lockdep: Introduce print_shortest_lock_dependencies
[linux-2.6.git] / kernel / lockdep.c
1 /*
2  * kernel/lockdep.c
3  *
4  * Runtime locking correctness validator
5  *
6  * Started by Ingo Molnar:
7  *
8  *  Copyright (C) 2006,2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
9  *  Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
10  *
11  * this code maps all the lock dependencies as they occur in a live kernel
12  * and will warn about the following classes of locking bugs:
13  *
14  * - lock inversion scenarios
15  * - circular lock dependencies
16  * - hardirq/softirq safe/unsafe locking bugs
17  *
18  * Bugs are reported even if the current locking scenario does not cause
19  * any deadlock at this point.
20  *
21  * I.e. if anytime in the past two locks were taken in a different order,
22  * even if it happened for another task, even if those were different
23  * locks (but of the same class as this lock), this code will detect it.
24  *
25  * Thanks to Arjan van de Ven for coming up with the initial idea of
26  * mapping lock dependencies runtime.
27  */
28 #define DISABLE_BRANCH_PROFILING
29 #include <linux/mutex.h>
30 #include <linux/sched.h>
31 #include <linux/delay.h>
32 #include <linux/module.h>
33 #include <linux/proc_fs.h>
34 #include <linux/seq_file.h>
35 #include <linux/spinlock.h>
36 #include <linux/kallsyms.h>
37 #include <linux/interrupt.h>
38 #include <linux/stacktrace.h>
39 #include <linux/debug_locks.h>
40 #include <linux/irqflags.h>
41 #include <linux/utsname.h>
42 #include <linux/hash.h>
43 #include <linux/ftrace.h>
44 #include <linux/stringify.h>
45 #include <linux/bitops.h>
46 #include <asm/sections.h>
47
48 #include "lockdep_internals.h"
49
50 #define CREATE_TRACE_POINTS
51 #include <trace/events/lockdep.h>
52
53 #ifdef CONFIG_PROVE_LOCKING
54 int prove_locking = 1;
55 module_param(prove_locking, int, 0644);
56 #else
57 #define prove_locking 0
58 #endif
59
60 #ifdef CONFIG_LOCK_STAT
61 int lock_stat = 1;
62 module_param(lock_stat, int, 0644);
63 #else
64 #define lock_stat 0
65 #endif
66
67 /*
68  * lockdep_lock: protects the lockdep graph, the hashes and the
69  *               class/list/hash allocators.
70  *
71  * This is one of the rare exceptions where it's justified
72  * to use a raw spinlock - we really dont want the spinlock
73  * code to recurse back into the lockdep code...
74  */
75 static raw_spinlock_t lockdep_lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
76
77 static int graph_lock(void)
78 {
79         __raw_spin_lock(&lockdep_lock);
80         /*
81          * Make sure that if another CPU detected a bug while
82          * walking the graph we dont change it (while the other
83          * CPU is busy printing out stuff with the graph lock
84          * dropped already)
85          */
86         if (!debug_locks) {
87                 __raw_spin_unlock(&lockdep_lock);
88                 return 0;
89         }
90         /* prevent any recursions within lockdep from causing deadlocks */
91         current->lockdep_recursion++;
92         return 1;
93 }
94
95 static inline int graph_unlock(void)
96 {
97         if (debug_locks && !__raw_spin_is_locked(&lockdep_lock))
98                 return DEBUG_LOCKS_WARN_ON(1);
99
100         current->lockdep_recursion--;
101         __raw_spin_unlock(&lockdep_lock);
102         return 0;
103 }
104
105 /*
106  * Turn lock debugging off and return with 0 if it was off already,
107  * and also release the graph lock:
108  */
109 static inline int debug_locks_off_graph_unlock(void)
110 {
111         int ret = debug_locks_off();
112
113         __raw_spin_unlock(&lockdep_lock);
114
115         return ret;
116 }
117
118 static int lockdep_initialized;
119
120 unsigned long nr_list_entries;
121 struct lock_list list_entries[MAX_LOCKDEP_ENTRIES];
122
123 /*
124  * All data structures here are protected by the global debug_lock.
125  *
126  * Mutex key structs only get allocated, once during bootup, and never
127  * get freed - this significantly simplifies the debugging code.
128  */
129 unsigned long nr_lock_classes;
130 static struct lock_class lock_classes[MAX_LOCKDEP_KEYS];
131
132 static inline struct lock_class *hlock_class(struct held_lock *hlock)
133 {
134         if (!hlock->class_idx) {
135                 DEBUG_LOCKS_WARN_ON(1);
136                 return NULL;
137         }
138         return lock_classes + hlock->class_idx - 1;
139 }
140
141 #ifdef CONFIG_LOCK_STAT
142 static DEFINE_PER_CPU(struct lock_class_stats[MAX_LOCKDEP_KEYS], lock_stats);
143
144 static int lock_point(unsigned long points[], unsigned long ip)
145 {
146         int i;
147
148         for (i = 0; i < LOCKSTAT_POINTS; i++) {
149                 if (points[i] == 0) {
150                         points[i] = ip;
151                         break;
152                 }
153                 if (points[i] == ip)
154                         break;
155         }
156
157         return i;
158 }
159
160 static void lock_time_inc(struct lock_time *lt, s64 time)
161 {
162         if (time > lt->max)
163                 lt->max = time;
164
165         if (time < lt->min || !lt->min)
166                 lt->min = time;
167
168         lt->total += time;
169         lt->nr++;
170 }
171
172 static inline void lock_time_add(struct lock_time *src, struct lock_time *dst)
173 {
174         dst->min += src->min;
175         dst->max += src->max;
176         dst->total += src->total;
177         dst->nr += src->nr;
178 }
179
180 struct lock_class_stats lock_stats(struct lock_class *class)
181 {
182         struct lock_class_stats stats;
183         int cpu, i;
184
185         memset(&stats, 0, sizeof(struct lock_class_stats));
186         for_each_possible_cpu(cpu) {
187                 struct lock_class_stats *pcs =
188                         &per_cpu(lock_stats, cpu)[class - lock_classes];
189
190                 for (i = 0; i < ARRAY_SIZE(stats.contention_point); i++)
191                         stats.contention_point[i] += pcs->contention_point[i];
192
193                 for (i = 0; i < ARRAY_SIZE(stats.contending_point); i++)
194                         stats.contending_point[i] += pcs->contending_point[i];
195
196                 lock_time_add(&pcs->read_waittime, &stats.read_waittime);
197                 lock_time_add(&pcs->write_waittime, &stats.write_waittime);
198
199                 lock_time_add(&pcs->read_holdtime, &stats.read_holdtime);
200                 lock_time_add(&pcs->write_holdtime, &stats.write_holdtime);
201
202                 for (i = 0; i < ARRAY_SIZE(stats.bounces); i++)
203                         stats.bounces[i] += pcs->bounces[i];
204         }
205
206         return stats;
207 }
208
209 void clear_lock_stats(struct lock_class *class)
210 {
211         int cpu;
212
213         for_each_possible_cpu(cpu) {
214                 struct lock_class_stats *cpu_stats =
215                         &per_cpu(lock_stats, cpu)[class - lock_classes];
216
217                 memset(cpu_stats, 0, sizeof(struct lock_class_stats));
218         }
219         memset(class->contention_point, 0, sizeof(class->contention_point));
220         memset(class->contending_point, 0, sizeof(class->contending_point));
221 }
222
223 static struct lock_class_stats *get_lock_stats(struct lock_class *class)
224 {
225         return &get_cpu_var(lock_stats)[class - lock_classes];
226 }
227
228 static void put_lock_stats(struct lock_class_stats *stats)
229 {
230         put_cpu_var(lock_stats);
231 }
232
233 static void lock_release_holdtime(struct held_lock *hlock)
234 {
235         struct lock_class_stats *stats;
236         s64 holdtime;
237
238         if (!lock_stat)
239                 return;
240
241         holdtime = sched_clock() - hlock->holdtime_stamp;
242
243         stats = get_lock_stats(hlock_class(hlock));
244         if (hlock->read)
245                 lock_time_inc(&stats->read_holdtime, holdtime);
246         else
247                 lock_time_inc(&stats->write_holdtime, holdtime);
248         put_lock_stats(stats);
249 }
250 #else
251 static inline void lock_release_holdtime(struct held_lock *hlock)
252 {
253 }
254 #endif
255
256 /*
257  * We keep a global list of all lock classes. The list only grows,
258  * never shrinks. The list is only accessed with the lockdep
259  * spinlock lock held.
260  */
261 LIST_HEAD(all_lock_classes);
262
263 /*
264  * The lockdep classes are in a hash-table as well, for fast lookup:
265  */
266 #define CLASSHASH_BITS          (MAX_LOCKDEP_KEYS_BITS - 1)
267 #define CLASSHASH_SIZE          (1UL << CLASSHASH_BITS)
268 #define __classhashfn(key)      hash_long((unsigned long)key, CLASSHASH_BITS)
269 #define classhashentry(key)     (classhash_table + __classhashfn((key)))
270
271 static struct list_head classhash_table[CLASSHASH_SIZE];
272
273 /*
274  * We put the lock dependency chains into a hash-table as well, to cache
275  * their existence:
276  */
277 #define CHAINHASH_BITS          (MAX_LOCKDEP_CHAINS_BITS-1)
278 #define CHAINHASH_SIZE          (1UL << CHAINHASH_BITS)
279 #define __chainhashfn(chain)    hash_long(chain, CHAINHASH_BITS)
280 #define chainhashentry(chain)   (chainhash_table + __chainhashfn((chain)))
281
282 static struct list_head chainhash_table[CHAINHASH_SIZE];
283
284 /*
285  * The hash key of the lock dependency chains is a hash itself too:
286  * it's a hash of all locks taken up to that lock, including that lock.
287  * It's a 64-bit hash, because it's important for the keys to be
288  * unique.
289  */
290 #define iterate_chain_key(key1, key2) \
291         (((key1) << MAX_LOCKDEP_KEYS_BITS) ^ \
292         ((key1) >> (64-MAX_LOCKDEP_KEYS_BITS)) ^ \
293         (key2))
294
295 void lockdep_off(void)
296 {
297         current->lockdep_recursion++;
298 }
299 EXPORT_SYMBOL(lockdep_off);
300
301 void lockdep_on(void)
302 {
303         current->lockdep_recursion--;
304 }
305 EXPORT_SYMBOL(lockdep_on);
306
307 /*
308  * Debugging switches:
309  */
310
311 #define VERBOSE                 0
312 #define VERY_VERBOSE            0
313
314 #if VERBOSE
315 # define HARDIRQ_VERBOSE        1
316 # define SOFTIRQ_VERBOSE        1
317 # define RECLAIM_VERBOSE        1
318 #else
319 # define HARDIRQ_VERBOSE        0
320 # define SOFTIRQ_VERBOSE        0
321 # define RECLAIM_VERBOSE        0
322 #endif
323
324 #if VERBOSE || HARDIRQ_VERBOSE || SOFTIRQ_VERBOSE || RECLAIM_VERBOSE
325 /*
326  * Quick filtering for interesting events:
327  */
328 static int class_filter(struct lock_class *class)
329 {
330 #if 0
331         /* Example */
332         if (class->name_version == 1 &&
333                         !strcmp(class->name, "lockname"))
334                 return 1;
335         if (class->name_version == 1 &&
336                         !strcmp(class->name, "&struct->lockfield"))
337                 return 1;
338 #endif
339         /* Filter everything else. 1 would be to allow everything else */
340         return 0;
341 }
342 #endif
343
344 static int verbose(struct lock_class *class)
345 {
346 #if VERBOSE
347         return class_filter(class);
348 #endif
349         return 0;
350 }
351
352 /*
353  * Stack-trace: tightly packed array of stack backtrace
354  * addresses. Protected by the graph_lock.
355  */
356 unsigned long nr_stack_trace_entries;
357 static unsigned long stack_trace[MAX_STACK_TRACE_ENTRIES];
358
359 static int save_trace(struct stack_trace *trace)
360 {
361         trace->nr_entries = 0;
362         trace->max_entries = MAX_STACK_TRACE_ENTRIES - nr_stack_trace_entries;
363         trace->entries = stack_trace + nr_stack_trace_entries;
364
365         trace->skip = 3;
366
367         save_stack_trace(trace);
368
369         trace->max_entries = trace->nr_entries;
370
371         nr_stack_trace_entries += trace->nr_entries;
372
373         if (nr_stack_trace_entries == MAX_STACK_TRACE_ENTRIES) {
374                 if (!debug_locks_off_graph_unlock())
375                         return 0;
376
377                 printk("BUG: MAX_STACK_TRACE_ENTRIES too low!\n");
378                 printk("turning off the locking correctness validator.\n");
379                 dump_stack();
380
381                 return 0;
382         }
383
384         return 1;
385 }
386
387 unsigned int nr_hardirq_chains;
388 unsigned int nr_softirq_chains;
389 unsigned int nr_process_chains;
390 unsigned int max_lockdep_depth;
391 unsigned int max_recursion_depth;
392
393 static unsigned int lockdep_dependency_gen_id;
394
395 static bool lockdep_dependency_visit(struct lock_class *source,
396                                      unsigned int depth)
397 {
398         if (!depth)
399                 lockdep_dependency_gen_id++;
400         if (source->dep_gen_id == lockdep_dependency_gen_id)
401                 return true;
402         source->dep_gen_id = lockdep_dependency_gen_id;
403         return false;
404 }
405
406 #ifdef CONFIG_DEBUG_LOCKDEP
407 /*
408  * We cannot printk in early bootup code. Not even early_printk()
409  * might work. So we mark any initialization errors and printk
410  * about it later on, in lockdep_info().
411  */
412 static int lockdep_init_error;
413 static unsigned long lockdep_init_trace_data[20];
414 static struct stack_trace lockdep_init_trace = {
415         .max_entries = ARRAY_SIZE(lockdep_init_trace_data),
416         .entries = lockdep_init_trace_data,
417 };
418
419 /*
420  * Various lockdep statistics:
421  */
422 atomic_t chain_lookup_hits;
423 atomic_t chain_lookup_misses;
424 atomic_t hardirqs_on_events;
425 atomic_t hardirqs_off_events;
426 atomic_t redundant_hardirqs_on;
427 atomic_t redundant_hardirqs_off;
428 atomic_t softirqs_on_events;
429 atomic_t softirqs_off_events;
430 atomic_t redundant_softirqs_on;
431 atomic_t redundant_softirqs_off;
432 atomic_t nr_unused_locks;
433 atomic_t nr_cyclic_checks;
434 atomic_t nr_cyclic_check_recursions;
435 atomic_t nr_find_usage_forwards_checks;
436 atomic_t nr_find_usage_forwards_recursions;
437 atomic_t nr_find_usage_backwards_checks;
438 atomic_t nr_find_usage_backwards_recursions;
439 #endif
440
441 /*
442  * Locking printouts:
443  */
444
445 #define __USAGE(__STATE)                                                \
446         [LOCK_USED_IN_##__STATE] = "IN-"__stringify(__STATE)"-W",       \
447         [LOCK_ENABLED_##__STATE] = __stringify(__STATE)"-ON-W",         \
448         [LOCK_USED_IN_##__STATE##_READ] = "IN-"__stringify(__STATE)"-R",\
449         [LOCK_ENABLED_##__STATE##_READ] = __stringify(__STATE)"-ON-R",
450
451 static const char *usage_str[] =
452 {
453 #define LOCKDEP_STATE(__STATE) __USAGE(__STATE)
454 #include "lockdep_states.h"
455 #undef LOCKDEP_STATE
456         [LOCK_USED] = "INITIAL USE",
457 };
458
459 const char * __get_key_name(struct lockdep_subclass_key *key, char *str)
460 {
461         return kallsyms_lookup((unsigned long)key, NULL, NULL, NULL, str);
462 }
463
464 static inline unsigned long lock_flag(enum lock_usage_bit bit)
465 {
466         return 1UL << bit;
467 }
468
469 static char get_usage_char(struct lock_class *class, enum lock_usage_bit bit)
470 {
471         char c = '.';
472
473         if (class->usage_mask & lock_flag(bit + 2))
474                 c = '+';
475         if (class->usage_mask & lock_flag(bit)) {
476                 c = '-';
477                 if (class->usage_mask & lock_flag(bit + 2))
478                         c = '?';
479         }
480
481         return c;
482 }
483
484 void get_usage_chars(struct lock_class *class, char usage[LOCK_USAGE_CHARS])
485 {
486         int i = 0;
487
488 #define LOCKDEP_STATE(__STATE)                                          \
489         usage[i++] = get_usage_char(class, LOCK_USED_IN_##__STATE);     \
490         usage[i++] = get_usage_char(class, LOCK_USED_IN_##__STATE##_READ);
491 #include "lockdep_states.h"
492 #undef LOCKDEP_STATE
493
494         usage[i] = '\0';
495 }
496
497 static void print_lock_name(struct lock_class *class)
498 {
499         char str[KSYM_NAME_LEN], usage[LOCK_USAGE_CHARS];
500         const char *name;
501
502         get_usage_chars(class, usage);
503
504         name = class->name;
505         if (!name) {
506                 name = __get_key_name(class->key, str);
507                 printk(" (%s", name);
508         } else {
509                 printk(" (%s", name);
510                 if (class->name_version > 1)
511                         printk("#%d", class->name_version);
512                 if (class->subclass)
513                         printk("/%d", class->subclass);
514         }
515         printk("){%s}", usage);
516 }
517
518 static void print_lockdep_cache(struct lockdep_map *lock)
519 {
520         const char *name;
521         char str[KSYM_NAME_LEN];
522
523         name = lock->name;
524         if (!name)
525                 name = __get_key_name(lock->key->subkeys, str);
526
527         printk("%s", name);
528 }
529
530 static void print_lock(struct held_lock *hlock)
531 {
532         print_lock_name(hlock_class(hlock));
533         printk(", at: ");
534         print_ip_sym(hlock->acquire_ip);
535 }
536
537 static void lockdep_print_held_locks(struct task_struct *curr)
538 {
539         int i, depth = curr->lockdep_depth;
540
541         if (!depth) {
542                 printk("no locks held by %s/%d.\n", curr->comm, task_pid_nr(curr));
543                 return;
544         }
545         printk("%d lock%s held by %s/%d:\n",
546                 depth, depth > 1 ? "s" : "", curr->comm, task_pid_nr(curr));
547
548         for (i = 0; i < depth; i++) {
549                 printk(" #%d: ", i);
550                 print_lock(curr->held_locks + i);
551         }
552 }
553
554 static void print_lock_class_header(struct lock_class *class, int depth)
555 {
556         int bit;
557
558         printk("%*s->", depth, "");
559         print_lock_name(class);
560         printk(" ops: %lu", class->ops);
561         printk(" {\n");
562
563         for (bit = 0; bit < LOCK_USAGE_STATES; bit++) {
564                 if (class->usage_mask & (1 << bit)) {
565                         int len = depth;
566
567                         len += printk("%*s   %s", depth, "", usage_str[bit]);
568                         len += printk(" at:\n");
569                         print_stack_trace(class->usage_traces + bit, len);
570                 }
571         }
572         printk("%*s }\n", depth, "");
573
574         printk("%*s ... key      at: ",depth,"");
575         print_ip_sym((unsigned long)class->key);
576 }
577
578 /*
579  * printk the shortest lock dependencies from @start to @end in reverse order:
580  */
581 static void __used
582 print_shortest_lock_dependencies(struct lock_list *leaf,
583                                 struct lock_list *root)
584 {
585         struct lock_list *entry = leaf;
586         int depth;
587
588         /*compute depth from generated tree by BFS*/
589         depth = get_lock_depth(leaf);
590
591         do {
592                 print_lock_class_header(entry->class, depth);
593                 printk("%*s ... acquired at:\n", depth, "");
594                 print_stack_trace(&entry->trace, 2);
595                 printk("\n");
596
597                 if (depth == 0 && (entry != root)) {
598                         printk("lockdep:%s bad BFS generated tree\n", __func__);
599                         break;
600                 }
601
602                 entry = get_lock_parent(entry);
603                 depth--;
604         } while (entry && (depth >= 0));
605
606         return;
607 }
608 /*
609  * printk all lock dependencies starting at <entry>:
610  */
611 static void __used
612 print_lock_dependencies(struct lock_class *class, int depth)
613 {
614         struct lock_list *entry;
615
616         if (lockdep_dependency_visit(class, depth))
617                 return;
618
619         if (DEBUG_LOCKS_WARN_ON(depth >= 20))
620                 return;
621
622         print_lock_class_header(class, depth);
623
624         list_for_each_entry(entry, &class->locks_after, entry) {
625                 if (DEBUG_LOCKS_WARN_ON(!entry->class))
626                         return;
627
628                 print_lock_dependencies(entry->class, depth + 1);
629
630                 printk("%*s ... acquired at:\n",depth,"");
631                 print_stack_trace(&entry->trace, 2);
632                 printk("\n");
633         }
634 }
635
636 static void print_kernel_version(void)
637 {
638         printk("%s %.*s\n", init_utsname()->release,
639                 (int)strcspn(init_utsname()->version, " "),
640                 init_utsname()->version);
641 }
642
643 static int very_verbose(struct lock_class *class)
644 {
645 #if VERY_VERBOSE
646         return class_filter(class);
647 #endif
648         return 0;
649 }
650
651 /*
652  * Is this the address of a static object:
653  */
654 static int static_obj(void *obj)
655 {
656         unsigned long start = (unsigned long) &_stext,
657                       end   = (unsigned long) &_end,
658                       addr  = (unsigned long) obj;
659 #ifdef CONFIG_SMP
660         int i;
661 #endif
662
663         /*
664          * static variable?
665          */
666         if ((addr >= start) && (addr < end))
667                 return 1;
668
669 #ifdef CONFIG_SMP
670         /*
671          * percpu var?
672          */
673         for_each_possible_cpu(i) {
674                 start = (unsigned long) &__per_cpu_start + per_cpu_offset(i);
675                 end   = (unsigned long) &__per_cpu_start + PERCPU_ENOUGH_ROOM
676                                         + per_cpu_offset(i);
677
678                 if ((addr >= start) && (addr < end))
679                         return 1;
680         }
681 #endif
682
683         /*
684          * module var?
685          */
686         return is_module_address(addr);
687 }
688
689 /*
690  * To make lock name printouts unique, we calculate a unique
691  * class->name_version generation counter:
692  */
693 static int count_matching_names(struct lock_class *new_class)
694 {
695         struct lock_class *class;
696         int count = 0;
697
698         if (!new_class->name)
699                 return 0;
700
701         list_for_each_entry(class, &all_lock_classes, lock_entry) {
702                 if (new_class->key - new_class->subclass == class->key)
703                         return class->name_version;
704                 if (class->name && !strcmp(class->name, new_class->name))
705                         count = max(count, class->name_version);
706         }
707
708         return count + 1;
709 }
710
711 /*
712  * Register a lock's class in the hash-table, if the class is not present
713  * yet. Otherwise we look it up. We cache the result in the lock object
714  * itself, so actual lookup of the hash should be once per lock object.
715  */
716 static inline struct lock_class *
717 look_up_lock_class(struct lockdep_map *lock, unsigned int subclass)
718 {
719         struct lockdep_subclass_key *key;
720         struct list_head *hash_head;
721         struct lock_class *class;
722
723 #ifdef CONFIG_DEBUG_LOCKDEP
724         /*
725          * If the architecture calls into lockdep before initializing
726          * the hashes then we'll warn about it later. (we cannot printk
727          * right now)
728          */
729         if (unlikely(!lockdep_initialized)) {
730                 lockdep_init();
731                 lockdep_init_error = 1;
732                 save_stack_trace(&lockdep_init_trace);
733         }
734 #endif
735
736         /*
737          * Static locks do not have their class-keys yet - for them the key
738          * is the lock object itself:
739          */
740         if (unlikely(!lock->key))
741                 lock->key = (void *)lock;
742
743         /*
744          * NOTE: the class-key must be unique. For dynamic locks, a static
745          * lock_class_key variable is passed in through the mutex_init()
746          * (or spin_lock_init()) call - which acts as the key. For static
747          * locks we use the lock object itself as the key.
748          */
749         BUILD_BUG_ON(sizeof(struct lock_class_key) >
750                         sizeof(struct lockdep_map));
751
752         key = lock->key->subkeys + subclass;
753
754         hash_head = classhashentry(key);
755
756         /*
757          * We can walk the hash lockfree, because the hash only
758          * grows, and we are careful when adding entries to the end:
759          */
760         list_for_each_entry(class, hash_head, hash_entry) {
761                 if (class->key == key) {
762                         WARN_ON_ONCE(class->name != lock->name);
763                         return class;
764                 }
765         }
766
767         return NULL;
768 }
769
770 /*
771  * Register a lock's class in the hash-table, if the class is not present
772  * yet. Otherwise we look it up. We cache the result in the lock object
773  * itself, so actual lookup of the hash should be once per lock object.
774  */
775 static inline struct lock_class *
776 register_lock_class(struct lockdep_map *lock, unsigned int subclass, int force)
777 {
778         struct lockdep_subclass_key *key;
779         struct list_head *hash_head;
780         struct lock_class *class;
781         unsigned long flags;
782
783         class = look_up_lock_class(lock, subclass);
784         if (likely(class))
785                 return class;
786
787         /*
788          * Debug-check: all keys must be persistent!
789          */
790         if (!static_obj(lock->key)) {
791                 debug_locks_off();
792                 printk("INFO: trying to register non-static key.\n");
793                 printk("the code is fine but needs lockdep annotation.\n");
794                 printk("turning off the locking correctness validator.\n");
795                 dump_stack();
796
797                 return NULL;
798         }
799
800         key = lock->key->subkeys + subclass;
801         hash_head = classhashentry(key);
802
803         raw_local_irq_save(flags);
804         if (!graph_lock()) {
805                 raw_local_irq_restore(flags);
806                 return NULL;
807         }
808         /*
809          * We have to do the hash-walk again, to avoid races
810          * with another CPU:
811          */
812         list_for_each_entry(class, hash_head, hash_entry)
813                 if (class->key == key)
814                         goto out_unlock_set;
815         /*
816          * Allocate a new key from the static array, and add it to
817          * the hash:
818          */
819         if (nr_lock_classes >= MAX_LOCKDEP_KEYS) {
820                 if (!debug_locks_off_graph_unlock()) {
821                         raw_local_irq_restore(flags);
822                         return NULL;
823                 }
824                 raw_local_irq_restore(flags);
825
826                 printk("BUG: MAX_LOCKDEP_KEYS too low!\n");
827                 printk("turning off the locking correctness validator.\n");
828                 dump_stack();
829                 return NULL;
830         }
831         class = lock_classes + nr_lock_classes++;
832         debug_atomic_inc(&nr_unused_locks);
833         class->key = key;
834         class->name = lock->name;
835         class->subclass = subclass;
836         INIT_LIST_HEAD(&class->lock_entry);
837         INIT_LIST_HEAD(&class->locks_before);
838         INIT_LIST_HEAD(&class->locks_after);
839         class->name_version = count_matching_names(class);
840         /*
841          * We use RCU's safe list-add method to make
842          * parallel walking of the hash-list safe:
843          */
844         list_add_tail_rcu(&class->hash_entry, hash_head);
845         /*
846          * Add it to the global list of classes:
847          */
848         list_add_tail_rcu(&class->lock_entry, &all_lock_classes);
849
850         if (verbose(class)) {
851                 graph_unlock();
852                 raw_local_irq_restore(flags);
853
854                 printk("\nnew class %p: %s", class->key, class->name);
855                 if (class->name_version > 1)
856                         printk("#%d", class->name_version);
857                 printk("\n");
858                 dump_stack();
859
860                 raw_local_irq_save(flags);
861                 if (!graph_lock()) {
862                         raw_local_irq_restore(flags);
863                         return NULL;
864                 }
865         }
866 out_unlock_set:
867         graph_unlock();
868         raw_local_irq_restore(flags);
869
870         if (!subclass || force)
871                 lock->class_cache = class;
872
873         if (DEBUG_LOCKS_WARN_ON(class->subclass != subclass))
874                 return NULL;
875
876         return class;
877 }
878
879 #ifdef CONFIG_PROVE_LOCKING
880 /*
881  * Allocate a lockdep entry. (assumes the graph_lock held, returns
882  * with NULL on failure)
883  */
884 static struct lock_list *alloc_list_entry(void)
885 {
886         if (nr_list_entries >= MAX_LOCKDEP_ENTRIES) {
887                 if (!debug_locks_off_graph_unlock())
888                         return NULL;
889
890                 printk("BUG: MAX_LOCKDEP_ENTRIES too low!\n");
891                 printk("turning off the locking correctness validator.\n");
892                 dump_stack();
893                 return NULL;
894         }
895         return list_entries + nr_list_entries++;
896 }
897
898 /*
899  * Add a new dependency to the head of the list:
900  */
901 static int add_lock_to_list(struct lock_class *class, struct lock_class *this,
902                             struct list_head *head, unsigned long ip, int distance)
903 {
904         struct lock_list *entry;
905         /*
906          * Lock not present yet - get a new dependency struct and
907          * add it to the list:
908          */
909         entry = alloc_list_entry();
910         if (!entry)
911                 return 0;
912
913         if (!save_trace(&entry->trace))
914                 return 0;
915
916         entry->class = this;
917         entry->distance = distance;
918         /*
919          * Since we never remove from the dependency list, the list can
920          * be walked lockless by other CPUs, it's only allocation
921          * that must be protected by the spinlock. But this also means
922          * we must make new entries visible only once writes to the
923          * entry become visible - hence the RCU op:
924          */
925         list_add_tail_rcu(&entry->entry, head);
926
927         return 1;
928 }
929
930 unsigned long bfs_accessed[BITS_TO_LONGS(MAX_LOCKDEP_ENTRIES)];
931 static struct circular_queue  lock_cq;
932
933 static int __bfs(struct lock_list *source_entry,
934                         void *data,
935                         int (*match)(struct lock_list *entry, void *data),
936                         struct lock_list **target_entry,
937                         int forward)
938 {
939         struct lock_list *entry;
940         struct list_head *head;
941         struct circular_queue *cq = &lock_cq;
942         int ret = 1;
943
944         if (match(source_entry, data)) {
945                 *target_entry = source_entry;
946                 ret = 0;
947                 goto exit;
948         }
949
950         if (forward)
951                 head = &source_entry->class->locks_after;
952         else
953                 head = &source_entry->class->locks_before;
954
955         if (list_empty(head))
956                 goto exit;
957
958         __cq_init(cq);
959         __cq_enqueue(cq, (unsigned long)source_entry);
960
961         while (!__cq_empty(cq)) {
962                 struct lock_list *lock;
963
964                 __cq_dequeue(cq, (unsigned long *)&lock);
965
966                 if (!lock->class) {
967                         ret = -2;
968                         goto exit;
969                 }
970
971                 if (forward)
972                         head = &lock->class->locks_after;
973                 else
974                         head = &lock->class->locks_before;
975
976                 list_for_each_entry(entry, head, entry) {
977                         if (!lock_accessed(entry)) {
978                                 mark_lock_accessed(entry, lock);
979                                 if (match(entry, data)) {
980                                         *target_entry = entry;
981                                         ret = 0;
982                                         goto exit;
983                                 }
984
985                                 if (__cq_enqueue(cq, (unsigned long)entry)) {
986                                         ret = -1;
987                                         goto exit;
988                                 }
989                         }
990                 }
991         }
992 exit:
993         return ret;
994 }
995
996 static inline int __bfs_forwards(struct lock_list *src_entry,
997                         void *data,
998                         int (*match)(struct lock_list *entry, void *data),
999                         struct lock_list **target_entry)
1000 {
1001         return __bfs(src_entry, data, match, target_entry, 1);
1002
1003 }
1004
1005 static inline int __bfs_backwards(struct lock_list *src_entry,
1006                         void *data,
1007                         int (*match)(struct lock_list *entry, void *data),
1008                         struct lock_list **target_entry)
1009 {
1010         return __bfs(src_entry, data, match, target_entry, 0);
1011
1012 }
1013
1014 /*
1015  * Recursive, forwards-direction lock-dependency checking, used for
1016  * both noncyclic checking and for hardirq-unsafe/softirq-unsafe
1017  * checking.
1018  */
1019
1020 /*
1021  * Print a dependency chain entry (this is only done when a deadlock
1022  * has been detected):
1023  */
1024 static noinline int
1025 print_circular_bug_entry(struct lock_list *target, int depth)
1026 {
1027         if (debug_locks_silent)
1028                 return 0;
1029         printk("\n-> #%u", depth);
1030         print_lock_name(target->class);
1031         printk(":\n");
1032         print_stack_trace(&target->trace, 6);
1033
1034         return 0;
1035 }
1036
1037 /*
1038  * When a circular dependency is detected, print the
1039  * header first:
1040  */
1041 static noinline int
1042 print_circular_bug_header(struct lock_list *entry, unsigned int depth,
1043                         struct held_lock *check_src,
1044                         struct held_lock *check_tgt)
1045 {
1046         struct task_struct *curr = current;
1047
1048         if (debug_locks_silent)
1049                 return 0;
1050
1051         printk("\n=======================================================\n");
1052         printk(  "[ INFO: possible circular locking dependency detected ]\n");
1053         print_kernel_version();
1054         printk(  "-------------------------------------------------------\n");
1055         printk("%s/%d is trying to acquire lock:\n",
1056                 curr->comm, task_pid_nr(curr));
1057         print_lock(check_src);
1058         printk("\nbut task is already holding lock:\n");
1059         print_lock(check_tgt);
1060         printk("\nwhich lock already depends on the new lock.\n\n");
1061         printk("\nthe existing dependency chain (in reverse order) is:\n");
1062
1063         print_circular_bug_entry(entry, depth);
1064
1065         return 0;
1066 }
1067
1068 static inline int class_equal(struct lock_list *entry, void *data)
1069 {
1070         return entry->class == data;
1071 }
1072
1073 static noinline int print_circular_bug(struct lock_list *this,
1074                                 struct lock_list *target,
1075                                 struct held_lock *check_src,
1076                                 struct held_lock *check_tgt)
1077 {
1078         struct task_struct *curr = current;
1079         struct lock_list *parent;
1080         int depth;
1081
1082         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1083                 return 0;
1084
1085         if (!save_trace(&this->trace))
1086                 return 0;
1087
1088         depth = get_lock_depth(target);
1089
1090         print_circular_bug_header(target, depth, check_src, check_tgt);
1091
1092         parent = get_lock_parent(target);
1093
1094         while (parent) {
1095                 print_circular_bug_entry(parent, --depth);
1096                 parent = get_lock_parent(parent);
1097         }
1098
1099         printk("\nother info that might help us debug this:\n\n");
1100         lockdep_print_held_locks(curr);
1101
1102         printk("\nstack backtrace:\n");
1103         dump_stack();
1104
1105         return 0;
1106 }
1107
1108 static noinline int print_bfs_bug(int ret)
1109 {
1110         if (!debug_locks_off_graph_unlock())
1111                 return 0;
1112
1113         WARN(1, "lockdep bfs error:%d\n", ret);
1114
1115         return 0;
1116 }
1117
1118 unsigned long __lockdep_count_forward_deps(struct lock_class *class,
1119                                            unsigned int depth)
1120 {
1121         struct lock_list *entry;
1122         unsigned long ret = 1;
1123
1124         if (lockdep_dependency_visit(class, depth))
1125                 return 0;
1126
1127         /*
1128          * Recurse this class's dependency list:
1129          */
1130         list_for_each_entry(entry, &class->locks_after, entry)
1131                 ret += __lockdep_count_forward_deps(entry->class, depth + 1);
1132
1133         return ret;
1134 }
1135
1136 unsigned long lockdep_count_forward_deps(struct lock_class *class)
1137 {
1138         unsigned long ret, flags;
1139
1140         local_irq_save(flags);
1141         __raw_spin_lock(&lockdep_lock);
1142         ret = __lockdep_count_forward_deps(class, 0);
1143         __raw_spin_unlock(&lockdep_lock);
1144         local_irq_restore(flags);
1145
1146         return ret;
1147 }
1148
1149 unsigned long __lockdep_count_backward_deps(struct lock_class *class,
1150                                             unsigned int depth)
1151 {
1152         struct lock_list *entry;
1153         unsigned long ret = 1;
1154
1155         if (lockdep_dependency_visit(class, depth))
1156                 return 0;
1157         /*
1158          * Recurse this class's dependency list:
1159          */
1160         list_for_each_entry(entry, &class->locks_before, entry)
1161                 ret += __lockdep_count_backward_deps(entry->class, depth + 1);
1162
1163         return ret;
1164 }
1165
1166 unsigned long lockdep_count_backward_deps(struct lock_class *class)
1167 {
1168         unsigned long ret, flags;
1169
1170         local_irq_save(flags);
1171         __raw_spin_lock(&lockdep_lock);
1172         ret = __lockdep_count_backward_deps(class, 0);
1173         __raw_spin_unlock(&lockdep_lock);
1174         local_irq_restore(flags);
1175
1176         return ret;
1177 }
1178
1179 /*
1180  * Prove that the dependency graph starting at <entry> can not
1181  * lead to <target>. Print an error and return 0 if it does.
1182  */
1183 static noinline int
1184 check_noncircular(struct lock_list *root, struct lock_class *target,
1185                 struct lock_list **target_entry)
1186 {
1187         int result;
1188
1189         debug_atomic_inc(&nr_cyclic_checks);
1190
1191         result = __bfs_forwards(root, target, class_equal, target_entry);
1192
1193         return result;
1194 }
1195
1196 #if defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING)
1197 /*
1198  * Forwards and backwards subgraph searching, for the purposes of
1199  * proving that two subgraphs can be connected by a new dependency
1200  * without creating any illegal irq-safe -> irq-unsafe lock dependency.
1201  */
1202
1203
1204 #define   BFS_PROCESS_RET(ret)  do { \
1205                                         if (ret < 0) \
1206                                                 return print_bfs_bug(ret); \
1207                                         if (ret == 1) \
1208                                                 return 1; \
1209                                 } while (0)
1210
1211 static inline int usage_match(struct lock_list *entry, void *bit)
1212 {
1213         return entry->class->usage_mask & (1 << (enum lock_usage_bit)bit);
1214 }
1215
1216
1217
1218 /*
1219  * Find a node in the forwards-direction dependency sub-graph starting
1220  * at @root->class that matches @bit.
1221  *
1222  * Return 0 if such a node exists in the subgraph, and put that node
1223  * into *@target_entry.
1224  *
1225  * Return 1 otherwise and keep *@target_entry unchanged.
1226  * Return <0 on error.
1227  */
1228 static int
1229 find_usage_forwards(struct lock_list *root, enum lock_usage_bit bit,
1230                         struct lock_list **target_entry)
1231 {
1232         int result;
1233
1234         debug_atomic_inc(&nr_find_usage_forwards_checks);
1235
1236         result = __bfs_forwards(root, (void *)bit, usage_match, target_entry);
1237
1238         return result;
1239 }
1240
1241 /*
1242  * Find a node in the backwards-direction dependency sub-graph starting
1243  * at @root->class that matches @bit.
1244  *
1245  * Return 0 if such a node exists in the subgraph, and put that node
1246  * into *@target_entry.
1247  *
1248  * Return 1 otherwise and keep *@target_entry unchanged.
1249  * Return <0 on error.
1250  */
1251 static int
1252 find_usage_backwards(struct lock_list *root, enum lock_usage_bit bit,
1253                         struct lock_list **target_entry)
1254 {
1255         int result;
1256
1257         debug_atomic_inc(&nr_find_usage_backwards_checks);
1258
1259         result = __bfs_backwards(root, (void *)bit, usage_match, target_entry);
1260
1261         return result;
1262 }
1263
1264
1265 static int
1266 print_bad_irq_dependency(struct task_struct *curr,
1267                          struct lock_list *prev_root,
1268                          struct lock_list *next_root,
1269                          struct lock_list *backwards_entry,
1270                          struct lock_list *forwards_entry,
1271                          struct held_lock *prev,
1272                          struct held_lock *next,
1273                          enum lock_usage_bit bit1,
1274                          enum lock_usage_bit bit2,
1275                          const char *irqclass)
1276 {
1277         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1278                 return 0;
1279
1280         printk("\n======================================================\n");
1281         printk(  "[ INFO: %s-safe -> %s-unsafe lock order detected ]\n",
1282                 irqclass, irqclass);
1283         print_kernel_version();
1284         printk(  "------------------------------------------------------\n");
1285         printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] is trying to acquire:\n",
1286                 curr->comm, task_pid_nr(curr),
1287                 curr->hardirq_context, hardirq_count() >> HARDIRQ_SHIFT,
1288                 curr->softirq_context, softirq_count() >> SOFTIRQ_SHIFT,
1289                 curr->hardirqs_enabled,
1290                 curr->softirqs_enabled);
1291         print_lock(next);
1292
1293         printk("\nand this task is already holding:\n");
1294         print_lock(prev);
1295         printk("which would create a new lock dependency:\n");
1296         print_lock_name(hlock_class(prev));
1297         printk(" ->");
1298         print_lock_name(hlock_class(next));
1299         printk("\n");
1300
1301         printk("\nbut this new dependency connects a %s-irq-safe lock:\n",
1302                 irqclass);
1303         print_lock_name(backwards_entry->class);
1304         printk("\n... which became %s-irq-safe at:\n", irqclass);
1305
1306         print_stack_trace(backwards_entry->class->usage_traces + bit1, 1);
1307
1308         printk("\nto a %s-irq-unsafe lock:\n", irqclass);
1309         print_lock_name(forwards_entry->class);
1310         printk("\n... which became %s-irq-unsafe at:\n", irqclass);
1311         printk("...");
1312
1313         print_stack_trace(forwards_entry->class->usage_traces + bit2, 1);
1314
1315         printk("\nother info that might help us debug this:\n\n");
1316         lockdep_print_held_locks(curr);
1317
1318         printk("\nthe dependencies between %s-irq-safe lock", irqclass);
1319         printk(" and the holding lock:\n");
1320         if (!save_trace(&prev_root->trace))
1321                 return 0;
1322         print_shortest_lock_dependencies(backwards_entry, prev_root);
1323
1324         printk("\nthe dependencies between the lock to be acquired");
1325         printk(" and %s-irq-unsafe lock:\n", irqclass);
1326         if (!save_trace(&next_root->trace))
1327                 return 0;
1328         print_shortest_lock_dependencies(forwards_entry, next_root);
1329
1330         printk("\nstack backtrace:\n");
1331         dump_stack();
1332
1333         return 0;
1334 }
1335
1336 static int
1337 check_usage(struct task_struct *curr, struct held_lock *prev,
1338             struct held_lock *next, enum lock_usage_bit bit_backwards,
1339             enum lock_usage_bit bit_forwards, const char *irqclass)
1340 {
1341         int ret;
1342         struct lock_list this, that;
1343         struct lock_list *uninitialized_var(target_entry);
1344         struct lock_list *uninitialized_var(target_entry1);
1345
1346         this.parent = NULL;
1347
1348         this.class = hlock_class(prev);
1349         ret = find_usage_backwards(&this, bit_backwards, &target_entry);
1350         BFS_PROCESS_RET(ret);
1351
1352         that.parent = NULL;
1353         that.class = hlock_class(next);
1354         ret = find_usage_forwards(&that, bit_forwards, &target_entry1);
1355         BFS_PROCESS_RET(ret);
1356
1357         return print_bad_irq_dependency(curr, &this, &that,
1358                         target_entry, target_entry1,
1359                         prev, next,
1360                         bit_backwards, bit_forwards, irqclass);
1361 }
1362
1363 static const char *state_names[] = {
1364 #define LOCKDEP_STATE(__STATE) \
1365         __stringify(__STATE),
1366 #include "lockdep_states.h"
1367 #undef LOCKDEP_STATE
1368 };
1369
1370 static const char *state_rnames[] = {
1371 #define LOCKDEP_STATE(__STATE) \
1372         __stringify(__STATE)"-READ",
1373 #include "lockdep_states.h"
1374 #undef LOCKDEP_STATE
1375 };
1376
1377 static inline const char *state_name(enum lock_usage_bit bit)
1378 {
1379         return (bit & 1) ? state_rnames[bit >> 2] : state_names[bit >> 2];
1380 }
1381
1382 static int exclusive_bit(int new_bit)
1383 {
1384         /*
1385          * USED_IN
1386          * USED_IN_READ
1387          * ENABLED
1388          * ENABLED_READ
1389          *
1390          * bit 0 - write/read
1391          * bit 1 - used_in/enabled
1392          * bit 2+  state
1393          */
1394
1395         int state = new_bit & ~3;
1396         int dir = new_bit & 2;
1397
1398         /*
1399          * keep state, bit flip the direction and strip read.
1400          */
1401         return state | (dir ^ 2);
1402 }
1403
1404 static int check_irq_usage(struct task_struct *curr, struct held_lock *prev,
1405                            struct held_lock *next, enum lock_usage_bit bit)
1406 {
1407         /*
1408          * Prove that the new dependency does not connect a hardirq-safe
1409          * lock with a hardirq-unsafe lock - to achieve this we search
1410          * the backwards-subgraph starting at <prev>, and the
1411          * forwards-subgraph starting at <next>:
1412          */
1413         if (!check_usage(curr, prev, next, bit,
1414                            exclusive_bit(bit), state_name(bit)))
1415                 return 0;
1416
1417         bit++; /* _READ */
1418
1419         /*
1420          * Prove that the new dependency does not connect a hardirq-safe-read
1421          * lock with a hardirq-unsafe lock - to achieve this we search
1422          * the backwards-subgraph starting at <prev>, and the
1423          * forwards-subgraph starting at <next>:
1424          */
1425         if (!check_usage(curr, prev, next, bit,
1426                            exclusive_bit(bit), state_name(bit)))
1427                 return 0;
1428
1429         return 1;
1430 }
1431
1432 static int
1433 check_prev_add_irq(struct task_struct *curr, struct held_lock *prev,
1434                 struct held_lock *next)
1435 {
1436 #define LOCKDEP_STATE(__STATE)                                          \
1437         if (!check_irq_usage(curr, prev, next, LOCK_USED_IN_##__STATE)) \
1438                 return 0;
1439 #include "lockdep_states.h"
1440 #undef LOCKDEP_STATE
1441
1442         return 1;
1443 }
1444
1445 static void inc_chains(void)
1446 {
1447         if (current->hardirq_context)
1448                 nr_hardirq_chains++;
1449         else {
1450                 if (current->softirq_context)
1451                         nr_softirq_chains++;
1452                 else
1453                         nr_process_chains++;
1454         }
1455 }
1456
1457 #else
1458
1459 static inline int
1460 check_prev_add_irq(struct task_struct *curr, struct held_lock *prev,
1461                 struct held_lock *next)
1462 {
1463         return 1;
1464 }
1465
1466 static inline void inc_chains(void)
1467 {
1468         nr_process_chains++;
1469 }
1470
1471 #endif
1472
1473 static int
1474 print_deadlock_bug(struct task_struct *curr, struct held_lock *prev,
1475                    struct held_lock *next)
1476 {
1477         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1478                 return 0;
1479
1480         printk("\n=============================================\n");
1481         printk(  "[ INFO: possible recursive locking detected ]\n");
1482         print_kernel_version();
1483         printk(  "---------------------------------------------\n");
1484         printk("%s/%d is trying to acquire lock:\n",
1485                 curr->comm, task_pid_nr(curr));
1486         print_lock(next);
1487         printk("\nbut task is already holding lock:\n");
1488         print_lock(prev);
1489
1490         printk("\nother info that might help us debug this:\n");
1491         lockdep_print_held_locks(curr);
1492
1493         printk("\nstack backtrace:\n");
1494         dump_stack();
1495
1496         return 0;
1497 }
1498
1499 /*
1500  * Check whether we are holding such a class already.
1501  *
1502  * (Note that this has to be done separately, because the graph cannot
1503  * detect such classes of deadlocks.)
1504  *
1505  * Returns: 0 on deadlock detected, 1 on OK, 2 on recursive read
1506  */
1507 static int
1508 check_deadlock(struct task_struct *curr, struct held_lock *next,
1509                struct lockdep_map *next_instance, int read)
1510 {
1511         struct held_lock *prev;
1512         struct held_lock *nest = NULL;
1513         int i;
1514
1515         for (i = 0; i < curr->lockdep_depth; i++) {
1516                 prev = curr->held_locks + i;
1517
1518                 if (prev->instance == next->nest_lock)
1519                         nest = prev;
1520
1521                 if (hlock_class(prev) != hlock_class(next))
1522                         continue;
1523
1524                 /*
1525                  * Allow read-after-read recursion of the same
1526                  * lock class (i.e. read_lock(lock)+read_lock(lock)):
1527                  */
1528                 if ((read == 2) && prev->read)
1529                         return 2;
1530
1531                 /*
1532                  * We're holding the nest_lock, which serializes this lock's
1533                  * nesting behaviour.
1534                  */
1535                 if (nest)
1536                         return 2;
1537
1538                 return print_deadlock_bug(curr, prev, next);
1539         }
1540         return 1;
1541 }
1542
1543 /*
1544  * There was a chain-cache miss, and we are about to add a new dependency
1545  * to a previous lock. We recursively validate the following rules:
1546  *
1547  *  - would the adding of the <prev> -> <next> dependency create a
1548  *    circular dependency in the graph? [== circular deadlock]
1549  *
1550  *  - does the new prev->next dependency connect any hardirq-safe lock
1551  *    (in the full backwards-subgraph starting at <prev>) with any
1552  *    hardirq-unsafe lock (in the full forwards-subgraph starting at
1553  *    <next>)? [== illegal lock inversion with hardirq contexts]
1554  *
1555  *  - does the new prev->next dependency connect any softirq-safe lock
1556  *    (in the full backwards-subgraph starting at <prev>) with any
1557  *    softirq-unsafe lock (in the full forwards-subgraph starting at
1558  *    <next>)? [== illegal lock inversion with softirq contexts]
1559  *
1560  * any of these scenarios could lead to a deadlock.
1561  *
1562  * Then if all the validations pass, we add the forwards and backwards
1563  * dependency.
1564  */
1565 static int
1566 check_prev_add(struct task_struct *curr, struct held_lock *prev,
1567                struct held_lock *next, int distance)
1568 {
1569         struct lock_list *entry;
1570         int ret;
1571         struct lock_list this;
1572         struct lock_list *uninitialized_var(target_entry);
1573
1574         /*
1575          * Prove that the new <prev> -> <next> dependency would not
1576          * create a circular dependency in the graph. (We do this by
1577          * forward-recursing into the graph starting at <next>, and
1578          * checking whether we can reach <prev>.)
1579          *
1580          * We are using global variables to control the recursion, to
1581          * keep the stackframe size of the recursive functions low:
1582          */
1583         this.class = hlock_class(next);
1584         this.parent = NULL;
1585         ret = check_noncircular(&this, hlock_class(prev), &target_entry);
1586         if (unlikely(!ret))
1587                 return print_circular_bug(&this, target_entry, next, prev);
1588         else if (unlikely(ret < 0))
1589                 return print_bfs_bug(ret);
1590
1591         if (!check_prev_add_irq(curr, prev, next))
1592                 return 0;
1593
1594         /*
1595          * For recursive read-locks we do all the dependency checks,
1596          * but we dont store read-triggered dependencies (only
1597          * write-triggered dependencies). This ensures that only the
1598          * write-side dependencies matter, and that if for example a
1599          * write-lock never takes any other locks, then the reads are
1600          * equivalent to a NOP.
1601          */
1602         if (next->read == 2 || prev->read == 2)
1603                 return 1;
1604         /*
1605          * Is the <prev> -> <next> dependency already present?
1606          *
1607          * (this may occur even though this is a new chain: consider
1608          *  e.g. the L1 -> L2 -> L3 -> L4 and the L5 -> L1 -> L2 -> L3
1609          *  chains - the second one will be new, but L1 already has
1610          *  L2 added to its dependency list, due to the first chain.)
1611          */
1612         list_for_each_entry(entry, &hlock_class(prev)->locks_after, entry) {
1613                 if (entry->class == hlock_class(next)) {
1614                         if (distance == 1)
1615                                 entry->distance = 1;
1616                         return 2;
1617                 }
1618         }
1619
1620         /*
1621          * Ok, all validations passed, add the new lock
1622          * to the previous lock's dependency list:
1623          */
1624         ret = add_lock_to_list(hlock_class(prev), hlock_class(next),
1625                                &hlock_class(prev)->locks_after,
1626                                next->acquire_ip, distance);
1627
1628         if (!ret)
1629                 return 0;
1630
1631         ret = add_lock_to_list(hlock_class(next), hlock_class(prev),
1632                                &hlock_class(next)->locks_before,
1633                                next->acquire_ip, distance);
1634         if (!ret)
1635                 return 0;
1636
1637         /*
1638          * Debugging printouts:
1639          */
1640         if (verbose(hlock_class(prev)) || verbose(hlock_class(next))) {
1641                 graph_unlock();
1642                 printk("\n new dependency: ");
1643                 print_lock_name(hlock_class(prev));
1644                 printk(" => ");
1645                 print_lock_name(hlock_class(next));
1646                 printk("\n");
1647                 dump_stack();
1648                 return graph_lock();
1649         }
1650         return 1;
1651 }
1652
1653 /*
1654  * Add the dependency to all directly-previous locks that are 'relevant'.
1655  * The ones that are relevant are (in increasing distance from curr):
1656  * all consecutive trylock entries and the final non-trylock entry - or
1657  * the end of this context's lock-chain - whichever comes first.
1658  */
1659 static int
1660 check_prevs_add(struct task_struct *curr, struct held_lock *next)
1661 {
1662         int depth = curr->lockdep_depth;
1663         struct held_lock *hlock;
1664
1665         /*
1666          * Debugging checks.
1667          *
1668          * Depth must not be zero for a non-head lock:
1669          */
1670         if (!depth)
1671                 goto out_bug;
1672         /*
1673          * At least two relevant locks must exist for this
1674          * to be a head:
1675          */
1676         if (curr->held_locks[depth].irq_context !=
1677                         curr->held_locks[depth-1].irq_context)
1678                 goto out_bug;
1679
1680         for (;;) {
1681                 int distance = curr->lockdep_depth - depth + 1;
1682                 hlock = curr->held_locks + depth-1;
1683                 /*
1684                  * Only non-recursive-read entries get new dependencies
1685                  * added:
1686                  */
1687                 if (hlock->read != 2) {
1688                         if (!check_prev_add(curr, hlock, next, distance))
1689                                 return 0;
1690                         /*
1691                          * Stop after the first non-trylock entry,
1692                          * as non-trylock entries have added their
1693                          * own direct dependencies already, so this
1694                          * lock is connected to them indirectly:
1695                          */
1696                         if (!hlock->trylock)
1697                                 break;
1698                 }
1699                 depth--;
1700                 /*
1701                  * End of lock-stack?
1702                  */
1703                 if (!depth)
1704                         break;
1705                 /*
1706                  * Stop the search if we cross into another context:
1707                  */
1708                 if (curr->held_locks[depth].irq_context !=
1709                                 curr->held_locks[depth-1].irq_context)
1710                         break;
1711         }
1712         return 1;
1713 out_bug:
1714         if (!debug_locks_off_graph_unlock())
1715                 return 0;
1716
1717         WARN_ON(1);
1718
1719         return 0;
1720 }
1721
1722 unsigned long nr_lock_chains;
1723 struct lock_chain lock_chains[MAX_LOCKDEP_CHAINS];
1724 int nr_chain_hlocks;
1725 static u16 chain_hlocks[MAX_LOCKDEP_CHAIN_HLOCKS];
1726
1727 struct lock_class *lock_chain_get_class(struct lock_chain *chain, int i)
1728 {
1729         return lock_classes + chain_hlocks[chain->base + i];
1730 }
1731
1732 /*
1733  * Look up a dependency chain. If the key is not present yet then
1734  * add it and return 1 - in this case the new dependency chain is
1735  * validated. If the key is already hashed, return 0.
1736  * (On return with 1 graph_lock is held.)
1737  */
1738 static inline int lookup_chain_cache(struct task_struct *curr,
1739                                      struct held_lock *hlock,
1740                                      u64 chain_key)
1741 {
1742         struct lock_class *class = hlock_class(hlock);
1743         struct list_head *hash_head = chainhashentry(chain_key);
1744         struct lock_chain *chain;
1745         struct held_lock *hlock_curr, *hlock_next;
1746         int i, j, n, cn;
1747
1748         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
1749                 return 0;
1750         /*
1751          * We can walk it lock-free, because entries only get added
1752          * to the hash:
1753          */
1754         list_for_each_entry(chain, hash_head, entry) {
1755                 if (chain->chain_key == chain_key) {
1756 cache_hit:
1757                         debug_atomic_inc(&chain_lookup_hits);
1758                         if (very_verbose(class))
1759                                 printk("\nhash chain already cached, key: "
1760                                         "%016Lx tail class: [%p] %s\n",
1761                                         (unsigned long long)chain_key,
1762                                         class->key, class->name);
1763                         return 0;
1764                 }
1765         }
1766         if (very_verbose(class))
1767                 printk("\nnew hash chain, key: %016Lx tail class: [%p] %s\n",
1768                         (unsigned long long)chain_key, class->key, class->name);
1769         /*
1770          * Allocate a new chain entry from the static array, and add
1771          * it to the hash:
1772          */
1773         if (!graph_lock())
1774                 return 0;
1775         /*
1776          * We have to walk the chain again locked - to avoid duplicates:
1777          */
1778         list_for_each_entry(chain, hash_head, entry) {
1779                 if (chain->chain_key == chain_key) {
1780                         graph_unlock();
1781                         goto cache_hit;
1782                 }
1783         }
1784         if (unlikely(nr_lock_chains >= MAX_LOCKDEP_CHAINS)) {
1785                 if (!debug_locks_off_graph_unlock())
1786                         return 0;
1787
1788                 printk("BUG: MAX_LOCKDEP_CHAINS too low!\n");
1789                 printk("turning off the locking correctness validator.\n");
1790                 dump_stack();
1791                 return 0;
1792         }
1793         chain = lock_chains + nr_lock_chains++;
1794         chain->chain_key = chain_key;
1795         chain->irq_context = hlock->irq_context;
1796         /* Find the first held_lock of current chain */
1797         hlock_next = hlock;
1798         for (i = curr->lockdep_depth - 1; i >= 0; i--) {
1799                 hlock_curr = curr->held_locks + i;
1800                 if (hlock_curr->irq_context != hlock_next->irq_context)
1801                         break;
1802                 hlock_next = hlock;
1803         }
1804         i++;
1805         chain->depth = curr->lockdep_depth + 1 - i;
1806         cn = nr_chain_hlocks;
1807         while (cn + chain->depth <= MAX_LOCKDEP_CHAIN_HLOCKS) {
1808                 n = cmpxchg(&nr_chain_hlocks, cn, cn + chain->depth);
1809                 if (n == cn)
1810                         break;
1811                 cn = n;
1812         }
1813         if (likely(cn + chain->depth <= MAX_LOCKDEP_CHAIN_HLOCKS)) {
1814                 chain->base = cn;
1815                 for (j = 0; j < chain->depth - 1; j++, i++) {
1816                         int lock_id = curr->held_locks[i].class_idx - 1;
1817                         chain_hlocks[chain->base + j] = lock_id;
1818                 }
1819                 chain_hlocks[chain->base + j] = class - lock_classes;
1820         }
1821         list_add_tail_rcu(&chain->entry, hash_head);
1822         debug_atomic_inc(&chain_lookup_misses);
1823         inc_chains();
1824
1825         return 1;
1826 }
1827
1828 static int validate_chain(struct task_struct *curr, struct lockdep_map *lock,
1829                 struct held_lock *hlock, int chain_head, u64 chain_key)
1830 {
1831         /*
1832          * Trylock needs to maintain the stack of held locks, but it
1833          * does not add new dependencies, because trylock can be done
1834          * in any order.
1835          *
1836          * We look up the chain_key and do the O(N^2) check and update of
1837          * the dependencies only if this is a new dependency chain.
1838          * (If lookup_chain_cache() returns with 1 it acquires
1839          * graph_lock for us)
1840          */
1841         if (!hlock->trylock && (hlock->check == 2) &&
1842             lookup_chain_cache(curr, hlock, chain_key)) {
1843                 /*
1844                  * Check whether last held lock:
1845                  *
1846                  * - is irq-safe, if this lock is irq-unsafe
1847                  * - is softirq-safe, if this lock is hardirq-unsafe
1848                  *
1849                  * And check whether the new lock's dependency graph
1850                  * could lead back to the previous lock.
1851                  *
1852                  * any of these scenarios could lead to a deadlock. If
1853                  * All validations
1854                  */
1855                 int ret = check_deadlock(curr, hlock, lock, hlock->read);
1856
1857                 if (!ret)
1858                         return 0;
1859                 /*
1860                  * Mark recursive read, as we jump over it when
1861                  * building dependencies (just like we jump over
1862                  * trylock entries):
1863                  */
1864                 if (ret == 2)
1865                         hlock->read = 2;
1866                 /*
1867                  * Add dependency only if this lock is not the head
1868                  * of the chain, and if it's not a secondary read-lock:
1869                  */
1870                 if (!chain_head && ret != 2)
1871                         if (!check_prevs_add(curr, hlock))
1872                                 return 0;
1873                 graph_unlock();
1874         } else
1875                 /* after lookup_chain_cache(): */
1876                 if (unlikely(!debug_locks))
1877                         return 0;
1878
1879         return 1;
1880 }
1881 #else
1882 static inline int validate_chain(struct task_struct *curr,
1883                 struct lockdep_map *lock, struct held_lock *hlock,
1884                 int chain_head, u64 chain_key)
1885 {
1886         return 1;
1887 }
1888 #endif
1889
1890 /*
1891  * We are building curr_chain_key incrementally, so double-check
1892  * it from scratch, to make sure that it's done correctly:
1893  */
1894 static void check_chain_key(struct task_struct *curr)
1895 {
1896 #ifdef CONFIG_DEBUG_LOCKDEP
1897         struct held_lock *hlock, *prev_hlock = NULL;
1898         unsigned int i, id;
1899         u64 chain_key = 0;
1900
1901         for (i = 0; i < curr->lockdep_depth; i++) {
1902                 hlock = curr->held_locks + i;
1903                 if (chain_key != hlock->prev_chain_key) {
1904                         debug_locks_off();
1905                         WARN(1, "hm#1, depth: %u [%u], %016Lx != %016Lx\n",
1906                                 curr->lockdep_depth, i,
1907                                 (unsigned long long)chain_key,
1908                                 (unsigned long long)hlock->prev_chain_key);
1909                         return;
1910                 }
1911                 id = hlock->class_idx - 1;
1912                 if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
1913                         return;
1914
1915                 if (prev_hlock && (prev_hlock->irq_context !=
1916                                                         hlock->irq_context))
1917                         chain_key = 0;
1918                 chain_key = iterate_chain_key(chain_key, id);
1919                 prev_hlock = hlock;
1920         }
1921         if (chain_key != curr->curr_chain_key) {
1922                 debug_locks_off();
1923                 WARN(1, "hm#2, depth: %u [%u], %016Lx != %016Lx\n",
1924                         curr->lockdep_depth, i,
1925                         (unsigned long long)chain_key,
1926                         (unsigned long long)curr->curr_chain_key);
1927         }
1928 #endif
1929 }
1930
1931 static int
1932 print_usage_bug(struct task_struct *curr, struct held_lock *this,
1933                 enum lock_usage_bit prev_bit, enum lock_usage_bit new_bit)
1934 {
1935         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1936                 return 0;
1937
1938         printk("\n=================================\n");
1939         printk(  "[ INFO: inconsistent lock state ]\n");
1940         print_kernel_version();
1941         printk(  "---------------------------------\n");
1942
1943         printk("inconsistent {%s} -> {%s} usage.\n",
1944                 usage_str[prev_bit], usage_str[new_bit]);
1945
1946         printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] takes:\n",
1947                 curr->comm, task_pid_nr(curr),
1948                 trace_hardirq_context(curr), hardirq_count() >> HARDIRQ_SHIFT,
1949                 trace_softirq_context(curr), softirq_count() >> SOFTIRQ_SHIFT,
1950                 trace_hardirqs_enabled(curr),
1951                 trace_softirqs_enabled(curr));
1952         print_lock(this);
1953
1954         printk("{%s} state was registered at:\n", usage_str[prev_bit]);
1955         print_stack_trace(hlock_class(this)->usage_traces + prev_bit, 1);
1956
1957         print_irqtrace_events(curr);
1958         printk("\nother info that might help us debug this:\n");
1959         lockdep_print_held_locks(curr);
1960
1961         printk("\nstack backtrace:\n");
1962         dump_stack();
1963
1964         return 0;
1965 }
1966
1967 /*
1968  * Print out an error if an invalid bit is set:
1969  */
1970 static inline int
1971 valid_state(struct task_struct *curr, struct held_lock *this,
1972             enum lock_usage_bit new_bit, enum lock_usage_bit bad_bit)
1973 {
1974         if (unlikely(hlock_class(this)->usage_mask & (1 << bad_bit)))
1975                 return print_usage_bug(curr, this, bad_bit, new_bit);
1976         return 1;
1977 }
1978
1979 static int mark_lock(struct task_struct *curr, struct held_lock *this,
1980                      enum lock_usage_bit new_bit);
1981
1982 #if defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING)
1983
1984 /*
1985  * print irq inversion bug:
1986  */
1987 static int
1988 print_irq_inversion_bug(struct task_struct *curr,
1989                         struct lock_list *root, struct lock_list *other,
1990                         struct held_lock *this, int forwards,
1991                         const char *irqclass)
1992 {
1993         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1994                 return 0;
1995
1996         printk("\n=========================================================\n");
1997         printk(  "[ INFO: possible irq lock inversion dependency detected ]\n");
1998         print_kernel_version();
1999         printk(  "---------------------------------------------------------\n");
2000         printk("%s/%d just changed the state of lock:\n",
2001                 curr->comm, task_pid_nr(curr));
2002         print_lock(this);
2003         if (forwards)
2004                 printk("but this lock took another, %s-unsafe lock in the past:\n", irqclass);
2005         else
2006                 printk("but this lock was taken by another, %s-safe lock in the past:\n", irqclass);
2007         print_lock_name(other->class);
2008         printk("\n\nand interrupts could create inverse lock ordering between them.\n\n");
2009
2010         printk("\nother info that might help us debug this:\n");
2011         lockdep_print_held_locks(curr);
2012
2013         printk("\nthe shortest dependencies between 2nd lock and 1st lock:\n");
2014         if (!save_trace(&root->trace))
2015                 return 0;
2016         print_shortest_lock_dependencies(other, root);
2017
2018         printk("\nstack backtrace:\n");
2019         dump_stack();
2020
2021         return 0;
2022 }
2023
2024 /*
2025  * Prove that in the forwards-direction subgraph starting at <this>
2026  * there is no lock matching <mask>:
2027  */
2028 static int
2029 check_usage_forwards(struct task_struct *curr, struct held_lock *this,
2030                      enum lock_usage_bit bit, const char *irqclass)
2031 {
2032         int ret;
2033         struct lock_list root;
2034         struct lock_list *uninitialized_var(target_entry);
2035
2036         root.parent = NULL;
2037         root.class = hlock_class(this);
2038         ret = find_usage_forwards(&root, bit, &target_entry);
2039         BFS_PROCESS_RET(ret);
2040
2041         return print_irq_inversion_bug(curr, &root, target_entry,
2042                                         this, 1, irqclass);
2043 }
2044
2045 /*
2046  * Prove that in the backwards-direction subgraph starting at <this>
2047  * there is no lock matching <mask>:
2048  */
2049 static int
2050 check_usage_backwards(struct task_struct *curr, struct held_lock *this,
2051                       enum lock_usage_bit bit, const char *irqclass)
2052 {
2053         int ret;
2054         struct lock_list root;
2055         struct lock_list *uninitialized_var(target_entry);
2056
2057         root.parent = NULL;
2058         root.class = hlock_class(this);
2059         ret = find_usage_backwards(&root, bit, &target_entry);
2060         BFS_PROCESS_RET(ret);
2061
2062         return print_irq_inversion_bug(curr, &root, target_entry,
2063                                         this, 1, irqclass);
2064 }
2065
2066 void print_irqtrace_events(struct task_struct *curr)
2067 {
2068         printk("irq event stamp: %u\n", curr->irq_events);
2069         printk("hardirqs last  enabled at (%u): ", curr->hardirq_enable_event);
2070         print_ip_sym(curr->hardirq_enable_ip);
2071         printk("hardirqs last disabled at (%u): ", curr->hardirq_disable_event);
2072         print_ip_sym(curr->hardirq_disable_ip);
2073         printk("softirqs last  enabled at (%u): ", curr->softirq_enable_event);
2074         print_ip_sym(curr->softirq_enable_ip);
2075         printk("softirqs last disabled at (%u): ", curr->softirq_disable_event);
2076         print_ip_sym(curr->softirq_disable_ip);
2077 }
2078
2079 static int HARDIRQ_verbose(struct lock_class *class)
2080 {
2081 #if HARDIRQ_VERBOSE
2082         return class_filter(class);
2083 #endif
2084         return 0;
2085 }
2086
2087 static int SOFTIRQ_verbose(struct lock_class *class)
2088 {
2089 #if SOFTIRQ_VERBOSE
2090         return class_filter(class);
2091 #endif
2092         return 0;
2093 }
2094
2095 static int RECLAIM_FS_verbose(struct lock_class *class)
2096 {
2097 #if RECLAIM_VERBOSE
2098         return class_filter(class);
2099 #endif
2100         return 0;
2101 }
2102
2103 #define STRICT_READ_CHECKS      1
2104
2105 static int (*state_verbose_f[])(struct lock_class *class) = {
2106 #define LOCKDEP_STATE(__STATE) \
2107         __STATE##_verbose,
2108 #include "lockdep_states.h"
2109 #undef LOCKDEP_STATE
2110 };
2111
2112 static inline int state_verbose(enum lock_usage_bit bit,
2113                                 struct lock_class *class)
2114 {
2115         return state_verbose_f[bit >> 2](class);
2116 }
2117
2118 typedef int (*check_usage_f)(struct task_struct *, struct held_lock *,
2119                              enum lock_usage_bit bit, const char *name);
2120
2121 static int
2122 mark_lock_irq(struct task_struct *curr, struct held_lock *this,
2123                 enum lock_usage_bit new_bit)
2124 {
2125         int excl_bit = exclusive_bit(new_bit);
2126         int read = new_bit & 1;
2127         int dir = new_bit & 2;
2128
2129         /*
2130          * mark USED_IN has to look forwards -- to ensure no dependency
2131          * has ENABLED state, which would allow recursion deadlocks.
2132          *
2133          * mark ENABLED has to look backwards -- to ensure no dependee
2134          * has USED_IN state, which, again, would allow  recursion deadlocks.
2135          */
2136         check_usage_f usage = dir ?
2137                 check_usage_backwards : check_usage_forwards;
2138
2139         /*
2140          * Validate that this particular lock does not have conflicting
2141          * usage states.
2142          */
2143         if (!valid_state(curr, this, new_bit, excl_bit))
2144                 return 0;
2145
2146         /*
2147          * Validate that the lock dependencies don't have conflicting usage
2148          * states.
2149          */
2150         if ((!read || !dir || STRICT_READ_CHECKS) &&
2151                         !usage(curr, this, excl_bit, state_name(new_bit & ~1)))
2152                 return 0;
2153
2154         /*
2155          * Check for read in write conflicts
2156          */
2157         if (!read) {
2158                 if (!valid_state(curr, this, new_bit, excl_bit + 1))
2159                         return 0;
2160
2161                 if (STRICT_READ_CHECKS &&
2162                         !usage(curr, this, excl_bit + 1,
2163                                 state_name(new_bit + 1)))
2164                         return 0;
2165         }
2166
2167         if (state_verbose(new_bit, hlock_class(this)))
2168                 return 2;
2169
2170         return 1;
2171 }
2172
2173 enum mark_type {
2174 #define LOCKDEP_STATE(__STATE)  __STATE,
2175 #include "lockdep_states.h"
2176 #undef LOCKDEP_STATE
2177 };
2178
2179 /*
2180  * Mark all held locks with a usage bit:
2181  */
2182 static int
2183 mark_held_locks(struct task_struct *curr, enum mark_type mark)
2184 {
2185         enum lock_usage_bit usage_bit;
2186         struct held_lock *hlock;
2187         int i;
2188
2189         for (i = 0; i < curr->lockdep_depth; i++) {
2190                 hlock = curr->held_locks + i;
2191
2192                 usage_bit = 2 + (mark << 2); /* ENABLED */
2193                 if (hlock->read)
2194                         usage_bit += 1; /* READ */
2195
2196                 BUG_ON(usage_bit >= LOCK_USAGE_STATES);
2197
2198                 if (!mark_lock(curr, hlock, usage_bit))
2199                         return 0;
2200         }
2201
2202         return 1;
2203 }
2204
2205 /*
2206  * Debugging helper: via this flag we know that we are in
2207  * 'early bootup code', and will warn about any invalid irqs-on event:
2208  */
2209 static int early_boot_irqs_enabled;
2210
2211 void early_boot_irqs_off(void)
2212 {
2213         early_boot_irqs_enabled = 0;
2214 }
2215
2216 void early_boot_irqs_on(void)
2217 {
2218         early_boot_irqs_enabled = 1;
2219 }
2220
2221 /*
2222  * Hardirqs will be enabled:
2223  */
2224 void trace_hardirqs_on_caller(unsigned long ip)
2225 {
2226         struct task_struct *curr = current;
2227
2228         time_hardirqs_on(CALLER_ADDR0, ip);
2229
2230         if (unlikely(!debug_locks || current->lockdep_recursion))
2231                 return;
2232
2233         if (DEBUG_LOCKS_WARN_ON(unlikely(!early_boot_irqs_enabled)))
2234                 return;
2235
2236         if (unlikely(curr->hardirqs_enabled)) {
2237                 debug_atomic_inc(&redundant_hardirqs_on);
2238                 return;
2239         }
2240         /* we'll do an OFF -> ON transition: */
2241         curr->hardirqs_enabled = 1;
2242
2243         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2244                 return;
2245         if (DEBUG_LOCKS_WARN_ON(current->hardirq_context))
2246                 return;
2247         /*
2248          * We are going to turn hardirqs on, so set the
2249          * usage bit for all held locks:
2250          */
2251         if (!mark_held_locks(curr, HARDIRQ))
2252                 return;
2253         /*
2254          * If we have softirqs enabled, then set the usage
2255          * bit for all held locks. (disabled hardirqs prevented
2256          * this bit from being set before)
2257          */
2258         if (curr->softirqs_enabled)
2259                 if (!mark_held_locks(curr, SOFTIRQ))
2260                         return;
2261
2262         curr->hardirq_enable_ip = ip;
2263         curr->hardirq_enable_event = ++curr->irq_events;
2264         debug_atomic_inc(&hardirqs_on_events);
2265 }
2266 EXPORT_SYMBOL(trace_hardirqs_on_caller);
2267
2268 void trace_hardirqs_on(void)
2269 {
2270         trace_hardirqs_on_caller(CALLER_ADDR0);
2271 }
2272 EXPORT_SYMBOL(trace_hardirqs_on);
2273
2274 /*
2275  * Hardirqs were disabled:
2276  */
2277 void trace_hardirqs_off_caller(unsigned long ip)
2278 {
2279         struct task_struct *curr = current;
2280
2281         time_hardirqs_off(CALLER_ADDR0, ip);
2282
2283         if (unlikely(!debug_locks || current->lockdep_recursion))
2284                 return;
2285
2286         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2287                 return;
2288
2289         if (curr->hardirqs_enabled) {
2290                 /*
2291                  * We have done an ON -> OFF transition:
2292                  */
2293                 curr->hardirqs_enabled = 0;
2294                 curr->hardirq_disable_ip = ip;
2295                 curr->hardirq_disable_event = ++curr->irq_events;
2296                 debug_atomic_inc(&hardirqs_off_events);
2297         } else
2298                 debug_atomic_inc(&redundant_hardirqs_off);
2299 }
2300 EXPORT_SYMBOL(trace_hardirqs_off_caller);
2301
2302 void trace_hardirqs_off(void)
2303 {
2304         trace_hardirqs_off_caller(CALLER_ADDR0);
2305 }
2306 EXPORT_SYMBOL(trace_hardirqs_off);
2307
2308 /*
2309  * Softirqs will be enabled:
2310  */
2311 void trace_softirqs_on(unsigned long ip)
2312 {
2313         struct task_struct *curr = current;
2314
2315         if (unlikely(!debug_locks))
2316                 return;
2317
2318         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2319                 return;
2320
2321         if (curr->softirqs_enabled) {
2322                 debug_atomic_inc(&redundant_softirqs_on);
2323                 return;
2324         }
2325
2326         /*
2327          * We'll do an OFF -> ON transition:
2328          */
2329         curr->softirqs_enabled = 1;
2330         curr->softirq_enable_ip = ip;
2331         curr->softirq_enable_event = ++curr->irq_events;
2332         debug_atomic_inc(&softirqs_on_events);
2333         /*
2334          * We are going to turn softirqs on, so set the
2335          * usage bit for all held locks, if hardirqs are
2336          * enabled too:
2337          */
2338         if (curr->hardirqs_enabled)
2339                 mark_held_locks(curr, SOFTIRQ);
2340 }
2341
2342 /*
2343  * Softirqs were disabled:
2344  */
2345 void trace_softirqs_off(unsigned long ip)
2346 {
2347         struct task_struct *curr = current;
2348
2349         if (unlikely(!debug_locks))
2350                 return;
2351
2352         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2353                 return;
2354
2355         if (curr->softirqs_enabled) {
2356                 /*
2357                  * We have done an ON -> OFF transition:
2358                  */
2359                 curr->softirqs_enabled = 0;
2360                 curr->softirq_disable_ip = ip;
2361                 curr->softirq_disable_event = ++curr->irq_events;
2362                 debug_atomic_inc(&softirqs_off_events);
2363                 DEBUG_LOCKS_WARN_ON(!softirq_count());
2364         } else
2365                 debug_atomic_inc(&redundant_softirqs_off);
2366 }
2367
2368 static void __lockdep_trace_alloc(gfp_t gfp_mask, unsigned long flags)
2369 {
2370         struct task_struct *curr = current;
2371
2372         if (unlikely(!debug_locks))
2373                 return;
2374
2375         /* no reclaim without waiting on it */
2376         if (!(gfp_mask & __GFP_WAIT))
2377                 return;
2378
2379         /* this guy won't enter reclaim */
2380         if ((curr->flags & PF_MEMALLOC) && !(gfp_mask & __GFP_NOMEMALLOC))
2381                 return;
2382
2383         /* We're only interested __GFP_FS allocations for now */
2384         if (!(gfp_mask & __GFP_FS))
2385                 return;
2386
2387         if (DEBUG_LOCKS_WARN_ON(irqs_disabled_flags(flags)))
2388                 return;
2389
2390         mark_held_locks(curr, RECLAIM_FS);
2391 }
2392
2393 static void check_flags(unsigned long flags);
2394
2395 void lockdep_trace_alloc(gfp_t gfp_mask)
2396 {
2397         unsigned long flags;
2398
2399         if (unlikely(current->lockdep_recursion))
2400                 return;
2401
2402         raw_local_irq_save(flags);
2403         check_flags(flags);
2404         current->lockdep_recursion = 1;
2405         __lockdep_trace_alloc(gfp_mask, flags);
2406         current->lockdep_recursion = 0;
2407         raw_local_irq_restore(flags);
2408 }
2409
2410 static int mark_irqflags(struct task_struct *curr, struct held_lock *hlock)
2411 {
2412         /*
2413          * If non-trylock use in a hardirq or softirq context, then
2414          * mark the lock as used in these contexts:
2415          */
2416         if (!hlock->trylock) {
2417                 if (hlock->read) {
2418                         if (curr->hardirq_context)
2419                                 if (!mark_lock(curr, hlock,
2420                                                 LOCK_USED_IN_HARDIRQ_READ))
2421                                         return 0;
2422                         if (curr->softirq_context)
2423                                 if (!mark_lock(curr, hlock,
2424                                                 LOCK_USED_IN_SOFTIRQ_READ))
2425                                         return 0;
2426                 } else {
2427                         if (curr->hardirq_context)
2428                                 if (!mark_lock(curr, hlock, LOCK_USED_IN_HARDIRQ))
2429                                         return 0;
2430                         if (curr->softirq_context)
2431                                 if (!mark_lock(curr, hlock, LOCK_USED_IN_SOFTIRQ))
2432                                         return 0;
2433                 }
2434         }
2435         if (!hlock->hardirqs_off) {
2436                 if (hlock->read) {
2437                         if (!mark_lock(curr, hlock,
2438                                         LOCK_ENABLED_HARDIRQ_READ))
2439                                 return 0;
2440                         if (curr->softirqs_enabled)
2441                                 if (!mark_lock(curr, hlock,
2442                                                 LOCK_ENABLED_SOFTIRQ_READ))
2443                                         return 0;
2444                 } else {
2445                         if (!mark_lock(curr, hlock,
2446                                         LOCK_ENABLED_HARDIRQ))
2447                                 return 0;
2448                         if (curr->softirqs_enabled)
2449                                 if (!mark_lock(curr, hlock,
2450                                                 LOCK_ENABLED_SOFTIRQ))
2451                                         return 0;
2452                 }
2453         }
2454
2455         /*
2456          * We reuse the irq context infrastructure more broadly as a general
2457          * context checking code. This tests GFP_FS recursion (a lock taken
2458          * during reclaim for a GFP_FS allocation is held over a GFP_FS
2459          * allocation).
2460          */
2461         if (!hlock->trylock && (curr->lockdep_reclaim_gfp & __GFP_FS)) {
2462                 if (hlock->read) {
2463                         if (!mark_lock(curr, hlock, LOCK_USED_IN_RECLAIM_FS_READ))
2464                                         return 0;
2465                 } else {
2466                         if (!mark_lock(curr, hlock, LOCK_USED_IN_RECLAIM_FS))
2467                                         return 0;
2468                 }
2469         }
2470
2471         return 1;
2472 }
2473
2474 static int separate_irq_context(struct task_struct *curr,
2475                 struct held_lock *hlock)
2476 {
2477         unsigned int depth = curr->lockdep_depth;
2478
2479         /*
2480          * Keep track of points where we cross into an interrupt context:
2481          */
2482         hlock->irq_context = 2*(curr->hardirq_context ? 1 : 0) +
2483                                 curr->softirq_context;
2484         if (depth) {
2485                 struct held_lock *prev_hlock;
2486
2487                 prev_hlock = curr->held_locks + depth-1;
2488                 /*
2489                  * If we cross into another context, reset the
2490                  * hash key (this also prevents the checking and the
2491                  * adding of the dependency to 'prev'):
2492                  */
2493                 if (prev_hlock->irq_context != hlock->irq_context)
2494                         return 1;
2495         }
2496         return 0;
2497 }
2498
2499 #else
2500
2501 static inline
2502 int mark_lock_irq(struct task_struct *curr, struct held_lock *this,
2503                 enum lock_usage_bit new_bit)
2504 {
2505         WARN_ON(1);
2506         return 1;
2507 }
2508
2509 static inline int mark_irqflags(struct task_struct *curr,
2510                 struct held_lock *hlock)
2511 {
2512         return 1;
2513 }
2514
2515 static inline int separate_irq_context(struct task_struct *curr,
2516                 struct held_lock *hlock)
2517 {
2518         return 0;
2519 }
2520
2521 void lockdep_trace_alloc(gfp_t gfp_mask)
2522 {
2523 }
2524
2525 #endif
2526
2527 /*
2528  * Mark a lock with a usage bit, and validate the state transition:
2529  */
2530 static int mark_lock(struct task_struct *curr, struct held_lock *this,
2531                              enum lock_usage_bit new_bit)
2532 {
2533         unsigned int new_mask = 1 << new_bit, ret = 1;
2534
2535         /*
2536          * If already set then do not dirty the cacheline,
2537          * nor do any checks:
2538          */
2539         if (likely(hlock_class(this)->usage_mask & new_mask))
2540                 return 1;
2541
2542         if (!graph_lock())
2543                 return 0;
2544         /*
2545          * Make sure we didnt race:
2546          */
2547         if (unlikely(hlock_class(this)->usage_mask & new_mask)) {
2548                 graph_unlock();
2549                 return 1;
2550         }
2551
2552         hlock_class(this)->usage_mask |= new_mask;
2553
2554         if (!save_trace(hlock_class(this)->usage_traces + new_bit))
2555                 return 0;
2556
2557         switch (new_bit) {
2558 #define LOCKDEP_STATE(__STATE)                  \
2559         case LOCK_USED_IN_##__STATE:            \
2560         case LOCK_USED_IN_##__STATE##_READ:     \
2561         case LOCK_ENABLED_##__STATE:            \
2562         case LOCK_ENABLED_##__STATE##_READ:
2563 #include "lockdep_states.h"
2564 #undef LOCKDEP_STATE
2565                 ret = mark_lock_irq(curr, this, new_bit);
2566                 if (!ret)
2567                         return 0;
2568                 break;
2569         case LOCK_USED:
2570                 debug_atomic_dec(&nr_unused_locks);
2571                 break;
2572         default:
2573                 if (!debug_locks_off_graph_unlock())
2574                         return 0;
2575                 WARN_ON(1);
2576                 return 0;
2577         }
2578
2579         graph_unlock();
2580
2581         /*
2582          * We must printk outside of the graph_lock:
2583          */
2584         if (ret == 2) {
2585                 printk("\nmarked lock as {%s}:\n", usage_str[new_bit]);
2586                 print_lock(this);
2587                 print_irqtrace_events(curr);
2588                 dump_stack();
2589         }
2590
2591         return ret;
2592 }
2593
2594 /*
2595  * Initialize a lock instance's lock-class mapping info:
2596  */
2597 void lockdep_init_map(struct lockdep_map *lock, const char *name,
2598                       struct lock_class_key *key, int subclass)
2599 {
2600         lock->class_cache = NULL;
2601 #ifdef CONFIG_LOCK_STAT
2602         lock->cpu = raw_smp_processor_id();
2603 #endif
2604
2605         if (DEBUG_LOCKS_WARN_ON(!name)) {
2606                 lock->name = "NULL";
2607                 return;
2608         }
2609
2610         lock->name = name;
2611
2612         if (DEBUG_LOCKS_WARN_ON(!key))
2613                 return;
2614         /*
2615          * Sanity check, the lock-class key must be persistent:
2616          */
2617         if (!static_obj(key)) {
2618                 printk("BUG: key %p not in .data!\n", key);
2619                 DEBUG_LOCKS_WARN_ON(1);
2620                 return;
2621         }
2622         lock->key = key;
2623
2624         if (unlikely(!debug_locks))
2625                 return;
2626
2627         if (subclass)
2628                 register_lock_class(lock, subclass, 1);
2629 }
2630 EXPORT_SYMBOL_GPL(lockdep_init_map);
2631
2632 /*
2633  * This gets called for every mutex_lock*()/spin_lock*() operation.
2634  * We maintain the dependency maps and validate the locking attempt:
2635  */
2636 static int __lock_acquire(struct lockdep_map *lock, unsigned int subclass,
2637                           int trylock, int read, int check, int hardirqs_off,
2638                           struct lockdep_map *nest_lock, unsigned long ip)
2639 {
2640         struct task_struct *curr = current;
2641         struct lock_class *class = NULL;
2642         struct held_lock *hlock;
2643         unsigned int depth, id;
2644         int chain_head = 0;
2645         u64 chain_key;
2646
2647         if (!prove_locking)
2648                 check = 1;
2649
2650         if (unlikely(!debug_locks))
2651                 return 0;
2652
2653         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2654                 return 0;
2655
2656         if (unlikely(subclass >= MAX_LOCKDEP_SUBCLASSES)) {
2657                 debug_locks_off();
2658                 printk("BUG: MAX_LOCKDEP_SUBCLASSES too low!\n");
2659                 printk("turning off the locking correctness validator.\n");
2660                 dump_stack();
2661                 return 0;
2662         }
2663
2664         if (!subclass)
2665                 class = lock->class_cache;
2666         /*
2667          * Not cached yet or subclass?
2668          */
2669         if (unlikely(!class)) {
2670                 class = register_lock_class(lock, subclass, 0);
2671                 if (!class)
2672                         return 0;
2673         }
2674         debug_atomic_inc((atomic_t *)&class->ops);
2675         if (very_verbose(class)) {
2676                 printk("\nacquire class [%p] %s", class->key, class->name);
2677                 if (class->name_version > 1)
2678                         printk("#%d", class->name_version);
2679                 printk("\n");
2680                 dump_stack();
2681         }
2682
2683         /*
2684          * Add the lock to the list of currently held locks.
2685          * (we dont increase the depth just yet, up until the
2686          * dependency checks are done)
2687          */
2688         depth = curr->lockdep_depth;
2689         if (DEBUG_LOCKS_WARN_ON(depth >= MAX_LOCK_DEPTH))
2690                 return 0;
2691
2692         hlock = curr->held_locks + depth;
2693         if (DEBUG_LOCKS_WARN_ON(!class))
2694                 return 0;
2695         hlock->class_idx = class - lock_classes + 1;
2696         hlock->acquire_ip = ip;
2697         hlock->instance = lock;
2698         hlock->nest_lock = nest_lock;
2699         hlock->trylock = trylock;
2700         hlock->read = read;
2701         hlock->check = check;
2702         hlock->hardirqs_off = !!hardirqs_off;
2703 #ifdef CONFIG_LOCK_STAT
2704         hlock->waittime_stamp = 0;
2705         hlock->holdtime_stamp = sched_clock();
2706 #endif
2707
2708         if (check == 2 && !mark_irqflags(curr, hlock))
2709                 return 0;
2710
2711         /* mark it as used: */
2712         if (!mark_lock(curr, hlock, LOCK_USED))
2713                 return 0;
2714
2715         /*
2716          * Calculate the chain hash: it's the combined hash of all the
2717          * lock keys along the dependency chain. We save the hash value
2718          * at every step so that we can get the current hash easily
2719          * after unlock. The chain hash is then used to cache dependency
2720          * results.
2721          *
2722          * The 'key ID' is what is the most compact key value to drive
2723          * the hash, not class->key.
2724          */
2725         id = class - lock_classes;
2726         if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
2727                 return 0;
2728
2729         chain_key = curr->curr_chain_key;
2730         if (!depth) {
2731                 if (DEBUG_LOCKS_WARN_ON(chain_key != 0))
2732                         return 0;
2733                 chain_head = 1;
2734         }
2735
2736         hlock->prev_chain_key = chain_key;
2737         if (separate_irq_context(curr, hlock)) {
2738                 chain_key = 0;
2739                 chain_head = 1;
2740         }
2741         chain_key = iterate_chain_key(chain_key, id);
2742
2743         if (!validate_chain(curr, lock, hlock, chain_head, chain_key))
2744                 return 0;
2745
2746         curr->curr_chain_key = chain_key;
2747         curr->lockdep_depth++;
2748         check_chain_key(curr);
2749 #ifdef CONFIG_DEBUG_LOCKDEP
2750         if (unlikely(!debug_locks))
2751                 return 0;
2752 #endif
2753         if (unlikely(curr->lockdep_depth >= MAX_LOCK_DEPTH)) {
2754                 debug_locks_off();
2755                 printk("BUG: MAX_LOCK_DEPTH too low!\n");
2756                 printk("turning off the locking correctness validator.\n");
2757                 dump_stack();
2758                 return 0;
2759         }
2760
2761         if (unlikely(curr->lockdep_depth > max_lockdep_depth))
2762                 max_lockdep_depth = curr->lockdep_depth;
2763
2764         return 1;
2765 }
2766
2767 static int
2768 print_unlock_inbalance_bug(struct task_struct *curr, struct lockdep_map *lock,
2769                            unsigned long ip)
2770 {
2771         if (!debug_locks_off())
2772                 return 0;
2773         if (debug_locks_silent)
2774                 return 0;
2775
2776         printk("\n=====================================\n");
2777         printk(  "[ BUG: bad unlock balance detected! ]\n");
2778         printk(  "-------------------------------------\n");
2779         printk("%s/%d is trying to release lock (",
2780                 curr->comm, task_pid_nr(curr));
2781         print_lockdep_cache(lock);
2782         printk(") at:\n");
2783         print_ip_sym(ip);
2784         printk("but there are no more locks to release!\n");
2785         printk("\nother info that might help us debug this:\n");
2786         lockdep_print_held_locks(curr);
2787
2788         printk("\nstack backtrace:\n");
2789         dump_stack();
2790
2791         return 0;
2792 }
2793
2794 /*
2795  * Common debugging checks for both nested and non-nested unlock:
2796  */
2797 static int check_unlock(struct task_struct *curr, struct lockdep_map *lock,
2798                         unsigned long ip)
2799 {
2800         if (unlikely(!debug_locks))
2801                 return 0;
2802         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2803                 return 0;
2804
2805         if (curr->lockdep_depth <= 0)
2806                 return print_unlock_inbalance_bug(curr, lock, ip);
2807
2808         return 1;
2809 }
2810
2811 static int
2812 __lock_set_class(struct lockdep_map *lock, const char *name,
2813                  struct lock_class_key *key, unsigned int subclass,
2814                  unsigned long ip)
2815 {
2816         struct task_struct *curr = current;
2817         struct held_lock *hlock, *prev_hlock;
2818         struct lock_class *class;
2819         unsigned int depth;
2820         int i;
2821
2822         depth = curr->lockdep_depth;
2823         if (DEBUG_LOCKS_WARN_ON(!depth))
2824                 return 0;
2825
2826         prev_hlock = NULL;
2827         for (i = depth-1; i >= 0; i--) {
2828                 hlock = curr->held_locks + i;
2829                 /*
2830                  * We must not cross into another context:
2831                  */
2832                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
2833                         break;
2834                 if (hlock->instance == lock)
2835                         goto found_it;
2836                 prev_hlock = hlock;
2837         }
2838         return print_unlock_inbalance_bug(curr, lock, ip);
2839
2840 found_it:
2841         lockdep_init_map(lock, name, key, 0);
2842         class = register_lock_class(lock, subclass, 0);
2843         hlock->class_idx = class - lock_classes + 1;
2844
2845         curr->lockdep_depth = i;
2846         curr->curr_chain_key = hlock->prev_chain_key;
2847
2848         for (; i < depth; i++) {
2849                 hlock = curr->held_locks + i;
2850                 if (!__lock_acquire(hlock->instance,
2851                         hlock_class(hlock)->subclass, hlock->trylock,
2852                                 hlock->read, hlock->check, hlock->hardirqs_off,
2853                                 hlock->nest_lock, hlock->acquire_ip))
2854                         return 0;
2855         }
2856
2857         if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth))
2858                 return 0;
2859         return 1;
2860 }
2861
2862 /*
2863  * Remove the lock to the list of currently held locks in a
2864  * potentially non-nested (out of order) manner. This is a
2865  * relatively rare operation, as all the unlock APIs default
2866  * to nested mode (which uses lock_release()):
2867  */
2868 static int
2869 lock_release_non_nested(struct task_struct *curr,
2870                         struct lockdep_map *lock, unsigned long ip)
2871 {
2872         struct held_lock *hlock, *prev_hlock;
2873         unsigned int depth;
2874         int i;
2875
2876         /*
2877          * Check whether the lock exists in the current stack
2878          * of held locks:
2879          */
2880         depth = curr->lockdep_depth;
2881         if (DEBUG_LOCKS_WARN_ON(!depth))
2882                 return 0;
2883
2884         prev_hlock = NULL;
2885         for (i = depth-1; i >= 0; i--) {
2886                 hlock = curr->held_locks + i;
2887                 /*
2888                  * We must not cross into another context:
2889                  */
2890                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
2891                         break;
2892                 if (hlock->instance == lock)
2893                         goto found_it;
2894                 prev_hlock = hlock;
2895         }
2896         return print_unlock_inbalance_bug(curr, lock, ip);
2897
2898 found_it:
2899         lock_release_holdtime(hlock);
2900
2901         /*
2902          * We have the right lock to unlock, 'hlock' points to it.
2903          * Now we remove it from the stack, and add back the other
2904          * entries (if any), recalculating the hash along the way:
2905          */
2906         curr->lockdep_depth = i;
2907         curr->curr_chain_key = hlock->prev_chain_key;
2908
2909         for (i++; i < depth; i++) {
2910                 hlock = curr->held_locks + i;
2911                 if (!__lock_acquire(hlock->instance,
2912                         hlock_class(hlock)->subclass, hlock->trylock,
2913                                 hlock->read, hlock->check, hlock->hardirqs_off,
2914                                 hlock->nest_lock, hlock->acquire_ip))
2915                         return 0;
2916         }
2917
2918         if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth - 1))
2919                 return 0;
2920         return 1;
2921 }
2922
2923 /*
2924  * Remove the lock to the list of currently held locks - this gets
2925  * called on mutex_unlock()/spin_unlock*() (or on a failed
2926  * mutex_lock_interruptible()). This is done for unlocks that nest
2927  * perfectly. (i.e. the current top of the lock-stack is unlocked)
2928  */
2929 static int lock_release_nested(struct task_struct *curr,
2930                                struct lockdep_map *lock, unsigned long ip)
2931 {
2932         struct held_lock *hlock;
2933         unsigned int depth;
2934
2935         /*
2936          * Pop off the top of the lock stack:
2937          */
2938         depth = curr->lockdep_depth - 1;
2939         hlock = curr->held_locks + depth;
2940
2941         /*
2942          * Is the unlock non-nested:
2943          */
2944         if (hlock->instance != lock)
2945                 return lock_release_non_nested(curr, lock, ip);
2946         curr->lockdep_depth--;
2947
2948         if (DEBUG_LOCKS_WARN_ON(!depth && (hlock->prev_chain_key != 0)))
2949                 return 0;
2950
2951         curr->curr_chain_key = hlock->prev_chain_key;
2952
2953         lock_release_holdtime(hlock);
2954
2955 #ifdef CONFIG_DEBUG_LOCKDEP
2956         hlock->prev_chain_key = 0;
2957         hlock->class_idx = 0;
2958         hlock->acquire_ip = 0;
2959         hlock->irq_context = 0;
2960 #endif
2961         return 1;
2962 }
2963
2964 /*
2965  * Remove the lock to the list of currently held locks - this gets
2966  * called on mutex_unlock()/spin_unlock*() (or on a failed
2967  * mutex_lock_interruptible()). This is done for unlocks that nest
2968  * perfectly. (i.e. the current top of the lock-stack is unlocked)
2969  */
2970 static void
2971 __lock_release(struct lockdep_map *lock, int nested, unsigned long ip)
2972 {
2973         struct task_struct *curr = current;
2974
2975         if (!check_unlock(curr, lock, ip))
2976                 return;
2977
2978         if (nested) {
2979                 if (!lock_release_nested(curr, lock, ip))
2980                         return;
2981         } else {
2982                 if (!lock_release_non_nested(curr, lock, ip))
2983                         return;
2984         }
2985
2986         check_chain_key(curr);
2987 }
2988
2989 /*
2990  * Check whether we follow the irq-flags state precisely:
2991  */
2992 static void check_flags(unsigned long flags)
2993 {
2994 #if defined(CONFIG_PROVE_LOCKING) && defined(CONFIG_DEBUG_LOCKDEP) && \
2995     defined(CONFIG_TRACE_IRQFLAGS)
2996         if (!debug_locks)
2997                 return;
2998
2999         if (irqs_disabled_flags(flags)) {
3000                 if (DEBUG_LOCKS_WARN_ON(current->hardirqs_enabled)) {
3001                         printk("possible reason: unannotated irqs-off.\n");
3002                 }
3003         } else {
3004                 if (DEBUG_LOCKS_WARN_ON(!current->hardirqs_enabled)) {
3005                         printk("possible reason: unannotated irqs-on.\n");
3006                 }
3007         }
3008
3009         /*
3010          * We dont accurately track softirq state in e.g.
3011          * hardirq contexts (such as on 4KSTACKS), so only
3012          * check if not in hardirq contexts:
3013          */
3014         if (!hardirq_count()) {
3015                 if (softirq_count())
3016                         DEBUG_LOCKS_WARN_ON(current->softirqs_enabled);
3017                 else
3018                         DEBUG_LOCKS_WARN_ON(!current->softirqs_enabled);
3019         }
3020
3021         if (!debug_locks)
3022                 print_irqtrace_events(current);
3023 #endif
3024 }
3025
3026 void lock_set_class(struct lockdep_map *lock, const char *name,
3027                     struct lock_class_key *key, unsigned int subclass,
3028                     unsigned long ip)
3029 {
3030         unsigned long flags;
3031
3032         if (unlikely(current->lockdep_recursion))
3033                 return;
3034
3035         raw_local_irq_save(flags);
3036         current->lockdep_recursion = 1;
3037         check_flags(flags);
3038         if (__lock_set_class(lock, name, key, subclass, ip))
3039                 check_chain_key(current);
3040         current->lockdep_recursion = 0;
3041         raw_local_irq_restore(flags);
3042 }
3043 EXPORT_SYMBOL_GPL(lock_set_class);
3044
3045 /*
3046  * We are not always called with irqs disabled - do that here,
3047  * and also avoid lockdep recursion:
3048  */
3049 void lock_acquire(struct lockdep_map *lock, unsigned int subclass,
3050                           int trylock, int read, int check,
3051                           struct lockdep_map *nest_lock, unsigned long ip)
3052 {
3053         unsigned long flags;
3054
3055         trace_lock_acquire(lock, subclass, trylock, read, check, nest_lock, ip);
3056
3057         if (unlikely(current->lockdep_recursion))
3058                 return;
3059
3060         raw_local_irq_save(flags);
3061         check_flags(flags);
3062
3063         current->lockdep_recursion = 1;
3064         __lock_acquire(lock, subclass, trylock, read, check,
3065                        irqs_disabled_flags(flags), nest_lock, ip);
3066         current->lockdep_recursion = 0;
3067         raw_local_irq_restore(flags);
3068 }
3069 EXPORT_SYMBOL_GPL(lock_acquire);
3070
3071 void lock_release(struct lockdep_map *lock, int nested,
3072                           unsigned long ip)
3073 {
3074         unsigned long flags;
3075
3076         trace_lock_release(lock, nested, ip);
3077
3078         if (unlikely(current->lockdep_recursion))
3079                 return;
3080
3081         raw_local_irq_save(flags);
3082         check_flags(flags);
3083         current->lockdep_recursion = 1;
3084         __lock_release(lock, nested, ip);
3085         current->lockdep_recursion = 0;
3086         raw_local_irq_restore(flags);
3087 }
3088 EXPORT_SYMBOL_GPL(lock_release);
3089
3090 void lockdep_set_current_reclaim_state(gfp_t gfp_mask)
3091 {
3092         current->lockdep_reclaim_gfp = gfp_mask;
3093 }
3094
3095 void lockdep_clear_current_reclaim_state(void)
3096 {
3097         current->lockdep_reclaim_gfp = 0;
3098 }
3099
3100 #ifdef CONFIG_LOCK_STAT
3101 static int
3102 print_lock_contention_bug(struct task_struct *curr, struct lockdep_map *lock,
3103                            unsigned long ip)
3104 {
3105         if (!debug_locks_off())
3106                 return 0;
3107         if (debug_locks_silent)
3108                 return 0;
3109
3110         printk("\n=================================\n");
3111         printk(  "[ BUG: bad contention detected! ]\n");
3112         printk(  "---------------------------------\n");
3113         printk("%s/%d is trying to contend lock (",
3114                 curr->comm, task_pid_nr(curr));
3115         print_lockdep_cache(lock);
3116         printk(") at:\n");
3117         print_ip_sym(ip);
3118         printk("but there are no locks held!\n");
3119         printk("\nother info that might help us debug this:\n");
3120         lockdep_print_held_locks(curr);
3121
3122         printk("\nstack backtrace:\n");
3123         dump_stack();
3124
3125         return 0;
3126 }
3127
3128 static void
3129 __lock_contended(struct lockdep_map *lock, unsigned long ip)
3130 {
3131         struct task_struct *curr = current;
3132         struct held_lock *hlock, *prev_hlock;
3133         struct lock_class_stats *stats;
3134         unsigned int depth;
3135         int i, contention_point, contending_point;
3136
3137         depth = curr->lockdep_depth;
3138         if (DEBUG_LOCKS_WARN_ON(!depth))
3139                 return;
3140
3141         prev_hlock = NULL;
3142         for (i = depth-1; i >= 0; i--) {
3143                 hlock = curr->held_locks + i;
3144                 /*
3145                  * We must not cross into another context:
3146                  */
3147                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3148                         break;
3149                 if (hlock->instance == lock)
3150                         goto found_it;
3151                 prev_hlock = hlock;
3152         }
3153         print_lock_contention_bug(curr, lock, ip);
3154         return;
3155
3156 found_it:
3157         hlock->waittime_stamp = sched_clock();
3158
3159         contention_point = lock_point(hlock_class(hlock)->contention_point, ip);
3160         contending_point = lock_point(hlock_class(hlock)->contending_point,
3161                                       lock->ip);
3162
3163         stats = get_lock_stats(hlock_class(hlock));
3164         if (contention_point < LOCKSTAT_POINTS)
3165                 stats->contention_point[contention_point]++;
3166         if (contending_point < LOCKSTAT_POINTS)
3167                 stats->contending_point[contending_point]++;
3168         if (lock->cpu != smp_processor_id())
3169                 stats->bounces[bounce_contended + !!hlock->read]++;
3170         put_lock_stats(stats);
3171 }
3172
3173 static void
3174 __lock_acquired(struct lockdep_map *lock, unsigned long ip)
3175 {
3176         struct task_struct *curr = current;
3177         struct held_lock *hlock, *prev_hlock;
3178         struct lock_class_stats *stats;
3179         unsigned int depth;
3180         u64 now;
3181         s64 waittime = 0;
3182         int i, cpu;
3183
3184         depth = curr->lockdep_depth;
3185         if (DEBUG_LOCKS_WARN_ON(!depth))
3186                 return;
3187
3188         prev_hlock = NULL;
3189         for (i = depth-1; i >= 0; i--) {
3190                 hlock = curr->held_locks + i;
3191                 /*
3192                  * We must not cross into another context:
3193                  */
3194                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3195                         break;
3196                 if (hlock->instance == lock)
3197                         goto found_it;
3198                 prev_hlock = hlock;
3199         }
3200         print_lock_contention_bug(curr, lock, _RET_IP_);
3201         return;
3202
3203 found_it:
3204         cpu = smp_processor_id();
3205         if (hlock->waittime_stamp) {
3206                 now = sched_clock();
3207                 waittime = now - hlock->waittime_stamp;
3208                 hlock->holdtime_stamp = now;
3209         }
3210
3211         trace_lock_acquired(lock, ip, waittime);
3212
3213         stats = get_lock_stats(hlock_class(hlock));
3214         if (waittime) {
3215                 if (hlock->read)
3216                         lock_time_inc(&stats->read_waittime, waittime);
3217                 else
3218                         lock_time_inc(&stats->write_waittime, waittime);
3219         }
3220         if (lock->cpu != cpu)
3221                 stats->bounces[bounce_acquired + !!hlock->read]++;
3222         put_lock_stats(stats);
3223
3224         lock->cpu = cpu;
3225         lock->ip = ip;
3226 }
3227
3228 void lock_contended(struct lockdep_map *lock, unsigned long ip)
3229 {
3230         unsigned long flags;
3231
3232         trace_lock_contended(lock, ip);
3233
3234         if (unlikely(!lock_stat))
3235                 return;
3236
3237         if (unlikely(current->lockdep_recursion))
3238                 return;
3239
3240         raw_local_irq_save(flags);
3241         check_flags(flags);
3242         current->lockdep_recursion = 1;
3243         __lock_contended(lock, ip);
3244         current->lockdep_recursion = 0;
3245         raw_local_irq_restore(flags);
3246 }
3247 EXPORT_SYMBOL_GPL(lock_contended);
3248
3249 void lock_acquired(struct lockdep_map *lock, unsigned long ip)
3250 {
3251         unsigned long flags;
3252
3253         if (unlikely(!lock_stat))
3254                 return;
3255
3256         if (unlikely(current->lockdep_recursion))
3257                 return;
3258
3259         raw_local_irq_save(flags);
3260         check_flags(flags);
3261         current->lockdep_recursion = 1;
3262         __lock_acquired(lock, ip);
3263         current->lockdep_recursion = 0;
3264         raw_local_irq_restore(flags);
3265 }
3266 EXPORT_SYMBOL_GPL(lock_acquired);
3267 #endif
3268
3269 /*
3270  * Used by the testsuite, sanitize the validator state
3271  * after a simulated failure:
3272  */
3273
3274 void lockdep_reset(void)
3275 {
3276         unsigned long flags;
3277         int i;
3278
3279         raw_local_irq_save(flags);
3280         current->curr_chain_key = 0;
3281         current->lockdep_depth = 0;
3282         current->lockdep_recursion = 0;
3283         memset(current->held_locks, 0, MAX_LOCK_DEPTH*sizeof(struct held_lock));
3284         nr_hardirq_chains = 0;
3285         nr_softirq_chains = 0;
3286         nr_process_chains = 0;
3287         debug_locks = 1;
3288         for (i = 0; i < CHAINHASH_SIZE; i++)
3289                 INIT_LIST_HEAD(chainhash_table + i);
3290         raw_local_irq_restore(flags);
3291 }
3292
3293 static void zap_class(struct lock_class *class)
3294 {
3295         int i;
3296
3297         /*
3298          * Remove all dependencies this lock is
3299          * involved in:
3300          */
3301         for (i = 0; i < nr_list_entries; i++) {
3302                 if (list_entries[i].class == class)
3303                         list_del_rcu(&list_entries[i].entry);
3304         }
3305         /*
3306          * Unhash the class and remove it from the all_lock_classes list:
3307          */
3308         list_del_rcu(&class->hash_entry);
3309         list_del_rcu(&class->lock_entry);
3310
3311         class->key = NULL;
3312 }
3313
3314 static inline int within(const void *addr, void *start, unsigned long size)
3315 {
3316         return addr >= start && addr < start + size;
3317 }
3318
3319 void lockdep_free_key_range(void *start, unsigned long size)
3320 {
3321         struct lock_class *class, *next;
3322         struct list_head *head;
3323         unsigned long flags;
3324         int i;
3325         int locked;
3326
3327         raw_local_irq_save(flags);
3328         locked = graph_lock();
3329
3330         /*
3331          * Unhash all classes that were created by this module:
3332          */
3333         for (i = 0; i < CLASSHASH_SIZE; i++) {
3334                 head = classhash_table + i;
3335                 if (list_empty(head))
3336                         continue;
3337                 list_for_each_entry_safe(class, next, head, hash_entry) {
3338                         if (within(class->key, start, size))
3339                                 zap_class(class);
3340                         else if (within(class->name, start, size))
3341                                 zap_class(class);
3342                 }
3343         }
3344
3345         if (locked)
3346                 graph_unlock();
3347         raw_local_irq_restore(flags);
3348 }
3349
3350 void lockdep_reset_lock(struct lockdep_map *lock)
3351 {
3352         struct lock_class *class, *next;
3353         struct list_head *head;
3354         unsigned long flags;
3355         int i, j;
3356         int locked;
3357
3358         raw_local_irq_save(flags);
3359
3360         /*
3361          * Remove all classes this lock might have:
3362          */
3363         for (j = 0; j < MAX_LOCKDEP_SUBCLASSES; j++) {
3364                 /*
3365                  * If the class exists we look it up and zap it:
3366                  */
3367                 class = look_up_lock_class(lock, j);
3368                 if (class)
3369                         zap_class(class);
3370         }
3371         /*
3372          * Debug check: in the end all mapped classes should
3373          * be gone.
3374          */
3375         locked = graph_lock();
3376         for (i = 0; i < CLASSHASH_SIZE; i++) {
3377                 head = classhash_table + i;
3378                 if (list_empty(head))
3379                         continue;
3380                 list_for_each_entry_safe(class, next, head, hash_entry) {
3381                         if (unlikely(class == lock->class_cache)) {
3382                                 if (debug_locks_off_graph_unlock())
3383                                         WARN_ON(1);
3384                                 goto out_restore;
3385                         }
3386                 }
3387         }
3388         if (locked)
3389                 graph_unlock();
3390
3391 out_restore:
3392         raw_local_irq_restore(flags);
3393 }
3394
3395 void lockdep_init(void)
3396 {
3397         int i;
3398
3399         /*
3400          * Some architectures have their own start_kernel()
3401          * code which calls lockdep_init(), while we also
3402          * call lockdep_init() from the start_kernel() itself,
3403          * and we want to initialize the hashes only once:
3404          */
3405         if (lockdep_initialized)
3406                 return;
3407
3408         for (i = 0; i < CLASSHASH_SIZE; i++)
3409                 INIT_LIST_HEAD(classhash_table + i);
3410
3411         for (i = 0; i < CHAINHASH_SIZE; i++)
3412                 INIT_LIST_HEAD(chainhash_table + i);
3413
3414         lockdep_initialized = 1;
3415 }
3416
3417 void __init lockdep_info(void)
3418 {
3419         printk("Lock dependency validator: Copyright (c) 2006 Red Hat, Inc., Ingo Molnar\n");
3420
3421         printk("... MAX_LOCKDEP_SUBCLASSES:  %lu\n", MAX_LOCKDEP_SUBCLASSES);
3422         printk("... MAX_LOCK_DEPTH:          %lu\n", MAX_LOCK_DEPTH);
3423         printk("... MAX_LOCKDEP_KEYS:        %lu\n", MAX_LOCKDEP_KEYS);
3424         printk("... CLASSHASH_SIZE:          %lu\n", CLASSHASH_SIZE);
3425         printk("... MAX_LOCKDEP_ENTRIES:     %lu\n", MAX_LOCKDEP_ENTRIES);
3426         printk("... MAX_LOCKDEP_CHAINS:      %lu\n", MAX_LOCKDEP_CHAINS);
3427         printk("... CHAINHASH_SIZE:          %lu\n", CHAINHASH_SIZE);
3428
3429         printk(" memory used by lock dependency info: %lu kB\n",
3430                 (sizeof(struct lock_class) * MAX_LOCKDEP_KEYS +
3431                 sizeof(struct list_head) * CLASSHASH_SIZE +
3432                 sizeof(struct lock_list) * MAX_LOCKDEP_ENTRIES +
3433                 sizeof(struct lock_chain) * MAX_LOCKDEP_CHAINS +
3434                 sizeof(struct list_head) * CHAINHASH_SIZE) / 1024);
3435
3436         printk(" per task-struct memory footprint: %lu bytes\n",
3437                 sizeof(struct held_lock) * MAX_LOCK_DEPTH);
3438
3439 #ifdef CONFIG_DEBUG_LOCKDEP
3440         if (lockdep_init_error) {
3441                 printk("WARNING: lockdep init error! Arch code didn't call lockdep_init() early enough?\n");
3442                 printk("Call stack leading to lockdep invocation was:\n");
3443                 print_stack_trace(&lockdep_init_trace, 0);
3444         }
3445 #endif
3446 }
3447
3448 static void
3449 print_freed_lock_bug(struct task_struct *curr, const void *mem_from,
3450                      const void *mem_to, struct held_lock *hlock)
3451 {
3452         if (!debug_locks_off())
3453                 return;
3454         if (debug_locks_silent)
3455                 return;
3456
3457         printk("\n=========================\n");
3458         printk(  "[ BUG: held lock freed! ]\n");
3459         printk(  "-------------------------\n");
3460         printk("%s/%d is freeing memory %p-%p, with a lock still held there!\n",
3461                 curr->comm, task_pid_nr(curr), mem_from, mem_to-1);
3462         print_lock(hlock);
3463         lockdep_print_held_locks(curr);
3464
3465         printk("\nstack backtrace:\n");
3466         dump_stack();
3467 }
3468
3469 static inline int not_in_range(const void* mem_from, unsigned long mem_len,
3470                                 const void* lock_from, unsigned long lock_len)
3471 {
3472         return lock_from + lock_len <= mem_from ||
3473                 mem_from + mem_len <= lock_from;
3474 }
3475
3476 /*
3477  * Called when kernel memory is freed (or unmapped), or if a lock
3478  * is destroyed or reinitialized - this code checks whether there is
3479  * any held lock in the memory range of <from> to <to>:
3480  */
3481 void debug_check_no_locks_freed(const void *mem_from, unsigned long mem_len)
3482 {
3483         struct task_struct *curr = current;
3484         struct held_lock *hlock;
3485         unsigned long flags;
3486         int i;
3487
3488         if (unlikely(!debug_locks))
3489                 return;
3490
3491         local_irq_save(flags);
3492         for (i = 0; i < curr->lockdep_depth; i++) {
3493                 hlock = curr->held_locks + i;
3494
3495                 if (not_in_range(mem_from, mem_len, hlock->instance,
3496                                         sizeof(*hlock->instance)))
3497                         continue;
3498
3499                 print_freed_lock_bug(curr, mem_from, mem_from + mem_len, hlock);
3500                 break;
3501         }
3502         local_irq_restore(flags);
3503 }
3504 EXPORT_SYMBOL_GPL(debug_check_no_locks_freed);
3505
3506 static void print_held_locks_bug(struct task_struct *curr)
3507 {
3508         if (!debug_locks_off())
3509                 return;
3510         if (debug_locks_silent)
3511                 return;
3512
3513         printk("\n=====================================\n");
3514         printk(  "[ BUG: lock held at task exit time! ]\n");
3515         printk(  "-------------------------------------\n");
3516         printk("%s/%d is exiting with locks still held!\n",
3517                 curr->comm, task_pid_nr(curr));
3518         lockdep_print_held_locks(curr);
3519
3520         printk("\nstack backtrace:\n");
3521         dump_stack();
3522 }
3523
3524 void debug_check_no_locks_held(struct task_struct *task)
3525 {
3526         if (unlikely(task->lockdep_depth > 0))
3527                 print_held_locks_bug(task);
3528 }
3529
3530 void debug_show_all_locks(void)
3531 {
3532         struct task_struct *g, *p;
3533         int count = 10;
3534         int unlock = 1;
3535
3536         if (unlikely(!debug_locks)) {
3537                 printk("INFO: lockdep is turned off.\n");
3538                 return;
3539         }
3540         printk("\nShowing all locks held in the system:\n");
3541
3542         /*
3543          * Here we try to get the tasklist_lock as hard as possible,
3544          * if not successful after 2 seconds we ignore it (but keep
3545          * trying). This is to enable a debug printout even if a
3546          * tasklist_lock-holding task deadlocks or crashes.
3547          */
3548 retry:
3549         if (!read_trylock(&tasklist_lock)) {
3550                 if (count == 10)
3551                         printk("hm, tasklist_lock locked, retrying... ");
3552                 if (count) {
3553                         count--;
3554                         printk(" #%d", 10-count);
3555                         mdelay(200);
3556                         goto retry;
3557                 }
3558                 printk(" ignoring it.\n");
3559                 unlock = 0;
3560         } else {
3561                 if (count != 10)
3562                         printk(KERN_CONT " locked it.\n");
3563         }
3564
3565         do_each_thread(g, p) {
3566                 /*
3567                  * It's not reliable to print a task's held locks
3568                  * if it's not sleeping (or if it's not the current
3569                  * task):
3570                  */
3571                 if (p->state == TASK_RUNNING && p != current)
3572                         continue;
3573                 if (p->lockdep_depth)
3574                         lockdep_print_held_locks(p);
3575                 if (!unlock)
3576                         if (read_trylock(&tasklist_lock))
3577                                 unlock = 1;
3578         } while_each_thread(g, p);
3579
3580         printk("\n");
3581         printk("=============================================\n\n");
3582
3583         if (unlock)
3584                 read_unlock(&tasklist_lock);
3585 }
3586 EXPORT_SYMBOL_GPL(debug_show_all_locks);
3587
3588 /*
3589  * Careful: only use this function if you are sure that
3590  * the task cannot run in parallel!
3591  */
3592 void __debug_show_held_locks(struct task_struct *task)
3593 {
3594         if (unlikely(!debug_locks)) {
3595                 printk("INFO: lockdep is turned off.\n");
3596                 return;
3597         }
3598         lockdep_print_held_locks(task);
3599 }
3600 EXPORT_SYMBOL_GPL(__debug_show_held_locks);
3601
3602 void debug_show_held_locks(struct task_struct *task)
3603 {
3604                 __debug_show_held_locks(task);
3605 }
3606 EXPORT_SYMBOL_GPL(debug_show_held_locks);
3607
3608 void lockdep_sys_exit(void)
3609 {
3610         struct task_struct *curr = current;
3611
3612         if (unlikely(curr->lockdep_depth)) {
3613                 if (!debug_locks_off())
3614                         return;
3615                 printk("\n================================================\n");
3616                 printk(  "[ BUG: lock held when returning to user space! ]\n");
3617                 printk(  "------------------------------------------------\n");
3618                 printk("%s/%d is leaving the kernel with locks still held!\n",
3619                                 curr->comm, curr->pid);
3620                 lockdep_print_held_locks(curr);
3621         }
3622 }