
3081                  * We must not cross into another context:
3082                  */
3083                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3084                         break;
3085                 if (hlock->instance == lock)
3086                         goto found_it;
3087                 prev_hlock = hlock;
3088         }
3089         print_lock_contention_bug(curr, lock, _RET_IP_);
3090         return;
3091
3092 found_it:
3093         cpu = smp_processor_id();
3094         if (hlock->waittime_stamp) {
3095                 now = sched_clock();
3096                 waittime = now - hlock->waittime_stamp;
3097                 hlock->holdtime_stamp = now;
3098         }
3099
3100         trace_lock_acquired(lock, ip, waittime);
3101
3102         stats = get_lock_stats(hlock_class(hlock));
3103         if (waittime) {
3104                 if (hlock->read)
3105                         lock_time_inc(&stats->read_waittime, waittime);
3106                 else
3107                         lock_time_inc(&stats->write_waittime, waittime);
3108         }
3109         if (lock->cpu != cpu)
3110                 stats->bounces[bounce_acquired + !!hlock->read]++;
3111         put_lock_stats(stats);
3112
3113         lock->cpu = cpu;
3114         lock->ip = ip;
3115 }
3116
3117 void lock_contended(struct lockdep_map *lock, unsigned long ip)
3118 {
3119         unsigned long flags;
3120
3121         trace_lock_contended(lock, ip);
3122
3123         if (unlikely(!lock_stat))
3124                 return;
3125
3126         if (unlikely(current->lockdep_recursion))
3127                 return;
3128
3129         raw_local_irq_save(flags);
3130         check_flags(flags);
3131         current->lockdep_recursion = 1;
3132         __lock_contended(lock, ip);
3133         current->lockdep_recursion = 0;
3134         raw_local_irq_restore(flags);
3135 }
3136 EXPORT_SYMBOL_GPL(lock_contended);
3137
3138 void lock_acquired(struct lockdep_map *lock, unsigned long ip)
3139 {
3140         unsigned long flags;
3141
3142         if (unlikely(!lock_stat))
3143                 return;
3144
3145         if (unlikely(current->lockdep_recursion))
3146                 return;
3147
3148         raw_local_irq_save(flags);
3149         check_flags(flags);
3150         current->lockdep_recursion = 1;
3151         __lock_acquired(lock, ip);
3152         current->lockdep_recursion = 0;
3153         raw_local_irq_restore(flags);
3154 }
3155 EXPORT_SYMBOL_GPL(lock_acquired);
3156 #endif
3157
3158 /*
3159  * Used by the testsuite, sanitize the validator state
3160  * after a simulated failure:
3161  */
3162
3163 void lockdep_reset(void)
3164 {
3165         unsigned long flags;
3166         int i;
3167
3168         raw_local_irq_save(flags);
3169         current->curr_chain_key = 0;
3170         current->lockdep_depth = 0;
3171         current->lockdep_recursion = 0;
3172         memset(current->held_locks, 0, MAX_LOCK_DEPTH*sizeof(struct held_lock));
3173         nr_hardirq_chains = 0;
3174         nr_softirq_chains = 0;
3175         nr_process_chains = 0;
3176         debug_locks = 1;
3177         for (i = 0; i < CHAINHASH_SIZE; i++)
3178                 INIT_LIST_HEAD(chainhash_table + i);
3179         raw_local_irq_restore(flags);
3180 }
3181
3182 static void zap_class(struct lock_class *class)
3183 {
3184         int i;
3185
3186         /*
3187          * Remove all dependencies this lock is
3188          * involved in:
3189          */
3190         for (i = 0; i < nr_list_entries; i++) {
3191                 if (list_entries[i].class == class)
3192                         list_del_rcu(&list_entries[i].entry);
3193         }
3194         /*
3195          * Unhash the class and remove it from the all_lock_classes list:
3196          */
3197         list_del_rcu(&class->hash_entry);
3198         list_del_rcu(&class->lock_entry);
3199
3200         class->key = NULL;
3201 }
3202
3203 static inline int within(const void *addr, void *start, unsigned long size)
3204 {
3205         return addr >= start && addr < start + size;
3206 }
3207
3208 void lockdep_free_key_range(void *start, unsigned long size)
3209 {
3210         struct lock_class *class, *next;
3211         struct list_head *head;
3212         unsigned long flags;
3213         int i;
3214         int locked;
3215
3216         raw_local_irq_save(flags);
3217         locked = graph_lock();
3218
3219         /*
3220          * Unhash all classes that were created by this module:
3221          */
3222         for (i = 0; i < CLASSHASH_SIZE; i++) {
3223                 head = classhash_table + i;
3224                 if (list_empty(head))
3225                         continue;
3226                 list_for_each_entry_safe(class, next, head, hash_entry) {
3227                         if (within(class->key, start, size))
3228                                 zap_class(class);
3229                         else if (within(class->name, start, size))
3230                                 zap_class(class);
3231                 }
3232         }
3233
3234         if (locked)
3235                 graph_unlock();
3236         raw_local_irq_restore(flags);
3237 }
3238
3239 void lockdep_reset_lock(struct lockdep_map *lock)
3240 {
3241         struct lock_class *class, *next;
3242         struct list_head *head;
3243         unsigned long flags;
3244         int i, j;
3245         int locked;
3246
3247         raw_local_irq_save(flags);
3248
3249         /*
3250          * Remove all classes this lock might have:
3251          */
3252         for (j = 0; j < MAX_LOCKDEP_SUBCLASSES; j++) {
3253                 /*
3254                  * If the class exists we look it up and zap it:
3255                  */
3256                 class = look_up_lock_class(lock, j);
3257                 if (class)
3258                         zap_class(class);
3259         }
3260         /*
3261          * Debug check: in the end all mapped classes should
3262          * be gone.
3263          */
3264         locked = graph_lock();
3265         for (i = 0; i < CLASSHASH_SIZE; i++) {
3266                 head = classhash_table + i;
3267                 if (list_empty(head))
3268                         continue;
3269                 list_for_each_entry_safe(class, next, head, hash_entry) {
3270                         if (unlikely(class == lock->class_cache)) {
3271                                 if (debug_locks_off_graph_unlock())
3272                                         WARN_ON(1);
3273                                 goto out_restore;
3274                         }
3275                 }
3276         }
3277         if (locked)
3278                 graph_unlock();
3279
3280 out_restore:
3281         raw_local_irq_restore(flags);
3282 }
3283
3284 void lockdep_init(void)
3285 {
3286         int i;
3287
3288         /*
3289          * Some architectures have their own start_kernel()
3290          * code which calls lockdep_init(), while we also
3291          * call lockdep_init() from the start_kernel() itself,
3292          * and we want to initialize the hashes only once:
3293          */
3294         if (lockdep_initialized)
3295                 return;
3296
3297         for (i = 0; i < CLASSHASH_SIZE; i++)
3298                 INIT_LIST_HEAD(classhash_table + i);
3299
3300         for (i = 0; i < CHAINHASH_SIZE; i++)
3301                 INIT_LIST_HEAD(chainhash_table + i);
3302
3303         lockdep_initialized = 1;
3304 }
3305
3306 void __init lockdep_info(void)
3307 {
3308         printk("Lock dependency validator: Copyright (c) 2006 Red Hat, Inc., Ingo Molnar\n");
3309
3310         printk("... MAX_LOCKDEP_SUBCLASSES:  %lu\n", MAX_LOCKDEP_SUBCLASSES);
3311         printk("... MAX_LOCK_DEPTH:          %lu\n", MAX_LOCK_DEPTH);
3312         printk("... MAX_LOCKDEP_KEYS:        %lu\n", MAX_LOCKDEP_KEYS);
3313         printk("... CLASSHASH_SIZE:          %lu\n", CLASSHASH_SIZE);
3314         printk("... MAX_LOCKDEP_ENTRIES:     %lu\n", MAX_LOCKDEP_ENTRIES);
3315         printk("... MAX_LOCKDEP_CHAINS:      %lu\n", MAX_LOCKDEP_CHAINS);
3316         printk("... CHAINHASH_SIZE:          %lu\n", CHAINHASH_SIZE);
3317
3318         printk(" memory used by lock dependency info: %lu kB\n",
3319                 (sizeof(struct lock_class) * MAX_LOCKDEP_KEYS +
3320                 sizeof(struct list_head) * CLASSHASH_SIZE +
3321                 sizeof(struct lock_list) * MAX_LOCKDEP_ENTRIES +
3322                 sizeof(struct lock_chain) * MAX_LOCKDEP_CHAINS +
3323                 sizeof(struct list_head) * CHAINHASH_SIZE) / 1024);
3324
3325         printk(" per task-struct memory footprint: %lu bytes\n",
3326                 sizeof(struct held_lock) * MAX_LOCK_DEPTH);
3327
3328 #ifdef CONFIG_DEBUG_LOCKDEP
3329         if (lockdep_init_error) {
3330                 printk("WARNING: lockdep init error! Arch code didn't call lockdep_init() early enough?\n");
3331                 printk("Call stack leading to lockdep invocation was:\n");
3332                 print_stack_trace(&lockdep_init_trace, 0);
3333         }
3334 #endif
3335 }
3336
3337 static void
3338 print_freed_lock_bug(struct task_struct *curr, const void *mem_from,
3339                      const void *mem_to, struct held_lock *hlock)
3340 {
3341         if (!debug_locks_off())
3342                 return;
3343         if (debug_locks_silent)
3344                 return;
3345
3346         printk("\n=========================\n");
3347         printk(  "[ BUG: held lock freed! ]\n");
3348         printk(  "-------------------------\n");
3349         printk("%s/%d is freeing memory %p-%p, with a lock still held there!\n",
3350                 curr->comm, task_pid_nr(curr), mem_from, mem_to-1);
3351         print_lock(hlock);
3352         lockdep_print_held_locks(curr);
3353
3354         printk("\nstack backtrace:\n");
3355         dump_stack();
3356 }
3357
3358 static inline int not_in_range(const void* mem_from, unsigned long mem_len,
3359                                 const void* lock_from, unsigned long lock_len)
3360 {
3361         return lock_from + lock_len <= mem_from ||
3362                 mem_from + mem_len <= lock_from;
3363 }
3364
3365 /*
3366  * Called when kernel memory is freed (or unmapped), or if a lock
3367  * is destroyed or reinitialized - this code checks whether there is
3368  * any held lock in the memory range of <from> to <to>:
3369  */
3370 void debug_check_no_locks_freed(const void *mem_from, unsigned long mem_len)
3371 {
3372         struct task_struct *curr = current;
3373         struct held_lock *hlock;
3374         unsigned long flags;
3375         int i;
3376
3377         if (unlikely(!debug_locks))
3378                 return;
3379
3380         local_irq_save(flags);
3381         for (i = 0; i < curr->lockdep_depth; i++) {
3382                 hlock = curr->held_locks + i;
3383
3384                 if (not_in_range(mem_from, mem_len, hlock->instance,
3385                                         sizeof(*hlock->instance)))
3386                         continue;
3387
3388                 print_freed_lock_bug(curr, mem_from, mem_from + mem_len, hlock);
3389                 break;
3390         }
3391         local_irq_restore(flags);
3392 }
3393 EXPORT_SYMBOL_GPL(debug_check_no_locks_freed);
3394
3395 static void print_held_locks_bug(struct task_struct *curr)
3396 {
3397         if (!debug_locks_off())
3398                 return;
3399         if (debug_locks_silent)
3400                 return;
3401
3402         printk("\n=====================================\n");
3403         printk(  "[ BUG: lock held at task exit time! ]\n");
3404         printk(  "-------------------------------------\n");
3405         printk("%s/%d is exiting with locks still held!\n",
3406                 curr->comm, task_pid_nr(curr));
3407         lockdep_print_held_locks(curr);
3408
3409         printk("\nstack backtrace:\n");
3410         dump_stack();
3411 }
3412
3413 void debug_check_no_locks_held(struct task_struct *task)
3414 {
3415         if (unlikely(task->lockdep_depth > 0))
3416                 print_held_locks_bug(task);
3417 }
3418
3419 void debug_show_all_locks(void)
3420 {
3421         struct task_struct *g, *p;
3422         int count = 10;
3423         int unlock = 1;
3424
3425         if (unlikely(!debug_locks)) {
3426                 printk("INFO: lockdep is turned off.\n");
3427                 return;
3428         }
3429         printk("\nShowing all locks held in the system:\n");
3430
3431         /*
3432          * Here we try to get the tasklist_lock as hard as possible,
3433          * if not successful after 2 seconds we ignore it (but keep
3434          * trying). This is to enable a debug printout even if a
3435          * tasklist_lock-holding task deadlocks or crashes.
3436          */
3437 retry:
3438         if (!read_trylock(&tasklist_lock)) {
3439                 if (count == 10)
3440                         printk("hm, tasklist_lock locked, retrying... ");
3441                 if (count) {
3442                         count--;
3443                         printk(" #%d", 10-count);
3444                         mdelay(200);
3445                         goto retry;
3446                 }
3447                 printk(" ignoring it.\n");
3448                 unlock = 0;
3449         } else {
3450                 if (count != 10)
3451                         printk(KERN_CONT " locked it.\n");
3452         }
3453
3454         do_each_thread(g, p) {
3455                 /*
3456                  * It's not reliable to print a task's held locks
3457                  * if it's not sleeping (or if it's not the current
3458                  * task):
3459                  */
3460                 if (p->state == TASK_RUNNING && p != current)
3461                         continue;
3462                 if (p->lockdep_depth)
3463                         lockdep_print_held_locks(p);
3464                 if (!unlock)
3465                         if (read_trylock(&tasklist_lock))
3466                                 unlock = 1;
3467         } while_each_thread(g, p);
3468
3469         printk("\n");
3470         printk("=============================================\n\n");
3471
3472         if (unlock)
3473                 read_unlock(&tasklist_lock);
3474 }
3475 EXPORT_SYMBOL_GPL(debug_show_all_locks);
3476
3477 /*
3478  * Careful: only use this function if you are sure that
3479  * the task cannot run in parallel!
3480  */
3481 void __debug_show_held_locks(struct task_struct *task)
3482 {
3483         if (unlikely(!debug_locks)) {
3484                 printk("INFO: lockdep is turned off.\n");
3485                 return;
3486         }
3487         lockdep_print_held_locks(task);
3488 }
3489 EXPORT_SYMBOL_GPL(__debug_show_held_locks);
3490
3491 void debug_show_held_locks(struct task_struct *task)
3492 {
3493                 __debug_show_held_locks(task);
3494 }
3495 EXPORT_SYMBOL_GPL(debug_show_held_locks);
3496
3497 void lockdep_sys_exit(void)
3498 {
3499         struct task_struct *curr = current;
3500
3501         if (unlikely(curr->lockdep_depth)) {
3502                 if (!debug_locks_off())
3503                         return;
3504                 printk("\n================================================\n");
3505                 printk(  "[ BUG: lock held when returning to user space! ]\n");
3506                 printk(  "------------------------------------------------\n");
3507                 printk("%s/%d is leaving the kernel with locks still held!\n",
3508                                 curr->comm, curr->pid);
3509                 lockdep_print_held_locks(curr);
3510         }
3511 }