lockdep: Improve implementation of BFS
[safe/jmp/linux-2.6] / kernel / lockdep.c
1 /*
2  * kernel/lockdep.c
3  *
4  * Runtime locking correctness validator
5  *
6  * Started by Ingo Molnar:
7  *
8  *  Copyright (C) 2006,2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
9  *  Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
10  *
11  * this code maps all the lock dependencies as they occur in a live kernel
12  * and will warn about the following classes of locking bugs:
13  *
14  * - lock inversion scenarios
15  * - circular lock dependencies
16  * - hardirq/softirq safe/unsafe locking bugs
17  *
18  * Bugs are reported even if the current locking scenario does not cause
19  * any deadlock at this point.
20  *
21  * I.e. if anytime in the past two locks were taken in a different order,
22  * even if it happened for another task, even if those were different
23  * locks (but of the same class as this lock), this code will detect it.
24  *
25  * Thanks to Arjan van de Ven for coming up with the initial idea of
26  * mapping lock dependencies runtime.
27  */
28 #define DISABLE_BRANCH_PROFILING
29 #include <linux/mutex.h>
30 #include <linux/sched.h>
31 #include <linux/delay.h>
32 #include <linux/module.h>
33 #include <linux/proc_fs.h>
34 #include <linux/seq_file.h>
35 #include <linux/spinlock.h>
36 #include <linux/kallsyms.h>
37 #include <linux/interrupt.h>
38 #include <linux/stacktrace.h>
39 #include <linux/debug_locks.h>
40 #include <linux/irqflags.h>
41 #include <linux/utsname.h>
42 #include <linux/hash.h>
43 #include <linux/ftrace.h>
44 #include <linux/stringify.h>
45 #include <linux/bitops.h>
46 #include <asm/sections.h>
47
48 #include "lockdep_internals.h"
49
50 #define CREATE_TRACE_POINTS
51 #include <trace/events/lockdep.h>
52
53 #ifdef CONFIG_PROVE_LOCKING
54 int prove_locking = 1;
55 module_param(prove_locking, int, 0644);
56 #else
57 #define prove_locking 0
58 #endif
59
60 #ifdef CONFIG_LOCK_STAT
61 int lock_stat = 1;
62 module_param(lock_stat, int, 0644);
63 #else
64 #define lock_stat 0
65 #endif
66
67 /*
68  * lockdep_lock: protects the lockdep graph, the hashes and the
69  *               class/list/hash allocators.
70  *
71  * This is one of the rare exceptions where it's justified
72  * to use a raw spinlock - we really dont want the spinlock
73  * code to recurse back into the lockdep code...
74  */
75 static raw_spinlock_t lockdep_lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
76
77 static int graph_lock(void)
78 {
79         __raw_spin_lock(&lockdep_lock);
80         /*
81          * Make sure that if another CPU detected a bug while
82          * walking the graph we dont change it (while the other
83          * CPU is busy printing out stuff with the graph lock
84          * dropped already)
85          */
86         if (!debug_locks) {
87                 __raw_spin_unlock(&lockdep_lock);
88                 return 0;
89         }
90         /* prevent any recursions within lockdep from causing deadlocks */
91         current->lockdep_recursion++;
92         return 1;
93 }
94
95 static inline int graph_unlock(void)
96 {
97         if (debug_locks && !__raw_spin_is_locked(&lockdep_lock))
98                 return DEBUG_LOCKS_WARN_ON(1);
99
100         current->lockdep_recursion--;
101         __raw_spin_unlock(&lockdep_lock);
102         return 0;
103 }
104
105 /*
106  * Turn lock debugging off and return with 0 if it was off already,
107  * and also release the graph lock:
108  */
109 static inline int debug_locks_off_graph_unlock(void)
110 {
111         int ret = debug_locks_off();
112
113         __raw_spin_unlock(&lockdep_lock);
114
115         return ret;
116 }
117
118 static int lockdep_initialized;
119
120 unsigned long nr_list_entries;
121 struct lock_list list_entries[MAX_LOCKDEP_ENTRIES];
122
123 /*
124  * All data structures here are protected by the global debug_lock.
125  *
126  * Mutex key structs only get allocated, once during bootup, and never
127  * get freed - this significantly simplifies the debugging code.
128  */
129 unsigned long nr_lock_classes;
130 static struct lock_class lock_classes[MAX_LOCKDEP_KEYS];
131
132 static inline struct lock_class *hlock_class(struct held_lock *hlock)
133 {
134         if (!hlock->class_idx) {
135                 DEBUG_LOCKS_WARN_ON(1);
136                 return NULL;
137         }
138         return lock_classes + hlock->class_idx - 1;
139 }
140
141 #ifdef CONFIG_LOCK_STAT
142 static DEFINE_PER_CPU(struct lock_class_stats[MAX_LOCKDEP_KEYS], lock_stats);
143
144 static int lock_point(unsigned long points[], unsigned long ip)
145 {
146         int i;
147
148         for (i = 0; i < LOCKSTAT_POINTS; i++) {
149                 if (points[i] == 0) {
150                         points[i] = ip;
151                         break;
152                 }
153                 if (points[i] == ip)
154                         break;
155         }
156
157         return i;
158 }
159
160 static void lock_time_inc(struct lock_time *lt, s64 time)
161 {
162         if (time > lt->max)
163                 lt->max = time;
164
165         if (time < lt->min || !lt->min)
166                 lt->min = time;
167
168         lt->total += time;
169         lt->nr++;
170 }
171
172 static inline void lock_time_add(struct lock_time *src, struct lock_time *dst)
173 {
174         dst->min += src->min;
175         dst->max += src->max;
176         dst->total += src->total;
177         dst->nr += src->nr;
178 }
179
180 struct lock_class_stats lock_stats(struct lock_class *class)
181 {
182         struct lock_class_stats stats;
183         int cpu, i;
184
185         memset(&stats, 0, sizeof(struct lock_class_stats));
186         for_each_possible_cpu(cpu) {
187                 struct lock_class_stats *pcs =
188                         &per_cpu(lock_stats, cpu)[class - lock_classes];
189
190                 for (i = 0; i < ARRAY_SIZE(stats.contention_point); i++)
191                         stats.contention_point[i] += pcs->contention_point[i];
192
193                 for (i = 0; i < ARRAY_SIZE(stats.contending_point); i++)
194                         stats.contending_point[i] += pcs->contending_point[i];
195
196                 lock_time_add(&pcs->read_waittime, &stats.read_waittime);
197                 lock_time_add(&pcs->write_waittime, &stats.write_waittime);
198
199                 lock_time_add(&pcs->read_holdtime, &stats.read_holdtime);
200                 lock_time_add(&pcs->write_holdtime, &stats.write_holdtime);
201
202                 for (i = 0; i < ARRAY_SIZE(stats.bounces); i++)
203                         stats.bounces[i] += pcs->bounces[i];
204         }
205
206         return stats;
207 }
208
209 void clear_lock_stats(struct lock_class *class)
210 {
211         int cpu;
212
213         for_each_possible_cpu(cpu) {
214                 struct lock_class_stats *cpu_stats =
215                         &per_cpu(lock_stats, cpu)[class - lock_classes];
216
217                 memset(cpu_stats, 0, sizeof(struct lock_class_stats));
218         }
219         memset(class->contention_point, 0, sizeof(class->contention_point));
220         memset(class->contending_point, 0, sizeof(class->contending_point));
221 }
222
223 static struct lock_class_stats *get_lock_stats(struct lock_class *class)
224 {
225         return &get_cpu_var(lock_stats)[class - lock_classes];
226 }
227
228 static void put_lock_stats(struct lock_class_stats *stats)
229 {
230         put_cpu_var(lock_stats);
231 }
232
233 static void lock_release_holdtime(struct held_lock *hlock)
234 {
235         struct lock_class_stats *stats;
236         s64 holdtime;
237
238         if (!lock_stat)
239                 return;
240
241         holdtime = sched_clock() - hlock->holdtime_stamp;
242
243         stats = get_lock_stats(hlock_class(hlock));
244         if (hlock->read)
245                 lock_time_inc(&stats->read_holdtime, holdtime);
246         else
247                 lock_time_inc(&stats->write_holdtime, holdtime);
248         put_lock_stats(stats);
249 }
250 #else
251 static inline void lock_release_holdtime(struct held_lock *hlock)
252 {
253 }
254 #endif
255
256 /*
257  * We keep a global list of all lock classes. The list only grows,
258  * never shrinks. The list is only accessed with the lockdep
259  * spinlock lock held.
260  */
261 LIST_HEAD(all_lock_classes);
262
263 /*
264  * The lockdep classes are in a hash-table as well, for fast lookup:
265  */
266 #define CLASSHASH_BITS          (MAX_LOCKDEP_KEYS_BITS - 1)
267 #define CLASSHASH_SIZE          (1UL << CLASSHASH_BITS)
268 #define __classhashfn(key)      hash_long((unsigned long)key, CLASSHASH_BITS)
269 #define classhashentry(key)     (classhash_table + __classhashfn((key)))
270
271 static struct list_head classhash_table[CLASSHASH_SIZE];
272
273 /*
274  * We put the lock dependency chains into a hash-table as well, to cache
275  * their existence:
276  */
277 #define CHAINHASH_BITS          (MAX_LOCKDEP_CHAINS_BITS-1)
278 #define CHAINHASH_SIZE          (1UL << CHAINHASH_BITS)
279 #define __chainhashfn(chain)    hash_long(chain, CHAINHASH_BITS)
280 #define chainhashentry(chain)   (chainhash_table + __chainhashfn((chain)))
281
282 static struct list_head chainhash_table[CHAINHASH_SIZE];
283
284 /*
285  * The hash key of the lock dependency chains is a hash itself too:
286  * it's a hash of all locks taken up to that lock, including that lock.
287  * It's a 64-bit hash, because it's important for the keys to be
288  * unique.
289  */
290 #define iterate_chain_key(key1, key2) \
291         (((key1) << MAX_LOCKDEP_KEYS_BITS) ^ \
292         ((key1) >> (64-MAX_LOCKDEP_KEYS_BITS)) ^ \
293         (key2))
294
295 void lockdep_off(void)
296 {
297         current->lockdep_recursion++;
298 }
299 EXPORT_SYMBOL(lockdep_off);
300
301 void lockdep_on(void)
302 {
303         current->lockdep_recursion--;
304 }
305 EXPORT_SYMBOL(lockdep_on);
306
307 /*
308  * Debugging switches:
309  */
310
311 #define VERBOSE                 0
312 #define VERY_VERBOSE            0
313
314 #if VERBOSE
315 # define HARDIRQ_VERBOSE        1
316 # define SOFTIRQ_VERBOSE        1
317 # define RECLAIM_VERBOSE        1
318 #else
319 # define HARDIRQ_VERBOSE        0
320 # define SOFTIRQ_VERBOSE        0
321 # define RECLAIM_VERBOSE        0
322 #endif
323
324 #if VERBOSE || HARDIRQ_VERBOSE || SOFTIRQ_VERBOSE || RECLAIM_VERBOSE
325 /*
326  * Quick filtering for interesting events:
327  */
328 static int class_filter(struct lock_class *class)
329 {
330 #if 0
331         /* Example */
332         if (class->name_version == 1 &&
333                         !strcmp(class->name, "lockname"))
334                 return 1;
335         if (class->name_version == 1 &&
336                         !strcmp(class->name, "&struct->lockfield"))
337                 return 1;
338 #endif
339         /* Filter everything else. 1 would be to allow everything else */
340         return 0;
341 }
342 #endif
343
344 static int verbose(struct lock_class *class)
345 {
346 #if VERBOSE
347         return class_filter(class);
348 #endif
349         return 0;
350 }
351
352 /*
353  * Stack-trace: tightly packed array of stack backtrace
354  * addresses. Protected by the graph_lock.
355  */
356 unsigned long nr_stack_trace_entries;
357 static unsigned long stack_trace[MAX_STACK_TRACE_ENTRIES];
358
359 static int save_trace(struct stack_trace *trace)
360 {
361         trace->nr_entries = 0;
362         trace->max_entries = MAX_STACK_TRACE_ENTRIES - nr_stack_trace_entries;
363         trace->entries = stack_trace + nr_stack_trace_entries;
364
365         trace->skip = 3;
366
367         save_stack_trace(trace);
368
369         trace->max_entries = trace->nr_entries;
370
371         nr_stack_trace_entries += trace->nr_entries;
372
373         if (nr_stack_trace_entries == MAX_STACK_TRACE_ENTRIES) {
374                 if (!debug_locks_off_graph_unlock())
375                         return 0;
376
377                 printk("BUG: MAX_STACK_TRACE_ENTRIES too low!\n");
378                 printk("turning off the locking correctness validator.\n");
379                 dump_stack();
380
381                 return 0;
382         }
383
384         return 1;
385 }
386
387 unsigned int nr_hardirq_chains;
388 unsigned int nr_softirq_chains;
389 unsigned int nr_process_chains;
390 unsigned int max_lockdep_depth;
391 unsigned int max_recursion_depth;
392
393 static unsigned int lockdep_dependency_gen_id;
394
395 static bool lockdep_dependency_visit(struct lock_class *source,
396                                      unsigned int depth)
397 {
398         if (!depth)
399                 lockdep_dependency_gen_id++;
400         if (source->dep_gen_id == lockdep_dependency_gen_id)
401                 return true;
402         source->dep_gen_id = lockdep_dependency_gen_id;
403         return false;
404 }
405
406 #ifdef CONFIG_DEBUG_LOCKDEP
407 /*
408  * We cannot printk in early bootup code. Not even early_printk()
409  * might work. So we mark any initialization errors and printk
410  * about it later on, in lockdep_info().
411  */
412 static int lockdep_init_error;
413 static unsigned long lockdep_init_trace_data[20];
414 static struct stack_trace lockdep_init_trace = {
415         .max_entries = ARRAY_SIZE(lockdep_init_trace_data),
416         .entries = lockdep_init_trace_data,
417 };
418
419 /*
420  * Various lockdep statistics:
421  */
422 atomic_t chain_lookup_hits;
423 atomic_t chain_lookup_misses;
424 atomic_t hardirqs_on_events;
425 atomic_t hardirqs_off_events;
426 atomic_t redundant_hardirqs_on;
427 atomic_t redundant_hardirqs_off;
428 atomic_t softirqs_on_events;
429 atomic_t softirqs_off_events;
430 atomic_t redundant_softirqs_on;
431 atomic_t redundant_softirqs_off;
432 atomic_t nr_unused_locks;
433 atomic_t nr_cyclic_checks;
434 atomic_t nr_cyclic_check_recursions;
435 atomic_t nr_find_usage_forwards_checks;
436 atomic_t nr_find_usage_forwards_recursions;
437 atomic_t nr_find_usage_backwards_checks;
438 atomic_t nr_find_usage_backwards_recursions;
439 #endif
440
441 /*
442  * Locking printouts:
443  */
444
445 #define __USAGE(__STATE)                                                \
446         [LOCK_USED_IN_##__STATE] = "IN-"__stringify(__STATE)"-W",       \
447         [LOCK_ENABLED_##__STATE] = __stringify(__STATE)"-ON-W",         \
448         [LOCK_USED_IN_##__STATE##_READ] = "IN-"__stringify(__STATE)"-R",\
449         [LOCK_ENABLED_##__STATE##_READ] = __stringify(__STATE)"-ON-R",
450
451 static const char *usage_str[] =
452 {
453 #define LOCKDEP_STATE(__STATE) __USAGE(__STATE)
454 #include "lockdep_states.h"
455 #undef LOCKDEP_STATE
456         [LOCK_USED] = "INITIAL USE",
457 };
458
459 const char * __get_key_name(struct lockdep_subclass_key *key, char *str)
460 {
461         return kallsyms_lookup((unsigned long)key, NULL, NULL, NULL, str);
462 }
463
464 static inline unsigned long lock_flag(enum lock_usage_bit bit)
465 {
466         return 1UL << bit;
467 }
468
469 static char get_usage_char(struct lock_class *class, enum lock_usage_bit bit)
470 {
471         char c = '.';
472
473         if (class->usage_mask & lock_flag(bit + 2))
474                 c = '+';
475         if (class->usage_mask & lock_flag(bit)) {
476                 c = '-';
477                 if (class->usage_mask & lock_flag(bit + 2))
478                         c = '?';
479         }
480
481         return c;
482 }
483
484 void get_usage_chars(struct lock_class *class, char usage[LOCK_USAGE_CHARS])
485 {
486         int i = 0;
487
488 #define LOCKDEP_STATE(__STATE)                                          \
489         usage[i++] = get_usage_char(class, LOCK_USED_IN_##__STATE);     \
490         usage[i++] = get_usage_char(class, LOCK_USED_IN_##__STATE##_READ);
491 #include "lockdep_states.h"
492 #undef LOCKDEP_STATE
493
494         usage[i] = '\0';
495 }
496
497 static void print_lock_name(struct lock_class *class)
498 {
499         char str[KSYM_NAME_LEN], usage[LOCK_USAGE_CHARS];
500         const char *name;
501
502         get_usage_chars(class, usage);
503
504         name = class->name;
505         if (!name) {
506                 name = __get_key_name(class->key, str);
507                 printk(" (%s", name);
508         } else {
509                 printk(" (%s", name);
510                 if (class->name_version > 1)
511                         printk("#%d", class->name_version);
512                 if (class->subclass)
513                         printk("/%d", class->subclass);
514         }
515         printk("){%s}", usage);
516 }
517
518 static void print_lockdep_cache(struct lockdep_map *lock)
519 {
520         const char *name;
521         char str[KSYM_NAME_LEN];
522
523         name = lock->name;
524         if (!name)
525                 name = __get_key_name(lock->key->subkeys, str);
526
527         printk("%s", name);
528 }
529
530 static void print_lock(struct held_lock *hlock)
531 {
532         print_lock_name(hlock_class(hlock));
533         printk(", at: ");
534         print_ip_sym(hlock->acquire_ip);
535 }
536
537 static void lockdep_print_held_locks(struct task_struct *curr)
538 {
539         int i, depth = curr->lockdep_depth;
540
541         if (!depth) {
542                 printk("no locks held by %s/%d.\n", curr->comm, task_pid_nr(curr));
543                 return;
544         }
545         printk("%d lock%s held by %s/%d:\n",
546                 depth, depth > 1 ? "s" : "", curr->comm, task_pid_nr(curr));
547
548         for (i = 0; i < depth; i++) {
549                 printk(" #%d: ", i);
550                 print_lock(curr->held_locks + i);
551         }
552 }
553
554 static void print_lock_class_header(struct lock_class *class, int depth)
555 {
556         int bit;
557
558         printk("%*s->", depth, "");
559         print_lock_name(class);
560         printk(" ops: %lu", class->ops);
561         printk(" {\n");
562
563         for (bit = 0; bit < LOCK_USAGE_STATES; bit++) {
564                 if (class->usage_mask & (1 << bit)) {
565                         int len = depth;
566
567                         len += printk("%*s   %s", depth, "", usage_str[bit]);
568                         len += printk(" at:\n");
569                         print_stack_trace(class->usage_traces + bit, len);
570                 }
571         }
572         printk("%*s }\n", depth, "");
573
574         printk("%*s ... key      at: ",depth,"");
575         print_ip_sym((unsigned long)class->key);
576 }
577
578 /*
579  * printk all lock dependencies starting at <entry>:
580  */
581 static void __used
582 print_lock_dependencies(struct lock_class *class, int depth)
583 {
584         struct lock_list *entry;
585
586         if (lockdep_dependency_visit(class, depth))
587                 return;
588
589         if (DEBUG_LOCKS_WARN_ON(depth >= 20))
590                 return;
591
592         print_lock_class_header(class, depth);
593
594         list_for_each_entry(entry, &class->locks_after, entry) {
595                 if (DEBUG_LOCKS_WARN_ON(!entry->class))
596                         return;
597
598                 print_lock_dependencies(entry->class, depth + 1);
599
600                 printk("%*s ... acquired at:\n",depth,"");
601                 print_stack_trace(&entry->trace, 2);
602                 printk("\n");
603         }
604 }
605
606 static void print_kernel_version(void)
607 {
608         printk("%s %.*s\n", init_utsname()->release,
609                 (int)strcspn(init_utsname()->version, " "),
610                 init_utsname()->version);
611 }
612
613 static int very_verbose(struct lock_class *class)
614 {
615 #if VERY_VERBOSE
616         return class_filter(class);
617 #endif
618         return 0;
619 }
620
621 /*
622  * Is this the address of a static object:
623  */
624 static int static_obj(void *obj)
625 {
626         unsigned long start = (unsigned long) &_stext,
627                       end   = (unsigned long) &_end,
628                       addr  = (unsigned long) obj;
629 #ifdef CONFIG_SMP
630         int i;
631 #endif
632
633         /*
634          * static variable?
635          */
636         if ((addr >= start) && (addr < end))
637                 return 1;
638
639 #ifdef CONFIG_SMP
640         /*
641          * percpu var?
642          */
643         for_each_possible_cpu(i) {
644                 start = (unsigned long) &__per_cpu_start + per_cpu_offset(i);
645                 end   = (unsigned long) &__per_cpu_start + PERCPU_ENOUGH_ROOM
646                                         + per_cpu_offset(i);
647
648                 if ((addr >= start) && (addr < end))
649                         return 1;
650         }
651 #endif
652
653         /*
654          * module var?
655          */
656         return is_module_address(addr);
657 }
658
659 /*
660  * To make lock name printouts unique, we calculate a unique
661  * class->name_version generation counter:
662  */
663 static int count_matching_names(struct lock_class *new_class)
664 {
665         struct lock_class *class;
666         int count = 0;
667
668         if (!new_class->name)
669                 return 0;
670
671         list_for_each_entry(class, &all_lock_classes, lock_entry) {
672                 if (new_class->key - new_class->subclass == class->key)
673                         return class->name_version;
674                 if (class->name && !strcmp(class->name, new_class->name))
675                         count = max(count, class->name_version);
676         }
677
678         return count + 1;
679 }
680
681 /*
682  * Register a lock's class in the hash-table, if the class is not present
683  * yet. Otherwise we look it up. We cache the result in the lock object
684  * itself, so actual lookup of the hash should be once per lock object.
685  */
686 static inline struct lock_class *
687 look_up_lock_class(struct lockdep_map *lock, unsigned int subclass)
688 {
689         struct lockdep_subclass_key *key;
690         struct list_head *hash_head;
691         struct lock_class *class;
692
693 #ifdef CONFIG_DEBUG_LOCKDEP
694         /*
695          * If the architecture calls into lockdep before initializing
696          * the hashes then we'll warn about it later. (we cannot printk
697          * right now)
698          */
699         if (unlikely(!lockdep_initialized)) {
700                 lockdep_init();
701                 lockdep_init_error = 1;
702                 save_stack_trace(&lockdep_init_trace);
703         }
704 #endif
705
706         /*
707          * Static locks do not have their class-keys yet - for them the key
708          * is the lock object itself:
709          */
710         if (unlikely(!lock->key))
711                 lock->key = (void *)lock;
712
713         /*
714          * NOTE: the class-key must be unique. For dynamic locks, a static
715          * lock_class_key variable is passed in through the mutex_init()
716          * (or spin_lock_init()) call - which acts as the key. For static
717          * locks we use the lock object itself as the key.
718          */
719         BUILD_BUG_ON(sizeof(struct lock_class_key) >
720                         sizeof(struct lockdep_map));
721
722         key = lock->key->subkeys + subclass;
723
724         hash_head = classhashentry(key);
725
726         /*
727          * We can walk the hash lockfree, because the hash only
728          * grows, and we are careful when adding entries to the end:
729          */
730         list_for_each_entry(class, hash_head, hash_entry) {
731                 if (class->key == key) {
732                         WARN_ON_ONCE(class->name != lock->name);
733                         return class;
734                 }
735         }
736
737         return NULL;
738 }
739
740 /*
741  * Register a lock's class in the hash-table, if the class is not present
742  * yet. Otherwise we look it up. We cache the result in the lock object
743  * itself, so actual lookup of the hash should be once per lock object.
744  */
745 static inline struct lock_class *
746 register_lock_class(struct lockdep_map *lock, unsigned int subclass, int force)
747 {
748         struct lockdep_subclass_key *key;
749         struct list_head *hash_head;
750         struct lock_class *class;
751         unsigned long flags;
752
753         class = look_up_lock_class(lock, subclass);
754         if (likely(class))
755                 return class;
756
757         /*
758          * Debug-check: all keys must be persistent!
759          */
760         if (!static_obj(lock->key)) {
761                 debug_locks_off();
762                 printk("INFO: trying to register non-static key.\n");
763                 printk("the code is fine but needs lockdep annotation.\n");
764                 printk("turning off the locking correctness validator.\n");
765                 dump_stack();
766
767                 return NULL;
768         }
769
770         key = lock->key->subkeys + subclass;
771         hash_head = classhashentry(key);
772
773         raw_local_irq_save(flags);
774         if (!graph_lock()) {
775                 raw_local_irq_restore(flags);
776                 return NULL;
777         }
778         /*
779          * We have to do the hash-walk again, to avoid races
780          * with another CPU:
781          */
782         list_for_each_entry(class, hash_head, hash_entry)
783                 if (class->key == key)
784                         goto out_unlock_set;
785         /*
786          * Allocate a new key from the static array, and add it to
787          * the hash:
788          */
789         if (nr_lock_classes >= MAX_LOCKDEP_KEYS) {
790                 if (!debug_locks_off_graph_unlock()) {
791                         raw_local_irq_restore(flags);
792                         return NULL;
793                 }
794                 raw_local_irq_restore(flags);
795
796                 printk("BUG: MAX_LOCKDEP_KEYS too low!\n");
797                 printk("turning off the locking correctness validator.\n");
798                 dump_stack();
799                 return NULL;
800         }
801         class = lock_classes + nr_lock_classes++;
802         debug_atomic_inc(&nr_unused_locks);
803         class->key = key;
804         class->name = lock->name;
805         class->subclass = subclass;
806         INIT_LIST_HEAD(&class->lock_entry);
807         INIT_LIST_HEAD(&class->locks_before);
808         INIT_LIST_HEAD(&class->locks_after);
809         class->name_version = count_matching_names(class);
810         /*
811          * We use RCU's safe list-add method to make
812          * parallel walking of the hash-list safe:
813          */
814         list_add_tail_rcu(&class->hash_entry, hash_head);
815         /*
816          * Add it to the global list of classes:
817          */
818         list_add_tail_rcu(&class->lock_entry, &all_lock_classes);
819
820         if (verbose(class)) {
821                 graph_unlock();
822                 raw_local_irq_restore(flags);
823
824                 printk("\nnew class %p: %s", class->key, class->name);
825                 if (class->name_version > 1)
826                         printk("#%d", class->name_version);
827                 printk("\n");
828                 dump_stack();
829
830                 raw_local_irq_save(flags);
831                 if (!graph_lock()) {
832                         raw_local_irq_restore(flags);
833                         return NULL;
834                 }
835         }
836 out_unlock_set:
837         graph_unlock();
838         raw_local_irq_restore(flags);
839
840         if (!subclass || force)
841                 lock->class_cache = class;
842
843         if (DEBUG_LOCKS_WARN_ON(class->subclass != subclass))
844                 return NULL;
845
846         return class;
847 }
848
849 #ifdef CONFIG_PROVE_LOCKING
850 /*
851  * Allocate a lockdep entry. (assumes the graph_lock held, returns
852  * with NULL on failure)
853  */
854 static struct lock_list *alloc_list_entry(void)
855 {
856         if (nr_list_entries >= MAX_LOCKDEP_ENTRIES) {
857                 if (!debug_locks_off_graph_unlock())
858                         return NULL;
859
860                 printk("BUG: MAX_LOCKDEP_ENTRIES too low!\n");
861                 printk("turning off the locking correctness validator.\n");
862                 dump_stack();
863                 return NULL;
864         }
865         return list_entries + nr_list_entries++;
866 }
867
868 /*
869  * Add a new dependency to the head of the list:
870  */
871 static int add_lock_to_list(struct lock_class *class, struct lock_class *this,
872                             struct list_head *head, unsigned long ip, int distance)
873 {
874         struct lock_list *entry;
875         /*
876          * Lock not present yet - get a new dependency struct and
877          * add it to the list:
878          */
879         entry = alloc_list_entry();
880         if (!entry)
881                 return 0;
882
883         if (!save_trace(&entry->trace))
884                 return 0;
885
886         entry->class = this;
887         entry->distance = distance;
888         /*
889          * Since we never remove from the dependency list, the list can
890          * be walked lockless by other CPUs, it's only allocation
891          * that must be protected by the spinlock. But this also means
892          * we must make new entries visible only once writes to the
893          * entry become visible - hence the RCU op:
894          */
895         list_add_tail_rcu(&entry->entry, head);
896
897         return 1;
898 }
899
900 unsigned long bfs_accessed[BITS_TO_LONGS(MAX_LOCKDEP_ENTRIES)];
901 static struct circular_queue  lock_cq;
902
903 static int __search_shortest_path(struct lock_list *source_entry,
904                                 struct lock_class *target,
905                                 struct lock_list **target_entry,
906                                 int forward)
907 {
908         struct lock_list *entry;
909         struct list_head *head;
910         struct circular_queue *cq = &lock_cq;
911         int ret = 1;
912
913         if (source_entry->class == target) {
914                 *target_entry = source_entry;
915                 ret = 0;
916                 goto exit;
917         }
918
919         if (forward)
920                 head = &source_entry->class->locks_after;
921         else
922                 head = &source_entry->class->locks_before;
923
924         if (list_empty(head))
925                 goto exit;
926
927         __cq_init(cq);
928         __cq_enqueue(cq, (unsigned long)source_entry);
929
930         while (!__cq_empty(cq)) {
931                 struct lock_list *lock;
932
933                 __cq_dequeue(cq, (unsigned long *)&lock);
934
935                 if (!lock->class) {
936                         ret = -2;
937                         goto exit;
938                 }
939
940                 if (forward)
941                         head = &lock->class->locks_after;
942                 else
943                         head = &lock->class->locks_before;
944
945                 list_for_each_entry(entry, head, entry) {
946                         if (!lock_accessed(entry)) {
947                                 mark_lock_accessed(entry, lock);
948                                 if (entry->class == target) {
949                                         *target_entry = entry;
950                                         ret = 0;
951                                         goto exit;
952                                 }
953
954                                 if (__cq_enqueue(cq, (unsigned long)entry)) {
955                                         ret = -1;
956                                         goto exit;
957                                 }
958                         }
959                 }
960         }
961 exit:
962         return ret;
963 }
964
965 static inline int __search_forward_shortest_path(struct lock_list *src_entry,
966                                 struct lock_class *target,
967                                 struct lock_list **target_entry)
968 {
969         return __search_shortest_path(src_entry, target, target_entry, 1);
970
971 }
972
973 static inline int __search_backward_shortest_path(struct lock_list *src_entry,
974                                 struct lock_class *target,
975                                 struct lock_list **target_entry)
976 {
977         return __search_shortest_path(src_entry, target, target_entry, 0);
978
979 }
980
981 /*
982  * Recursive, forwards-direction lock-dependency checking, used for
983  * both noncyclic checking and for hardirq-unsafe/softirq-unsafe
984  * checking.
985  *
986  * (to keep the stackframe of the recursive functions small we
987  *  use these global variables, and we also mark various helper
988  *  functions as noinline.)
989  */
990 static struct held_lock *check_source, *check_target;
991
992 /*
993  * Print a dependency chain entry (this is only done when a deadlock
994  * has been detected):
995  */
996 static noinline int
997 print_circular_bug_entry(struct lock_list *target, unsigned int depth)
998 {
999         if (debug_locks_silent)
1000                 return 0;
1001         printk("\n-> #%u", depth);
1002         print_lock_name(target->class);
1003         printk(":\n");
1004         print_stack_trace(&target->trace, 6);
1005
1006         return 0;
1007 }
1008
1009 /*
1010  * When a circular dependency is detected, print the
1011  * header first:
1012  */
1013 static noinline int
1014 print_circular_bug_header(struct lock_list *entry, unsigned int depth)
1015 {
1016         struct task_struct *curr = current;
1017
1018         if (debug_locks_silent)
1019                 return 0;
1020
1021         printk("\n=======================================================\n");
1022         printk(  "[ INFO: possible circular locking dependency detected ]\n");
1023         print_kernel_version();
1024         printk(  "-------------------------------------------------------\n");
1025         printk("%s/%d is trying to acquire lock:\n",
1026                 curr->comm, task_pid_nr(curr));
1027         print_lock(check_source);
1028         printk("\nbut task is already holding lock:\n");
1029         print_lock(check_target);
1030         printk("\nwhich lock already depends on the new lock.\n\n");
1031         printk("\nthe existing dependency chain (in reverse order) is:\n");
1032
1033         print_circular_bug_entry(entry, depth);
1034
1035         return 0;
1036 }
1037
1038 static noinline int print_circular_bug(void)
1039 {
1040         struct task_struct *curr = current;
1041         struct lock_list this;
1042         struct lock_list *target;
1043         struct lock_list *parent;
1044         int result;
1045         unsigned long depth;
1046
1047         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1048                 return 0;
1049
1050         this.class = hlock_class(check_source);
1051         this.parent = NULL;
1052         if (!save_trace(&this.trace))
1053                 return 0;
1054
1055         result = __search_forward_shortest_path(&this,
1056                                                 hlock_class(check_target),
1057                                                 &target);
1058         if (result) {
1059                 printk("\n%s:search shortest path failed:%d\n", __func__,
1060                         result);
1061                 return 0;
1062         }
1063
1064         depth = get_lock_depth(target);
1065
1066         print_circular_bug_header(target, depth);
1067
1068         parent = get_lock_parent(target);
1069
1070         while (parent) {
1071                 print_circular_bug_entry(parent, --depth);
1072                 parent = get_lock_parent(parent);
1073         }
1074
1075         printk("\nother info that might help us debug this:\n\n");
1076         lockdep_print_held_locks(curr);
1077
1078         printk("\nstack backtrace:\n");
1079         dump_stack();
1080
1081         return 0;
1082 }
1083
1084 #define RECURSION_LIMIT 40
1085
1086 static int noinline print_infinite_recursion_bug(void)
1087 {
1088         if (!debug_locks_off_graph_unlock())
1089                 return 0;
1090
1091         WARN_ON(1);
1092
1093         return 0;
1094 }
1095
1096 unsigned long __lockdep_count_forward_deps(struct lock_class *class,
1097                                            unsigned int depth)
1098 {
1099         struct lock_list *entry;
1100         unsigned long ret = 1;
1101
1102         if (lockdep_dependency_visit(class, depth))
1103                 return 0;
1104
1105         /*
1106          * Recurse this class's dependency list:
1107          */
1108         list_for_each_entry(entry, &class->locks_after, entry)
1109                 ret += __lockdep_count_forward_deps(entry->class, depth + 1);
1110
1111         return ret;
1112 }
1113
1114 unsigned long lockdep_count_forward_deps(struct lock_class *class)
1115 {
1116         unsigned long ret, flags;
1117
1118         local_irq_save(flags);
1119         __raw_spin_lock(&lockdep_lock);
1120         ret = __lockdep_count_forward_deps(class, 0);
1121         __raw_spin_unlock(&lockdep_lock);
1122         local_irq_restore(flags);
1123
1124         return ret;
1125 }
1126
1127 unsigned long __lockdep_count_backward_deps(struct lock_class *class,
1128                                             unsigned int depth)
1129 {
1130         struct lock_list *entry;
1131         unsigned long ret = 1;
1132
1133         if (lockdep_dependency_visit(class, depth))
1134                 return 0;
1135         /*
1136          * Recurse this class's dependency list:
1137          */
1138         list_for_each_entry(entry, &class->locks_before, entry)
1139                 ret += __lockdep_count_backward_deps(entry->class, depth + 1);
1140
1141         return ret;
1142 }
1143
1144 unsigned long lockdep_count_backward_deps(struct lock_class *class)
1145 {
1146         unsigned long ret, flags;
1147
1148         local_irq_save(flags);
1149         __raw_spin_lock(&lockdep_lock);
1150         ret = __lockdep_count_backward_deps(class, 0);
1151         __raw_spin_unlock(&lockdep_lock);
1152         local_irq_restore(flags);
1153
1154         return ret;
1155 }
1156
1157 /*
1158  * Prove that the dependency graph starting at <entry> can not
1159  * lead to <target>. Print an error and return 0 if it does.
1160  */
1161 static noinline int
1162 check_noncircular(struct lock_class *source, unsigned int depth)
1163 {
1164         struct lock_list *entry;
1165
1166         if (lockdep_dependency_visit(source, depth))
1167                 return 1;
1168
1169         debug_atomic_inc(&nr_cyclic_check_recursions);
1170         if (depth > max_recursion_depth)
1171                 max_recursion_depth = depth;
1172         if (depth >= RECURSION_LIMIT)
1173                 return print_infinite_recursion_bug();
1174         /*
1175          * Check this lock's dependency list:
1176          */
1177         list_for_each_entry(entry, &source->locks_after, entry) {
1178                 if (entry->class == hlock_class(check_target))
1179                         return 2;
1180                 debug_atomic_inc(&nr_cyclic_checks);
1181                 if (check_noncircular(entry->class, depth+1) == 2)
1182                         return 2;
1183         }
1184         return 1;
1185 }
1186
1187
1188 #if defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING)
1189 /*
1190  * Forwards and backwards subgraph searching, for the purposes of
1191  * proving that two subgraphs can be connected by a new dependency
1192  * without creating any illegal irq-safe -> irq-unsafe lock dependency.
1193  */
1194 static enum lock_usage_bit find_usage_bit;
1195 static struct lock_class *forwards_match, *backwards_match;
1196
1197 /*
1198  * Find a node in the forwards-direction dependency sub-graph starting
1199  * at <source> that matches <find_usage_bit>.
1200  *
1201  * Return 2 if such a node exists in the subgraph, and put that node
1202  * into <forwards_match>.
1203  *
1204  * Return 1 otherwise and keep <forwards_match> unchanged.
1205  * Return 0 on error.
1206  */
1207 static noinline int
1208 find_usage_forwards(struct lock_class *source, unsigned int depth)
1209 {
1210         struct lock_list *entry;
1211         int ret;
1212
1213         if (lockdep_dependency_visit(source, depth))
1214                 return 1;
1215
1216         if (depth > max_recursion_depth)
1217                 max_recursion_depth = depth;
1218         if (depth >= RECURSION_LIMIT)
1219                 return print_infinite_recursion_bug();
1220
1221         debug_atomic_inc(&nr_find_usage_forwards_checks);
1222         if (source->usage_mask & (1 << find_usage_bit)) {
1223                 forwards_match = source;
1224                 return 2;
1225         }
1226
1227         /*
1228          * Check this lock's dependency list:
1229          */
1230         list_for_each_entry(entry, &source->locks_after, entry) {
1231                 debug_atomic_inc(&nr_find_usage_forwards_recursions);
1232                 ret = find_usage_forwards(entry->class, depth+1);
1233                 if (ret == 2 || ret == 0)
1234                         return ret;
1235         }
1236         return 1;
1237 }
1238
1239 /*
1240  * Find a node in the backwards-direction dependency sub-graph starting
1241  * at <source> that matches <find_usage_bit>.
1242  *
1243  * Return 2 if such a node exists in the subgraph, and put that node
1244  * into <backwards_match>.
1245  *
1246  * Return 1 otherwise and keep <backwards_match> unchanged.
1247  * Return 0 on error.
1248  */
1249 static noinline int
1250 find_usage_backwards(struct lock_class *source, unsigned int depth)
1251 {
1252         struct lock_list *entry;
1253         int ret;
1254
1255         if (lockdep_dependency_visit(source, depth))
1256                 return 1;
1257
1258         if (!__raw_spin_is_locked(&lockdep_lock))
1259                 return DEBUG_LOCKS_WARN_ON(1);
1260
1261         if (depth > max_recursion_depth)
1262                 max_recursion_depth = depth;
1263         if (depth >= RECURSION_LIMIT)
1264                 return print_infinite_recursion_bug();
1265
1266         debug_atomic_inc(&nr_find_usage_backwards_checks);
1267         if (source->usage_mask & (1 << find_usage_bit)) {
1268                 backwards_match = source;
1269                 return 2;
1270         }
1271
1272         if (!source && debug_locks_off_graph_unlock()) {
1273                 WARN_ON(1);
1274                 return 0;
1275         }
1276
1277         /*
1278          * Check this lock's dependency list:
1279          */
1280         list_for_each_entry(entry, &source->locks_before, entry) {
1281                 debug_atomic_inc(&nr_find_usage_backwards_recursions);
1282                 ret = find_usage_backwards(entry->class, depth+1);
1283                 if (ret == 2 || ret == 0)
1284                         return ret;
1285         }
1286         return 1;
1287 }
1288
1289 static int
1290 print_bad_irq_dependency(struct task_struct *curr,
1291                          struct held_lock *prev,
1292                          struct held_lock *next,
1293                          enum lock_usage_bit bit1,
1294                          enum lock_usage_bit bit2,
1295                          const char *irqclass)
1296 {
1297         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1298                 return 0;
1299
1300         printk("\n======================================================\n");
1301         printk(  "[ INFO: %s-safe -> %s-unsafe lock order detected ]\n",
1302                 irqclass, irqclass);
1303         print_kernel_version();
1304         printk(  "------------------------------------------------------\n");
1305         printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] is trying to acquire:\n",
1306                 curr->comm, task_pid_nr(curr),
1307                 curr->hardirq_context, hardirq_count() >> HARDIRQ_SHIFT,
1308                 curr->softirq_context, softirq_count() >> SOFTIRQ_SHIFT,
1309                 curr->hardirqs_enabled,
1310                 curr->softirqs_enabled);
1311         print_lock(next);
1312
1313         printk("\nand this task is already holding:\n");
1314         print_lock(prev);
1315         printk("which would create a new lock dependency:\n");
1316         print_lock_name(hlock_class(prev));
1317         printk(" ->");
1318         print_lock_name(hlock_class(next));
1319         printk("\n");
1320
1321         printk("\nbut this new dependency connects a %s-irq-safe lock:\n",
1322                 irqclass);
1323         print_lock_name(backwards_match);
1324         printk("\n... which became %s-irq-safe at:\n", irqclass);
1325
1326         print_stack_trace(backwards_match->usage_traces + bit1, 1);
1327
1328         printk("\nto a %s-irq-unsafe lock:\n", irqclass);
1329         print_lock_name(forwards_match);
1330         printk("\n... which became %s-irq-unsafe at:\n", irqclass);
1331         printk("...");
1332
1333         print_stack_trace(forwards_match->usage_traces + bit2, 1);
1334
1335         printk("\nother info that might help us debug this:\n\n");
1336         lockdep_print_held_locks(curr);
1337
1338         printk("\nthe %s-irq-safe lock's dependencies:\n", irqclass);
1339         print_lock_dependencies(backwards_match, 0);
1340
1341         printk("\nthe %s-irq-unsafe lock's dependencies:\n", irqclass);
1342         print_lock_dependencies(forwards_match, 0);
1343
1344         printk("\nstack backtrace:\n");
1345         dump_stack();
1346
1347         return 0;
1348 }
1349
1350 static int
1351 check_usage(struct task_struct *curr, struct held_lock *prev,
1352             struct held_lock *next, enum lock_usage_bit bit_backwards,
1353             enum lock_usage_bit bit_forwards, const char *irqclass)
1354 {
1355         int ret;
1356
1357         find_usage_bit = bit_backwards;
1358         /* fills in <backwards_match> */
1359         ret = find_usage_backwards(hlock_class(prev), 0);
1360         if (!ret || ret == 1)
1361                 return ret;
1362
1363         find_usage_bit = bit_forwards;
1364         ret = find_usage_forwards(hlock_class(next), 0);
1365         if (!ret || ret == 1)
1366                 return ret;
1367         /* ret == 2 */
1368         return print_bad_irq_dependency(curr, prev, next,
1369                         bit_backwards, bit_forwards, irqclass);
1370 }
1371
1372 static const char *state_names[] = {
1373 #define LOCKDEP_STATE(__STATE) \
1374         __stringify(__STATE),
1375 #include "lockdep_states.h"
1376 #undef LOCKDEP_STATE
1377 };
1378
1379 static const char *state_rnames[] = {
1380 #define LOCKDEP_STATE(__STATE) \
1381         __stringify(__STATE)"-READ",
1382 #include "lockdep_states.h"
1383 #undef LOCKDEP_STATE
1384 };
1385
1386 static inline const char *state_name(enum lock_usage_bit bit)
1387 {
1388         return (bit & 1) ? state_rnames[bit >> 2] : state_names[bit >> 2];
1389 }
1390
1391 static int exclusive_bit(int new_bit)
1392 {
1393         /*
1394          * USED_IN
1395          * USED_IN_READ
1396          * ENABLED
1397          * ENABLED_READ
1398          *
1399          * bit 0 - write/read
1400          * bit 1 - used_in/enabled
1401          * bit 2+  state
1402          */
1403
1404         int state = new_bit & ~3;
1405         int dir = new_bit & 2;
1406
1407         /*
1408          * keep state, bit flip the direction and strip read.
1409          */
1410         return state | (dir ^ 2);
1411 }
1412
1413 static int check_irq_usage(struct task_struct *curr, struct held_lock *prev,
1414                            struct held_lock *next, enum lock_usage_bit bit)
1415 {
1416         /*
1417          * Prove that the new dependency does not connect a hardirq-safe
1418          * lock with a hardirq-unsafe lock - to achieve this we search
1419          * the backwards-subgraph starting at <prev>, and the
1420          * forwards-subgraph starting at <next>:
1421          */
1422         if (!check_usage(curr, prev, next, bit,
1423                            exclusive_bit(bit), state_name(bit)))
1424                 return 0;
1425
1426         bit++; /* _READ */
1427
1428         /*
1429          * Prove that the new dependency does not connect a hardirq-safe-read
1430          * lock with a hardirq-unsafe lock - to achieve this we search
1431          * the backwards-subgraph starting at <prev>, and the
1432          * forwards-subgraph starting at <next>:
1433          */
1434         if (!check_usage(curr, prev, next, bit,
1435                            exclusive_bit(bit), state_name(bit)))
1436                 return 0;
1437
1438         return 1;
1439 }
1440
1441 static int
1442 check_prev_add_irq(struct task_struct *curr, struct held_lock *prev,
1443                 struct held_lock *next)
1444 {
1445 #define LOCKDEP_STATE(__STATE)                                          \
1446         if (!check_irq_usage(curr, prev, next, LOCK_USED_IN_##__STATE)) \
1447                 return 0;
1448 #include "lockdep_states.h"
1449 #undef LOCKDEP_STATE
1450
1451         return 1;
1452 }
1453
1454 static void inc_chains(void)
1455 {
1456         if (current->hardirq_context)
1457                 nr_hardirq_chains++;
1458         else {
1459                 if (current->softirq_context)
1460                         nr_softirq_chains++;
1461                 else
1462                         nr_process_chains++;
1463         }
1464 }
1465
1466 #else
1467
1468 static inline int
1469 check_prev_add_irq(struct task_struct *curr, struct held_lock *prev,
1470                 struct held_lock *next)
1471 {
1472         return 1;
1473 }
1474
1475 static inline void inc_chains(void)
1476 {
1477         nr_process_chains++;
1478 }
1479
1480 #endif
1481
1482 static int
1483 print_deadlock_bug(struct task_struct *curr, struct held_lock *prev,
1484                    struct held_lock *next)
1485 {
1486         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1487                 return 0;
1488
1489         printk("\n=============================================\n");
1490         printk(  "[ INFO: possible recursive locking detected ]\n");
1491         print_kernel_version();
1492         printk(  "---------------------------------------------\n");
1493         printk("%s/%d is trying to acquire lock:\n",
1494                 curr->comm, task_pid_nr(curr));
1495         print_lock(next);
1496         printk("\nbut task is already holding lock:\n");
1497         print_lock(prev);
1498
1499         printk("\nother info that might help us debug this:\n");
1500         lockdep_print_held_locks(curr);
1501
1502         printk("\nstack backtrace:\n");
1503         dump_stack();
1504
1505         return 0;
1506 }
1507
1508 /*
1509  * Check whether we are holding such a class already.
1510  *
1511  * (Note that this has to be done separately, because the graph cannot
1512  * detect such classes of deadlocks.)
1513  *
1514  * Returns: 0 on deadlock detected, 1 on OK, 2 on recursive read
1515  */
1516 static int
1517 check_deadlock(struct task_struct *curr, struct held_lock *next,
1518                struct lockdep_map *next_instance, int read)
1519 {
1520         struct held_lock *prev;
1521         struct held_lock *nest = NULL;
1522         int i;
1523
1524         for (i = 0; i < curr->lockdep_depth; i++) {
1525                 prev = curr->held_locks + i;
1526
1527                 if (prev->instance == next->nest_lock)
1528                         nest = prev;
1529
1530                 if (hlock_class(prev) != hlock_class(next))
1531                         continue;
1532
1533                 /*
1534                  * Allow read-after-read recursion of the same
1535                  * lock class (i.e. read_lock(lock)+read_lock(lock)):
1536                  */
1537                 if ((read == 2) && prev->read)
1538                         return 2;
1539
1540                 /*
1541                  * We're holding the nest_lock, which serializes this lock's
1542                  * nesting behaviour.
1543                  */
1544                 if (nest)
1545                         return 2;
1546
1547                 return print_deadlock_bug(curr, prev, next);
1548         }
1549         return 1;
1550 }
1551
1552 /*
1553  * There was a chain-cache miss, and we are about to add a new dependency
1554  * to a previous lock. We recursively validate the following rules:
1555  *
1556  *  - would the adding of the <prev> -> <next> dependency create a
1557  *    circular dependency in the graph? [== circular deadlock]
1558  *
1559  *  - does the new prev->next dependency connect any hardirq-safe lock
1560  *    (in the full backwards-subgraph starting at <prev>) with any
1561  *    hardirq-unsafe lock (in the full forwards-subgraph starting at
1562  *    <next>)? [== illegal lock inversion with hardirq contexts]
1563  *
1564  *  - does the new prev->next dependency connect any softirq-safe lock
1565  *    (in the full backwards-subgraph starting at <prev>) with any
1566  *    softirq-unsafe lock (in the full forwards-subgraph starting at
1567  *    <next>)? [== illegal lock inversion with softirq contexts]
1568  *
1569  * any of these scenarios could lead to a deadlock.
1570  *
1571  * Then if all the validations pass, we add the forwards and backwards
1572  * dependency.
1573  */
1574 static int
1575 check_prev_add(struct task_struct *curr, struct held_lock *prev,
1576                struct held_lock *next, int distance)
1577 {
1578         struct lock_list *entry;
1579         int ret;
1580
1581         /*
1582          * Prove that the new <prev> -> <next> dependency would not
1583          * create a circular dependency in the graph. (We do this by
1584          * forward-recursing into the graph starting at <next>, and
1585          * checking whether we can reach <prev>.)
1586          *
1587          * We are using global variables to control the recursion, to
1588          * keep the stackframe size of the recursive functions low:
1589          */
1590         check_source = next;
1591         check_target = prev;
1592
1593         if (check_noncircular(hlock_class(next), 0) == 2)
1594                 return print_circular_bug();
1595
1596         if (!check_prev_add_irq(curr, prev, next))
1597                 return 0;
1598
1599         /*
1600          * For recursive read-locks we do all the dependency checks,
1601          * but we dont store read-triggered dependencies (only
1602          * write-triggered dependencies). This ensures that only the
1603          * write-side dependencies matter, and that if for example a
1604          * write-lock never takes any other locks, then the reads are
1605          * equivalent to a NOP.
1606          */
1607         if (next->read == 2 || prev->read == 2)
1608                 return 1;
1609         /*
1610          * Is the <prev> -> <next> dependency already present?
1611          *
1612          * (this may occur even though this is a new chain: consider
1613          *  e.g. the L1 -> L2 -> L3 -> L4 and the L5 -> L1 -> L2 -> L3
1614          *  chains - the second one will be new, but L1 already has
1615          *  L2 added to its dependency list, due to the first chain.)
1616          */
1617         list_for_each_entry(entry, &hlock_class(prev)->locks_after, entry) {
1618                 if (entry->class == hlock_class(next)) {
1619                         if (distance == 1)
1620                                 entry->distance = 1;
1621                         return 2;
1622                 }
1623         }
1624
1625         /*
1626          * Ok, all validations passed, add the new lock
1627          * to the previous lock's dependency list:
1628          */
1629         ret = add_lock_to_list(hlock_class(prev), hlock_class(next),
1630                                &hlock_class(prev)->locks_after,
1631                                next->acquire_ip, distance);
1632
1633         if (!ret)
1634                 return 0;
1635
1636         ret = add_lock_to_list(hlock_class(next), hlock_class(prev),
1637                                &hlock_class(next)->locks_before,
1638                                next->acquire_ip, distance);
1639         if (!ret)
1640                 return 0;
1641
1642         /*
1643          * Debugging printouts:
1644          */
1645         if (verbose(hlock_class(prev)) || verbose(hlock_class(next))) {
1646                 graph_unlock();
1647                 printk("\n new dependency: ");
1648                 print_lock_name(hlock_class(prev));
1649                 printk(" => ");
1650                 print_lock_name(hlock_class(next));
1651                 printk("\n");
1652                 dump_stack();
1653                 return graph_lock();
1654         }
1655         return 1;
1656 }
1657
1658 /*
1659  * Add the dependency to all directly-previous locks that are 'relevant'.
1660  * The ones that are relevant are (in increasing distance from curr):
1661  * all consecutive trylock entries and the final non-trylock entry - or
1662  * the end of this context's lock-chain - whichever comes first.
1663  */
1664 static int
1665 check_prevs_add(struct task_struct *curr, struct held_lock *next)
1666 {
1667         int depth = curr->lockdep_depth;
1668         struct held_lock *hlock;
1669
1670         /*
1671          * Debugging checks.
1672          *
1673          * Depth must not be zero for a non-head lock:
1674          */
1675         if (!depth)
1676                 goto out_bug;
1677         /*
1678          * At least two relevant locks must exist for this
1679          * to be a head:
1680          */
1681         if (curr->held_locks[depth].irq_context !=
1682                         curr->held_locks[depth-1].irq_context)
1683                 goto out_bug;
1684
1685         for (;;) {
1686                 int distance = curr->lockdep_depth - depth + 1;
1687                 hlock = curr->held_locks + depth-1;
1688                 /*
1689                  * Only non-recursive-read entries get new dependencies
1690                  * added:
1691                  */
1692                 if (hlock->read != 2) {
1693                         if (!check_prev_add(curr, hlock, next, distance))
1694                                 return 0;
1695                         /*
1696                          * Stop after the first non-trylock entry,
1697                          * as non-trylock entries have added their
1698                          * own direct dependencies already, so this
1699                          * lock is connected to them indirectly:
1700                          */
1701                         if (!hlock->trylock)
1702                                 break;
1703                 }
1704                 depth--;
1705                 /*
1706                  * End of lock-stack?
1707                  */
1708                 if (!depth)
1709                         break;
1710                 /*
1711                  * Stop the search if we cross into another context:
1712                  */
1713                 if (curr->held_locks[depth].irq_context !=
1714                                 curr->held_locks[depth-1].irq_context)
1715                         break;
1716         }
1717         return 1;
1718 out_bug:
1719         if (!debug_locks_off_graph_unlock())
1720                 return 0;
1721
1722         WARN_ON(1);
1723
1724         return 0;
1725 }
1726
1727 unsigned long nr_lock_chains;
1728 struct lock_chain lock_chains[MAX_LOCKDEP_CHAINS];
1729 int nr_chain_hlocks;
1730 static u16 chain_hlocks[MAX_LOCKDEP_CHAIN_HLOCKS];
1731
1732 struct lock_class *lock_chain_get_class(struct lock_chain *chain, int i)
1733 {
1734         return lock_classes + chain_hlocks[chain->base + i];
1735 }
1736
1737 /*
1738  * Look up a dependency chain. If the key is not present yet then
1739  * add it and return 1 - in this case the new dependency chain is
1740  * validated. If the key is already hashed, return 0.
1741  * (On return with 1 graph_lock is held.)
1742  */
1743 static inline int lookup_chain_cache(struct task_struct *curr,
1744                                      struct held_lock *hlock,
1745                                      u64 chain_key)
1746 {
1747         struct lock_class *class = hlock_class(hlock);
1748         struct list_head *hash_head = chainhashentry(chain_key);
1749         struct lock_chain *chain;
1750         struct held_lock *hlock_curr, *hlock_next;
1751         int i, j, n, cn;
1752
1753         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
1754                 return 0;
1755         /*
1756          * We can walk it lock-free, because entries only get added
1757          * to the hash:
1758          */
1759         list_for_each_entry(chain, hash_head, entry) {
1760                 if (chain->chain_key == chain_key) {
1761 cache_hit:
1762                         debug_atomic_inc(&chain_lookup_hits);
1763                         if (very_verbose(class))
1764                                 printk("\nhash chain already cached, key: "
1765                                         "%016Lx tail class: [%p] %s\n",
1766                                         (unsigned long long)chain_key,
1767                                         class->key, class->name);
1768                         return 0;
1769                 }
1770         }
1771         if (very_verbose(class))
1772                 printk("\nnew hash chain, key: %016Lx tail class: [%p] %s\n",
1773                         (unsigned long long)chain_key, class->key, class->name);
1774         /*
1775          * Allocate a new chain entry from the static array, and add
1776          * it to the hash:
1777          */
1778         if (!graph_lock())
1779                 return 0;
1780         /*
1781          * We have to walk the chain again locked - to avoid duplicates:
1782          */
1783         list_for_each_entry(chain, hash_head, entry) {
1784                 if (chain->chain_key == chain_key) {
1785                         graph_unlock();
1786                         goto cache_hit;
1787                 }
1788         }
1789         if (unlikely(nr_lock_chains >= MAX_LOCKDEP_CHAINS)) {
1790                 if (!debug_locks_off_graph_unlock())
1791                         return 0;
1792
1793                 printk("BUG: MAX_LOCKDEP_CHAINS too low!\n");
1794                 printk("turning off the locking correctness validator.\n");
1795                 dump_stack();
1796                 return 0;
1797         }
1798         chain = lock_chains + nr_lock_chains++;
1799         chain->chain_key = chain_key;
1800         chain->irq_context = hlock->irq_context;
1801         /* Find the first held_lock of current chain */
1802         hlock_next = hlock;
1803         for (i = curr->lockdep_depth - 1; i >= 0; i--) {
1804                 hlock_curr = curr->held_locks + i;
1805                 if (hlock_curr->irq_context != hlock_next->irq_context)
1806                         break;
1807                 hlock_next = hlock;
1808         }
1809         i++;
1810         chain->depth = curr->lockdep_depth + 1 - i;
1811         cn = nr_chain_hlocks;
1812         while (cn + chain->depth <= MAX_LOCKDEP_CHAIN_HLOCKS) {
1813                 n = cmpxchg(&nr_chain_hlocks, cn, cn + chain->depth);
1814                 if (n == cn)
1815                         break;
1816                 cn = n;
1817         }
1818         if (likely(cn + chain->depth <= MAX_LOCKDEP_CHAIN_HLOCKS)) {
1819                 chain->base = cn;
1820                 for (j = 0; j < chain->depth - 1; j++, i++) {
1821                         int lock_id = curr->held_locks[i].class_idx - 1;
1822                         chain_hlocks[chain->base + j] = lock_id;
1823                 }
1824                 chain_hlocks[chain->base + j] = class - lock_classes;
1825         }
1826         list_add_tail_rcu(&chain->entry, hash_head);
1827         debug_atomic_inc(&chain_lookup_misses);
1828         inc_chains();
1829
1830         return 1;
1831 }
1832
1833 static int validate_chain(struct task_struct *curr, struct lockdep_map *lock,
1834                 struct held_lock *hlock, int chain_head, u64 chain_key)
1835 {
1836         /*
1837          * Trylock needs to maintain the stack of held locks, but it
1838          * does not add new dependencies, because trylock can be done
1839          * in any order.
1840          *
1841          * We look up the chain_key and do the O(N^2) check and update of
1842          * the dependencies only if this is a new dependency chain.
1843          * (If lookup_chain_cache() returns with 1 it acquires
1844          * graph_lock for us)
1845          */
1846         if (!hlock->trylock && (hlock->check == 2) &&
1847             lookup_chain_cache(curr, hlock, chain_key)) {
1848                 /*
1849                  * Check whether last held lock:
1850                  *
1851                  * - is irq-safe, if this lock is irq-unsafe
1852                  * - is softirq-safe, if this lock is hardirq-unsafe
1853                  *
1854                  * And check whether the new lock's dependency graph
1855                  * could lead back to the previous lock.
1856                  *
1857                  * any of these scenarios could lead to a deadlock. If
1858                  * All validations
1859                  */
1860                 int ret = check_deadlock(curr, hlock, lock, hlock->read);
1861
1862                 if (!ret)
1863                         return 0;
1864                 /*
1865                  * Mark recursive read, as we jump over it when
1866                  * building dependencies (just like we jump over
1867                  * trylock entries):
1868                  */
1869                 if (ret == 2)
1870                         hlock->read = 2;
1871                 /*
1872                  * Add dependency only if this lock is not the head
1873                  * of the chain, and if it's not a secondary read-lock:
1874                  */
1875                 if (!chain_head && ret != 2)
1876                         if (!check_prevs_add(curr, hlock))
1877                                 return 0;
1878                 graph_unlock();
1879         } else
1880                 /* after lookup_chain_cache(): */
1881                 if (unlikely(!debug_locks))
1882                         return 0;
1883
1884         return 1;
1885 }
1886 #else
1887 static inline int validate_chain(struct task_struct *curr,
1888                 struct lockdep_map *lock, struct held_lock *hlock,
1889                 int chain_head, u64 chain_key)
1890 {
1891         return 1;
1892 }
1893 #endif
1894
1895 /*
1896  * We are building curr_chain_key incrementally, so double-check
1897  * it from scratch, to make sure that it's done correctly:
1898  */
1899 static void check_chain_key(struct task_struct *curr)
1900 {
1901 #ifdef CONFIG_DEBUG_LOCKDEP
1902         struct held_lock *hlock, *prev_hlock = NULL;
1903         unsigned int i, id;
1904         u64 chain_key = 0;
1905
1906         for (i = 0; i < curr->lockdep_depth; i++) {
1907                 hlock = curr->held_locks + i;
1908                 if (chain_key != hlock->prev_chain_key) {
1909                         debug_locks_off();
1910                         WARN(1, "hm#1, depth: %u [%u], %016Lx != %016Lx\n",
1911                                 curr->lockdep_depth, i,
1912                                 (unsigned long long)chain_key,
1913                                 (unsigned long long)hlock->prev_chain_key);
1914                         return;
1915                 }
1916                 id = hlock->class_idx - 1;
1917                 if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
1918                         return;
1919
1920                 if (prev_hlock && (prev_hlock->irq_context !=
1921                                                         hlock->irq_context))
1922                         chain_key = 0;
1923                 chain_key = iterate_chain_key(chain_key, id);
1924                 prev_hlock = hlock;
1925         }
1926         if (chain_key != curr->curr_chain_key) {
1927                 debug_locks_off();
1928                 WARN(1, "hm#2, depth: %u [%u], %016Lx != %016Lx\n",
1929                         curr->lockdep_depth, i,
1930                         (unsigned long long)chain_key,
1931                         (unsigned long long)curr->curr_chain_key);
1932         }
1933 #endif
1934 }
1935
1936 static int
1937 print_usage_bug(struct task_struct *curr, struct held_lock *this,
1938                 enum lock_usage_bit prev_bit, enum lock_usage_bit new_bit)
1939 {
1940         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1941                 return 0;
1942
1943         printk("\n=================================\n");
1944         printk(  "[ INFO: inconsistent lock state ]\n");
1945         print_kernel_version();
1946         printk(  "---------------------------------\n");
1947
1948         printk("inconsistent {%s} -> {%s} usage.\n",
1949                 usage_str[prev_bit], usage_str[new_bit]);
1950
1951         printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] takes:\n",
1952                 curr->comm, task_pid_nr(curr),
1953                 trace_hardirq_context(curr), hardirq_count() >> HARDIRQ_SHIFT,
1954                 trace_softirq_context(curr), softirq_count() >> SOFTIRQ_SHIFT,
1955                 trace_hardirqs_enabled(curr),
1956                 trace_softirqs_enabled(curr));
1957         print_lock(this);
1958
1959         printk("{%s} state was registered at:\n", usage_str[prev_bit]);
1960         print_stack_trace(hlock_class(this)->usage_traces + prev_bit, 1);
1961
1962         print_irqtrace_events(curr);
1963         printk("\nother info that might help us debug this:\n");
1964         lockdep_print_held_locks(curr);
1965
1966         printk("\nstack backtrace:\n");
1967         dump_stack();
1968
1969         return 0;
1970 }
1971
1972 /*
1973  * Print out an error if an invalid bit is set:
1974  */
1975 static inline int
1976 valid_state(struct task_struct *curr, struct held_lock *this,
1977             enum lock_usage_bit new_bit, enum lock_usage_bit bad_bit)
1978 {
1979         if (unlikely(hlock_class(this)->usage_mask & (1 << bad_bit)))
1980                 return print_usage_bug(curr, this, bad_bit, new_bit);
1981         return 1;
1982 }
1983
1984 static int mark_lock(struct task_struct *curr, struct held_lock *this,
1985                      enum lock_usage_bit new_bit);
1986
1987 #if defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING)
1988
1989 /*
1990  * print irq inversion bug:
1991  */
1992 static int
1993 print_irq_inversion_bug(struct task_struct *curr, struct lock_class *other,
1994                         struct held_lock *this, int forwards,
1995                         const char *irqclass)
1996 {
1997         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1998                 return 0;
1999
2000         printk("\n=========================================================\n");
2001         printk(  "[ INFO: possible irq lock inversion dependency detected ]\n");
2002         print_kernel_version();
2003         printk(  "---------------------------------------------------------\n");
2004         printk("%s/%d just changed the state of lock:\n",
2005                 curr->comm, task_pid_nr(curr));
2006         print_lock(this);
2007         if (forwards)
2008                 printk("but this lock took another, %s-unsafe lock in the past:\n", irqclass);
2009         else
2010                 printk("but this lock was taken by another, %s-safe lock in the past:\n", irqclass);
2011         print_lock_name(other);
2012         printk("\n\nand interrupts could create inverse lock ordering between them.\n\n");
2013
2014         printk("\nother info that might help us debug this:\n");
2015         lockdep_print_held_locks(curr);
2016
2017         printk("\nthe first lock's dependencies:\n");
2018         print_lock_dependencies(hlock_class(this), 0);
2019
2020         printk("\nthe second lock's dependencies:\n");
2021         print_lock_dependencies(other, 0);
2022
2023         printk("\nstack backtrace:\n");
2024         dump_stack();
2025
2026         return 0;
2027 }
2028
2029 /*
2030  * Prove that in the forwards-direction subgraph starting at <this>
2031  * there is no lock matching <mask>:
2032  */
2033 static int
2034 check_usage_forwards(struct task_struct *curr, struct held_lock *this,
2035                      enum lock_usage_bit bit, const char *irqclass)
2036 {
2037         int ret;
2038
2039         find_usage_bit = bit;
2040         /* fills in <forwards_match> */
2041         ret = find_usage_forwards(hlock_class(this), 0);
2042         if (!ret || ret == 1)
2043                 return ret;
2044
2045         return print_irq_inversion_bug(curr, forwards_match, this, 1, irqclass);
2046 }
2047
2048 /*
2049  * Prove that in the backwards-direction subgraph starting at <this>
2050  * there is no lock matching <mask>:
2051  */
2052 static int
2053 check_usage_backwards(struct task_struct *curr, struct held_lock *this,
2054                       enum lock_usage_bit bit, const char *irqclass)
2055 {
2056         int ret;
2057
2058         find_usage_bit = bit;
2059         /* fills in <backwards_match> */
2060         ret = find_usage_backwards(hlock_class(this), 0);
2061         if (!ret || ret == 1)
2062                 return ret;
2063
2064         return print_irq_inversion_bug(curr, backwards_match, this, 0, irqclass);
2065 }
2066
2067 void print_irqtrace_events(struct task_struct *curr)
2068 {
2069         printk("irq event stamp: %u\n", curr->irq_events);
2070         printk("hardirqs last  enabled at (%u): ", curr->hardirq_enable_event);
2071         print_ip_sym(curr->hardirq_enable_ip);
2072         printk("hardirqs last disabled at (%u): ", curr->hardirq_disable_event);
2073         print_ip_sym(curr->hardirq_disable_ip);
2074         printk("softirqs last  enabled at (%u): ", curr->softirq_enable_event);
2075         print_ip_sym(curr->softirq_enable_ip);
2076         printk("softirqs last disabled at (%u): ", curr->softirq_disable_event);
2077         print_ip_sym(curr->softirq_disable_ip);
2078 }
2079
2080 static int HARDIRQ_verbose(struct lock_class *class)
2081 {
2082 #if HARDIRQ_VERBOSE
2083         return class_filter(class);
2084 #endif
2085         return 0;
2086 }
2087
2088 static int SOFTIRQ_verbose(struct lock_class *class)
2089 {
2090 #if SOFTIRQ_VERBOSE
2091         return class_filter(class);
2092 #endif
2093         return 0;
2094 }
2095
2096 static int RECLAIM_FS_verbose(struct lock_class *class)
2097 {
2098 #if RECLAIM_VERBOSE
2099         return class_filter(class);
2100 #endif
2101         return 0;
2102 }
2103
2104 #define STRICT_READ_CHECKS      1
2105
2106 static int (*state_verbose_f[])(struct lock_class *class) = {
2107 #define LOCKDEP_STATE(__STATE) \
2108         __STATE##_verbose,
2109 #include "lockdep_states.h"
2110 #undef LOCKDEP_STATE
2111 };
2112
2113 static inline int state_verbose(enum lock_usage_bit bit,
2114                                 struct lock_class *class)
2115 {
2116         return state_verbose_f[bit >> 2](class);
2117 }
2118
2119 typedef int (*check_usage_f)(struct task_struct *, struct held_lock *,
2120                              enum lock_usage_bit bit, const char *name);
2121
2122 static int
2123 mark_lock_irq(struct task_struct *curr, struct held_lock *this,
2124                 enum lock_usage_bit new_bit)
2125 {
2126         int excl_bit = exclusive_bit(new_bit);
2127         int read = new_bit & 1;
2128         int dir = new_bit & 2;
2129
2130         /*
2131          * mark USED_IN has to look forwards -- to ensure no dependency
2132          * has ENABLED state, which would allow recursion deadlocks.
2133          *
2134          * mark ENABLED has to look backwards -- to ensure no dependee
2135          * has USED_IN state, which, again, would allow  recursion deadlocks.
2136          */
2137         check_usage_f usage = dir ?
2138                 check_usage_backwards : check_usage_forwards;
2139
2140         /*
2141          * Validate that this particular lock does not have conflicting
2142          * usage states.
2143          */
2144         if (!valid_state(curr, this, new_bit, excl_bit))
2145                 return 0;
2146
2147         /*
2148          * Validate that the lock dependencies don't have conflicting usage
2149          * states.
2150          */
2151         if ((!read || !dir || STRICT_READ_CHECKS) &&
2152                         !usage(curr, this, excl_bit, state_name(new_bit & ~1)))
2153                 return 0;
2154
2155         /*
2156          * Check for read in write conflicts
2157          */
2158         if (!read) {
2159                 if (!valid_state(curr, this, new_bit, excl_bit + 1))
2160                         return 0;
2161
2162                 if (STRICT_READ_CHECKS &&
2163                         !usage(curr, this, excl_bit + 1,
2164                                 state_name(new_bit + 1)))
2165                         return 0;
2166         }
2167
2168         if (state_verbose(new_bit, hlock_class(this)))
2169                 return 2;
2170
2171         return 1;
2172 }
2173
2174 enum mark_type {
2175 #define LOCKDEP_STATE(__STATE)  __STATE,
2176 #include "lockdep_states.h"
2177 #undef LOCKDEP_STATE
2178 };
2179
2180 /*
2181  * Mark all held locks with a usage bit:
2182  */
2183 static int
2184 mark_held_locks(struct task_struct *curr, enum mark_type mark)
2185 {
2186         enum lock_usage_bit usage_bit;
2187         struct held_lock *hlock;
2188         int i;
2189
2190         for (i = 0; i < curr->lockdep_depth; i++) {
2191                 hlock = curr->held_locks + i;
2192
2193                 usage_bit = 2 + (mark << 2); /* ENABLED */
2194                 if (hlock->read)
2195                         usage_bit += 1; /* READ */
2196
2197                 BUG_ON(usage_bit >= LOCK_USAGE_STATES);
2198
2199                 if (!mark_lock(curr, hlock, usage_bit))
2200                         return 0;
2201         }
2202
2203         return 1;
2204 }
2205
2206 /*
2207  * Debugging helper: via this flag we know that we are in
2208  * 'early bootup code', and will warn about any invalid irqs-on event:
2209  */
2210 static int early_boot_irqs_enabled;
2211
2212 void early_boot_irqs_off(void)
2213 {
2214         early_boot_irqs_enabled = 0;
2215 }
2216
2217 void early_boot_irqs_on(void)
2218 {
2219         early_boot_irqs_enabled = 1;
2220 }
2221
2222 /*
2223  * Hardirqs will be enabled:
2224  */
2225 void trace_hardirqs_on_caller(unsigned long ip)
2226 {
2227         struct task_struct *curr = current;
2228
2229         time_hardirqs_on(CALLER_ADDR0, ip);
2230
2231         if (unlikely(!debug_locks || current->lockdep_recursion))
2232                 return;
2233
2234         if (DEBUG_LOCKS_WARN_ON(unlikely(!early_boot_irqs_enabled)))
2235                 return;
2236
2237         if (unlikely(curr->hardirqs_enabled)) {
2238                 debug_atomic_inc(&redundant_hardirqs_on);
2239                 return;
2240         }
2241         /* we'll do an OFF -> ON transition: */
2242         curr->hardirqs_enabled = 1;
2243
2244         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2245                 return;
2246         if (DEBUG_LOCKS_WARN_ON(current->hardirq_context))
2247                 return;
2248         /*
2249          * We are going to turn hardirqs on, so set the
2250          * usage bit for all held locks:
2251          */
2252         if (!mark_held_locks(curr, HARDIRQ))
2253                 return;
2254         /*
2255          * If we have softirqs enabled, then set the usage
2256          * bit for all held locks. (disabled hardirqs prevented
2257          * this bit from being set before)
2258          */
2259         if (curr->softirqs_enabled)
2260                 if (!mark_held_locks(curr, SOFTIRQ))
2261                         return;
2262
2263         curr->hardirq_enable_ip = ip;
2264         curr->hardirq_enable_event = ++curr->irq_events;
2265         debug_atomic_inc(&hardirqs_on_events);
2266 }
2267 EXPORT_SYMBOL(trace_hardirqs_on_caller);
2268
2269 void trace_hardirqs_on(void)
2270 {
2271         trace_hardirqs_on_caller(CALLER_ADDR0);
2272 }
2273 EXPORT_SYMBOL(trace_hardirqs_on);
2274
2275 /*
2276  * Hardirqs were disabled:
2277  */
2278 void trace_hardirqs_off_caller(unsigned long ip)
2279 {
2280         struct task_struct *curr = current;
2281
2282         time_hardirqs_off(CALLER_ADDR0, ip);
2283
2284         if (unlikely(!debug_locks || current->lockdep_recursion))
2285                 return;
2286
2287         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2288                 return;
2289
2290         if (curr->hardirqs_enabled) {
2291                 /*
2292                  * We have done an ON -> OFF transition:
2293                  */
2294                 curr->hardirqs_enabled = 0;
2295                 curr->hardirq_disable_ip = ip;
2296                 curr->hardirq_disable_event = ++curr->irq_events;
2297                 debug_atomic_inc(&hardirqs_off_events);
2298         } else
2299                 debug_atomic_inc(&redundant_hardirqs_off);
2300 }
2301 EXPORT_SYMBOL(trace_hardirqs_off_caller);
2302
2303 void trace_hardirqs_off(void)
2304 {
2305         trace_hardirqs_off_caller(CALLER_ADDR0);
2306 }
2307 EXPORT_SYMBOL(trace_hardirqs_off);
2308
2309 /*
2310  * Softirqs will be enabled:
2311  */
2312 void trace_softirqs_on(unsigned long ip)
2313 {
2314         struct task_struct *curr = current;
2315
2316         if (unlikely(!debug_locks))
2317                 return;
2318
2319         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2320                 return;
2321
2322         if (curr->softirqs_enabled) {
2323                 debug_atomic_inc(&redundant_softirqs_on);
2324                 return;
2325         }
2326
2327         /*
2328          * We'll do an OFF -> ON transition:
2329          */
2330         curr->softirqs_enabled = 1;
2331         curr->softirq_enable_ip = ip;
2332         curr->softirq_enable_event = ++curr->irq_events;
2333         debug_atomic_inc(&softirqs_on_events);
2334         /*
2335          * We are going to turn softirqs on, so set the
2336          * usage bit for all held locks, if hardirqs are
2337          * enabled too:
2338          */
2339         if (curr->hardirqs_enabled)
2340                 mark_held_locks(curr, SOFTIRQ);
2341 }
2342
2343 /*
2344  * Softirqs were disabled:
2345  */
2346 void trace_softirqs_off(unsigned long ip)
2347 {
2348         struct task_struct *curr = current;
2349
2350         if (unlikely(!debug_locks))
2351                 return;
2352
2353         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2354                 return;
2355
2356         if (curr->softirqs_enabled) {
2357                 /*
2358                  * We have done an ON -> OFF transition:
2359                  */
2360                 curr->softirqs_enabled = 0;
2361                 curr->softirq_disable_ip = ip;
2362                 curr->softirq_disable_event = ++curr->irq_events;
2363                 debug_atomic_inc(&softirqs_off_events);
2364                 DEBUG_LOCKS_WARN_ON(!softirq_count());
2365         } else
2366                 debug_atomic_inc(&redundant_softirqs_off);
2367 }
2368
2369 static void __lockdep_trace_alloc(gfp_t gfp_mask, unsigned long flags)
2370 {
2371         struct task_struct *curr = current;
2372
2373         if (unlikely(!debug_locks))
2374                 return;
2375
2376         /* no reclaim without waiting on it */
2377         if (!(gfp_mask & __GFP_WAIT))
2378                 return;
2379
2380         /* this guy won't enter reclaim */
2381         if ((curr->flags & PF_MEMALLOC) && !(gfp_mask & __GFP_NOMEMALLOC))
2382                 return;
2383
2384         /* We're only interested __GFP_FS allocations for now */
2385         if (!(gfp_mask & __GFP_FS))
2386                 return;
2387
2388         if (DEBUG_LOCKS_WARN_ON(irqs_disabled_flags(flags)))
2389                 return;
2390
2391         mark_held_locks(curr, RECLAIM_FS);
2392 }
2393
2394 static void check_flags(unsigned long flags);
2395
2396 void lockdep_trace_alloc(gfp_t gfp_mask)
2397 {
2398         unsigned long flags;
2399
2400         if (unlikely(current->lockdep_recursion))
2401                 return;
2402
2403         raw_local_irq_save(flags);
2404         check_flags(flags);
2405         current->lockdep_recursion = 1;
2406         __lockdep_trace_alloc(gfp_mask, flags);
2407         current->lockdep_recursion = 0;
2408         raw_local_irq_restore(flags);
2409 }
2410
2411 static int mark_irqflags(struct task_struct *curr, struct held_lock *hlock)
2412 {
2413         /*
2414          * If non-trylock use in a hardirq or softirq context, then
2415          * mark the lock as used in these contexts:
2416          */
2417         if (!hlock->trylock) {
2418                 if (hlock->read) {
2419                         if (curr->hardirq_context)
2420                                 if (!mark_lock(curr, hlock,
2421                                                 LOCK_USED_IN_HARDIRQ_READ))
2422                                         return 0;
2423                         if (curr->softirq_context)
2424                                 if (!mark_lock(curr, hlock,
2425                                                 LOCK_USED_IN_SOFTIRQ_READ))
2426                                         return 0;
2427                 } else {
2428                         if (curr->hardirq_context)
2429                                 if (!mark_lock(curr, hlock, LOCK_USED_IN_HARDIRQ))
2430                                         return 0;
2431                         if (curr->softirq_context)
2432                                 if (!mark_lock(curr, hlock, LOCK_USED_IN_SOFTIRQ))
2433                                         return 0;
2434                 }
2435         }
2436         if (!hlock->hardirqs_off) {
2437                 if (hlock->read) {
2438                         if (!mark_lock(curr, hlock,
2439                                         LOCK_ENABLED_HARDIRQ_READ))
2440                                 return 0;
2441                         if (curr->softirqs_enabled)
2442                                 if (!mark_lock(curr, hlock,
2443                                                 LOCK_ENABLED_SOFTIRQ_READ))
2444                                         return 0;
2445                 } else {
2446                         if (!mark_lock(curr, hlock,
2447                                         LOCK_ENABLED_HARDIRQ))
2448                                 return 0;
2449                         if (curr->softirqs_enabled)
2450                                 if (!mark_lock(curr, hlock,
2451                                                 LOCK_ENABLED_SOFTIRQ))
2452                                         return 0;
2453                 }
2454         }
2455
2456         /*
2457          * We reuse the irq context infrastructure more broadly as a general
2458          * context checking code. This tests GFP_FS recursion (a lock taken
2459          * during reclaim for a GFP_FS allocation is held over a GFP_FS
2460          * allocation).
2461          */
2462         if (!hlock->trylock && (curr->lockdep_reclaim_gfp & __GFP_FS)) {
2463                 if (hlock->read) {
2464                         if (!mark_lock(curr, hlock, LOCK_USED_IN_RECLAIM_FS_READ))
2465                                         return 0;
2466                 } else {
2467                         if (!mark_lock(curr, hlock, LOCK_USED_IN_RECLAIM_FS))
2468                                         return 0;
2469                 }
2470         }
2471
2472         return 1;
2473 }
2474
2475 static int separate_irq_context(struct task_struct *curr,
2476                 struct held_lock *hlock)
2477 {
2478         unsigned int depth = curr->lockdep_depth;
2479
2480         /*
2481          * Keep track of points where we cross into an interrupt context:
2482          */
2483         hlock->irq_context = 2*(curr->hardirq_context ? 1 : 0) +
2484                                 curr->softirq_context;
2485         if (depth) {
2486                 struct held_lock *prev_hlock;
2487
2488                 prev_hlock = curr->held_locks + depth-1;
2489                 /*
2490                  * If we cross into another context, reset the
2491                  * hash key (this also prevents the checking and the
2492                  * adding of the dependency to 'prev'):
2493                  */
2494                 if (prev_hlock->irq_context != hlock->irq_context)
2495                         return 1;
2496         }
2497         return 0;
2498 }
2499
2500 #else
2501
2502 static inline
2503 int mark_lock_irq(struct task_struct *curr, struct held_lock *this,
2504                 enum lock_usage_bit new_bit)
2505 {
2506         WARN_ON(1);
2507         return 1;
2508 }
2509
2510 static inline int mark_irqflags(struct task_struct *curr,
2511                 struct held_lock *hlock)
2512 {
2513         return 1;
2514 }
2515
2516 static inline int separate_irq_context(struct task_struct *curr,
2517                 struct held_lock *hlock)
2518 {
2519         return 0;
2520 }
2521
2522 void lockdep_trace_alloc(gfp_t gfp_mask)
2523 {
2524 }
2525
2526 #endif
2527
2528 /*
2529  * Mark a lock with a usage bit, and validate the state transition:
2530  */
2531 static int mark_lock(struct task_struct *curr, struct held_lock *this,
2532                              enum lock_usage_bit new_bit)
2533 {
2534         unsigned int new_mask = 1 << new_bit, ret = 1;
2535
2536         /*
2537          * If already set then do not dirty the cacheline,
2538          * nor do any checks:
2539          */
2540         if (likely(hlock_class(this)->usage_mask & new_mask))
2541                 return 1;
2542
2543         if (!graph_lock())
2544                 return 0;
2545         /*
2546          * Make sure we didnt race:
2547          */
2548         if (unlikely(hlock_class(this)->usage_mask & new_mask)) {
2549                 graph_unlock();
2550                 return 1;
2551         }
2552
2553         hlock_class(this)->usage_mask |= new_mask;
2554
2555         if (!save_trace(hlock_class(this)->usage_traces + new_bit))
2556                 return 0;
2557
2558         switch (new_bit) {
2559 #define LOCKDEP_STATE(__STATE)                  \
2560         case LOCK_USED_IN_##__STATE:            \
2561         case LOCK_USED_IN_##__STATE##_READ:     \
2562         case LOCK_ENABLED_##__STATE:            \
2563         case LOCK_ENABLED_##__STATE##_READ:
2564 #include "lockdep_states.h"
2565 #undef LOCKDEP_STATE
2566                 ret = mark_lock_irq(curr, this, new_bit);
2567                 if (!ret)
2568                         return 0;
2569                 break;
2570         case LOCK_USED:
2571                 debug_atomic_dec(&nr_unused_locks);
2572                 break;
2573         default:
2574                 if (!debug_locks_off_graph_unlock())
2575                         return 0;
2576                 WARN_ON(1);
2577                 return 0;
2578         }
2579
2580         graph_unlock();
2581
2582         /*
2583          * We must printk outside of the graph_lock:
2584          */
2585         if (ret == 2) {
2586                 printk("\nmarked lock as {%s}:\n", usage_str[new_bit]);
2587                 print_lock(this);
2588                 print_irqtrace_events(curr);
2589                 dump_stack();
2590         }
2591
2592         return ret;
2593 }
2594
2595 /*
2596  * Initialize a lock instance's lock-class mapping info:
2597  */
2598 void lockdep_init_map(struct lockdep_map *lock, const char *name,
2599                       struct lock_class_key *key, int subclass)
2600 {
2601         lock->class_cache = NULL;
2602 #ifdef CONFIG_LOCK_STAT
2603         lock->cpu = raw_smp_processor_id();
2604 #endif
2605
2606         if (DEBUG_LOCKS_WARN_ON(!name)) {
2607                 lock->name = "NULL";
2608                 return;
2609         }
2610
2611         lock->name = name;
2612
2613         if (DEBUG_LOCKS_WARN_ON(!key))
2614                 return;
2615         /*
2616          * Sanity check, the lock-class key must be persistent:
2617          */
2618         if (!static_obj(key)) {
2619                 printk("BUG: key %p not in .data!\n", key);
2620                 DEBUG_LOCKS_WARN_ON(1);
2621                 return;
2622         }
2623         lock->key = key;
2624
2625         if (unlikely(!debug_locks))
2626                 return;
2627
2628         if (subclass)
2629                 register_lock_class(lock, subclass, 1);
2630 }
2631 EXPORT_SYMBOL_GPL(lockdep_init_map);
2632
2633 /*
2634  * This gets called for every mutex_lock*()/spin_lock*() operation.
2635  * We maintain the dependency maps and validate the locking attempt:
2636  */
2637 static int __lock_acquire(struct lockdep_map *lock, unsigned int subclass,
2638                           int trylock, int read, int check, int hardirqs_off,
2639                           struct lockdep_map *nest_lock, unsigned long ip)
2640 {
2641         struct task_struct *curr = current;
2642         struct lock_class *class = NULL;
2643         struct held_lock *hlock;
2644         unsigned int depth, id;
2645         int chain_head = 0;
2646         u64 chain_key;
2647
2648         if (!prove_locking)
2649                 check = 1;
2650
2651         if (unlikely(!debug_locks))
2652                 return 0;
2653
2654         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2655                 return 0;
2656
2657         if (unlikely(subclass >= MAX_LOCKDEP_SUBCLASSES)) {
2658                 debug_locks_off();
2659                 printk("BUG: MAX_LOCKDEP_SUBCLASSES too low!\n");
2660                 printk("turning off the locking correctness validator.\n");
2661                 dump_stack();
2662                 return 0;
2663         }
2664
2665         if (!subclass)
2666                 class = lock->class_cache;
2667         /*
2668          * Not cached yet or subclass?
2669          */
2670         if (unlikely(!class)) {
2671                 class = register_lock_class(lock, subclass, 0);
2672                 if (!class)
2673                         return 0;
2674         }
2675         debug_atomic_inc((atomic_t *)&class->ops);
2676         if (very_verbose(class)) {
2677                 printk("\nacquire class [%p] %s", class->key, class->name);
2678                 if (class->name_version > 1)
2679                         printk("#%d", class->name_version);
2680                 printk("\n");
2681                 dump_stack();
2682         }
2683
2684         /*
2685          * Add the lock to the list of currently held locks.
2686          * (we dont increase the depth just yet, up until the
2687          * dependency checks are done)
2688          */
2689         depth = curr->lockdep_depth;
2690         if (DEBUG_LOCKS_WARN_ON(depth >= MAX_LOCK_DEPTH))
2691                 return 0;
2692
2693         hlock = curr->held_locks + depth;
2694         if (DEBUG_LOCKS_WARN_ON(!class))
2695                 return 0;
2696         hlock->class_idx = class - lock_classes + 1;
2697         hlock->acquire_ip = ip;
2698         hlock->instance = lock;
2699         hlock->nest_lock = nest_lock;
2700         hlock->trylock = trylock;
2701         hlock->read = read;
2702         hlock->check = check;
2703         hlock->hardirqs_off = !!hardirqs_off;
2704 #ifdef CONFIG_LOCK_STAT
2705         hlock->waittime_stamp = 0;
2706         hlock->holdtime_stamp = sched_clock();
2707 #endif
2708
2709         if (check == 2 && !mark_irqflags(curr, hlock))
2710                 return 0;
2711
2712         /* mark it as used: */
2713         if (!mark_lock(curr, hlock, LOCK_USED))
2714                 return 0;
2715
2716         /*
2717          * Calculate the chain hash: it's the combined hash of all the
2718          * lock keys along the dependency chain. We save the hash value
2719          * at every step so that we can get the current hash easily
2720          * after unlock. The chain hash is then used to cache dependency
2721          * results.
2722          *
2723          * The 'key ID' is what is the most compact key value to drive
2724          * the hash, not class->key.
2725          */
2726         id = class - lock_classes;
2727         if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
2728                 return 0;
2729
2730         chain_key = curr->curr_chain_key;
2731         if (!depth) {
2732                 if (DEBUG_LOCKS_WARN_ON(chain_key != 0))
2733                         return 0;
2734                 chain_head = 1;
2735         }
2736
2737         hlock->prev_chain_key = chain_key;
2738         if (separate_irq_context(curr, hlock)) {
2739                 chain_key = 0;
2740                 chain_head = 1;
2741         }
2742         chain_key = iterate_chain_key(chain_key, id);
2743
2744         if (!validate_chain(curr, lock, hlock, chain_head, chain_key))
2745                 return 0;
2746
2747         curr->curr_chain_key = chain_key;
2748         curr->lockdep_depth++;
2749         check_chain_key(curr);
2750 #ifdef CONFIG_DEBUG_LOCKDEP
2751         if (unlikely(!debug_locks))
2752                 return 0;
2753 #endif
2754         if (unlikely(curr->lockdep_depth >= MAX_LOCK_DEPTH)) {
2755                 debug_locks_off();
2756                 printk("BUG: MAX_LOCK_DEPTH too low!\n");
2757                 printk("turning off the locking correctness validator.\n");
2758                 dump_stack();
2759                 return 0;
2760         }
2761
2762         if (unlikely(curr->lockdep_depth > max_lockdep_depth))
2763                 max_lockdep_depth = curr->lockdep_depth;
2764
2765         return 1;
2766 }
2767
2768 static int
2769 print_unlock_inbalance_bug(struct task_struct *curr, struct lockdep_map *lock,
2770                            unsigned long ip)
2771 {
2772         if (!debug_locks_off())
2773                 return 0;
2774         if (debug_locks_silent)
2775                 return 0;
2776
2777         printk("\n=====================================\n");
2778         printk(  "[ BUG: bad unlock balance detected! ]\n");
2779         printk(  "-------------------------------------\n");
2780         printk("%s/%d is trying to release lock (",
2781                 curr->comm, task_pid_nr(curr));
2782         print_lockdep_cache(lock);
2783         printk(") at:\n");
2784         print_ip_sym(ip);
2785         printk("but there are no more locks to release!\n");
2786         printk("\nother info that might help us debug this:\n");
2787         lockdep_print_held_locks(curr);
2788
2789         printk("\nstack backtrace:\n");
2790         dump_stack();
2791
2792         return 0;
2793 }
2794
2795 /*
2796  * Common debugging checks for both nested and non-nested unlock:
2797  */
2798 static int check_unlock(struct task_struct *curr, struct lockdep_map *lock,
2799                         unsigned long ip)
2800 {
2801         if (unlikely(!debug_locks))
2802                 return 0;
2803         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2804                 return 0;
2805
2806         if (curr->lockdep_depth <= 0)
2807                 return print_unlock_inbalance_bug(curr, lock, ip);
2808
2809         return 1;
2810 }
2811
2812 static int
2813 __lock_set_class(struct lockdep_map *lock, const char *name,
2814                  struct lock_class_key *key, unsigned int subclass,
2815                  unsigned long ip)
2816 {
2817         struct task_struct *curr = current;
2818         struct held_lock *hlock, *prev_hlock;
2819         struct lock_class *class;
2820         unsigned int depth;
2821         int i;
2822
2823         depth = curr->lockdep_depth;
2824         if (DEBUG_LOCKS_WARN_ON(!depth))
2825                 return 0;
2826
2827         prev_hlock = NULL;
2828         for (i = depth-1; i >= 0; i--) {
2829                 hlock = curr->held_locks + i;
2830                 /*
2831                  * We must not cross into another context:
2832                  */
2833                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
2834                         break;
2835                 if (hlock->instance == lock)
2836                         goto found_it;
2837                 prev_hlock = hlock;
2838         }
2839         return print_unlock_inbalance_bug(curr, lock, ip);
2840
2841 found_it:
2842         lockdep_init_map(lock, name, key, 0);
2843         class = register_lock_class(lock, subclass, 0);
2844         hlock->class_idx = class - lock_classes + 1;
2845
2846         curr->lockdep_depth = i;
2847         curr->curr_chain_key = hlock->prev_chain_key;
2848
2849         for (; i < depth; i++) {
2850                 hlock = curr->held_locks + i;
2851                 if (!__lock_acquire(hlock->instance,
2852                         hlock_class(hlock)->subclass, hlock->trylock,
2853                                 hlock->read, hlock->check, hlock->hardirqs_off,
2854                                 hlock->nest_lock, hlock->acquire_ip))
2855                         return 0;
2856         }
2857
2858         if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth))
2859                 return 0;
2860         return 1;
2861 }
2862
2863 /*
2864  * Remove the lock to the list of currently held locks in a
2865  * potentially non-nested (out of order) manner. This is a
2866  * relatively rare operation, as all the unlock APIs default
2867  * to nested mode (which uses lock_release()):
2868  */
2869 static int
2870 lock_release_non_nested(struct task_struct *curr,
2871                         struct lockdep_map *lock, unsigned long ip)
2872 {
2873         struct held_lock *hlock, *prev_hlock;
2874         unsigned int depth;
2875         int i;
2876
2877         /*
2878          * Check whether the lock exists in the current stack
2879          * of held locks:
2880          */
2881         depth = curr->lockdep_depth;
2882         if (DEBUG_LOCKS_WARN_ON(!depth))
2883                 return 0;
2884
2885         prev_hlock = NULL;
2886         for (i = depth-1; i >= 0; i--) {
2887                 hlock = curr->held_locks + i;
2888                 /*
2889                  * We must not cross into another context:
2890                  */
2891                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
2892                         break;
2893                 if (hlock->instance == lock)
2894                         goto found_it;
2895                 prev_hlock = hlock;
2896         }
2897         return print_unlock_inbalance_bug(curr, lock, ip);
2898
2899 found_it:
2900         lock_release_holdtime(hlock);
2901
2902         /*
2903          * We have the right lock to unlock, 'hlock' points to it.
2904          * Now we remove it from the stack, and add back the other
2905          * entries (if any), recalculating the hash along the way:
2906          */
2907         curr->lockdep_depth = i;
2908         curr->curr_chain_key = hlock->prev_chain_key;
2909
2910         for (i++; i < depth; i++) {
2911                 hlock = curr->held_locks + i;
2912                 if (!__lock_acquire(hlock->instance,
2913                         hlock_class(hlock)->subclass, hlock->trylock,
2914                                 hlock->read, hlock->check, hlock->hardirqs_off,
2915                                 hlock->nest_lock, hlock->acquire_ip))
2916                         return 0;
2917         }
2918
2919         if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth - 1))
2920                 return 0;
2921         return 1;
2922 }
2923
2924 /*
2925  * Remove the lock to the list of currently held locks - this gets
2926  * called on mutex_unlock()/spin_unlock*() (or on a failed
2927  * mutex_lock_interruptible()). This is done for unlocks that nest
2928  * perfectly. (i.e. the current top of the lock-stack is unlocked)
2929  */
2930 static int lock_release_nested(struct task_struct *curr,
2931                                struct lockdep_map *lock, unsigned long ip)
2932 {
2933         struct held_lock *hlock;
2934         unsigned int depth;
2935
2936         /*
2937          * Pop off the top of the lock stack:
2938          */
2939         depth = curr->lockdep_depth - 1;
2940         hlock = curr->held_locks + depth;
2941
2942         /*
2943          * Is the unlock non-nested:
2944          */
2945         if (hlock->instance != lock)
2946                 return lock_release_non_nested(curr, lock, ip);
2947         curr->lockdep_depth--;
2948
2949         if (DEBUG_LOCKS_WARN_ON(!depth && (hlock->prev_chain_key != 0)))
2950                 return 0;
2951
2952         curr->curr_chain_key = hlock->prev_chain_key;
2953
2954         lock_release_holdtime(hlock);
2955
2956 #ifdef CONFIG_DEBUG_LOCKDEP
2957         hlock->prev_chain_key = 0;
2958         hlock->class_idx = 0;
2959         hlock->acquire_ip = 0;
2960         hlock->irq_context = 0;
2961 #endif
2962         return 1;
2963 }
2964
2965 /*
2966  * Remove the lock to the list of currently held locks - this gets
2967  * called on mutex_unlock()/spin_unlock*() (or on a failed
2968  * mutex_lock_interruptible()). This is done for unlocks that nest
2969  * perfectly. (i.e. the current top of the lock-stack is unlocked)
2970  */
2971 static void
2972 __lock_release(struct lockdep_map *lock, int nested, unsigned long ip)
2973 {
2974         struct task_struct *curr = current;
2975
2976         if (!check_unlock(curr, lock, ip))
2977                 return;
2978
2979         if (nested) {
2980                 if (!lock_release_nested(curr, lock, ip))
2981                         return;
2982         } else {
2983                 if (!lock_release_non_nested(curr, lock, ip))
2984                         return;
2985         }
2986
2987         check_chain_key(curr);
2988 }
2989
2990 /*
2991  * Check whether we follow the irq-flags state precisely:
2992  */
2993 static void check_flags(unsigned long flags)
2994 {
2995 #if defined(CONFIG_PROVE_LOCKING) && defined(CONFIG_DEBUG_LOCKDEP) && \
2996     defined(CONFIG_TRACE_IRQFLAGS)
2997         if (!debug_locks)
2998                 return;
2999
3000         if (irqs_disabled_flags(flags)) {
3001                 if (DEBUG_LOCKS_WARN_ON(current->hardirqs_enabled)) {
3002                         printk("possible reason: unannotated irqs-off.\n");
3003                 }
3004         } else {
3005                 if (DEBUG_LOCKS_WARN_ON(!current->hardirqs_enabled)) {
3006                         printk("possible reason: unannotated irqs-on.\n");
3007                 }
3008         }
3009
3010         /*
3011          * We dont accurately track softirq state in e.g.
3012          * hardirq contexts (such as on 4KSTACKS), so only
3013          * check if not in hardirq contexts:
3014          */
3015         if (!hardirq_count()) {
3016                 if (softirq_count())
3017                         DEBUG_LOCKS_WARN_ON(current->softirqs_enabled);
3018                 else
3019                         DEBUG_LOCKS_WARN_ON(!current->softirqs_enabled);
3020         }
3021
3022         if (!debug_locks)
3023                 print_irqtrace_events(current);
3024 #endif
3025 }
3026
3027 void lock_set_class(struct lockdep_map *lock, const char *name,
3028                     struct lock_class_key *key, unsigned int subclass,
3029                     unsigned long ip)
3030 {
3031         unsigned long flags;
3032
3033         if (unlikely(current->lockdep_recursion))
3034                 return;
3035
3036         raw_local_irq_save(flags);
3037         current->lockdep_recursion = 1;
3038         check_flags(flags);
3039         if (__lock_set_class(lock, name, key, subclass, ip))
3040                 check_chain_key(current);
3041         current->lockdep_recursion = 0;
3042         raw_local_irq_restore(flags);
3043 }
3044 EXPORT_SYMBOL_GPL(lock_set_class);
3045
3046 /*
3047  * We are not always called with irqs disabled - do that here,
3048  * and also avoid lockdep recursion:
3049  */
3050 void lock_acquire(struct lockdep_map *lock, unsigned int subclass,
3051                           int trylock, int read, int check,
3052                           struct lockdep_map *nest_lock, unsigned long ip)
3053 {
3054         unsigned long flags;
3055
3056         trace_lock_acquire(lock, subclass, trylock, read, check, nest_lock, ip);
3057
3058         if (unlikely(current->lockdep_recursion))
3059                 return;
3060
3061         raw_local_irq_save(flags);
3062         check_flags(flags);
3063
3064         current->lockdep_recursion = 1;
3065         __lock_acquire(lock, subclass, trylock, read, check,
3066                        irqs_disabled_flags(flags), nest_lock, ip);
3067         current->lockdep_recursion = 0;
3068         raw_local_irq_restore(flags);
3069 }
3070 EXPORT_SYMBOL_GPL(lock_acquire);
3071
3072 void lock_release(struct lockdep_map *lock, int nested,
3073                           unsigned long ip)
3074 {
3075         unsigned long flags;
3076
3077         trace_lock_release(lock, nested, ip);
3078
3079         if (unlikely(current->lockdep_recursion))
3080                 return;
3081
3082         raw_local_irq_save(flags);
3083         check_flags(flags);
3084         current->lockdep_recursion = 1;
3085         __lock_release(lock, nested, ip);
3086         current->lockdep_recursion = 0;
3087         raw_local_irq_restore(flags);
3088 }
3089 EXPORT_SYMBOL_GPL(lock_release);
3090
3091 void lockdep_set_current_reclaim_state(gfp_t gfp_mask)
3092 {
3093         current->lockdep_reclaim_gfp = gfp_mask;
3094 }
3095
3096 void lockdep_clear_current_reclaim_state(void)
3097 {
3098         current->lockdep_reclaim_gfp = 0;
3099 }
3100
3101 #ifdef CONFIG_LOCK_STAT
3102 static int
3103 print_lock_contention_bug(struct task_struct *curr, struct lockdep_map *lock,
3104                            unsigned long ip)
3105 {
3106         if (!debug_locks_off())
3107                 return 0;
3108         if (debug_locks_silent)
3109                 return 0;
3110
3111         printk("\n=================================\n");
3112         printk(  "[ BUG: bad contention detected! ]\n");
3113         printk(  "---------------------------------\n");
3114         printk("%s/%d is trying to contend lock (",
3115                 curr->comm, task_pid_nr(curr));
3116         print_lockdep_cache(lock);
3117         printk(") at:\n");
3118         print_ip_sym(ip);
3119         printk("but there are no locks held!\n");
3120         printk("\nother info that might help us debug this:\n");
3121         lockdep_print_held_locks(curr);
3122
3123         printk("\nstack backtrace:\n");
3124         dump_stack();
3125
3126         return 0;
3127 }
3128
3129 static void
3130 __lock_contended(struct lockdep_map *lock, unsigned long ip)
3131 {
3132         struct task_struct *curr = current;
3133         struct held_lock *hlock, *prev_hlock;
3134         struct lock_class_stats *stats;
3135         unsigned int depth;
3136         int i, contention_point, contending_point;
3137
3138         depth = curr->lockdep_depth;
3139         if (DEBUG_LOCKS_WARN_ON(!depth))
3140                 return;
3141
3142         prev_hlock = NULL;
3143         for (i = depth-1; i >= 0; i--) {
3144                 hlock = curr->held_locks + i;
3145                 /*
3146                  * We must not cross into another context:
3147                  */
3148                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3149                         break;
3150                 if (hlock->instance == lock)
3151                         goto found_it;
3152                 prev_hlock = hlock;
3153         }
3154         print_lock_contention_bug(curr, lock, ip);
3155         return;
3156
3157 found_it:
3158         hlock->waittime_stamp = sched_clock();
3159
3160         contention_point = lock_point(hlock_class(hlock)->contention_point, ip);
3161         contending_point = lock_point(hlock_class(hlock)->contending_point,
3162                                       lock->ip);
3163
3164         stats = get_lock_stats(hlock_class(hlock));
3165         if (contention_point < LOCKSTAT_POINTS)
3166                 stats->contention_point[contention_point]++;
3167         if (contending_point < LOCKSTAT_POINTS)
3168                 stats->contending_point[contending_point]++;
3169         if (lock->cpu != smp_processor_id())
3170                 stats->bounces[bounce_contended + !!hlock->read]++;
3171         put_lock_stats(stats);
3172 }
3173
3174 static void
3175 __lock_acquired(struct lockdep_map *lock, unsigned long ip)
3176 {
3177         struct task_struct *curr = current;
3178         struct held_lock *hlock, *prev_hlock;
3179         struct lock_class_stats *stats;
3180         unsigned int depth;
3181         u64 now;
3182         s64 waittime = 0;
3183         int i, cpu;
3184
3185         depth = curr->lockdep_depth;
3186         if (DEBUG_LOCKS_WARN_ON(!depth))
3187                 return;
3188
3189         prev_hlock = NULL;
3190         for (i = depth-1; i >= 0; i--) {
3191                 hlock = curr->held_locks + i;
3192                 /*
3193                  * We must not cross into another context:
3194                  */
3195                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3196                         break;
3197                 if (hlock->instance == lock)
3198                         goto found_it;
3199                 prev_hlock = hlock;
3200         }
3201         print_lock_contention_bug(curr, lock, _RET_IP_);
3202         return;
3203
3204 found_it:
3205         cpu = smp_processor_id();
3206         if (hlock->waittime_stamp) {
3207                 now = sched_clock();
3208                 waittime = now - hlock->waittime_stamp;
3209                 hlock->holdtime_stamp = now;
3210         }
3211
3212         trace_lock_acquired(lock, ip, waittime);
3213
3214         stats = get_lock_stats(hlock_class(hlock));
3215         if (waittime) {
3216                 if (hlock->read)
3217                         lock_time_inc(&stats->read_waittime, waittime);
3218                 else
3219                         lock_time_inc(&stats->write_waittime, waittime);
3220         }
3221         if (lock->cpu != cpu)
3222                 stats->bounces[bounce_acquired + !!hlock->read]++;
3223         put_lock_stats(stats);
3224
3225         lock->cpu = cpu;
3226         lock->ip = ip;
3227 }
3228
3229 void lock_contended(struct lockdep_map *lock, unsigned long ip)
3230 {
3231         unsigned long flags;
3232
3233         trace_lock_contended(lock, ip);
3234
3235         if (unlikely(!lock_stat))
3236                 return;
3237
3238         if (unlikely(current->lockdep_recursion))
3239                 return;
3240
3241         raw_local_irq_save(flags);
3242         check_flags(flags);
3243         current->lockdep_recursion = 1;
3244         __lock_contended(lock, ip);
3245         current->lockdep_recursion = 0;
3246         raw_local_irq_restore(flags);
3247 }
3248 EXPORT_SYMBOL_GPL(lock_contended);
3249
3250 void lock_acquired(struct lockdep_map *lock, unsigned long ip)
3251 {
3252         unsigned long flags;
3253
3254         if (unlikely(!lock_stat))
3255                 return;
3256
3257         if (unlikely(current->lockdep_recursion))
3258                 return;
3259
3260         raw_local_irq_save(flags);
3261         check_flags(flags);
3262         current->lockdep_recursion = 1;
3263         __lock_acquired(lock, ip);
3264         current->lockdep_recursion = 0;
3265         raw_local_irq_restore(flags);
3266 }
3267 EXPORT_SYMBOL_GPL(lock_acquired);
3268 #endif
3269
3270 /*
3271  * Used by the testsuite, sanitize the validator state
3272  * after a simulated failure:
3273  */
3274
3275 void lockdep_reset(void)
3276 {
3277         unsigned long flags;
3278         int i;
3279
3280         raw_local_irq_save(flags);
3281         current->curr_chain_key = 0;
3282         current->lockdep_depth = 0;
3283         current->lockdep_recursion = 0;
3284         memset(current->held_locks, 0, MAX_LOCK_DEPTH*sizeof(struct held_lock));
3285         nr_hardirq_chains = 0;
3286         nr_softirq_chains = 0;
3287         nr_process_chains = 0;
3288         debug_locks = 1;
3289         for (i = 0; i < CHAINHASH_SIZE; i++)
3290                 INIT_LIST_HEAD(chainhash_table + i);
3291         raw_local_irq_restore(flags);
3292 }
3293
3294 static void zap_class(struct lock_class *class)
3295 {
3296         int i;
3297
3298         /*
3299          * Remove all dependencies this lock is
3300          * involved in:
3301          */
3302         for (i = 0; i < nr_list_entries; i++) {
3303                 if (list_entries[i].class == class)
3304                         list_del_rcu(&list_entries[i].entry);
3305         }
3306         /*
3307          * Unhash the class and remove it from the all_lock_classes list:
3308          */
3309         list_del_rcu(&class->hash_entry);
3310         list_del_rcu(&class->lock_entry);
3311
3312         class->key = NULL;
3313 }
3314
3315 static inline int within(const void *addr, void *start, unsigned long size)
3316 {
3317         return addr >= start && addr < start + size;
3318 }
3319
3320 void lockdep_free_key_range(void *start, unsigned long size)
3321 {
3322         struct lock_class *class, *next;
3323         struct list_head *head;
3324         unsigned long flags;
3325         int i;
3326         int locked;
3327
3328         raw_local_irq_save(flags);
3329         locked = graph_lock();
3330
3331         /*
3332          * Unhash all classes that were created by this module:
3333          */
3334         for (i = 0; i < CLASSHASH_SIZE; i++) {
3335                 head = classhash_table + i;
3336                 if (list_empty(head))
3337                         continue;
3338                 list_for_each_entry_safe(class, next, head, hash_entry) {
3339                         if (within(class->key, start, size))
3340                                 zap_class(class);
3341                         else if (within(class->name, start, size))
3342                                 zap_class(class);
3343                 }
3344         }
3345
3346         if (locked)
3347                 graph_unlock();
3348         raw_local_irq_restore(flags);
3349 }
3350
3351 void lockdep_reset_lock(struct lockdep_map *lock)
3352 {
3353         struct lock_class *class, *next;
3354         struct list_head *head;
3355         unsigned long flags;
3356         int i, j;
3357         int locked;
3358
3359         raw_local_irq_save(flags);
3360
3361         /*
3362          * Remove all classes this lock might have:
3363          */
3364         for (j = 0; j < MAX_LOCKDEP_SUBCLASSES; j++) {
3365                 /*
3366                  * If the class exists we look it up and zap it:
3367                  */
3368                 class = look_up_lock_class(lock, j);
3369                 if (class)
3370                         zap_class(class);
3371         }
3372         /*
3373          * Debug check: in the end all mapped classes should
3374          * be gone.
3375          */
3376         locked = graph_lock();
3377         for (i = 0; i < CLASSHASH_SIZE; i++) {
3378                 head = classhash_table + i;
3379                 if (list_empty(head))
3380                         continue;
3381                 list_for_each_entry_safe(class, next, head, hash_entry) {
3382                         if (unlikely(class == lock->class_cache)) {
3383                                 if (debug_locks_off_graph_unlock())
3384                                         WARN_ON(1);
3385                                 goto out_restore;
3386                         }
3387                 }
3388         }
3389         if (locked)
3390                 graph_unlock();
3391
3392 out_restore:
3393         raw_local_irq_restore(flags);
3394 }
3395
3396 void lockdep_init(void)
3397 {
3398         int i;
3399
3400         /*
3401          * Some architectures have their own start_kernel()
3402          * code which calls lockdep_init(), while we also
3403          * call lockdep_init() from the start_kernel() itself,
3404          * and we want to initialize the hashes only once:
3405          */
3406         if (lockdep_initialized)
3407                 return;
3408
3409         for (i = 0; i < CLASSHASH_SIZE; i++)
3410                 INIT_LIST_HEAD(classhash_table + i);
3411
3412         for (i = 0; i < CHAINHASH_SIZE; i++)
3413                 INIT_LIST_HEAD(chainhash_table + i);
3414
3415         lockdep_initialized = 1;
3416 }
3417
3418 void __init lockdep_info(void)
3419 {
3420         printk("Lock dependency validator: Copyright (c) 2006 Red Hat, Inc., Ingo Molnar\n");
3421
3422         printk("... MAX_LOCKDEP_SUBCLASSES:  %lu\n", MAX_LOCKDEP_SUBCLASSES);
3423         printk("... MAX_LOCK_DEPTH:          %lu\n", MAX_LOCK_DEPTH);
3424         printk("... MAX_LOCKDEP_KEYS:        %lu\n", MAX_LOCKDEP_KEYS);
3425         printk("... CLASSHASH_SIZE:          %lu\n", CLASSHASH_SIZE);
3426         printk("... MAX_LOCKDEP_ENTRIES:     %lu\n", MAX_LOCKDEP_ENTRIES);
3427         printk("... MAX_LOCKDEP_CHAINS:      %lu\n", MAX_LOCKDEP_CHAINS);
3428         printk("... CHAINHASH_SIZE:          %lu\n", CHAINHASH_SIZE);
3429
3430         printk(" memory used by lock dependency info: %lu kB\n",
3431                 (sizeof(struct lock_class) * MAX_LOCKDEP_KEYS +
3432                 sizeof(struct list_head) * CLASSHASH_SIZE +
3433                 sizeof(struct lock_list) * MAX_LOCKDEP_ENTRIES +
3434                 sizeof(struct lock_chain) * MAX_LOCKDEP_CHAINS +
3435                 sizeof(struct list_head) * CHAINHASH_SIZE) / 1024);
3436
3437         printk(" per task-struct memory footprint: %lu bytes\n",
3438                 sizeof(struct held_lock) * MAX_LOCK_DEPTH);
3439
3440 #ifdef CONFIG_DEBUG_LOCKDEP
3441         if (lockdep_init_error) {
3442                 printk("WARNING: lockdep init error! Arch code didn't call lockdep_init() early enough?\n");
3443                 printk("Call stack leading to lockdep invocation was:\n");
3444                 print_stack_trace(&lockdep_init_trace, 0);
3445         }
3446 #endif
3447 }
3448
3449 static void
3450 print_freed_lock_bug(struct task_struct *curr, const void *mem_from,
3451                      const void *mem_to, struct held_lock *hlock)
3452 {
3453         if (!debug_locks_off())
3454                 return;
3455         if (debug_locks_silent)
3456                 return;
3457
3458         printk("\n=========================\n");
3459         printk(  "[ BUG: held lock freed! ]\n");
3460         printk(  "-------------------------\n");
3461         printk("%s/%d is freeing memory %p-%p, with a lock still held there!\n",
3462                 curr->comm, task_pid_nr(curr), mem_from, mem_to-1);
3463         print_lock(hlock);
3464         lockdep_print_held_locks(curr);
3465
3466         printk("\nstack backtrace:\n");
3467         dump_stack();
3468 }
3469
3470 static inline int not_in_range(const void* mem_from, unsigned long mem_len,
3471                                 const void* lock_from, unsigned long lock_len)
3472 {
3473         return lock_from + lock_len <= mem_from ||
3474                 mem_from + mem_len <= lock_from;
3475 }
3476
3477 /*
3478  * Called when kernel memory is freed (or unmapped), or if a lock
3479  * is destroyed or reinitialized - this code checks whether there is
3480  * any held lock in the memory range of <from> to <to>:
3481  */
3482 void debug_check_no_locks_freed(const void *mem_from, unsigned long mem_len)
3483 {
3484         struct task_struct *curr = current;
3485         struct held_lock *hlock;
3486         unsigned long flags;
3487         int i;
3488
3489         if (unlikely(!debug_locks))
3490                 return;
3491
3492         local_irq_save(flags);
3493         for (i = 0; i < curr->lockdep_depth; i++) {
3494                 hlock = curr->held_locks + i;
3495
3496                 if (not_in_range(mem_from, mem_len, hlock->instance,
3497                                         sizeof(*hlock->instance)))
3498                         continue;
3499
3500                 print_freed_lock_bug(curr, mem_from, mem_from + mem_len, hlock);
3501                 break;
3502         }
3503         local_irq_restore(flags);
3504 }
3505 EXPORT_SYMBOL_GPL(debug_check_no_locks_freed);
3506
3507 static void print_held_locks_bug(struct task_struct *curr)
3508 {
3509         if (!debug_locks_off())
3510                 return;
3511         if (debug_locks_silent)
3512                 return;
3513
3514         printk("\n=====================================\n");
3515         printk(  "[ BUG: lock held at task exit time! ]\n");
3516         printk(  "-------------------------------------\n");
3517         printk("%s/%d is exiting with locks still held!\n",
3518                 curr->comm, task_pid_nr(curr));
3519         lockdep_print_held_locks(curr);
3520
3521         printk("\nstack backtrace:\n");
3522         dump_stack();
3523 }
3524
3525 void debug_check_no_locks_held(struct task_struct *task)
3526 {
3527         if (unlikely(task->lockdep_depth > 0))
3528                 print_held_locks_bug(task);
3529 }
3530
3531 void debug_show_all_locks(void)
3532 {
3533         struct task_struct *g, *p;
3534         int count = 10;
3535         int unlock = 1;
3536
3537         if (unlikely(!debug_locks)) {
3538                 printk("INFO: lockdep is turned off.\n");
3539                 return;
3540         }
3541         printk("\nShowing all locks held in the system:\n");
3542
3543         /*
3544          * Here we try to get the tasklist_lock as hard as possible,
3545          * if not successful after 2 seconds we ignore it (but keep
3546          * trying). This is to enable a debug printout even if a
3547          * tasklist_lock-holding task deadlocks or crashes.
3548          */
3549 retry:
3550         if (!read_trylock(&tasklist_lock)) {
3551                 if (count == 10)
3552                         printk("hm, tasklist_lock locked, retrying... ");
3553                 if (count) {
3554                         count--;
3555                         printk(" #%d", 10-count);
3556                         mdelay(200);
3557                         goto retry;
3558                 }
3559                 printk(" ignoring it.\n");
3560                 unlock = 0;
3561         } else {
3562                 if (count != 10)
3563                         printk(KERN_CONT " locked it.\n");
3564         }
3565
3566         do_each_thread(g, p) {
3567                 /*
3568                  * It's not reliable to print a task's held locks
3569                  * if it's not sleeping (or if it's not the current
3570                  * task):
3571                  */
3572                 if (p->state == TASK_RUNNING && p != current)
3573                         continue;
3574                 if (p->lockdep_depth)
3575                         lockdep_print_held_locks(p);
3576                 if (!unlock)
3577                         if (read_trylock(&tasklist_lock))
3578                                 unlock = 1;
3579         } while_each_thread(g, p);
3580
3581         printk("\n");
3582         printk("=============================================\n\n");
3583
3584         if (unlock)
3585                 read_unlock(&tasklist_lock);
3586 }
3587 EXPORT_SYMBOL_GPL(debug_show_all_locks);
3588
3589 /*
3590  * Careful: only use this function if you are sure that
3591  * the task cannot run in parallel!
3592  */
3593 void __debug_show_held_locks(struct task_struct *task)
3594 {
3595         if (unlikely(!debug_locks)) {
3596                 printk("INFO: lockdep is turned off.\n");
3597                 return;
3598         }
3599         lockdep_print_held_locks(task);
3600 }
3601 EXPORT_SYMBOL_GPL(__debug_show_held_locks);
3602
3603 void debug_show_held_locks(struct task_struct *task)
3604 {
3605                 __debug_show_held_locks(task);
3606 }
3607 EXPORT_SYMBOL_GPL(debug_show_held_locks);
3608
3609 void lockdep_sys_exit(void)
3610 {
3611         struct task_struct *curr = current;
3612
3613         if (unlikely(curr->lockdep_depth)) {
3614                 if (!debug_locks_off())
3615                         return;
3616                 printk("\n================================================\n");
3617                 printk(  "[ BUG: lock held when returning to user space! ]\n");
3618                 printk(  "------------------------------------------------\n");
3619                 printk("%s/%d is leaving the kernel with locks still held!\n",
3620                                 curr->comm, curr->pid);
3621                 lockdep_print_held_locks(curr);
3622         }
3623 }