lockdep: more robust lockdep_map init sequence
[linux-3.10.git] / kernel / lockdep.c
1 /*
2  * kernel/lockdep.c
3  *
4  * Runtime locking correctness validator
5  *
6  * Started by Ingo Molnar:
7  *
8  *  Copyright (C) 2006,2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
9  *  Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
10  *
11  * this code maps all the lock dependencies as they occur in a live kernel
12  * and will warn about the following classes of locking bugs:
13  *
14  * - lock inversion scenarios
15  * - circular lock dependencies
16  * - hardirq/softirq safe/unsafe locking bugs
17  *
18  * Bugs are reported even if the current locking scenario does not cause
19  * any deadlock at this point.
20  *
21  * I.e. if anytime in the past two locks were taken in a different order,
22  * even if it happened for another task, even if those were different
23  * locks (but of the same class as this lock), this code will detect it.
24  *
25  * Thanks to Arjan van de Ven for coming up with the initial idea of
26  * mapping lock dependencies runtime.
27  */
28 #define DISABLE_BRANCH_PROFILING
29 #include <linux/mutex.h>
30 #include <linux/sched.h>
31 #include <linux/delay.h>
32 #include <linux/module.h>
33 #include <linux/proc_fs.h>
34 #include <linux/seq_file.h>
35 #include <linux/spinlock.h>
36 #include <linux/kallsyms.h>
37 #include <linux/interrupt.h>
38 #include <linux/stacktrace.h>
39 #include <linux/debug_locks.h>
40 #include <linux/irqflags.h>
41 #include <linux/utsname.h>
42 #include <linux/hash.h>
43 #include <linux/ftrace.h>
44 #include <linux/stringify.h>
45 #include <trace/lockdep.h>
46
47 #include <asm/sections.h>
48
49 #include "lockdep_internals.h"
50
51 #ifdef CONFIG_PROVE_LOCKING
52 int prove_locking = 1;
53 module_param(prove_locking, int, 0644);
54 #else
55 #define prove_locking 0
56 #endif
57
58 #ifdef CONFIG_LOCK_STAT
59 int lock_stat = 1;
60 module_param(lock_stat, int, 0644);
61 #else
62 #define lock_stat 0
63 #endif
64
65 /*
66  * lockdep_lock: protects the lockdep graph, the hashes and the
67  *               class/list/hash allocators.
68  *
69  * This is one of the rare exceptions where it's justified
70  * to use a raw spinlock - we really dont want the spinlock
71  * code to recurse back into the lockdep code...
72  */
73 static raw_spinlock_t lockdep_lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
74
75 static int graph_lock(void)
76 {
77         __raw_spin_lock(&lockdep_lock);
78         /*
79          * Make sure that if another CPU detected a bug while
80          * walking the graph we dont change it (while the other
81          * CPU is busy printing out stuff with the graph lock
82          * dropped already)
83          */
84         if (!debug_locks) {
85                 __raw_spin_unlock(&lockdep_lock);
86                 return 0;
87         }
88         /* prevent any recursions within lockdep from causing deadlocks */
89         current->lockdep_recursion++;
90         return 1;
91 }
92
93 static inline int graph_unlock(void)
94 {
95         if (debug_locks && !__raw_spin_is_locked(&lockdep_lock))
96                 return DEBUG_LOCKS_WARN_ON(1);
97
98         current->lockdep_recursion--;
99         __raw_spin_unlock(&lockdep_lock);
100         return 0;
101 }
102
103 /*
104  * Turn lock debugging off and return with 0 if it was off already,
105  * and also release the graph lock:
106  */
107 static inline int debug_locks_off_graph_unlock(void)
108 {
109         int ret = debug_locks_off();
110
111         __raw_spin_unlock(&lockdep_lock);
112
113         return ret;
114 }
115
116 static int lockdep_initialized;
117
118 unsigned long nr_list_entries;
119 static struct lock_list list_entries[MAX_LOCKDEP_ENTRIES];
120
121 /*
122  * All data structures here are protected by the global debug_lock.
123  *
124  * Mutex key structs only get allocated, once during bootup, and never
125  * get freed - this significantly simplifies the debugging code.
126  */
127 unsigned long nr_lock_classes;
128 static struct lock_class lock_classes[MAX_LOCKDEP_KEYS];
129
130 static inline struct lock_class *hlock_class(struct held_lock *hlock)
131 {
132         if (!hlock->class_idx) {
133                 DEBUG_LOCKS_WARN_ON(1);
134                 return NULL;
135         }
136         return lock_classes + hlock->class_idx - 1;
137 }
138
139 #ifdef CONFIG_LOCK_STAT
140 static DEFINE_PER_CPU(struct lock_class_stats[MAX_LOCKDEP_KEYS], lock_stats);
141
142 static int lock_point(unsigned long points[], unsigned long ip)
143 {
144         int i;
145
146         for (i = 0; i < LOCKSTAT_POINTS; i++) {
147                 if (points[i] == 0) {
148                         points[i] = ip;
149                         break;
150                 }
151                 if (points[i] == ip)
152                         break;
153         }
154
155         return i;
156 }
157
158 static void lock_time_inc(struct lock_time *lt, s64 time)
159 {
160         if (time > lt->max)
161                 lt->max = time;
162
163         if (time < lt->min || !lt->min)
164                 lt->min = time;
165
166         lt->total += time;
167         lt->nr++;
168 }
169
170 static inline void lock_time_add(struct lock_time *src, struct lock_time *dst)
171 {
172         dst->min += src->min;
173         dst->max += src->max;
174         dst->total += src->total;
175         dst->nr += src->nr;
176 }
177
178 struct lock_class_stats lock_stats(struct lock_class *class)
179 {
180         struct lock_class_stats stats;
181         int cpu, i;
182
183         memset(&stats, 0, sizeof(struct lock_class_stats));
184         for_each_possible_cpu(cpu) {
185                 struct lock_class_stats *pcs =
186                         &per_cpu(lock_stats, cpu)[class - lock_classes];
187
188                 for (i = 0; i < ARRAY_SIZE(stats.contention_point); i++)
189                         stats.contention_point[i] += pcs->contention_point[i];
190
191                 for (i = 0; i < ARRAY_SIZE(stats.contending_point); i++)
192                         stats.contending_point[i] += pcs->contending_point[i];
193
194                 lock_time_add(&pcs->read_waittime, &stats.read_waittime);
195                 lock_time_add(&pcs->write_waittime, &stats.write_waittime);
196
197                 lock_time_add(&pcs->read_holdtime, &stats.read_holdtime);
198                 lock_time_add(&pcs->write_holdtime, &stats.write_holdtime);
199
200                 for (i = 0; i < ARRAY_SIZE(stats.bounces); i++)
201                         stats.bounces[i] += pcs->bounces[i];
202         }
203
204         return stats;
205 }
206
207 void clear_lock_stats(struct lock_class *class)
208 {
209         int cpu;
210
211         for_each_possible_cpu(cpu) {
212                 struct lock_class_stats *cpu_stats =
213                         &per_cpu(lock_stats, cpu)[class - lock_classes];
214
215                 memset(cpu_stats, 0, sizeof(struct lock_class_stats));
216         }
217         memset(class->contention_point, 0, sizeof(class->contention_point));
218         memset(class->contending_point, 0, sizeof(class->contending_point));
219 }
220
221 static struct lock_class_stats *get_lock_stats(struct lock_class *class)
222 {
223         return &get_cpu_var(lock_stats)[class - lock_classes];
224 }
225
226 static void put_lock_stats(struct lock_class_stats *stats)
227 {
228         put_cpu_var(lock_stats);
229 }
230
231 static void lock_release_holdtime(struct held_lock *hlock)
232 {
233         struct lock_class_stats *stats;
234         s64 holdtime;
235
236         if (!lock_stat)
237                 return;
238
239         holdtime = sched_clock() - hlock->holdtime_stamp;
240
241         stats = get_lock_stats(hlock_class(hlock));
242         if (hlock->read)
243                 lock_time_inc(&stats->read_holdtime, holdtime);
244         else
245                 lock_time_inc(&stats->write_holdtime, holdtime);
246         put_lock_stats(stats);
247 }
248 #else
249 static inline void lock_release_holdtime(struct held_lock *hlock)
250 {
251 }
252 #endif
253
254 /*
255  * We keep a global list of all lock classes. The list only grows,
256  * never shrinks. The list is only accessed with the lockdep
257  * spinlock lock held.
258  */
259 LIST_HEAD(all_lock_classes);
260
261 /*
262  * The lockdep classes are in a hash-table as well, for fast lookup:
263  */
264 #define CLASSHASH_BITS          (MAX_LOCKDEP_KEYS_BITS - 1)
265 #define CLASSHASH_SIZE          (1UL << CLASSHASH_BITS)
266 #define __classhashfn(key)      hash_long((unsigned long)key, CLASSHASH_BITS)
267 #define classhashentry(key)     (classhash_table + __classhashfn((key)))
268
269 static struct list_head classhash_table[CLASSHASH_SIZE];
270
271 /*
272  * We put the lock dependency chains into a hash-table as well, to cache
273  * their existence:
274  */
275 #define CHAINHASH_BITS          (MAX_LOCKDEP_CHAINS_BITS-1)
276 #define CHAINHASH_SIZE          (1UL << CHAINHASH_BITS)
277 #define __chainhashfn(chain)    hash_long(chain, CHAINHASH_BITS)
278 #define chainhashentry(chain)   (chainhash_table + __chainhashfn((chain)))
279
280 static struct list_head chainhash_table[CHAINHASH_SIZE];
281
282 /*
283  * The hash key of the lock dependency chains is a hash itself too:
284  * it's a hash of all locks taken up to that lock, including that lock.
285  * It's a 64-bit hash, because it's important for the keys to be
286  * unique.
287  */
288 #define iterate_chain_key(key1, key2) \
289         (((key1) << MAX_LOCKDEP_KEYS_BITS) ^ \
290         ((key1) >> (64-MAX_LOCKDEP_KEYS_BITS)) ^ \
291         (key2))
292
293 void lockdep_off(void)
294 {
295         current->lockdep_recursion++;
296 }
297 EXPORT_SYMBOL(lockdep_off);
298
299 void lockdep_on(void)
300 {
301         current->lockdep_recursion--;
302 }
303 EXPORT_SYMBOL(lockdep_on);
304
305 /*
306  * Debugging switches:
307  */
308
309 #define VERBOSE                 0
310 #define VERY_VERBOSE            0
311
312 #if VERBOSE
313 # define HARDIRQ_VERBOSE        1
314 # define SOFTIRQ_VERBOSE        1
315 # define RECLAIM_VERBOSE        1
316 #else
317 # define HARDIRQ_VERBOSE        0
318 # define SOFTIRQ_VERBOSE        0
319 # define RECLAIM_VERBOSE        0
320 #endif
321
322 #if VERBOSE || HARDIRQ_VERBOSE || SOFTIRQ_VERBOSE || RECLAIM_VERBOSE
323 /*
324  * Quick filtering for interesting events:
325  */
326 static int class_filter(struct lock_class *class)
327 {
328 #if 0
329         /* Example */
330         if (class->name_version == 1 &&
331                         !strcmp(class->name, "lockname"))
332                 return 1;
333         if (class->name_version == 1 &&
334                         !strcmp(class->name, "&struct->lockfield"))
335                 return 1;
336 #endif
337         /* Filter everything else. 1 would be to allow everything else */
338         return 0;
339 }
340 #endif
341
342 static int verbose(struct lock_class *class)
343 {
344 #if VERBOSE
345         return class_filter(class);
346 #endif
347         return 0;
348 }
349
350 /*
351  * Stack-trace: tightly packed array of stack backtrace
352  * addresses. Protected by the graph_lock.
353  */
354 unsigned long nr_stack_trace_entries;
355 static unsigned long stack_trace[MAX_STACK_TRACE_ENTRIES];
356
357 static int save_trace(struct stack_trace *trace)
358 {
359         trace->nr_entries = 0;
360         trace->max_entries = MAX_STACK_TRACE_ENTRIES - nr_stack_trace_entries;
361         trace->entries = stack_trace + nr_stack_trace_entries;
362
363         trace->skip = 3;
364
365         save_stack_trace(trace);
366
367         trace->max_entries = trace->nr_entries;
368
369         nr_stack_trace_entries += trace->nr_entries;
370
371         if (nr_stack_trace_entries == MAX_STACK_TRACE_ENTRIES) {
372                 if (!debug_locks_off_graph_unlock())
373                         return 0;
374
375                 printk("BUG: MAX_STACK_TRACE_ENTRIES too low!\n");
376                 printk("turning off the locking correctness validator.\n");
377                 dump_stack();
378
379                 return 0;
380         }
381
382         return 1;
383 }
384
385 unsigned int nr_hardirq_chains;
386 unsigned int nr_softirq_chains;
387 unsigned int nr_process_chains;
388 unsigned int max_lockdep_depth;
389 unsigned int max_recursion_depth;
390
391 static unsigned int lockdep_dependency_gen_id;
392
393 static bool lockdep_dependency_visit(struct lock_class *source,
394                                      unsigned int depth)
395 {
396         if (!depth)
397                 lockdep_dependency_gen_id++;
398         if (source->dep_gen_id == lockdep_dependency_gen_id)
399                 return true;
400         source->dep_gen_id = lockdep_dependency_gen_id;
401         return false;
402 }
403
404 #ifdef CONFIG_DEBUG_LOCKDEP
405 /*
406  * We cannot printk in early bootup code. Not even early_printk()
407  * might work. So we mark any initialization errors and printk
408  * about it later on, in lockdep_info().
409  */
410 static int lockdep_init_error;
411 static unsigned long lockdep_init_trace_data[20];
412 static struct stack_trace lockdep_init_trace = {
413         .max_entries = ARRAY_SIZE(lockdep_init_trace_data),
414         .entries = lockdep_init_trace_data,
415 };
416
417 /*
418  * Various lockdep statistics:
419  */
420 atomic_t chain_lookup_hits;
421 atomic_t chain_lookup_misses;
422 atomic_t hardirqs_on_events;
423 atomic_t hardirqs_off_events;
424 atomic_t redundant_hardirqs_on;
425 atomic_t redundant_hardirqs_off;
426 atomic_t softirqs_on_events;
427 atomic_t softirqs_off_events;
428 atomic_t redundant_softirqs_on;
429 atomic_t redundant_softirqs_off;
430 atomic_t nr_unused_locks;
431 atomic_t nr_cyclic_checks;
432 atomic_t nr_cyclic_check_recursions;
433 atomic_t nr_find_usage_forwards_checks;
434 atomic_t nr_find_usage_forwards_recursions;
435 atomic_t nr_find_usage_backwards_checks;
436 atomic_t nr_find_usage_backwards_recursions;
437 #endif
438
439 /*
440  * Locking printouts:
441  */
442
443 #define __USAGE(__STATE)                                                \
444         [LOCK_USED_IN_##__STATE] = "IN-"__stringify(__STATE)"-W",       \
445         [LOCK_ENABLED_##__STATE] = __stringify(__STATE)"-ON-W",         \
446         [LOCK_USED_IN_##__STATE##_READ] = "IN-"__stringify(__STATE)"-R",\
447         [LOCK_ENABLED_##__STATE##_READ] = __stringify(__STATE)"-ON-R",
448
449 static const char *usage_str[] =
450 {
451 #define LOCKDEP_STATE(__STATE) __USAGE(__STATE)
452 #include "lockdep_states.h"
453 #undef LOCKDEP_STATE
454         [LOCK_USED] = "INITIAL USE",
455 };
456
457 const char * __get_key_name(struct lockdep_subclass_key *key, char *str)
458 {
459         return kallsyms_lookup((unsigned long)key, NULL, NULL, NULL, str);
460 }
461
462 static inline unsigned long lock_flag(enum lock_usage_bit bit)
463 {
464         return 1UL << bit;
465 }
466
467 static char get_usage_char(struct lock_class *class, enum lock_usage_bit bit)
468 {
469         char c = '.';
470
471         if (class->usage_mask & lock_flag(bit + 2))
472                 c = '+';
473         if (class->usage_mask & lock_flag(bit)) {
474                 c = '-';
475                 if (class->usage_mask & lock_flag(bit + 2))
476                         c = '?';
477         }
478
479         return c;
480 }
481
482 void get_usage_chars(struct lock_class *class, char usage[LOCK_USAGE_CHARS])
483 {
484         int i = 0;
485
486 #define LOCKDEP_STATE(__STATE)                                          \
487         usage[i++] = get_usage_char(class, LOCK_USED_IN_##__STATE);     \
488         usage[i++] = get_usage_char(class, LOCK_USED_IN_##__STATE##_READ);
489 #include "lockdep_states.h"
490 #undef LOCKDEP_STATE
491
492         usage[i] = '\0';
493 }
494
495 static void print_lock_name(struct lock_class *class)
496 {
497         char str[KSYM_NAME_LEN], usage[LOCK_USAGE_CHARS];
498         const char *name;
499
500         get_usage_chars(class, usage);
501
502         name = class->name;
503         if (!name) {
504                 name = __get_key_name(class->key, str);
505                 printk(" (%s", name);
506         } else {
507                 printk(" (%s", name);
508                 if (class->name_version > 1)
509                         printk("#%d", class->name_version);
510                 if (class->subclass)
511                         printk("/%d", class->subclass);
512         }
513         printk("){%s}", usage);
514 }
515
516 static void print_lockdep_cache(struct lockdep_map *lock)
517 {
518         const char *name;
519         char str[KSYM_NAME_LEN];
520
521         name = lock->name;
522         if (!name)
523                 name = __get_key_name(lock->key->subkeys, str);
524
525         printk("%s", name);
526 }
527
528 static void print_lock(struct held_lock *hlock)
529 {
530         print_lock_name(hlock_class(hlock));
531         printk(", at: ");
532         print_ip_sym(hlock->acquire_ip);
533 }
534
535 static void lockdep_print_held_locks(struct task_struct *curr)
536 {
537         int i, depth = curr->lockdep_depth;
538
539         if (!depth) {
540                 printk("no locks held by %s/%d.\n", curr->comm, task_pid_nr(curr));
541                 return;
542         }
543         printk("%d lock%s held by %s/%d:\n",
544                 depth, depth > 1 ? "s" : "", curr->comm, task_pid_nr(curr));
545
546         for (i = 0; i < depth; i++) {
547                 printk(" #%d: ", i);
548                 print_lock(curr->held_locks + i);
549         }
550 }
551
552 static void print_lock_class_header(struct lock_class *class, int depth)
553 {
554         int bit;
555
556         printk("%*s->", depth, "");
557         print_lock_name(class);
558         printk(" ops: %lu", class->ops);
559         printk(" {\n");
560
561         for (bit = 0; bit < LOCK_USAGE_STATES; bit++) {
562                 if (class->usage_mask & (1 << bit)) {
563                         int len = depth;
564
565                         len += printk("%*s   %s", depth, "", usage_str[bit]);
566                         len += printk(" at:\n");
567                         print_stack_trace(class->usage_traces + bit, len);
568                 }
569         }
570         printk("%*s }\n", depth, "");
571
572         printk("%*s ... key      at: ",depth,"");
573         print_ip_sym((unsigned long)class->key);
574 }
575
576 /*
577  * printk all lock dependencies starting at <entry>:
578  */
579 static void __used
580 print_lock_dependencies(struct lock_class *class, int depth)
581 {
582         struct lock_list *entry;
583
584         if (lockdep_dependency_visit(class, depth))
585                 return;
586
587         if (DEBUG_LOCKS_WARN_ON(depth >= 20))
588                 return;
589
590         print_lock_class_header(class, depth);
591
592         list_for_each_entry(entry, &class->locks_after, entry) {
593                 if (DEBUG_LOCKS_WARN_ON(!entry->class))
594                         return;
595
596                 print_lock_dependencies(entry->class, depth + 1);
597
598                 printk("%*s ... acquired at:\n",depth,"");
599                 print_stack_trace(&entry->trace, 2);
600                 printk("\n");
601         }
602 }
603
604 static void print_kernel_version(void)
605 {
606         printk("%s %.*s\n", init_utsname()->release,
607                 (int)strcspn(init_utsname()->version, " "),
608                 init_utsname()->version);
609 }
610
611 static int very_verbose(struct lock_class *class)
612 {
613 #if VERY_VERBOSE
614         return class_filter(class);
615 #endif
616         return 0;
617 }
618
619 /*
620  * Is this the address of a static object:
621  */
622 static int static_obj(void *obj)
623 {
624         unsigned long start = (unsigned long) &_stext,
625                       end   = (unsigned long) &_end,
626                       addr  = (unsigned long) obj;
627 #ifdef CONFIG_SMP
628         int i;
629 #endif
630
631         /*
632          * static variable?
633          */
634         if ((addr >= start) && (addr < end))
635                 return 1;
636
637 #ifdef CONFIG_SMP
638         /*
639          * percpu var?
640          */
641         for_each_possible_cpu(i) {
642                 start = (unsigned long) &__per_cpu_start + per_cpu_offset(i);
643                 end   = (unsigned long) &__per_cpu_start + PERCPU_ENOUGH_ROOM
644                                         + per_cpu_offset(i);
645
646                 if ((addr >= start) && (addr < end))
647                         return 1;
648         }
649 #endif
650
651         /*
652          * module var?
653          */
654         return is_module_address(addr);
655 }
656
657 /*
658  * To make lock name printouts unique, we calculate a unique
659  * class->name_version generation counter:
660  */
661 static int count_matching_names(struct lock_class *new_class)
662 {
663         struct lock_class *class;
664         int count = 0;
665
666         if (!new_class->name)
667                 return 0;
668
669         list_for_each_entry(class, &all_lock_classes, lock_entry) {
670                 if (new_class->key - new_class->subclass == class->key)
671                         return class->name_version;
672                 if (class->name && !strcmp(class->name, new_class->name))
673                         count = max(count, class->name_version);
674         }
675
676         return count + 1;
677 }
678
679 /*
680  * Register a lock's class in the hash-table, if the class is not present
681  * yet. Otherwise we look it up. We cache the result in the lock object
682  * itself, so actual lookup of the hash should be once per lock object.
683  */
684 static inline struct lock_class *
685 look_up_lock_class(struct lockdep_map *lock, unsigned int subclass)
686 {
687         struct lockdep_subclass_key *key;
688         struct list_head *hash_head;
689         struct lock_class *class;
690
691 #ifdef CONFIG_DEBUG_LOCKDEP
692         /*
693          * If the architecture calls into lockdep before initializing
694          * the hashes then we'll warn about it later. (we cannot printk
695          * right now)
696          */
697         if (unlikely(!lockdep_initialized)) {
698                 lockdep_init();
699                 lockdep_init_error = 1;
700                 save_stack_trace(&lockdep_init_trace);
701         }
702 #endif
703
704         /*
705          * Static locks do not have their class-keys yet - for them the key
706          * is the lock object itself:
707          */
708         if (unlikely(!lock->key))
709                 lock->key = (void *)lock;
710
711         /*
712          * NOTE: the class-key must be unique. For dynamic locks, a static
713          * lock_class_key variable is passed in through the mutex_init()
714          * (or spin_lock_init()) call - which acts as the key. For static
715          * locks we use the lock object itself as the key.
716          */
717         BUILD_BUG_ON(sizeof(struct lock_class_key) >
718                         sizeof(struct lockdep_map));
719
720         key = lock->key->subkeys + subclass;
721
722         hash_head = classhashentry(key);
723
724         /*
725          * We can walk the hash lockfree, because the hash only
726          * grows, and we are careful when adding entries to the end:
727          */
728         list_for_each_entry(class, hash_head, hash_entry) {
729                 if (class->key == key) {
730                         WARN_ON_ONCE(class->name != lock->name);
731                         return class;
732                 }
733         }
734
735         return NULL;
736 }
737
738 /*
739  * Register a lock's class in the hash-table, if the class is not present
740  * yet. Otherwise we look it up. We cache the result in the lock object
741  * itself, so actual lookup of the hash should be once per lock object.
742  */
743 static inline struct lock_class *
744 register_lock_class(struct lockdep_map *lock, unsigned int subclass, int force)
745 {
746         struct lockdep_subclass_key *key;
747         struct list_head *hash_head;
748         struct lock_class *class;
749         unsigned long flags;
750
751         class = look_up_lock_class(lock, subclass);
752         if (likely(class))
753                 return class;
754
755         /*
756          * Debug-check: all keys must be persistent!
757          */
758         if (!static_obj(lock->key)) {
759                 debug_locks_off();
760                 printk("INFO: trying to register non-static key.\n");
761                 printk("the code is fine but needs lockdep annotation.\n");
762                 printk("turning off the locking correctness validator.\n");
763                 dump_stack();
764
765                 return NULL;
766         }
767
768         key = lock->key->subkeys + subclass;
769         hash_head = classhashentry(key);
770
771         raw_local_irq_save(flags);
772         if (!graph_lock()) {
773                 raw_local_irq_restore(flags);
774                 return NULL;
775         }
776         /*
777          * We have to do the hash-walk again, to avoid races
778          * with another CPU:
779          */
780         list_for_each_entry(class, hash_head, hash_entry)
781                 if (class->key == key)
782                         goto out_unlock_set;
783         /*
784          * Allocate a new key from the static array, and add it to
785          * the hash:
786          */
787         if (nr_lock_classes >= MAX_LOCKDEP_KEYS) {
788                 if (!debug_locks_off_graph_unlock()) {
789                         raw_local_irq_restore(flags);
790                         return NULL;
791                 }
792                 raw_local_irq_restore(flags);
793
794                 printk("BUG: MAX_LOCKDEP_KEYS too low!\n");
795                 printk("turning off the locking correctness validator.\n");
796                 dump_stack();
797                 return NULL;
798         }
799         class = lock_classes + nr_lock_classes++;
800         debug_atomic_inc(&nr_unused_locks);
801         class->key = key;
802         class->name = lock->name;
803         class->subclass = subclass;
804         INIT_LIST_HEAD(&class->lock_entry);
805         INIT_LIST_HEAD(&class->locks_before);
806         INIT_LIST_HEAD(&class->locks_after);
807         class->name_version = count_matching_names(class);
808         /*
809          * We use RCU's safe list-add method to make
810          * parallel walking of the hash-list safe:
811          */
812         list_add_tail_rcu(&class->hash_entry, hash_head);
813         /*
814          * Add it to the global list of classes:
815          */
816         list_add_tail_rcu(&class->lock_entry, &all_lock_classes);
817
818         if (verbose(class)) {
819                 graph_unlock();
820                 raw_local_irq_restore(flags);
821
822                 printk("\nnew class %p: %s", class->key, class->name);
823                 if (class->name_version > 1)
824                         printk("#%d", class->name_version);
825                 printk("\n");
826                 dump_stack();
827
828                 raw_local_irq_save(flags);
829                 if (!graph_lock()) {
830                         raw_local_irq_restore(flags);
831                         return NULL;
832                 }
833         }
834 out_unlock_set:
835         graph_unlock();
836         raw_local_irq_restore(flags);
837
838         if (!subclass || force)
839                 lock->class_cache = class;
840
841         if (DEBUG_LOCKS_WARN_ON(class->subclass != subclass))
842                 return NULL;
843
844         return class;
845 }
846
847 #ifdef CONFIG_PROVE_LOCKING
848 /*
849  * Allocate a lockdep entry. (assumes the graph_lock held, returns
850  * with NULL on failure)
851  */
852 static struct lock_list *alloc_list_entry(void)
853 {
854         if (nr_list_entries >= MAX_LOCKDEP_ENTRIES) {
855                 if (!debug_locks_off_graph_unlock())
856                         return NULL;
857
858                 printk("BUG: MAX_LOCKDEP_ENTRIES too low!\n");
859                 printk("turning off the locking correctness validator.\n");
860                 dump_stack();
861                 return NULL;
862         }
863         return list_entries + nr_list_entries++;
864 }
865
866 /*
867  * Add a new dependency to the head of the list:
868  */
869 static int add_lock_to_list(struct lock_class *class, struct lock_class *this,
870                             struct list_head *head, unsigned long ip, int distance)
871 {
872         struct lock_list *entry;
873         /*
874          * Lock not present yet - get a new dependency struct and
875          * add it to the list:
876          */
877         entry = alloc_list_entry();
878         if (!entry)
879                 return 0;
880
881         if (!save_trace(&entry->trace))
882                 return 0;
883
884         entry->class = this;
885         entry->distance = distance;
886         /*
887          * Since we never remove from the dependency list, the list can
888          * be walked lockless by other CPUs, it's only allocation
889          * that must be protected by the spinlock. But this also means
890          * we must make new entries visible only once writes to the
891          * entry become visible - hence the RCU op:
892          */
893         list_add_tail_rcu(&entry->entry, head);
894
895         return 1;
896 }
897
898 /*
899  * Recursive, forwards-direction lock-dependency checking, used for
900  * both noncyclic checking and for hardirq-unsafe/softirq-unsafe
901  * checking.
902  *
903  * (to keep the stackframe of the recursive functions small we
904  *  use these global variables, and we also mark various helper
905  *  functions as noinline.)
906  */
907 static struct held_lock *check_source, *check_target;
908
909 /*
910  * Print a dependency chain entry (this is only done when a deadlock
911  * has been detected):
912  */
913 static noinline int
914 print_circular_bug_entry(struct lock_list *target, unsigned int depth)
915 {
916         if (debug_locks_silent)
917                 return 0;
918         printk("\n-> #%u", depth);
919         print_lock_name(target->class);
920         printk(":\n");
921         print_stack_trace(&target->trace, 6);
922
923         return 0;
924 }
925
926 /*
927  * When a circular dependency is detected, print the
928  * header first:
929  */
930 static noinline int
931 print_circular_bug_header(struct lock_list *entry, unsigned int depth)
932 {
933         struct task_struct *curr = current;
934
935         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
936                 return 0;
937
938         printk("\n=======================================================\n");
939         printk(  "[ INFO: possible circular locking dependency detected ]\n");
940         print_kernel_version();
941         printk(  "-------------------------------------------------------\n");
942         printk("%s/%d is trying to acquire lock:\n",
943                 curr->comm, task_pid_nr(curr));
944         print_lock(check_source);
945         printk("\nbut task is already holding lock:\n");
946         print_lock(check_target);
947         printk("\nwhich lock already depends on the new lock.\n\n");
948         printk("\nthe existing dependency chain (in reverse order) is:\n");
949
950         print_circular_bug_entry(entry, depth);
951
952         return 0;
953 }
954
955 static noinline int print_circular_bug_tail(void)
956 {
957         struct task_struct *curr = current;
958         struct lock_list this;
959
960         if (debug_locks_silent)
961                 return 0;
962
963         this.class = hlock_class(check_source);
964         if (!save_trace(&this.trace))
965                 return 0;
966
967         print_circular_bug_entry(&this, 0);
968
969         printk("\nother info that might help us debug this:\n\n");
970         lockdep_print_held_locks(curr);
971
972         printk("\nstack backtrace:\n");
973         dump_stack();
974
975         return 0;
976 }
977
978 #define RECURSION_LIMIT 40
979
980 static int noinline print_infinite_recursion_bug(void)
981 {
982         if (!debug_locks_off_graph_unlock())
983                 return 0;
984
985         WARN_ON(1);
986
987         return 0;
988 }
989
990 unsigned long __lockdep_count_forward_deps(struct lock_class *class,
991                                            unsigned int depth)
992 {
993         struct lock_list *entry;
994         unsigned long ret = 1;
995
996         if (lockdep_dependency_visit(class, depth))
997                 return 0;
998
999         /*
1000          * Recurse this class's dependency list:
1001          */
1002         list_for_each_entry(entry, &class->locks_after, entry)
1003                 ret += __lockdep_count_forward_deps(entry->class, depth + 1);
1004
1005         return ret;
1006 }
1007
1008 unsigned long lockdep_count_forward_deps(struct lock_class *class)
1009 {
1010         unsigned long ret, flags;
1011
1012         local_irq_save(flags);
1013         __raw_spin_lock(&lockdep_lock);
1014         ret = __lockdep_count_forward_deps(class, 0);
1015         __raw_spin_unlock(&lockdep_lock);
1016         local_irq_restore(flags);
1017
1018         return ret;
1019 }
1020
1021 unsigned long __lockdep_count_backward_deps(struct lock_class *class,
1022                                             unsigned int depth)
1023 {
1024         struct lock_list *entry;
1025         unsigned long ret = 1;
1026
1027         if (lockdep_dependency_visit(class, depth))
1028                 return 0;
1029         /*
1030          * Recurse this class's dependency list:
1031          */
1032         list_for_each_entry(entry, &class->locks_before, entry)
1033                 ret += __lockdep_count_backward_deps(entry->class, depth + 1);
1034
1035         return ret;
1036 }
1037
1038 unsigned long lockdep_count_backward_deps(struct lock_class *class)
1039 {
1040         unsigned long ret, flags;
1041
1042         local_irq_save(flags);
1043         __raw_spin_lock(&lockdep_lock);
1044         ret = __lockdep_count_backward_deps(class, 0);
1045         __raw_spin_unlock(&lockdep_lock);
1046         local_irq_restore(flags);
1047
1048         return ret;
1049 }
1050
1051 /*
1052  * Prove that the dependency graph starting at <entry> can not
1053  * lead to <target>. Print an error and return 0 if it does.
1054  */
1055 static noinline int
1056 check_noncircular(struct lock_class *source, unsigned int depth)
1057 {
1058         struct lock_list *entry;
1059
1060         if (lockdep_dependency_visit(source, depth))
1061                 return 1;
1062
1063         debug_atomic_inc(&nr_cyclic_check_recursions);
1064         if (depth > max_recursion_depth)
1065                 max_recursion_depth = depth;
1066         if (depth >= RECURSION_LIMIT)
1067                 return print_infinite_recursion_bug();
1068         /*
1069          * Check this lock's dependency list:
1070          */
1071         list_for_each_entry(entry, &source->locks_after, entry) {
1072                 if (entry->class == hlock_class(check_target))
1073                         return print_circular_bug_header(entry, depth+1);
1074                 debug_atomic_inc(&nr_cyclic_checks);
1075                 if (!check_noncircular(entry->class, depth+1))
1076                         return print_circular_bug_entry(entry, depth+1);
1077         }
1078         return 1;
1079 }
1080
1081 #if defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING)
1082 /*
1083  * Forwards and backwards subgraph searching, for the purposes of
1084  * proving that two subgraphs can be connected by a new dependency
1085  * without creating any illegal irq-safe -> irq-unsafe lock dependency.
1086  */
1087 static enum lock_usage_bit find_usage_bit;
1088 static struct lock_class *forwards_match, *backwards_match;
1089
1090 /*
1091  * Find a node in the forwards-direction dependency sub-graph starting
1092  * at <source> that matches <find_usage_bit>.
1093  *
1094  * Return 2 if such a node exists in the subgraph, and put that node
1095  * into <forwards_match>.
1096  *
1097  * Return 1 otherwise and keep <forwards_match> unchanged.
1098  * Return 0 on error.
1099  */
1100 static noinline int
1101 find_usage_forwards(struct lock_class *source, unsigned int depth)
1102 {
1103         struct lock_list *entry;
1104         int ret;
1105
1106         if (lockdep_dependency_visit(source, depth))
1107                 return 1;
1108
1109         if (depth > max_recursion_depth)
1110                 max_recursion_depth = depth;
1111         if (depth >= RECURSION_LIMIT)
1112                 return print_infinite_recursion_bug();
1113
1114         debug_atomic_inc(&nr_find_usage_forwards_checks);
1115         if (source->usage_mask & (1 << find_usage_bit)) {
1116                 forwards_match = source;
1117                 return 2;
1118         }
1119
1120         /*
1121          * Check this lock's dependency list:
1122          */
1123         list_for_each_entry(entry, &source->locks_after, entry) {
1124                 debug_atomic_inc(&nr_find_usage_forwards_recursions);
1125                 ret = find_usage_forwards(entry->class, depth+1);
1126                 if (ret == 2 || ret == 0)
1127                         return ret;
1128         }
1129         return 1;
1130 }
1131
1132 /*
1133  * Find a node in the backwards-direction dependency sub-graph starting
1134  * at <source> that matches <find_usage_bit>.
1135  *
1136  * Return 2 if such a node exists in the subgraph, and put that node
1137  * into <backwards_match>.
1138  *
1139  * Return 1 otherwise and keep <backwards_match> unchanged.
1140  * Return 0 on error.
1141  */
1142 static noinline int
1143 find_usage_backwards(struct lock_class *source, unsigned int depth)
1144 {
1145         struct lock_list *entry;
1146         int ret;
1147
1148         if (lockdep_dependency_visit(source, depth))
1149                 return 1;
1150
1151         if (!__raw_spin_is_locked(&lockdep_lock))
1152                 return DEBUG_LOCKS_WARN_ON(1);
1153
1154         if (depth > max_recursion_depth)
1155                 max_recursion_depth = depth;
1156         if (depth >= RECURSION_LIMIT)
1157                 return print_infinite_recursion_bug();
1158
1159         debug_atomic_inc(&nr_find_usage_backwards_checks);
1160         if (source->usage_mask & (1 << find_usage_bit)) {
1161                 backwards_match = source;
1162                 return 2;
1163         }
1164
1165         if (!source && debug_locks_off_graph_unlock()) {
1166                 WARN_ON(1);
1167                 return 0;
1168         }
1169
1170         /*
1171          * Check this lock's dependency list:
1172          */
1173         list_for_each_entry(entry, &source->locks_before, entry) {
1174                 debug_atomic_inc(&nr_find_usage_backwards_recursions);
1175                 ret = find_usage_backwards(entry->class, depth+1);
1176                 if (ret == 2 || ret == 0)
1177                         return ret;
1178         }
1179         return 1;
1180 }
1181
1182 static int
1183 print_bad_irq_dependency(struct task_struct *curr,
1184                          struct held_lock *prev,
1185                          struct held_lock *next,
1186                          enum lock_usage_bit bit1,
1187                          enum lock_usage_bit bit2,
1188                          const char *irqclass)
1189 {
1190         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1191                 return 0;
1192
1193         printk("\n======================================================\n");
1194         printk(  "[ INFO: %s-safe -> %s-unsafe lock order detected ]\n",
1195                 irqclass, irqclass);
1196         print_kernel_version();
1197         printk(  "------------------------------------------------------\n");
1198         printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] is trying to acquire:\n",
1199                 curr->comm, task_pid_nr(curr),
1200                 curr->hardirq_context, hardirq_count() >> HARDIRQ_SHIFT,
1201                 curr->softirq_context, softirq_count() >> SOFTIRQ_SHIFT,
1202                 curr->hardirqs_enabled,
1203                 curr->softirqs_enabled);
1204         print_lock(next);
1205
1206         printk("\nand this task is already holding:\n");
1207         print_lock(prev);
1208         printk("which would create a new lock dependency:\n");
1209         print_lock_name(hlock_class(prev));
1210         printk(" ->");
1211         print_lock_name(hlock_class(next));
1212         printk("\n");
1213
1214         printk("\nbut this new dependency connects a %s-irq-safe lock:\n",
1215                 irqclass);
1216         print_lock_name(backwards_match);
1217         printk("\n... which became %s-irq-safe at:\n", irqclass);
1218
1219         print_stack_trace(backwards_match->usage_traces + bit1, 1);
1220
1221         printk("\nto a %s-irq-unsafe lock:\n", irqclass);
1222         print_lock_name(forwards_match);
1223         printk("\n... which became %s-irq-unsafe at:\n", irqclass);
1224         printk("...");
1225
1226         print_stack_trace(forwards_match->usage_traces + bit2, 1);
1227
1228         printk("\nother info that might help us debug this:\n\n");
1229         lockdep_print_held_locks(curr);
1230
1231         printk("\nthe %s-irq-safe lock's dependencies:\n", irqclass);
1232         print_lock_dependencies(backwards_match, 0);
1233
1234         printk("\nthe %s-irq-unsafe lock's dependencies:\n", irqclass);
1235         print_lock_dependencies(forwards_match, 0);
1236
1237         printk("\nstack backtrace:\n");
1238         dump_stack();
1239
1240         return 0;
1241 }
1242
1243 static int
1244 check_usage(struct task_struct *curr, struct held_lock *prev,
1245             struct held_lock *next, enum lock_usage_bit bit_backwards,
1246             enum lock_usage_bit bit_forwards, const char *irqclass)
1247 {
1248         int ret;
1249
1250         find_usage_bit = bit_backwards;
1251         /* fills in <backwards_match> */
1252         ret = find_usage_backwards(hlock_class(prev), 0);
1253         if (!ret || ret == 1)
1254                 return ret;
1255
1256         find_usage_bit = bit_forwards;
1257         ret = find_usage_forwards(hlock_class(next), 0);
1258         if (!ret || ret == 1)
1259                 return ret;
1260         /* ret == 2 */
1261         return print_bad_irq_dependency(curr, prev, next,
1262                         bit_backwards, bit_forwards, irqclass);
1263 }
1264
1265 static const char *state_names[] = {
1266 #define LOCKDEP_STATE(__STATE) \
1267         __stringify(__STATE),
1268 #include "lockdep_states.h"
1269 #undef LOCKDEP_STATE
1270 };
1271
1272 static const char *state_rnames[] = {
1273 #define LOCKDEP_STATE(__STATE) \
1274         __stringify(__STATE)"-READ",
1275 #include "lockdep_states.h"
1276 #undef LOCKDEP_STATE
1277 };
1278
1279 static inline const char *state_name(enum lock_usage_bit bit)
1280 {
1281         return (bit & 1) ? state_rnames[bit >> 2] : state_names[bit >> 2];
1282 }
1283
1284 static int exclusive_bit(int new_bit)
1285 {
1286         /*
1287          * USED_IN
1288          * USED_IN_READ
1289          * ENABLED
1290          * ENABLED_READ
1291          *
1292          * bit 0 - write/read
1293          * bit 1 - used_in/enabled
1294          * bit 2+  state
1295          */
1296
1297         int state = new_bit & ~3;
1298         int dir = new_bit & 2;
1299
1300         /*
1301          * keep state, bit flip the direction and strip read.
1302          */
1303         return state | (dir ^ 2);
1304 }
1305
1306 static int check_irq_usage(struct task_struct *curr, struct held_lock *prev,
1307                            struct held_lock *next, enum lock_usage_bit bit)
1308 {
1309         /*
1310          * Prove that the new dependency does not connect a hardirq-safe
1311          * lock with a hardirq-unsafe lock - to achieve this we search
1312          * the backwards-subgraph starting at <prev>, and the
1313          * forwards-subgraph starting at <next>:
1314          */
1315         if (!check_usage(curr, prev, next, bit,
1316                            exclusive_bit(bit), state_name(bit)))
1317                 return 0;
1318
1319         bit++; /* _READ */
1320
1321         /*
1322          * Prove that the new dependency does not connect a hardirq-safe-read
1323          * lock with a hardirq-unsafe lock - to achieve this we search
1324          * the backwards-subgraph starting at <prev>, and the
1325          * forwards-subgraph starting at <next>:
1326          */
1327         if (!check_usage(curr, prev, next, bit,
1328                            exclusive_bit(bit), state_name(bit)))
1329                 return 0;
1330
1331         return 1;
1332 }
1333
1334 static int
1335 check_prev_add_irq(struct task_struct *curr, struct held_lock *prev,
1336                 struct held_lock *next)
1337 {
1338 #define LOCKDEP_STATE(__STATE)                                          \
1339         if (!check_irq_usage(curr, prev, next, LOCK_USED_IN_##__STATE)) \
1340                 return 0;
1341 #include "lockdep_states.h"
1342 #undef LOCKDEP_STATE
1343
1344         return 1;
1345 }
1346
1347 static void inc_chains(void)
1348 {
1349         if (current->hardirq_context)
1350                 nr_hardirq_chains++;
1351         else {
1352                 if (current->softirq_context)
1353                         nr_softirq_chains++;
1354                 else
1355                         nr_process_chains++;
1356         }
1357 }
1358
1359 #else
1360
1361 static inline int
1362 check_prev_add_irq(struct task_struct *curr, struct held_lock *prev,
1363                 struct held_lock *next)
1364 {
1365         return 1;
1366 }
1367
1368 static inline void inc_chains(void)
1369 {
1370         nr_process_chains++;
1371 }
1372
1373 #endif
1374
1375 static int
1376 print_deadlock_bug(struct task_struct *curr, struct held_lock *prev,
1377                    struct held_lock *next)
1378 {
1379         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1380                 return 0;
1381
1382         printk("\n=============================================\n");
1383         printk(  "[ INFO: possible recursive locking detected ]\n");
1384         print_kernel_version();
1385         printk(  "---------------------------------------------\n");
1386         printk("%s/%d is trying to acquire lock:\n",
1387                 curr->comm, task_pid_nr(curr));
1388         print_lock(next);
1389         printk("\nbut task is already holding lock:\n");
1390         print_lock(prev);
1391
1392         printk("\nother info that might help us debug this:\n");
1393         lockdep_print_held_locks(curr);
1394
1395         printk("\nstack backtrace:\n");
1396         dump_stack();
1397
1398         return 0;
1399 }
1400
1401 /*
1402  * Check whether we are holding such a class already.
1403  *
1404  * (Note that this has to be done separately, because the graph cannot
1405  * detect such classes of deadlocks.)
1406  *
1407  * Returns: 0 on deadlock detected, 1 on OK, 2 on recursive read
1408  */
1409 static int
1410 check_deadlock(struct task_struct *curr, struct held_lock *next,
1411                struct lockdep_map *next_instance, int read)
1412 {
1413         struct held_lock *prev;
1414         struct held_lock *nest = NULL;
1415         int i;
1416
1417         for (i = 0; i < curr->lockdep_depth; i++) {
1418                 prev = curr->held_locks + i;
1419
1420                 if (prev->instance == next->nest_lock)
1421                         nest = prev;
1422
1423                 if (hlock_class(prev) != hlock_class(next))
1424                         continue;
1425
1426                 /*
1427                  * Allow read-after-read recursion of the same
1428                  * lock class (i.e. read_lock(lock)+read_lock(lock)):
1429                  */
1430                 if ((read == 2) && prev->read)
1431                         return 2;
1432
1433                 /*
1434                  * We're holding the nest_lock, which serializes this lock's
1435                  * nesting behaviour.
1436                  */
1437                 if (nest)
1438                         return 2;
1439
1440                 return print_deadlock_bug(curr, prev, next);
1441         }
1442         return 1;
1443 }
1444
1445 /*
1446  * There was a chain-cache miss, and we are about to add a new dependency
1447  * to a previous lock. We recursively validate the following rules:
1448  *
1449  *  - would the adding of the <prev> -> <next> dependency create a
1450  *    circular dependency in the graph? [== circular deadlock]
1451  *
1452  *  - does the new prev->next dependency connect any hardirq-safe lock
1453  *    (in the full backwards-subgraph starting at <prev>) with any
1454  *    hardirq-unsafe lock (in the full forwards-subgraph starting at
1455  *    <next>)? [== illegal lock inversion with hardirq contexts]
1456  *
1457  *  - does the new prev->next dependency connect any softirq-safe lock
1458  *    (in the full backwards-subgraph starting at <prev>) with any
1459  *    softirq-unsafe lock (in the full forwards-subgraph starting at
1460  *    <next>)? [== illegal lock inversion with softirq contexts]
1461  *
1462  * any of these scenarios could lead to a deadlock.
1463  *
1464  * Then if all the validations pass, we add the forwards and backwards
1465  * dependency.
1466  */
1467 static int
1468 check_prev_add(struct task_struct *curr, struct held_lock *prev,
1469                struct held_lock *next, int distance)
1470 {
1471         struct lock_list *entry;
1472         int ret;
1473
1474         /*
1475          * Prove that the new <prev> -> <next> dependency would not
1476          * create a circular dependency in the graph. (We do this by
1477          * forward-recursing into the graph starting at <next>, and
1478          * checking whether we can reach <prev>.)
1479          *
1480          * We are using global variables to control the recursion, to
1481          * keep the stackframe size of the recursive functions low:
1482          */
1483         check_source = next;
1484         check_target = prev;
1485         if (!(check_noncircular(hlock_class(next), 0)))
1486                 return print_circular_bug_tail();
1487
1488         if (!check_prev_add_irq(curr, prev, next))
1489                 return 0;
1490
1491         /*
1492          * For recursive read-locks we do all the dependency checks,
1493          * but we dont store read-triggered dependencies (only
1494          * write-triggered dependencies). This ensures that only the
1495          * write-side dependencies matter, and that if for example a
1496          * write-lock never takes any other locks, then the reads are
1497          * equivalent to a NOP.
1498          */
1499         if (next->read == 2 || prev->read == 2)
1500                 return 1;
1501         /*
1502          * Is the <prev> -> <next> dependency already present?
1503          *
1504          * (this may occur even though this is a new chain: consider
1505          *  e.g. the L1 -> L2 -> L3 -> L4 and the L5 -> L1 -> L2 -> L3
1506          *  chains - the second one will be new, but L1 already has
1507          *  L2 added to its dependency list, due to the first chain.)
1508          */
1509         list_for_each_entry(entry, &hlock_class(prev)->locks_after, entry) {
1510                 if (entry->class == hlock_class(next)) {
1511                         if (distance == 1)
1512                                 entry->distance = 1;
1513                         return 2;
1514                 }
1515         }
1516
1517         /*
1518          * Ok, all validations passed, add the new lock
1519          * to the previous lock's dependency list:
1520          */
1521         ret = add_lock_to_list(hlock_class(prev), hlock_class(next),
1522                                &hlock_class(prev)->locks_after,
1523                                next->acquire_ip, distance);
1524
1525         if (!ret)
1526                 return 0;
1527
1528         ret = add_lock_to_list(hlock_class(next), hlock_class(prev),
1529                                &hlock_class(next)->locks_before,
1530                                next->acquire_ip, distance);
1531         if (!ret)
1532                 return 0;
1533
1534         /*
1535          * Debugging printouts:
1536          */
1537         if (verbose(hlock_class(prev)) || verbose(hlock_class(next))) {
1538                 graph_unlock();
1539                 printk("\n new dependency: ");
1540                 print_lock_name(hlock_class(prev));
1541                 printk(" => ");
1542                 print_lock_name(hlock_class(next));
1543                 printk("\n");
1544                 dump_stack();
1545                 return graph_lock();
1546         }
1547         return 1;
1548 }
1549
1550 /*
1551  * Add the dependency to all directly-previous locks that are 'relevant'.
1552  * The ones that are relevant are (in increasing distance from curr):
1553  * all consecutive trylock entries and the final non-trylock entry - or
1554  * the end of this context's lock-chain - whichever comes first.
1555  */
1556 static int
1557 check_prevs_add(struct task_struct *curr, struct held_lock *next)
1558 {
1559         int depth = curr->lockdep_depth;
1560         struct held_lock *hlock;
1561
1562         /*
1563          * Debugging checks.
1564          *
1565          * Depth must not be zero for a non-head lock:
1566          */
1567         if (!depth)
1568                 goto out_bug;
1569         /*
1570          * At least two relevant locks must exist for this
1571          * to be a head:
1572          */
1573         if (curr->held_locks[depth].irq_context !=
1574                         curr->held_locks[depth-1].irq_context)
1575                 goto out_bug;
1576
1577         for (;;) {
1578                 int distance = curr->lockdep_depth - depth + 1;
1579                 hlock = curr->held_locks + depth-1;
1580                 /*
1581                  * Only non-recursive-read entries get new dependencies
1582                  * added:
1583                  */
1584                 if (hlock->read != 2) {
1585                         if (!check_prev_add(curr, hlock, next, distance))
1586                                 return 0;
1587                         /*
1588                          * Stop after the first non-trylock entry,
1589                          * as non-trylock entries have added their
1590                          * own direct dependencies already, so this
1591                          * lock is connected to them indirectly:
1592                          */
1593                         if (!hlock->trylock)
1594                                 break;
1595                 }
1596                 depth--;
1597                 /*
1598                  * End of lock-stack?
1599                  */
1600                 if (!depth)
1601                         break;
1602                 /*
1603                  * Stop the search if we cross into another context:
1604                  */
1605                 if (curr->held_locks[depth].irq_context !=
1606                                 curr->held_locks[depth-1].irq_context)
1607                         break;
1608         }
1609         return 1;
1610 out_bug:
1611         if (!debug_locks_off_graph_unlock())
1612                 return 0;
1613
1614         WARN_ON(1);
1615
1616         return 0;
1617 }
1618
1619 unsigned long nr_lock_chains;
1620 struct lock_chain lock_chains[MAX_LOCKDEP_CHAINS];
1621 int nr_chain_hlocks;
1622 static u16 chain_hlocks[MAX_LOCKDEP_CHAIN_HLOCKS];
1623
1624 struct lock_class *lock_chain_get_class(struct lock_chain *chain, int i)
1625 {
1626         return lock_classes + chain_hlocks[chain->base + i];
1627 }
1628
1629 /*
1630  * Look up a dependency chain. If the key is not present yet then
1631  * add it and return 1 - in this case the new dependency chain is
1632  * validated. If the key is already hashed, return 0.
1633  * (On return with 1 graph_lock is held.)
1634  */
1635 static inline int lookup_chain_cache(struct task_struct *curr,
1636                                      struct held_lock *hlock,
1637                                      u64 chain_key)
1638 {
1639         struct lock_class *class = hlock_class(hlock);
1640         struct list_head *hash_head = chainhashentry(chain_key);
1641         struct lock_chain *chain;
1642         struct held_lock *hlock_curr, *hlock_next;
1643         int i, j, n, cn;
1644
1645         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
1646                 return 0;
1647         /*
1648          * We can walk it lock-free, because entries only get added
1649          * to the hash:
1650          */
1651         list_for_each_entry(chain, hash_head, entry) {
1652                 if (chain->chain_key == chain_key) {
1653 cache_hit:
1654                         debug_atomic_inc(&chain_lookup_hits);
1655                         if (very_verbose(class))
1656                                 printk("\nhash chain already cached, key: "
1657                                         "%016Lx tail class: [%p] %s\n",
1658                                         (unsigned long long)chain_key,
1659                                         class->key, class->name);
1660                         return 0;
1661                 }
1662         }
1663         if (very_verbose(class))
1664                 printk("\nnew hash chain, key: %016Lx tail class: [%p] %s\n",
1665                         (unsigned long long)chain_key, class->key, class->name);
1666         /*
1667          * Allocate a new chain entry from the static array, and add
1668          * it to the hash:
1669          */
1670         if (!graph_lock())
1671                 return 0;
1672         /*
1673          * We have to walk the chain again locked - to avoid duplicates:
1674          */
1675         list_for_each_entry(chain, hash_head, entry) {
1676                 if (chain->chain_key == chain_key) {
1677                         graph_unlock();
1678                         goto cache_hit;
1679                 }
1680         }
1681         if (unlikely(nr_lock_chains >= MAX_LOCKDEP_CHAINS)) {
1682                 if (!debug_locks_off_graph_unlock())
1683                         return 0;
1684
1685                 printk("BUG: MAX_LOCKDEP_CHAINS too low!\n");
1686                 printk("turning off the locking correctness validator.\n");
1687                 dump_stack();
1688                 return 0;
1689         }
1690         chain = lock_chains + nr_lock_chains++;
1691         chain->chain_key = chain_key;
1692         chain->irq_context = hlock->irq_context;
1693         /* Find the first held_lock of current chain */
1694         hlock_next = hlock;
1695         for (i = curr->lockdep_depth - 1; i >= 0; i--) {
1696                 hlock_curr = curr->held_locks + i;
1697                 if (hlock_curr->irq_context != hlock_next->irq_context)
1698                         break;
1699                 hlock_next = hlock;
1700         }
1701         i++;
1702         chain->depth = curr->lockdep_depth + 1 - i;
1703         cn = nr_chain_hlocks;
1704         while (cn + chain->depth <= MAX_LOCKDEP_CHAIN_HLOCKS) {
1705                 n = cmpxchg(&nr_chain_hlocks, cn, cn + chain->depth);
1706                 if (n == cn)
1707                         break;
1708                 cn = n;
1709         }
1710         if (likely(cn + chain->depth <= MAX_LOCKDEP_CHAIN_HLOCKS)) {
1711                 chain->base = cn;
1712                 for (j = 0; j < chain->depth - 1; j++, i++) {
1713                         int lock_id = curr->held_locks[i].class_idx - 1;
1714                         chain_hlocks[chain->base + j] = lock_id;
1715                 }
1716                 chain_hlocks[chain->base + j] = class - lock_classes;
1717         }
1718         list_add_tail_rcu(&chain->entry, hash_head);
1719         debug_atomic_inc(&chain_lookup_misses);
1720         inc_chains();
1721
1722         return 1;
1723 }
1724
1725 static int validate_chain(struct task_struct *curr, struct lockdep_map *lock,
1726                 struct held_lock *hlock, int chain_head, u64 chain_key)
1727 {
1728         /*
1729          * Trylock needs to maintain the stack of held locks, but it
1730          * does not add new dependencies, because trylock can be done
1731          * in any order.
1732          *
1733          * We look up the chain_key and do the O(N^2) check and update of
1734          * the dependencies only if this is a new dependency chain.
1735          * (If lookup_chain_cache() returns with 1 it acquires
1736          * graph_lock for us)
1737          */
1738         if (!hlock->trylock && (hlock->check == 2) &&
1739             lookup_chain_cache(curr, hlock, chain_key)) {
1740                 /*
1741                  * Check whether last held lock:
1742                  *
1743                  * - is irq-safe, if this lock is irq-unsafe
1744                  * - is softirq-safe, if this lock is hardirq-unsafe
1745                  *
1746                  * And check whether the new lock's dependency graph
1747                  * could lead back to the previous lock.
1748                  *
1749                  * any of these scenarios could lead to a deadlock. If
1750                  * All validations
1751                  */
1752                 int ret = check_deadlock(curr, hlock, lock, hlock->read);
1753
1754                 if (!ret)
1755                         return 0;
1756                 /*
1757                  * Mark recursive read, as we jump over it when
1758                  * building dependencies (just like we jump over
1759                  * trylock entries):
1760                  */
1761                 if (ret == 2)
1762                         hlock->read = 2;
1763                 /*
1764                  * Add dependency only if this lock is not the head
1765                  * of the chain, and if it's not a secondary read-lock:
1766                  */
1767                 if (!chain_head && ret != 2)
1768                         if (!check_prevs_add(curr, hlock))
1769                                 return 0;
1770                 graph_unlock();
1771         } else
1772                 /* after lookup_chain_cache(): */
1773                 if (unlikely(!debug_locks))
1774                         return 0;
1775
1776         return 1;
1777 }
1778 #else
1779 static inline int validate_chain(struct task_struct *curr,
1780                 struct lockdep_map *lock, struct held_lock *hlock,
1781                 int chain_head, u64 chain_key)
1782 {
1783         return 1;
1784 }
1785 #endif
1786
1787 /*
1788  * We are building curr_chain_key incrementally, so double-check
1789  * it from scratch, to make sure that it's done correctly:
1790  */
1791 static void check_chain_key(struct task_struct *curr)
1792 {
1793 #ifdef CONFIG_DEBUG_LOCKDEP
1794         struct held_lock *hlock, *prev_hlock = NULL;
1795         unsigned int i, id;
1796         u64 chain_key = 0;
1797
1798         for (i = 0; i < curr->lockdep_depth; i++) {
1799                 hlock = curr->held_locks + i;
1800                 if (chain_key != hlock->prev_chain_key) {
1801                         debug_locks_off();
1802                         WARN(1, "hm#1, depth: %u [%u], %016Lx != %016Lx\n",
1803                                 curr->lockdep_depth, i,
1804                                 (unsigned long long)chain_key,
1805                                 (unsigned long long)hlock->prev_chain_key);
1806                         return;
1807                 }
1808                 id = hlock->class_idx - 1;
1809                 if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
1810                         return;
1811
1812                 if (prev_hlock && (prev_hlock->irq_context !=
1813                                                         hlock->irq_context))
1814                         chain_key = 0;
1815                 chain_key = iterate_chain_key(chain_key, id);
1816                 prev_hlock = hlock;
1817         }
1818         if (chain_key != curr->curr_chain_key) {
1819                 debug_locks_off();
1820                 WARN(1, "hm#2, depth: %u [%u], %016Lx != %016Lx\n",
1821                         curr->lockdep_depth, i,
1822                         (unsigned long long)chain_key,
1823                         (unsigned long long)curr->curr_chain_key);
1824         }
1825 #endif
1826 }
1827
1828 static int
1829 print_usage_bug(struct task_struct *curr, struct held_lock *this,
1830                 enum lock_usage_bit prev_bit, enum lock_usage_bit new_bit)
1831 {
1832         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1833                 return 0;
1834
1835         printk("\n=================================\n");
1836         printk(  "[ INFO: inconsistent lock state ]\n");
1837         print_kernel_version();
1838         printk(  "---------------------------------\n");
1839
1840         printk("inconsistent {%s} -> {%s} usage.\n",
1841                 usage_str[prev_bit], usage_str[new_bit]);
1842
1843         printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] takes:\n",
1844                 curr->comm, task_pid_nr(curr),
1845                 trace_hardirq_context(curr), hardirq_count() >> HARDIRQ_SHIFT,
1846                 trace_softirq_context(curr), softirq_count() >> SOFTIRQ_SHIFT,
1847                 trace_hardirqs_enabled(curr),
1848                 trace_softirqs_enabled(curr));
1849         print_lock(this);
1850
1851         printk("{%s} state was registered at:\n", usage_str[prev_bit]);
1852         print_stack_trace(hlock_class(this)->usage_traces + prev_bit, 1);
1853
1854         print_irqtrace_events(curr);
1855         printk("\nother info that might help us debug this:\n");
1856         lockdep_print_held_locks(curr);
1857
1858         printk("\nstack backtrace:\n");
1859         dump_stack();
1860
1861         return 0;
1862 }
1863
1864 /*
1865  * Print out an error if an invalid bit is set:
1866  */
1867 static inline int
1868 valid_state(struct task_struct *curr, struct held_lock *this,
1869             enum lock_usage_bit new_bit, enum lock_usage_bit bad_bit)
1870 {
1871         if (unlikely(hlock_class(this)->usage_mask & (1 << bad_bit)))
1872                 return print_usage_bug(curr, this, bad_bit, new_bit);
1873         return 1;
1874 }
1875
1876 static int mark_lock(struct task_struct *curr, struct held_lock *this,
1877                      enum lock_usage_bit new_bit);
1878
1879 #if defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING)
1880
1881 /*
1882  * print irq inversion bug:
1883  */
1884 static int
1885 print_irq_inversion_bug(struct task_struct *curr, struct lock_class *other,
1886                         struct held_lock *this, int forwards,
1887                         const char *irqclass)
1888 {
1889         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1890                 return 0;
1891
1892         printk("\n=========================================================\n");
1893         printk(  "[ INFO: possible irq lock inversion dependency detected ]\n");
1894         print_kernel_version();
1895         printk(  "---------------------------------------------------------\n");
1896         printk("%s/%d just changed the state of lock:\n",
1897                 curr->comm, task_pid_nr(curr));
1898         print_lock(this);
1899         if (forwards)
1900                 printk("but this lock took another, %s-unsafe lock in the past:\n", irqclass);
1901         else
1902                 printk("but this lock was taken by another, %s-safe lock in the past:\n", irqclass);
1903         print_lock_name(other);
1904         printk("\n\nand interrupts could create inverse lock ordering between them.\n\n");
1905
1906         printk("\nother info that might help us debug this:\n");
1907         lockdep_print_held_locks(curr);
1908
1909         printk("\nthe first lock's dependencies:\n");
1910         print_lock_dependencies(hlock_class(this), 0);
1911
1912         printk("\nthe second lock's dependencies:\n");
1913         print_lock_dependencies(other, 0);
1914
1915         printk("\nstack backtrace:\n");
1916         dump_stack();
1917
1918         return 0;
1919 }
1920
1921 /*
1922  * Prove that in the forwards-direction subgraph starting at <this>
1923  * there is no lock matching <mask>:
1924  */
1925 static int
1926 check_usage_forwards(struct task_struct *curr, struct held_lock *this,
1927                      enum lock_usage_bit bit, const char *irqclass)
1928 {
1929         int ret;
1930
1931         find_usage_bit = bit;
1932         /* fills in <forwards_match> */
1933         ret = find_usage_forwards(hlock_class(this), 0);
1934         if (!ret || ret == 1)
1935                 return ret;
1936
1937         return print_irq_inversion_bug(curr, forwards_match, this, 1, irqclass);
1938 }
1939
1940 /*
1941  * Prove that in the backwards-direction subgraph starting at <this>
1942  * there is no lock matching <mask>:
1943  */
1944 static int
1945 check_usage_backwards(struct task_struct *curr, struct held_lock *this,
1946                       enum lock_usage_bit bit, const char *irqclass)
1947 {
1948         int ret;
1949
1950         find_usage_bit = bit;
1951         /* fills in <backwards_match> */
1952         ret = find_usage_backwards(hlock_class(this), 0);
1953         if (!ret || ret == 1)
1954                 return ret;
1955
1956         return print_irq_inversion_bug(curr, backwards_match, this, 0, irqclass);
1957 }
1958
1959 void print_irqtrace_events(struct task_struct *curr)
1960 {
1961         printk("irq event stamp: %u\n", curr->irq_events);
1962         printk("hardirqs last  enabled at (%u): ", curr->hardirq_enable_event);
1963         print_ip_sym(curr->hardirq_enable_ip);
1964         printk("hardirqs last disabled at (%u): ", curr->hardirq_disable_event);
1965         print_ip_sym(curr->hardirq_disable_ip);
1966         printk("softirqs last  enabled at (%u): ", curr->softirq_enable_event);
1967         print_ip_sym(curr->softirq_enable_ip);
1968         printk("softirqs last disabled at (%u): ", curr->softirq_disable_event);
1969         print_ip_sym(curr->softirq_disable_ip);
1970 }
1971
1972 static int HARDIRQ_verbose(struct lock_class *class)
1973 {
1974 #if HARDIRQ_VERBOSE
1975         return class_filter(class);
1976 #endif
1977         return 0;
1978 }
1979
1980 static int SOFTIRQ_verbose(struct lock_class *class)
1981 {
1982 #if SOFTIRQ_VERBOSE
1983         return class_filter(class);
1984 #endif
1985         return 0;
1986 }
1987
1988 static int RECLAIM_FS_verbose(struct lock_class *class)
1989 {
1990 #if RECLAIM_VERBOSE
1991         return class_filter(class);
1992 #endif
1993         return 0;
1994 }
1995
1996 #define STRICT_READ_CHECKS      1
1997
1998 static int (*state_verbose_f[])(struct lock_class *class) = {
1999 #define LOCKDEP_STATE(__STATE) \
2000         __STATE##_verbose,
2001 #include "lockdep_states.h"
2002 #undef LOCKDEP_STATE
2003 };
2004
2005 static inline int state_verbose(enum lock_usage_bit bit,
2006                                 struct lock_class *class)
2007 {
2008         return state_verbose_f[bit >> 2](class);
2009 }
2010
2011 typedef int (*check_usage_f)(struct task_struct *, struct held_lock *,
2012                              enum lock_usage_bit bit, const char *name);
2013
2014 static int
2015 mark_lock_irq(struct task_struct *curr, struct held_lock *this,
2016                 enum lock_usage_bit new_bit)
2017 {
2018         int excl_bit = exclusive_bit(new_bit);
2019         int read = new_bit & 1;
2020         int dir = new_bit & 2;
2021
2022         /*
2023          * mark USED_IN has to look forwards -- to ensure no dependency
2024          * has ENABLED state, which would allow recursion deadlocks.
2025          *
2026          * mark ENABLED has to look backwards -- to ensure no dependee
2027          * has USED_IN state, which, again, would allow  recursion deadlocks.
2028          */
2029         check_usage_f usage = dir ?
2030                 check_usage_backwards : check_usage_forwards;
2031
2032         /*
2033          * Validate that this particular lock does not have conflicting
2034          * usage states.
2035          */
2036         if (!valid_state(curr, this, new_bit, excl_bit))
2037                 return 0;
2038
2039         /*
2040          * Validate that the lock dependencies don't have conflicting usage
2041          * states.
2042          */
2043         if ((!read || !dir || STRICT_READ_CHECKS) &&
2044                         !usage(curr, this, excl_bit, state_name(new_bit & ~1)))
2045                 return 0;
2046
2047         /*
2048          * Check for read in write conflicts
2049          */
2050         if (!read) {
2051                 if (!valid_state(curr, this, new_bit, excl_bit + 1))
2052                         return 0;
2053
2054                 if (STRICT_READ_CHECKS &&
2055                         !usage(curr, this, excl_bit + 1,
2056                                 state_name(new_bit + 1)))
2057                         return 0;
2058         }
2059
2060         if (state_verbose(new_bit, hlock_class(this)))
2061                 return 2;
2062
2063         return 1;
2064 }
2065
2066 enum mark_type {
2067 #define LOCKDEP_STATE(__STATE)  __STATE,
2068 #include "lockdep_states.h"
2069 #undef LOCKDEP_STATE
2070 };
2071
2072 /*
2073  * Mark all held locks with a usage bit:
2074  */
2075 static int
2076 mark_held_locks(struct task_struct *curr, enum mark_type mark)
2077 {
2078         enum lock_usage_bit usage_bit;
2079         struct held_lock *hlock;
2080         int i;
2081
2082         for (i = 0; i < curr->lockdep_depth; i++) {
2083                 hlock = curr->held_locks + i;
2084
2085                 usage_bit = 2 + (mark << 2); /* ENABLED */
2086                 if (hlock->read)
2087                         usage_bit += 1; /* READ */
2088
2089                 BUG_ON(usage_bit >= LOCK_USAGE_STATES);
2090
2091                 if (!mark_lock(curr, hlock, usage_bit))
2092                         return 0;
2093         }
2094
2095         return 1;
2096 }
2097
2098 /*
2099  * Debugging helper: via this flag we know that we are in
2100  * 'early bootup code', and will warn about any invalid irqs-on event:
2101  */
2102 static int early_boot_irqs_enabled;
2103
2104 void early_boot_irqs_off(void)
2105 {
2106         early_boot_irqs_enabled = 0;
2107 }
2108
2109 void early_boot_irqs_on(void)
2110 {
2111         early_boot_irqs_enabled = 1;
2112 }
2113
2114 /*
2115  * Hardirqs will be enabled:
2116  */
2117 void trace_hardirqs_on_caller(unsigned long ip)
2118 {
2119         struct task_struct *curr = current;
2120
2121         time_hardirqs_on(CALLER_ADDR0, ip);
2122
2123         if (unlikely(!debug_locks || current->lockdep_recursion))
2124                 return;
2125
2126         if (DEBUG_LOCKS_WARN_ON(unlikely(!early_boot_irqs_enabled)))
2127                 return;
2128
2129         if (unlikely(curr->hardirqs_enabled)) {
2130                 debug_atomic_inc(&redundant_hardirqs_on);
2131                 return;
2132         }
2133         /* we'll do an OFF -> ON transition: */
2134         curr->hardirqs_enabled = 1;
2135
2136         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2137                 return;
2138         if (DEBUG_LOCKS_WARN_ON(current->hardirq_context))
2139                 return;
2140         /*
2141          * We are going to turn hardirqs on, so set the
2142          * usage bit for all held locks:
2143          */
2144         if (!mark_held_locks(curr, HARDIRQ))
2145                 return;
2146         /*
2147          * If we have softirqs enabled, then set the usage
2148          * bit for all held locks. (disabled hardirqs prevented
2149          * this bit from being set before)
2150          */
2151         if (curr->softirqs_enabled)
2152                 if (!mark_held_locks(curr, SOFTIRQ))
2153                         return;
2154
2155         curr->hardirq_enable_ip = ip;
2156         curr->hardirq_enable_event = ++curr->irq_events;
2157         debug_atomic_inc(&hardirqs_on_events);
2158 }
2159 EXPORT_SYMBOL(trace_hardirqs_on_caller);
2160
2161 void trace_hardirqs_on(void)
2162 {
2163         trace_hardirqs_on_caller(CALLER_ADDR0);
2164 }
2165 EXPORT_SYMBOL(trace_hardirqs_on);
2166
2167 /*
2168  * Hardirqs were disabled:
2169  */
2170 void trace_hardirqs_off_caller(unsigned long ip)
2171 {
2172         struct task_struct *curr = current;
2173
2174         time_hardirqs_off(CALLER_ADDR0, ip);
2175
2176         if (unlikely(!debug_locks || current->lockdep_recursion))
2177                 return;
2178
2179         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2180                 return;
2181
2182         if (curr->hardirqs_enabled) {
2183                 /*
2184                  * We have done an ON -> OFF transition:
2185                  */
2186                 curr->hardirqs_enabled = 0;
2187                 curr->hardirq_disable_ip = ip;
2188                 curr->hardirq_disable_event = ++curr->irq_events;
2189                 debug_atomic_inc(&hardirqs_off_events);
2190         } else
2191                 debug_atomic_inc(&redundant_hardirqs_off);
2192 }
2193 EXPORT_SYMBOL(trace_hardirqs_off_caller);
2194
2195 void trace_hardirqs_off(void)
2196 {
2197         trace_hardirqs_off_caller(CALLER_ADDR0);
2198 }
2199 EXPORT_SYMBOL(trace_hardirqs_off);
2200
2201 /*
2202  * Softirqs will be enabled:
2203  */
2204 void trace_softirqs_on(unsigned long ip)
2205 {
2206         struct task_struct *curr = current;
2207
2208         if (unlikely(!debug_locks))
2209                 return;
2210
2211         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2212                 return;
2213
2214         if (curr->softirqs_enabled) {
2215                 debug_atomic_inc(&redundant_softirqs_on);
2216                 return;
2217         }
2218
2219         /*
2220          * We'll do an OFF -> ON transition:
2221          */
2222         curr->softirqs_enabled = 1;
2223         curr->softirq_enable_ip = ip;
2224         curr->softirq_enable_event = ++curr->irq_events;
2225         debug_atomic_inc(&softirqs_on_events);
2226         /*
2227          * We are going to turn softirqs on, so set the
2228          * usage bit for all held locks, if hardirqs are
2229          * enabled too:
2230          */
2231         if (curr->hardirqs_enabled)
2232                 mark_held_locks(curr, SOFTIRQ);
2233 }
2234
2235 /*
2236  * Softirqs were disabled:
2237  */
2238 void trace_softirqs_off(unsigned long ip)
2239 {
2240         struct task_struct *curr = current;
2241
2242         if (unlikely(!debug_locks))
2243                 return;
2244
2245         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2246                 return;
2247
2248         if (curr->softirqs_enabled) {
2249                 /*
2250                  * We have done an ON -> OFF transition:
2251                  */
2252                 curr->softirqs_enabled = 0;
2253                 curr->softirq_disable_ip = ip;
2254                 curr->softirq_disable_event = ++curr->irq_events;
2255                 debug_atomic_inc(&softirqs_off_events);
2256                 DEBUG_LOCKS_WARN_ON(!softirq_count());
2257         } else
2258                 debug_atomic_inc(&redundant_softirqs_off);
2259 }
2260
2261 static void __lockdep_trace_alloc(gfp_t gfp_mask, unsigned long flags)
2262 {
2263         struct task_struct *curr = current;
2264
2265         if (unlikely(!debug_locks))
2266                 return;
2267
2268         /* no reclaim without waiting on it */
2269         if (!(gfp_mask & __GFP_WAIT))
2270                 return;
2271
2272         /* this guy won't enter reclaim */
2273         if ((curr->flags & PF_MEMALLOC) && !(gfp_mask & __GFP_NOMEMALLOC))
2274                 return;
2275
2276         /* We're only interested __GFP_FS allocations for now */
2277         if (!(gfp_mask & __GFP_FS))
2278                 return;
2279
2280         if (DEBUG_LOCKS_WARN_ON(irqs_disabled_flags(flags)))
2281                 return;
2282
2283         mark_held_locks(curr, RECLAIM_FS);
2284 }
2285
2286 static void check_flags(unsigned long flags);
2287
2288 void lockdep_trace_alloc(gfp_t gfp_mask)
2289 {
2290         unsigned long flags;
2291
2292         if (unlikely(current->lockdep_recursion))
2293                 return;
2294
2295         raw_local_irq_save(flags);
2296         check_flags(flags);
2297         current->lockdep_recursion = 1;
2298         __lockdep_trace_alloc(gfp_mask, flags);
2299         current->lockdep_recursion = 0;
2300         raw_local_irq_restore(flags);
2301 }
2302
2303 static int mark_irqflags(struct task_struct *curr, struct held_lock *hlock)
2304 {
2305         /*
2306          * If non-trylock use in a hardirq or softirq context, then
2307          * mark the lock as used in these contexts:
2308          */
2309         if (!hlock->trylock) {
2310                 if (hlock->read) {
2311                         if (curr->hardirq_context)
2312                                 if (!mark_lock(curr, hlock,
2313                                                 LOCK_USED_IN_HARDIRQ_READ))
2314                                         return 0;
2315                         if (curr->softirq_context)
2316                                 if (!mark_lock(curr, hlock,
2317                                                 LOCK_USED_IN_SOFTIRQ_READ))
2318                                         return 0;
2319                 } else {
2320                         if (curr->hardirq_context)
2321                                 if (!mark_lock(curr, hlock, LOCK_USED_IN_HARDIRQ))
2322                                         return 0;
2323                         if (curr->softirq_context)
2324                                 if (!mark_lock(curr, hlock, LOCK_USED_IN_SOFTIRQ))
2325                                         return 0;
2326                 }
2327         }
2328         if (!hlock->hardirqs_off) {
2329                 if (hlock->read) {
2330                         if (!mark_lock(curr, hlock,
2331                                         LOCK_ENABLED_HARDIRQ_READ))
2332                                 return 0;
2333                         if (curr->softirqs_enabled)
2334                                 if (!mark_lock(curr, hlock,
2335                                                 LOCK_ENABLED_SOFTIRQ_READ))
2336                                         return 0;
2337                 } else {
2338                         if (!mark_lock(curr, hlock,
2339                                         LOCK_ENABLED_HARDIRQ))
2340                                 return 0;
2341                         if (curr->softirqs_enabled)
2342                                 if (!mark_lock(curr, hlock,
2343                                                 LOCK_ENABLED_SOFTIRQ))
2344                                         return 0;
2345                 }
2346         }
2347
2348         /*
2349          * We reuse the irq context infrastructure more broadly as a general
2350          * context checking code. This tests GFP_FS recursion (a lock taken
2351          * during reclaim for a GFP_FS allocation is held over a GFP_FS
2352          * allocation).
2353          */
2354         if (!hlock->trylock && (curr->lockdep_reclaim_gfp & __GFP_FS)) {
2355                 if (hlock->read) {
2356                         if (!mark_lock(curr, hlock, LOCK_USED_IN_RECLAIM_FS_READ))
2357                                         return 0;
2358                 } else {
2359                         if (!mark_lock(curr, hlock, LOCK_USED_IN_RECLAIM_FS))
2360                                         return 0;
2361                 }
2362         }
2363
2364         return 1;
2365 }
2366
2367 static int separate_irq_context(struct task_struct *curr,
2368                 struct held_lock *hlock)
2369 {
2370         unsigned int depth = curr->lockdep_depth;
2371
2372         /*
2373          * Keep track of points where we cross into an interrupt context:
2374          */
2375         hlock->irq_context = 2*(curr->hardirq_context ? 1 : 0) +
2376                                 curr->softirq_context;
2377         if (depth) {
2378                 struct held_lock *prev_hlock;
2379
2380                 prev_hlock = curr->held_locks + depth-1;
2381                 /*
2382                  * If we cross into another context, reset the
2383                  * hash key (this also prevents the checking and the
2384                  * adding of the dependency to 'prev'):
2385                  */
2386                 if (prev_hlock->irq_context != hlock->irq_context)
2387                         return 1;
2388         }
2389         return 0;
2390 }
2391
2392 #else
2393
2394 static inline
2395 int mark_lock_irq(struct task_struct *curr, struct held_lock *this,
2396                 enum lock_usage_bit new_bit)
2397 {
2398         WARN_ON(1);
2399         return 1;
2400 }
2401
2402 static inline int mark_irqflags(struct task_struct *curr,
2403                 struct held_lock *hlock)
2404 {
2405         return 1;
2406 }
2407
2408 static inline int separate_irq_context(struct task_struct *curr,
2409                 struct held_lock *hlock)
2410 {
2411         return 0;
2412 }
2413
2414 void lockdep_trace_alloc(gfp_t gfp_mask)
2415 {
2416 }
2417
2418 #endif
2419
2420 /*
2421  * Mark a lock with a usage bit, and validate the state transition:
2422  */
2423 static int mark_lock(struct task_struct *curr, struct held_lock *this,
2424                              enum lock_usage_bit new_bit)
2425 {
2426         unsigned int new_mask = 1 << new_bit, ret = 1;
2427
2428         /*
2429          * If already set then do not dirty the cacheline,
2430          * nor do any checks:
2431          */
2432         if (likely(hlock_class(this)->usage_mask & new_mask))
2433                 return 1;
2434
2435         if (!graph_lock())
2436                 return 0;
2437         /*
2438          * Make sure we didnt race:
2439          */
2440         if (unlikely(hlock_class(this)->usage_mask & new_mask)) {
2441                 graph_unlock();
2442                 return 1;
2443         }
2444
2445         hlock_class(this)->usage_mask |= new_mask;
2446
2447         if (!save_trace(hlock_class(this)->usage_traces + new_bit))
2448                 return 0;
2449
2450         switch (new_bit) {
2451 #define LOCKDEP_STATE(__STATE)                  \
2452         case LOCK_USED_IN_##__STATE:            \
2453         case LOCK_USED_IN_##__STATE##_READ:     \
2454         case LOCK_ENABLED_##__STATE:            \
2455         case LOCK_ENABLED_##__STATE##_READ:
2456 #include "lockdep_states.h"
2457 #undef LOCKDEP_STATE
2458                 ret = mark_lock_irq(curr, this, new_bit);
2459                 if (!ret)
2460                         return 0;
2461                 break;
2462         case LOCK_USED:
2463                 debug_atomic_dec(&nr_unused_locks);
2464                 break;
2465         default:
2466                 if (!debug_locks_off_graph_unlock())
2467                         return 0;
2468                 WARN_ON(1);
2469                 return 0;
2470         }
2471
2472         graph_unlock();
2473
2474         /*
2475          * We must printk outside of the graph_lock:
2476          */
2477         if (ret == 2) {
2478                 printk("\nmarked lock as {%s}:\n", usage_str[new_bit]);
2479                 print_lock(this);
2480                 print_irqtrace_events(curr);
2481                 dump_stack();
2482         }
2483
2484         return ret;
2485 }
2486
2487 /*
2488  * Initialize a lock instance's lock-class mapping info:
2489  */
2490 void lockdep_init_map(struct lockdep_map *lock, const char *name,
2491                       struct lock_class_key *key, int subclass)
2492 {
2493         lock->class_cache = NULL;
2494 #ifdef CONFIG_LOCK_STAT
2495         lock->cpu = raw_smp_processor_id();
2496 #endif
2497
2498         if (DEBUG_LOCKS_WARN_ON(!name)) {
2499                 lock->name = "NULL";
2500                 return;
2501         }
2502
2503         lock->name = name;
2504
2505         if (DEBUG_LOCKS_WARN_ON(!key))
2506                 return;
2507         /*
2508          * Sanity check, the lock-class key must be persistent:
2509          */
2510         if (!static_obj(key)) {
2511                 printk("BUG: key %p not in .data!\n", key);
2512                 DEBUG_LOCKS_WARN_ON(1);
2513                 return;
2514         }
2515         lock->key = key;
2516
2517         if (unlikely(!debug_locks))
2518                 return;
2519
2520         if (subclass)
2521                 register_lock_class(lock, subclass, 1);
2522 }
2523 EXPORT_SYMBOL_GPL(lockdep_init_map);
2524
2525 /*
2526  * This gets called for every mutex_lock*()/spin_lock*() operation.
2527  * We maintain the dependency maps and validate the locking attempt:
2528  */
2529 static int __lock_acquire(struct lockdep_map *lock, unsigned int subclass,
2530                           int trylock, int read, int check, int hardirqs_off,
2531                           struct lockdep_map *nest_lock, unsigned long ip)
2532 {
2533         struct task_struct *curr = current;
2534         struct lock_class *class = NULL;
2535         struct held_lock *hlock;
2536         unsigned int depth, id;
2537         int chain_head = 0;
2538         u64 chain_key;
2539
2540         if (!prove_locking)
2541                 check = 1;
2542
2543         if (unlikely(!debug_locks))
2544                 return 0;
2545
2546         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2547                 return 0;
2548
2549         if (unlikely(subclass >= MAX_LOCKDEP_SUBCLASSES)) {
2550                 debug_locks_off();
2551                 printk("BUG: MAX_LOCKDEP_SUBCLASSES too low!\n");
2552                 printk("turning off the locking correctness validator.\n");
2553                 dump_stack();
2554                 return 0;
2555         }
2556
2557         if (!subclass)
2558                 class = lock->class_cache;
2559         /*
2560          * Not cached yet or subclass?
2561          */
2562         if (unlikely(!class)) {
2563                 class = register_lock_class(lock, subclass, 0);
2564                 if (!class)
2565                         return 0;
2566         }
2567         debug_atomic_inc((atomic_t *)&class->ops);
2568         if (very_verbose(class)) {
2569                 printk("\nacquire class [%p] %s", class->key, class->name);
2570                 if (class->name_version > 1)
2571                         printk("#%d", class->name_version);
2572                 printk("\n");
2573                 dump_stack();
2574         }
2575
2576         /*
2577          * Add the lock to the list of currently held locks.
2578          * (we dont increase the depth just yet, up until the
2579          * dependency checks are done)
2580          */
2581         depth = curr->lockdep_depth;
2582         if (DEBUG_LOCKS_WARN_ON(depth >= MAX_LOCK_DEPTH))
2583                 return 0;
2584
2585         hlock = curr->held_locks + depth;
2586         if (DEBUG_LOCKS_WARN_ON(!class))
2587                 return 0;
2588         hlock->class_idx = class - lock_classes + 1;
2589         hlock->acquire_ip = ip;
2590         hlock->instance = lock;
2591         hlock->nest_lock = nest_lock;
2592         hlock->trylock = trylock;
2593         hlock->read = read;
2594         hlock->check = check;
2595         hlock->hardirqs_off = !!hardirqs_off;
2596 #ifdef CONFIG_LOCK_STAT
2597         hlock->waittime_stamp = 0;
2598         hlock->holdtime_stamp = sched_clock();
2599 #endif
2600
2601         if (check == 2 && !mark_irqflags(curr, hlock))
2602                 return 0;
2603
2604         /* mark it as used: */
2605         if (!mark_lock(curr, hlock, LOCK_USED))
2606                 return 0;
2607
2608         /*
2609          * Calculate the chain hash: it's the combined hash of all the
2610          * lock keys along the dependency chain. We save the hash value
2611          * at every step so that we can get the current hash easily
2612          * after unlock. The chain hash is then used to cache dependency
2613          * results.
2614          *
2615          * The 'key ID' is what is the most compact key value to drive
2616          * the hash, not class->key.
2617          */
2618         id = class - lock_classes;
2619         if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
2620                 return 0;
2621
2622         chain_key = curr->curr_chain_key;
2623         if (!depth) {
2624                 if (DEBUG_LOCKS_WARN_ON(chain_key != 0))
2625                         return 0;
2626                 chain_head = 1;
2627         }
2628
2629         hlock->prev_chain_key = chain_key;
2630         if (separate_irq_context(curr, hlock)) {
2631                 chain_key = 0;
2632                 chain_head = 1;
2633         }
2634         chain_key = iterate_chain_key(chain_key, id);
2635
2636         if (!validate_chain(curr, lock, hlock, chain_head, chain_key))
2637                 return 0;
2638
2639         curr->curr_chain_key = chain_key;
2640         curr->lockdep_depth++;
2641         check_chain_key(curr);
2642 #ifdef CONFIG_DEBUG_LOCKDEP
2643         if (unlikely(!debug_locks))
2644                 return 0;
2645 #endif
2646         if (unlikely(curr->lockdep_depth >= MAX_LOCK_DEPTH)) {
2647                 debug_locks_off();
2648                 printk("BUG: MAX_LOCK_DEPTH too low!\n");
2649                 printk("turning off the locking correctness validator.\n");
2650                 dump_stack();
2651                 return 0;
2652         }
2653
2654         if (unlikely(curr->lockdep_depth > max_lockdep_depth))
2655                 max_lockdep_depth = curr->lockdep_depth;
2656
2657         return 1;
2658 }
2659
2660 static int
2661 print_unlock_inbalance_bug(struct task_struct *curr, struct lockdep_map *lock,
2662                            unsigned long ip)
2663 {
2664         if (!debug_locks_off())
2665                 return 0;
2666         if (debug_locks_silent)
2667                 return 0;
2668
2669         printk("\n=====================================\n");
2670         printk(  "[ BUG: bad unlock balance detected! ]\n");
2671         printk(  "-------------------------------------\n");
2672         printk("%s/%d is trying to release lock (",
2673                 curr->comm, task_pid_nr(curr));
2674         print_lockdep_cache(lock);
2675         printk(") at:\n");
2676         print_ip_sym(ip);
2677         printk("but there are no more locks to release!\n");
2678         printk("\nother info that might help us debug this:\n");
2679         lockdep_print_held_locks(curr);
2680
2681         printk("\nstack backtrace:\n");
2682         dump_stack();
2683
2684         return 0;
2685 }
2686
2687 /*
2688  * Common debugging checks for both nested and non-nested unlock:
2689  */
2690 static int check_unlock(struct task_struct *curr, struct lockdep_map *lock,
2691                         unsigned long ip)
2692 {
2693         if (unlikely(!debug_locks))
2694                 return 0;
2695         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2696                 return 0;
2697
2698         if (curr->lockdep_depth <= 0)
2699                 return print_unlock_inbalance_bug(curr, lock, ip);
2700
2701         return 1;
2702 }
2703
2704 static int
2705 __lock_set_class(struct lockdep_map *lock, const char *name,
2706                  struct lock_class_key *key, unsigned int subclass,
2707                  unsigned long ip)
2708 {
2709         struct task_struct *curr = current;
2710         struct held_lock *hlock, *prev_hlock;
2711         struct lock_class *class;
2712         unsigned int depth;
2713         int i;
2714
2715         depth = curr->lockdep_depth;
2716         if (DEBUG_LOCKS_WARN_ON(!depth))
2717                 return 0;
2718
2719         prev_hlock = NULL;
2720         for (i = depth-1; i >= 0; i--) {
2721                 hlock = curr->held_locks + i;
2722                 /*
2723                  * We must not cross into another context:
2724                  */
2725                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
2726                         break;
2727                 if (hlock->instance == lock)
2728                         goto found_it;
2729                 prev_hlock = hlock;
2730         }
2731         return print_unlock_inbalance_bug(curr, lock, ip);
2732
2733 found_it:
2734         lockdep_init_map(lock, name, key, 0);
2735         class = register_lock_class(lock, subclass, 0);
2736         hlock->class_idx = class - lock_classes + 1;
2737
2738         curr->lockdep_depth = i;
2739         curr->curr_chain_key = hlock->prev_chain_key;
2740
2741         for (; i < depth; i++) {
2742                 hlock = curr->held_locks + i;
2743                 if (!__lock_acquire(hlock->instance,
2744                         hlock_class(hlock)->subclass, hlock->trylock,
2745                                 hlock->read, hlock->check, hlock->hardirqs_off,
2746                                 hlock->nest_lock, hlock->acquire_ip))
2747                         return 0;
2748         }
2749
2750         if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth))
2751                 return 0;
2752         return 1;
2753 }
2754
2755 /*
2756  * Remove the lock to the list of currently held locks in a
2757  * potentially non-nested (out of order) manner. This is a
2758  * relatively rare operation, as all the unlock APIs default
2759  * to nested mode (which uses lock_release()):
2760  */
2761 static int
2762 lock_release_non_nested(struct task_struct *curr,
2763                         struct lockdep_map *lock, unsigned long ip)
2764 {
2765         struct held_lock *hlock, *prev_hlock;
2766         unsigned int depth;
2767         int i;
2768
2769         /*
2770          * Check whether the lock exists in the current stack
2771          * of held locks:
2772          */
2773         depth = curr->lockdep_depth;
2774         if (DEBUG_LOCKS_WARN_ON(!depth))
2775                 return 0;
2776
2777         prev_hlock = NULL;
2778         for (i = depth-1; i >= 0; i--) {
2779                 hlock = curr->held_locks + i;
2780                 /*
2781                  * We must not cross into another context:
2782                  */
2783                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
2784                         break;
2785                 if (hlock->instance == lock)
2786                         goto found_it;
2787                 prev_hlock = hlock;
2788         }
2789         return print_unlock_inbalance_bug(curr, lock, ip);
2790
2791 found_it:
2792         lock_release_holdtime(hlock);
2793
2794         /*
2795          * We have the right lock to unlock, 'hlock' points to it.
2796          * Now we remove it from the stack, and add back the other
2797          * entries (if any), recalculating the hash along the way:
2798          */
2799         curr->lockdep_depth = i;
2800         curr->curr_chain_key = hlock->prev_chain_key;
2801
2802         for (i++; i < depth; i++) {
2803                 hlock = curr->held_locks + i;
2804                 if (!__lock_acquire(hlock->instance,
2805                         hlock_class(hlock)->subclass, hlock->trylock,
2806                                 hlock->read, hlock->check, hlock->hardirqs_off,
2807                                 hlock->nest_lock, hlock->acquire_ip))
2808                         return 0;
2809         }
2810
2811         if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth - 1))
2812                 return 0;
2813         return 1;
2814 }
2815
2816 /*
2817  * Remove the lock to the list of currently held locks - this gets
2818  * called on mutex_unlock()/spin_unlock*() (or on a failed
2819  * mutex_lock_interruptible()). This is done for unlocks that nest
2820  * perfectly. (i.e. the current top of the lock-stack is unlocked)
2821  */
2822 static int lock_release_nested(struct task_struct *curr,
2823                                struct lockdep_map *lock, unsigned long ip)
2824 {
2825         struct held_lock *hlock;
2826         unsigned int depth;
2827
2828         /*
2829          * Pop off the top of the lock stack:
2830          */
2831         depth = curr->lockdep_depth - 1;
2832         hlock = curr->held_locks + depth;
2833
2834         /*
2835          * Is the unlock non-nested:
2836          */
2837         if (hlock->instance != lock)
2838                 return lock_release_non_nested(curr, lock, ip);
2839         curr->lockdep_depth--;
2840
2841         if (DEBUG_LOCKS_WARN_ON(!depth && (hlock->prev_chain_key != 0)))
2842                 return 0;
2843
2844         curr->curr_chain_key = hlock->prev_chain_key;
2845
2846         lock_release_holdtime(hlock);
2847
2848 #ifdef CONFIG_DEBUG_LOCKDEP
2849         hlock->prev_chain_key = 0;
2850         hlock->class_idx = 0;
2851         hlock->acquire_ip = 0;
2852         hlock->irq_context = 0;
2853 #endif
2854         return 1;
2855 }
2856
2857 /*
2858  * Remove the lock to the list of currently held locks - this gets
2859  * called on mutex_unlock()/spin_unlock*() (or on a failed
2860  * mutex_lock_interruptible()). This is done for unlocks that nest
2861  * perfectly. (i.e. the current top of the lock-stack is unlocked)
2862  */
2863 static void
2864 __lock_release(struct lockdep_map *lock, int nested, unsigned long ip)
2865 {
2866         struct task_struct *curr = current;
2867
2868         if (!check_unlock(curr, lock, ip))
2869                 return;
2870
2871         if (nested) {
2872                 if (!lock_release_nested(curr, lock, ip))
2873                         return;
2874         } else {
2875                 if (!lock_release_non_nested(curr, lock, ip))
2876                         return;
2877         }
2878
2879         check_chain_key(curr);
2880 }
2881
2882 /*
2883  * Check whether we follow the irq-flags state precisely:
2884  */
2885 static void check_flags(unsigned long flags)
2886 {
2887 #if defined(CONFIG_PROVE_LOCKING) && defined(CONFIG_DEBUG_LOCKDEP) && \
2888     defined(CONFIG_TRACE_IRQFLAGS)
2889         if (!debug_locks)
2890                 return;
2891
2892         if (irqs_disabled_flags(flags)) {
2893                 if (DEBUG_LOCKS_WARN_ON(current->hardirqs_enabled)) {
2894                         printk("possible reason: unannotated irqs-off.\n");
2895                 }
2896         } else {
2897                 if (DEBUG_LOCKS_WARN_ON(!current->hardirqs_enabled)) {
2898                         printk("possible reason: unannotated irqs-on.\n");
2899                 }
2900         }
2901
2902         /*
2903          * We dont accurately track softirq state in e.g.
2904          * hardirq contexts (such as on 4KSTACKS), so only
2905          * check if not in hardirq contexts:
2906          */
2907         if (!hardirq_count()) {
2908                 if (softirq_count())
2909                         DEBUG_LOCKS_WARN_ON(current->softirqs_enabled);
2910                 else
2911                         DEBUG_LOCKS_WARN_ON(!current->softirqs_enabled);
2912         }
2913
2914         if (!debug_locks)
2915                 print_irqtrace_events(current);
2916 #endif
2917 }
2918
2919 void lock_set_class(struct lockdep_map *lock, const char *name,
2920                     struct lock_class_key *key, unsigned int subclass,
2921                     unsigned long ip)
2922 {
2923         unsigned long flags;
2924
2925         if (unlikely(current->lockdep_recursion))
2926                 return;
2927
2928         raw_local_irq_save(flags);
2929         current->lockdep_recursion = 1;
2930         check_flags(flags);
2931         if (__lock_set_class(lock, name, key, subclass, ip))
2932                 check_chain_key(current);
2933         current->lockdep_recursion = 0;
2934         raw_local_irq_restore(flags);
2935 }
2936 EXPORT_SYMBOL_GPL(lock_set_class);
2937
2938 DEFINE_TRACE(lock_acquire);
2939
2940 /*
2941  * We are not always called with irqs disabled - do that here,
2942  * and also avoid lockdep recursion:
2943  */
2944 void lock_acquire(struct lockdep_map *lock, unsigned int subclass,
2945                           int trylock, int read, int check,
2946                           struct lockdep_map *nest_lock, unsigned long ip)
2947 {
2948         unsigned long flags;
2949
2950         trace_lock_acquire(lock, subclass, trylock, read, check, nest_lock, ip);
2951
2952         if (unlikely(current->lockdep_recursion))
2953                 return;
2954
2955         raw_local_irq_save(flags);
2956         check_flags(flags);
2957
2958         current->lockdep_recursion = 1;
2959         __lock_acquire(lock, subclass, trylock, read, check,
2960                        irqs_disabled_flags(flags), nest_lock, ip);
2961         current->lockdep_recursion = 0;
2962         raw_local_irq_restore(flags);
2963 }
2964 EXPORT_SYMBOL_GPL(lock_acquire);
2965
2966 DEFINE_TRACE(lock_release);
2967
2968 void lock_release(struct lockdep_map *lock, int nested,
2969                           unsigned long ip)
2970 {
2971         unsigned long flags;
2972
2973         trace_lock_release(lock, nested, ip);
2974
2975         if (unlikely(current->lockdep_recursion))
2976                 return;
2977
2978         raw_local_irq_save(flags);
2979         check_flags(flags);
2980         current->lockdep_recursion = 1;
2981         __lock_release(lock, nested, ip);
2982         current->lockdep_recursion = 0;
2983         raw_local_irq_restore(flags);
2984 }
2985 EXPORT_SYMBOL_GPL(lock_release);
2986
2987 void lockdep_set_current_reclaim_state(gfp_t gfp_mask)
2988 {
2989         current->lockdep_reclaim_gfp = gfp_mask;
2990 }
2991
2992 void lockdep_clear_current_reclaim_state(void)
2993 {
2994         current->lockdep_reclaim_gfp = 0;
2995 }
2996
2997 #ifdef CONFIG_LOCK_STAT
2998 static int
2999 print_lock_contention_bug(struct task_struct *curr, struct lockdep_map *lock,
3000                            unsigned long ip)
3001 {
3002         if (!debug_locks_off())
3003                 return 0;
3004         if (debug_locks_silent)
3005                 return 0;
3006
3007         printk("\n=================================\n");
3008         printk(  "[ BUG: bad contention detected! ]\n");
3009         printk(  "---------------------------------\n");
3010         printk("%s/%d is trying to contend lock (",
3011                 curr->comm, task_pid_nr(curr));
3012         print_lockdep_cache(lock);
3013         printk(") at:\n");
3014         print_ip_sym(ip);
3015         printk("but there are no locks held!\n");
3016         printk("\nother info that might help us debug this:\n");
3017         lockdep_print_held_locks(curr);
3018
3019         printk("\nstack backtrace:\n");
3020         dump_stack();
3021
3022         return 0;
3023 }
3024
3025 static void
3026 __lock_contended(struct lockdep_map *lock, unsigned long ip)
3027 {
3028         struct task_struct *curr = current;
3029         struct held_lock *hlock, *prev_hlock;
3030         struct lock_class_stats *stats;
3031         unsigned int depth;
3032         int i, contention_point, contending_point;
3033
3034         depth = curr->lockdep_depth;
3035         if (DEBUG_LOCKS_WARN_ON(!depth))
3036                 return;
3037
3038         prev_hlock = NULL;
3039         for (i = depth-1; i >= 0; i--) {
3040                 hlock = curr->held_locks + i;
3041                 /*
3042                  * We must not cross into another context:
3043                  */
3044                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3045                         break;
3046                 if (hlock->instance == lock)
3047                         goto found_it;
3048                 prev_hlock = hlock;
3049         }
3050         print_lock_contention_bug(curr, lock, ip);
3051         return;
3052
3053 found_it:
3054         hlock->waittime_stamp = sched_clock();
3055
3056         contention_point = lock_point(hlock_class(hlock)->contention_point, ip);
3057         contending_point = lock_point(hlock_class(hlock)->contending_point,
3058                                       lock->ip);
3059
3060         stats = get_lock_stats(hlock_class(hlock));
3061         if (contention_point < LOCKSTAT_POINTS)
3062                 stats->contention_point[contention_point]++;
3063         if (contending_point < LOCKSTAT_POINTS)
3064                 stats->contending_point[contending_point]++;
3065         if (lock->cpu != smp_processor_id())
3066                 stats->bounces[bounce_contended + !!hlock->read]++;
3067         put_lock_stats(stats);
3068 }
3069
3070 static void
3071 __lock_acquired(struct lockdep_map *lock, unsigned long ip)
3072 {
3073         struct task_struct *curr = current;
3074         struct held_lock *hlock, *prev_hlock;
3075         struct lock_class_stats *stats;
3076         unsigned int depth;
3077         u64 now;
3078         s64 waittime = 0;
3079         int i, cpu;
3080
3081         depth = curr->lockdep_depth;
3082         if (DEBUG_LOCKS_WARN_ON(!depth))
3083                 return;
3084
3085         prev_hlock = NULL;
3086         for (i = depth-1; i >= 0; i--) {
3087                 hlock = curr->held_locks + i;
3088                 /*
3089                  * We must not cross into another context:
3090                  */
3091                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3092                         break;
3093                 if (hlock->instance == lock)
3094                         goto found_it;
3095                 prev_hlock = hlock;
3096         }
3097         print_lock_contention_bug(curr, lock, _RET_IP_);
3098         return;
3099
3100 found_it:
3101         cpu = smp_processor_id();
3102         if (hlock->waittime_stamp) {
3103                 now = sched_clock();
3104                 waittime = now - hlock->waittime_stamp;
3105                 hlock->holdtime_stamp = now;
3106         }
3107
3108         stats = get_lock_stats(hlock_class(hlock));
3109         if (waittime) {
3110                 if (hlock->read)
3111                         lock_time_inc(&stats->read_waittime, waittime);
3112                 else
3113                         lock_time_inc(&stats->write_waittime, waittime);
3114         }
3115         if (lock->cpu != cpu)
3116                 stats->bounces[bounce_acquired + !!hlock->read]++;
3117         put_lock_stats(stats);
3118
3119         lock->cpu = cpu;
3120         lock->ip = ip;
3121 }
3122
3123 DEFINE_TRACE(lock_contended);
3124
3125 void lock_contended(struct lockdep_map *lock, unsigned long ip)
3126 {
3127         unsigned long flags;
3128
3129         trace_lock_contended(lock, ip);
3130
3131         if (unlikely(!lock_stat))
3132                 return;
3133
3134         if (unlikely(current->lockdep_recursion))
3135                 return;
3136
3137         raw_local_irq_save(flags);
3138         check_flags(flags);
3139         current->lockdep_recursion = 1;
3140         __lock_contended(lock, ip);
3141         current->lockdep_recursion = 0;
3142         raw_local_irq_restore(flags);
3143 }
3144 EXPORT_SYMBOL_GPL(lock_contended);
3145
3146 DEFINE_TRACE(lock_acquired);
3147
3148 void lock_acquired(struct lockdep_map *lock, unsigned long ip)
3149 {
3150         unsigned long flags;
3151
3152         trace_lock_acquired(lock, ip);
3153
3154         if (unlikely(!lock_stat))
3155                 return;
3156
3157         if (unlikely(current->lockdep_recursion))
3158                 return;
3159
3160         raw_local_irq_save(flags);
3161         check_flags(flags);
3162         current->lockdep_recursion = 1;
3163         __lock_acquired(lock, ip);
3164         current->lockdep_recursion = 0;
3165         raw_local_irq_restore(flags);
3166 }
3167 EXPORT_SYMBOL_GPL(lock_acquired);
3168 #endif
3169
3170 /*
3171  * Used by the testsuite, sanitize the validator state
3172  * after a simulated failure:
3173  */
3174
3175 void lockdep_reset(void)
3176 {
3177         unsigned long flags;
3178         int i;
3179
3180         raw_local_irq_save(flags);
3181         current->curr_chain_key = 0;
3182         current->lockdep_depth = 0;
3183         current->lockdep_recursion = 0;
3184         memset(current->held_locks, 0, MAX_LOCK_DEPTH*sizeof(struct held_lock));
3185         nr_hardirq_chains = 0;
3186         nr_softirq_chains = 0;
3187         nr_process_chains = 0;
3188         debug_locks = 1;
3189         for (i = 0; i < CHAINHASH_SIZE; i++)
3190                 INIT_LIST_HEAD(chainhash_table + i);
3191         raw_local_irq_restore(flags);
3192 }
3193
3194 static void zap_class(struct lock_class *class)
3195 {
3196         int i;
3197
3198         /*
3199          * Remove all dependencies this lock is
3200          * involved in:
3201          */
3202         for (i = 0; i < nr_list_entries; i++) {
3203                 if (list_entries[i].class == class)
3204                         list_del_rcu(&list_entries[i].entry);
3205         }
3206         /*
3207          * Unhash the class and remove it from the all_lock_classes list:
3208          */
3209         list_del_rcu(&class->hash_entry);
3210         list_del_rcu(&class->lock_entry);
3211
3212         class->key = NULL;
3213 }
3214
3215 static inline int within(const void *addr, void *start, unsigned long size)
3216 {
3217         return addr >= start && addr < start + size;
3218 }
3219
3220 void lockdep_free_key_range(void *start, unsigned long size)
3221 {
3222         struct lock_class *class, *next;
3223         struct list_head *head;
3224         unsigned long flags;
3225         int i;
3226         int locked;
3227
3228         raw_local_irq_save(flags);
3229         locked = graph_lock();
3230
3231         /*
3232          * Unhash all classes that were created by this module:
3233          */
3234         for (i = 0; i < CLASSHASH_SIZE; i++) {
3235                 head = classhash_table + i;
3236                 if (list_empty(head))
3237                         continue;
3238                 list_for_each_entry_safe(class, next, head, hash_entry) {
3239                         if (within(class->key, start, size))
3240                                 zap_class(class);
3241                         else if (within(class->name, start, size))
3242                                 zap_class(class);
3243                 }
3244         }
3245
3246         if (locked)
3247                 graph_unlock();
3248         raw_local_irq_restore(flags);
3249 }
3250
3251 void lockdep_reset_lock(struct lockdep_map *lock)
3252 {
3253         struct lock_class *class, *next;
3254         struct list_head *head;
3255         unsigned long flags;
3256         int i, j;
3257         int locked;
3258
3259         raw_local_irq_save(flags);
3260
3261         /*
3262          * Remove all classes this lock might have:
3263          */
3264         for (j = 0; j < MAX_LOCKDEP_SUBCLASSES; j++) {
3265                 /*
3266                  * If the class exists we look it up and zap it:
3267                  */
3268                 class = look_up_lock_class(lock, j);
3269                 if (class)
3270                         zap_class(class);
3271         }
3272         /*
3273          * Debug check: in the end all mapped classes should
3274          * be gone.
3275          */
3276         locked = graph_lock();
3277         for (i = 0; i < CLASSHASH_SIZE; i++) {
3278                 head = classhash_table + i;
3279                 if (list_empty(head))
3280                         continue;
3281                 list_for_each_entry_safe(class, next, head, hash_entry) {
3282                         if (unlikely(class == lock->class_cache)) {
3283                                 if (debug_locks_off_graph_unlock())
3284                                         WARN_ON(1);
3285                                 goto out_restore;
3286                         }
3287                 }
3288         }
3289         if (locked)
3290                 graph_unlock();
3291
3292 out_restore:
3293         raw_local_irq_restore(flags);
3294 }
3295
3296 void lockdep_init(void)
3297 {
3298         int i;
3299
3300         /*
3301          * Some architectures have their own start_kernel()
3302          * code which calls lockdep_init(), while we also
3303          * call lockdep_init() from the start_kernel() itself,
3304          * and we want to initialize the hashes only once:
3305          */
3306         if (lockdep_initialized)
3307                 return;
3308
3309         for (i = 0; i < CLASSHASH_SIZE; i++)
3310                 INIT_LIST_HEAD(classhash_table + i);
3311
3312         for (i = 0; i < CHAINHASH_SIZE; i++)
3313                 INIT_LIST_HEAD(chainhash_table + i);
3314
3315         lockdep_initialized = 1;
3316 }
3317
3318 void __init lockdep_info(void)
3319 {
3320         printk("Lock dependency validator: Copyright (c) 2006 Red Hat, Inc., Ingo Molnar\n");
3321
3322         printk("... MAX_LOCKDEP_SUBCLASSES:  %lu\n", MAX_LOCKDEP_SUBCLASSES);
3323         printk("... MAX_LOCK_DEPTH:          %lu\n", MAX_LOCK_DEPTH);
3324         printk("... MAX_LOCKDEP_KEYS:        %lu\n", MAX_LOCKDEP_KEYS);
3325         printk("... CLASSHASH_SIZE:          %lu\n", CLASSHASH_SIZE);
3326         printk("... MAX_LOCKDEP_ENTRIES:     %lu\n", MAX_LOCKDEP_ENTRIES);
3327         printk("... MAX_LOCKDEP_CHAINS:      %lu\n", MAX_LOCKDEP_CHAINS);
3328         printk("... CHAINHASH_SIZE:          %lu\n", CHAINHASH_SIZE);
3329
3330         printk(" memory used by lock dependency info: %lu kB\n",
3331                 (sizeof(struct lock_class) * MAX_LOCKDEP_KEYS +
3332                 sizeof(struct list_head) * CLASSHASH_SIZE +
3333                 sizeof(struct lock_list) * MAX_LOCKDEP_ENTRIES +
3334                 sizeof(struct lock_chain) * MAX_LOCKDEP_CHAINS +
3335                 sizeof(struct list_head) * CHAINHASH_SIZE) / 1024);
3336
3337         printk(" per task-struct memory footprint: %lu bytes\n",
3338                 sizeof(struct held_lock) * MAX_LOCK_DEPTH);
3339
3340 #ifdef CONFIG_DEBUG_LOCKDEP
3341         if (lockdep_init_error) {
3342                 printk("WARNING: lockdep init error! Arch code didn't call lockdep_init() early enough?\n");
3343                 printk("Call stack leading to lockdep invocation was:\n");
3344                 print_stack_trace(&lockdep_init_trace, 0);
3345         }
3346 #endif
3347 }
3348
3349 static void
3350 print_freed_lock_bug(struct task_struct *curr, const void *mem_from,
3351                      const void *mem_to, struct held_lock *hlock)
3352 {
3353         if (!debug_locks_off())
3354                 return;
3355         if (debug_locks_silent)
3356                 return;
3357
3358         printk("\n=========================\n");
3359         printk(  "[ BUG: held lock freed! ]\n");
3360         printk(  "-------------------------\n");
3361         printk("%s/%d is freeing memory %p-%p, with a lock still held there!\n",
3362                 curr->comm, task_pid_nr(curr), mem_from, mem_to-1);
3363         print_lock(hlock);
3364         lockdep_print_held_locks(curr);
3365
3366         printk("\nstack backtrace:\n");
3367         dump_stack();
3368 }
3369
3370 static inline int not_in_range(const void* mem_from, unsigned long mem_len,
3371                                 const void* lock_from, unsigned long lock_len)
3372 {
3373         return lock_from + lock_len <= mem_from ||
3374                 mem_from + mem_len <= lock_from;
3375 }
3376
3377 /*
3378  * Called when kernel memory is freed (or unmapped), or if a lock
3379  * is destroyed or reinitialized - this code checks whether there is
3380  * any held lock in the memory range of <from> to <to>:
3381  */
3382 void debug_check_no_locks_freed(const void *mem_from, unsigned long mem_len)
3383 {
3384         struct task_struct *curr = current;
3385         struct held_lock *hlock;
3386         unsigned long flags;
3387         int i;
3388
3389         if (unlikely(!debug_locks))
3390                 return;
3391
3392         local_irq_save(flags);
3393         for (i = 0; i < curr->lockdep_depth; i++) {
3394                 hlock = curr->held_locks + i;
3395
3396                 if (not_in_range(mem_from, mem_len, hlock->instance,
3397                                         sizeof(*hlock->instance)))
3398                         continue;
3399
3400                 print_freed_lock_bug(curr, mem_from, mem_from + mem_len, hlock);
3401                 break;
3402         }
3403         local_irq_restore(flags);
3404 }
3405 EXPORT_SYMBOL_GPL(debug_check_no_locks_freed);
3406
3407 static void print_held_locks_bug(struct task_struct *curr)
3408 {
3409         if (!debug_locks_off())
3410                 return;
3411         if (debug_locks_silent)
3412                 return;
3413
3414         printk("\n=====================================\n");
3415         printk(  "[ BUG: lock held at task exit time! ]\n");
3416         printk(  "-------------------------------------\n");
3417         printk("%s/%d is exiting with locks still held!\n",
3418                 curr->comm, task_pid_nr(curr));
3419         lockdep_print_held_locks(curr);
3420
3421         printk("\nstack backtrace:\n");
3422         dump_stack();
3423 }
3424
3425 void debug_check_no_locks_held(struct task_struct *task)
3426 {
3427         if (unlikely(task->lockdep_depth > 0))
3428                 print_held_locks_bug(task);
3429 }
3430
3431 void debug_show_all_locks(void)
3432 {
3433         struct task_struct *g, *p;
3434         int count = 10;
3435         int unlock = 1;
3436
3437         if (unlikely(!debug_locks)) {
3438                 printk("INFO: lockdep is turned off.\n");
3439                 return;
3440         }
3441         printk("\nShowing all locks held in the system:\n");
3442
3443         /*
3444          * Here we try to get the tasklist_lock as hard as possible,
3445          * if not successful after 2 seconds we ignore it (but keep
3446          * trying). This is to enable a debug printout even if a
3447          * tasklist_lock-holding task deadlocks or crashes.
3448          */
3449 retry:
3450         if (!read_trylock(&tasklist_lock)) {
3451                 if (count == 10)
3452                         printk("hm, tasklist_lock locked, retrying... ");
3453                 if (count) {
3454                         count--;
3455                         printk(" #%d", 10-count);
3456                         mdelay(200);
3457                         goto retry;
3458                 }
3459                 printk(" ignoring it.\n");
3460                 unlock = 0;
3461         } else {
3462                 if (count != 10)
3463                         printk(KERN_CONT " locked it.\n");
3464         }
3465
3466         do_each_thread(g, p) {
3467                 /*
3468                  * It's not reliable to print a task's held locks
3469                  * if it's not sleeping (or if it's not the current
3470                  * task):
3471                  */
3472                 if (p->state == TASK_RUNNING && p != current)
3473                         continue;
3474                 if (p->lockdep_depth)
3475                         lockdep_print_held_locks(p);
3476                 if (!unlock)
3477                         if (read_trylock(&tasklist_lock))
3478                                 unlock = 1;
3479         } while_each_thread(g, p);
3480
3481         printk("\n");
3482         printk("=============================================\n\n");
3483
3484         if (unlock)
3485                 read_unlock(&tasklist_lock);
3486 }
3487 EXPORT_SYMBOL_GPL(debug_show_all_locks);
3488
3489 /*
3490  * Careful: only use this function if you are sure that
3491  * the task cannot run in parallel!
3492  */
3493 void __debug_show_held_locks(struct task_struct *task)
3494 {
3495         if (unlikely(!debug_locks)) {
3496                 printk("INFO: lockdep is turned off.\n");
3497                 return;
3498         }
3499         lockdep_print_held_locks(task);
3500 }
3501 EXPORT_SYMBOL_GPL(__debug_show_held_locks);
3502
3503 void debug_show_held_locks(struct task_struct *task)
3504 {
3505                 __debug_show_held_locks(task);
3506 }
3507 EXPORT_SYMBOL_GPL(debug_show_held_locks);
3508
3509 void lockdep_sys_exit(void)
3510 {
3511         struct task_struct *curr = current;
3512
3513         if (unlikely(curr->lockdep_depth)) {
3514                 if (!debug_locks_off())
3515                         return;
3516                 printk("\n================================================\n");
3517                 printk(  "[ BUG: lock held when returning to user space! ]\n");
3518                 printk(  "------------------------------------------------\n");
3519                 printk("%s/%d is leaving the kernel with locks still held!\n",
3520                                 curr->comm, curr->pid);
3521                 lockdep_print_held_locks(curr);
3522         }
3523 }