a288ae107b50b24101e5ad193b8fdb4c5aa8a44d
[linux-2.6.git] / kernel / lockdep.c
1 /*
2  * kernel/lockdep.c
3  *
4  * Runtime locking correctness validator
5  *
6  * Started by Ingo Molnar:
7  *
8  *  Copyright (C) 2006,2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
9  *  Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
10  *
11  * this code maps all the lock dependencies as they occur in a live kernel
12  * and will warn about the following classes of locking bugs:
13  *
14  * - lock inversion scenarios
15  * - circular lock dependencies
16  * - hardirq/softirq safe/unsafe locking bugs
17  *
18  * Bugs are reported even if the current locking scenario does not cause
19  * any deadlock at this point.
20  *
21  * I.e. if anytime in the past two locks were taken in a different order,
22  * even if it happened for another task, even if those were different
23  * locks (but of the same class as this lock), this code will detect it.
24  *
25  * Thanks to Arjan van de Ven for coming up with the initial idea of
26  * mapping lock dependencies runtime.
27  */
28 #define DISABLE_BRANCH_PROFILING
29 #include <linux/mutex.h>
30 #include <linux/sched.h>
31 #include <linux/delay.h>
32 #include <linux/module.h>
33 #include <linux/proc_fs.h>
34 #include <linux/seq_file.h>
35 #include <linux/spinlock.h>
36 #include <linux/kallsyms.h>
37 #include <linux/interrupt.h>
38 #include <linux/stacktrace.h>
39 #include <linux/debug_locks.h>
40 #include <linux/irqflags.h>
41 #include <linux/utsname.h>
42 #include <linux/hash.h>
43 #include <linux/ftrace.h>
44 #include <linux/stringify.h>
45
46 #include <asm/sections.h>
47
48 #include "lockdep_internals.h"
49
50 #ifdef CONFIG_PROVE_LOCKING
51 int prove_locking = 1;
52 module_param(prove_locking, int, 0644);
53 #else
54 #define prove_locking 0
55 #endif
56
57 #ifdef CONFIG_LOCK_STAT
58 int lock_stat = 1;
59 module_param(lock_stat, int, 0644);
60 #else
61 #define lock_stat 0
62 #endif
63
64 /*
65  * lockdep_lock: protects the lockdep graph, the hashes and the
66  *               class/list/hash allocators.
67  *
68  * This is one of the rare exceptions where it's justified
69  * to use a raw spinlock - we really dont want the spinlock
70  * code to recurse back into the lockdep code...
71  */
72 static raw_spinlock_t lockdep_lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
73
74 static int graph_lock(void)
75 {
76         __raw_spin_lock(&lockdep_lock);
77         /*
78          * Make sure that if another CPU detected a bug while
79          * walking the graph we dont change it (while the other
80          * CPU is busy printing out stuff with the graph lock
81          * dropped already)
82          */
83         if (!debug_locks) {
84                 __raw_spin_unlock(&lockdep_lock);
85                 return 0;
86         }
87         /* prevent any recursions within lockdep from causing deadlocks */
88         current->lockdep_recursion++;
89         return 1;
90 }
91
92 static inline int graph_unlock(void)
93 {
94         if (debug_locks && !__raw_spin_is_locked(&lockdep_lock))
95                 return DEBUG_LOCKS_WARN_ON(1);
96
97         current->lockdep_recursion--;
98         __raw_spin_unlock(&lockdep_lock);
99         return 0;
100 }
101
102 /*
103  * Turn lock debugging off and return with 0 if it was off already,
104  * and also release the graph lock:
105  */
106 static inline int debug_locks_off_graph_unlock(void)
107 {
108         int ret = debug_locks_off();
109
110         __raw_spin_unlock(&lockdep_lock);
111
112         return ret;
113 }
114
115 static int lockdep_initialized;
116
117 unsigned long nr_list_entries;
118 static struct lock_list list_entries[MAX_LOCKDEP_ENTRIES];
119
120 /*
121  * All data structures here are protected by the global debug_lock.
122  *
123  * Mutex key structs only get allocated, once during bootup, and never
124  * get freed - this significantly simplifies the debugging code.
125  */
126 unsigned long nr_lock_classes;
127 static struct lock_class lock_classes[MAX_LOCKDEP_KEYS];
128
129 static inline struct lock_class *hlock_class(struct held_lock *hlock)
130 {
131         if (!hlock->class_idx) {
132                 DEBUG_LOCKS_WARN_ON(1);
133                 return NULL;
134         }
135         return lock_classes + hlock->class_idx - 1;
136 }
137
138 #ifdef CONFIG_LOCK_STAT
139 static DEFINE_PER_CPU(struct lock_class_stats[MAX_LOCKDEP_KEYS], lock_stats);
140
141 static int lock_point(unsigned long points[], unsigned long ip)
142 {
143         int i;
144
145         for (i = 0; i < LOCKSTAT_POINTS; i++) {
146                 if (points[i] == 0) {
147                         points[i] = ip;
148                         break;
149                 }
150                 if (points[i] == ip)
151                         break;
152         }
153
154         return i;
155 }
156
157 static void lock_time_inc(struct lock_time *lt, s64 time)
158 {
159         if (time > lt->max)
160                 lt->max = time;
161
162         if (time < lt->min || !lt->min)
163                 lt->min = time;
164
165         lt->total += time;
166         lt->nr++;
167 }
168
169 static inline void lock_time_add(struct lock_time *src, struct lock_time *dst)
170 {
171         dst->min += src->min;
172         dst->max += src->max;
173         dst->total += src->total;
174         dst->nr += src->nr;
175 }
176
177 struct lock_class_stats lock_stats(struct lock_class *class)
178 {
179         struct lock_class_stats stats;
180         int cpu, i;
181
182         memset(&stats, 0, sizeof(struct lock_class_stats));
183         for_each_possible_cpu(cpu) {
184                 struct lock_class_stats *pcs =
185                         &per_cpu(lock_stats, cpu)[class - lock_classes];
186
187                 for (i = 0; i < ARRAY_SIZE(stats.contention_point); i++)
188                         stats.contention_point[i] += pcs->contention_point[i];
189
190                 for (i = 0; i < ARRAY_SIZE(stats.contending_point); i++)
191                         stats.contending_point[i] += pcs->contending_point[i];
192
193                 lock_time_add(&pcs->read_waittime, &stats.read_waittime);
194                 lock_time_add(&pcs->write_waittime, &stats.write_waittime);
195
196                 lock_time_add(&pcs->read_holdtime, &stats.read_holdtime);
197                 lock_time_add(&pcs->write_holdtime, &stats.write_holdtime);
198
199                 for (i = 0; i < ARRAY_SIZE(stats.bounces); i++)
200                         stats.bounces[i] += pcs->bounces[i];
201         }
202
203         return stats;
204 }
205
206 void clear_lock_stats(struct lock_class *class)
207 {
208         int cpu;
209
210         for_each_possible_cpu(cpu) {
211                 struct lock_class_stats *cpu_stats =
212                         &per_cpu(lock_stats, cpu)[class - lock_classes];
213
214                 memset(cpu_stats, 0, sizeof(struct lock_class_stats));
215         }
216         memset(class->contention_point, 0, sizeof(class->contention_point));
217         memset(class->contending_point, 0, sizeof(class->contending_point));
218 }
219
220 static struct lock_class_stats *get_lock_stats(struct lock_class *class)
221 {
222         return &get_cpu_var(lock_stats)[class - lock_classes];
223 }
224
225 static void put_lock_stats(struct lock_class_stats *stats)
226 {
227         put_cpu_var(lock_stats);
228 }
229
230 static void lock_release_holdtime(struct held_lock *hlock)
231 {
232         struct lock_class_stats *stats;
233         s64 holdtime;
234
235         if (!lock_stat)
236                 return;
237
238         holdtime = sched_clock() - hlock->holdtime_stamp;
239
240         stats = get_lock_stats(hlock_class(hlock));
241         if (hlock->read)
242                 lock_time_inc(&stats->read_holdtime, holdtime);
243         else
244                 lock_time_inc(&stats->write_holdtime, holdtime);
245         put_lock_stats(stats);
246 }
247 #else
248 static inline void lock_release_holdtime(struct held_lock *hlock)
249 {
250 }
251 #endif
252
253 /*
254  * We keep a global list of all lock classes. The list only grows,
255  * never shrinks. The list is only accessed with the lockdep
256  * spinlock lock held.
257  */
258 LIST_HEAD(all_lock_classes);
259
260 /*
261  * The lockdep classes are in a hash-table as well, for fast lookup:
262  */
263 #define CLASSHASH_BITS          (MAX_LOCKDEP_KEYS_BITS - 1)
264 #define CLASSHASH_SIZE          (1UL << CLASSHASH_BITS)
265 #define __classhashfn(key)      hash_long((unsigned long)key, CLASSHASH_BITS)
266 #define classhashentry(key)     (classhash_table + __classhashfn((key)))
267
268 static struct list_head classhash_table[CLASSHASH_SIZE];
269
270 /*
271  * We put the lock dependency chains into a hash-table as well, to cache
272  * their existence:
273  */
274 #define CHAINHASH_BITS          (MAX_LOCKDEP_CHAINS_BITS-1)
275 #define CHAINHASH_SIZE          (1UL << CHAINHASH_BITS)
276 #define __chainhashfn(chain)    hash_long(chain, CHAINHASH_BITS)
277 #define chainhashentry(chain)   (chainhash_table + __chainhashfn((chain)))
278
279 static struct list_head chainhash_table[CHAINHASH_SIZE];
280
281 /*
282  * The hash key of the lock dependency chains is a hash itself too:
283  * it's a hash of all locks taken up to that lock, including that lock.
284  * It's a 64-bit hash, because it's important for the keys to be
285  * unique.
286  */
287 #define iterate_chain_key(key1, key2) \
288         (((key1) << MAX_LOCKDEP_KEYS_BITS) ^ \
289         ((key1) >> (64-MAX_LOCKDEP_KEYS_BITS)) ^ \
290         (key2))
291
292 void lockdep_off(void)
293 {
294         current->lockdep_recursion++;
295 }
296 EXPORT_SYMBOL(lockdep_off);
297
298 void lockdep_on(void)
299 {
300         current->lockdep_recursion--;
301 }
302 EXPORT_SYMBOL(lockdep_on);
303
304 /*
305  * Debugging switches:
306  */
307
308 #define VERBOSE                 0
309 #define VERY_VERBOSE            0
310
311 #if VERBOSE
312 # define HARDIRQ_VERBOSE        1
313 # define SOFTIRQ_VERBOSE        1
314 # define RECLAIM_VERBOSE        1
315 #else
316 # define HARDIRQ_VERBOSE        0
317 # define SOFTIRQ_VERBOSE        0
318 # define RECLAIM_VERBOSE        0
319 #endif
320
321 #if VERBOSE || HARDIRQ_VERBOSE || SOFTIRQ_VERBOSE || RECLAIM_VERBOSE
322 /*
323  * Quick filtering for interesting events:
324  */
325 static int class_filter(struct lock_class *class)
326 {
327 #if 0
328         /* Example */
329         if (class->name_version == 1 &&
330                         !strcmp(class->name, "lockname"))
331                 return 1;
332         if (class->name_version == 1 &&
333                         !strcmp(class->name, "&struct->lockfield"))
334                 return 1;
335 #endif
336         /* Filter everything else. 1 would be to allow everything else */
337         return 0;
338 }
339 #endif
340
341 static int verbose(struct lock_class *class)
342 {
343 #if VERBOSE
344         return class_filter(class);
345 #endif
346         return 0;
347 }
348
349 /*
350  * Stack-trace: tightly packed array of stack backtrace
351  * addresses. Protected by the graph_lock.
352  */
353 unsigned long nr_stack_trace_entries;
354 static unsigned long stack_trace[MAX_STACK_TRACE_ENTRIES];
355
356 static int save_trace(struct stack_trace *trace)
357 {
358         trace->nr_entries = 0;
359         trace->max_entries = MAX_STACK_TRACE_ENTRIES - nr_stack_trace_entries;
360         trace->entries = stack_trace + nr_stack_trace_entries;
361
362         trace->skip = 3;
363
364         save_stack_trace(trace);
365
366         trace->max_entries = trace->nr_entries;
367
368         nr_stack_trace_entries += trace->nr_entries;
369
370         if (nr_stack_trace_entries == MAX_STACK_TRACE_ENTRIES) {
371                 if (!debug_locks_off_graph_unlock())
372                         return 0;
373
374                 printk("BUG: MAX_STACK_TRACE_ENTRIES too low!\n");
375                 printk("turning off the locking correctness validator.\n");
376                 dump_stack();
377
378                 return 0;
379         }
380
381         return 1;
382 }
383
384 unsigned int nr_hardirq_chains;
385 unsigned int nr_softirq_chains;
386 unsigned int nr_process_chains;
387 unsigned int max_lockdep_depth;
388 unsigned int max_recursion_depth;
389
390 static unsigned int lockdep_dependency_gen_id;
391
392 static bool lockdep_dependency_visit(struct lock_class *source,
393                                      unsigned int depth)
394 {
395         if (!depth)
396                 lockdep_dependency_gen_id++;
397         if (source->dep_gen_id == lockdep_dependency_gen_id)
398                 return true;
399         source->dep_gen_id = lockdep_dependency_gen_id;
400         return false;
401 }
402
403 #ifdef CONFIG_DEBUG_LOCKDEP
404 /*
405  * We cannot printk in early bootup code. Not even early_printk()
406  * might work. So we mark any initialization errors and printk
407  * about it later on, in lockdep_info().
408  */
409 static int lockdep_init_error;
410 static unsigned long lockdep_init_trace_data[20];
411 static struct stack_trace lockdep_init_trace = {
412         .max_entries = ARRAY_SIZE(lockdep_init_trace_data),
413         .entries = lockdep_init_trace_data,
414 };
415
416 /*
417  * Various lockdep statistics:
418  */
419 atomic_t chain_lookup_hits;
420 atomic_t chain_lookup_misses;
421 atomic_t hardirqs_on_events;
422 atomic_t hardirqs_off_events;
423 atomic_t redundant_hardirqs_on;
424 atomic_t redundant_hardirqs_off;
425 atomic_t softirqs_on_events;
426 atomic_t softirqs_off_events;
427 atomic_t redundant_softirqs_on;
428 atomic_t redundant_softirqs_off;
429 atomic_t nr_unused_locks;
430 atomic_t nr_cyclic_checks;
431 atomic_t nr_cyclic_check_recursions;
432 atomic_t nr_find_usage_forwards_checks;
433 atomic_t nr_find_usage_forwards_recursions;
434 atomic_t nr_find_usage_backwards_checks;
435 atomic_t nr_find_usage_backwards_recursions;
436 #endif
437
438 /*
439  * Locking printouts:
440  */
441
442 #define __USAGE(__STATE)                                                \
443         [LOCK_USED_IN_##__STATE] = "IN-"__stringify(__STATE)"-W",       \
444         [LOCK_ENABLED_##__STATE] = __stringify(__STATE)"-ON-W",         \
445         [LOCK_USED_IN_##__STATE##_READ] = "IN-"__stringify(__STATE)"-R",\
446         [LOCK_ENABLED_##__STATE##_READ] = __stringify(__STATE)"-ON-R",
447
448 static const char *usage_str[] =
449 {
450 #define LOCKDEP_STATE(__STATE) __USAGE(__STATE)
451 #include "lockdep_states.h"
452 #undef LOCKDEP_STATE
453         [LOCK_USED] = "INITIAL USE",
454 };
455
456 const char * __get_key_name(struct lockdep_subclass_key *key, char *str)
457 {
458         return kallsyms_lookup((unsigned long)key, NULL, NULL, NULL, str);
459 }
460
461 static inline unsigned long lock_flag(enum lock_usage_bit bit)
462 {
463         return 1UL << bit;
464 }
465
466 static char get_usage_char(struct lock_class *class, enum lock_usage_bit bit)
467 {
468         char c = '.';
469
470         if (class->usage_mask & lock_flag(bit + 2))
471                 c = '+';
472         if (class->usage_mask & lock_flag(bit)) {
473                 c = '-';
474                 if (class->usage_mask & lock_flag(bit + 2))
475                         c = '?';
476         }
477
478         return c;
479 }
480
481 void get_usage_chars(struct lock_class *class, char usage[LOCK_USAGE_CHARS])
482 {
483         int i = 0;
484
485 #define LOCKDEP_STATE(__STATE)                                          \
486         usage[i++] = get_usage_char(class, LOCK_USED_IN_##__STATE);     \
487         usage[i++] = get_usage_char(class, LOCK_USED_IN_##__STATE##_READ);
488 #include "lockdep_states.h"
489 #undef LOCKDEP_STATE
490
491         usage[i] = '\0';
492 }
493
494 static void print_lock_name(struct lock_class *class)
495 {
496         char str[KSYM_NAME_LEN], usage[LOCK_USAGE_CHARS];
497         const char *name;
498
499         get_usage_chars(class, usage);
500
501         name = class->name;
502         if (!name) {
503                 name = __get_key_name(class->key, str);
504                 printk(" (%s", name);
505         } else {
506                 printk(" (%s", name);
507                 if (class->name_version > 1)
508                         printk("#%d", class->name_version);
509                 if (class->subclass)
510                         printk("/%d", class->subclass);
511         }
512         printk("){%s}", usage);
513 }
514
515 static void print_lockdep_cache(struct lockdep_map *lock)
516 {
517         const char *name;
518         char str[KSYM_NAME_LEN];
519
520         name = lock->name;
521         if (!name)
522                 name = __get_key_name(lock->key->subkeys, str);
523
524         printk("%s", name);
525 }
526
527 static void print_lock(struct held_lock *hlock)
528 {
529         print_lock_name(hlock_class(hlock));
530         printk(", at: ");
531         print_ip_sym(hlock->acquire_ip);
532 }
533
534 static void lockdep_print_held_locks(struct task_struct *curr)
535 {
536         int i, depth = curr->lockdep_depth;
537
538         if (!depth) {
539                 printk("no locks held by %s/%d.\n", curr->comm, task_pid_nr(curr));
540                 return;
541         }
542         printk("%d lock%s held by %s/%d:\n",
543                 depth, depth > 1 ? "s" : "", curr->comm, task_pid_nr(curr));
544
545         for (i = 0; i < depth; i++) {
546                 printk(" #%d: ", i);
547                 print_lock(curr->held_locks + i);
548         }
549 }
550
551 static void print_lock_class_header(struct lock_class *class, int depth)
552 {
553         int bit;
554
555         printk("%*s->", depth, "");
556         print_lock_name(class);
557         printk(" ops: %lu", class->ops);
558         printk(" {\n");
559
560         for (bit = 0; bit < LOCK_USAGE_STATES; bit++) {
561                 if (class->usage_mask & (1 << bit)) {
562                         int len = depth;
563
564                         len += printk("%*s   %s", depth, "", usage_str[bit]);
565                         len += printk(" at:\n");
566                         print_stack_trace(class->usage_traces + bit, len);
567                 }
568         }
569         printk("%*s }\n", depth, "");
570
571         printk("%*s ... key      at: ",depth,"");
572         print_ip_sym((unsigned long)class->key);
573 }
574
575 /*
576  * printk all lock dependencies starting at <entry>:
577  */
578 static void __used
579 print_lock_dependencies(struct lock_class *class, int depth)
580 {
581         struct lock_list *entry;
582
583         if (lockdep_dependency_visit(class, depth))
584                 return;
585
586         if (DEBUG_LOCKS_WARN_ON(depth >= 20))
587                 return;
588
589         print_lock_class_header(class, depth);
590
591         list_for_each_entry(entry, &class->locks_after, entry) {
592                 if (DEBUG_LOCKS_WARN_ON(!entry->class))
593                         return;
594
595                 print_lock_dependencies(entry->class, depth + 1);
596
597                 printk("%*s ... acquired at:\n",depth,"");
598                 print_stack_trace(&entry->trace, 2);
599                 printk("\n");
600         }
601 }
602
603 static void print_kernel_version(void)
604 {
605         printk("%s %.*s\n", init_utsname()->release,
606                 (int)strcspn(init_utsname()->version, " "),
607                 init_utsname()->version);
608 }
609
610 static int very_verbose(struct lock_class *class)
611 {
612 #if VERY_VERBOSE
613         return class_filter(class);
614 #endif
615         return 0;
616 }
617
618 /*
619  * Is this the address of a static object:
620  */
621 static int static_obj(void *obj)
622 {
623         unsigned long start = (unsigned long) &_stext,
624                       end   = (unsigned long) &_end,
625                       addr  = (unsigned long) obj;
626 #ifdef CONFIG_SMP
627         int i;
628 #endif
629
630         /*
631          * static variable?
632          */
633         if ((addr >= start) && (addr < end))
634                 return 1;
635
636 #ifdef CONFIG_SMP
637         /*
638          * percpu var?
639          */
640         for_each_possible_cpu(i) {
641                 start = (unsigned long) &__per_cpu_start + per_cpu_offset(i);
642                 end   = (unsigned long) &__per_cpu_start + PERCPU_ENOUGH_ROOM
643                                         + per_cpu_offset(i);
644
645                 if ((addr >= start) && (addr < end))
646                         return 1;
647         }
648 #endif
649
650         /*
651          * module var?
652          */
653         return is_module_address(addr);
654 }
655
656 /*
657  * To make lock name printouts unique, we calculate a unique
658  * class->name_version generation counter:
659  */
660 static int count_matching_names(struct lock_class *new_class)
661 {
662         struct lock_class *class;
663         int count = 0;
664
665         if (!new_class->name)
666                 return 0;
667
668         list_for_each_entry(class, &all_lock_classes, lock_entry) {
669                 if (new_class->key - new_class->subclass == class->key)
670                         return class->name_version;
671                 if (class->name && !strcmp(class->name, new_class->name))
672                         count = max(count, class->name_version);
673         }
674
675         return count + 1;
676 }
677
678 /*
679  * Register a lock's class in the hash-table, if the class is not present
680  * yet. Otherwise we look it up. We cache the result in the lock object
681  * itself, so actual lookup of the hash should be once per lock object.
682  */
683 static inline struct lock_class *
684 look_up_lock_class(struct lockdep_map *lock, unsigned int subclass)
685 {
686         struct lockdep_subclass_key *key;
687         struct list_head *hash_head;
688         struct lock_class *class;
689
690 #ifdef CONFIG_DEBUG_LOCKDEP
691         /*
692          * If the architecture calls into lockdep before initializing
693          * the hashes then we'll warn about it later. (we cannot printk
694          * right now)
695          */
696         if (unlikely(!lockdep_initialized)) {
697                 lockdep_init();
698                 lockdep_init_error = 1;
699                 save_stack_trace(&lockdep_init_trace);
700         }
701 #endif
702
703         /*
704          * Static locks do not have their class-keys yet - for them the key
705          * is the lock object itself:
706          */
707         if (unlikely(!lock->key))
708                 lock->key = (void *)lock;
709
710         /*
711          * NOTE: the class-key must be unique. For dynamic locks, a static
712          * lock_class_key variable is passed in through the mutex_init()
713          * (or spin_lock_init()) call - which acts as the key. For static
714          * locks we use the lock object itself as the key.
715          */
716         BUILD_BUG_ON(sizeof(struct lock_class_key) >
717                         sizeof(struct lockdep_map));
718
719         key = lock->key->subkeys + subclass;
720
721         hash_head = classhashentry(key);
722
723         /*
724          * We can walk the hash lockfree, because the hash only
725          * grows, and we are careful when adding entries to the end:
726          */
727         list_for_each_entry(class, hash_head, hash_entry) {
728                 if (class->key == key) {
729                         WARN_ON_ONCE(class->name != lock->name);
730                         return class;
731                 }
732         }
733
734         return NULL;
735 }
736
737 /*
738  * Register a lock's class in the hash-table, if the class is not present
739  * yet. Otherwise we look it up. We cache the result in the lock object
740  * itself, so actual lookup of the hash should be once per lock object.
741  */
742 static inline struct lock_class *
743 register_lock_class(struct lockdep_map *lock, unsigned int subclass, int force)
744 {
745         struct lockdep_subclass_key *key;
746         struct list_head *hash_head;
747         struct lock_class *class;
748         unsigned long flags;
749
750         class = look_up_lock_class(lock, subclass);
751         if (likely(class))
752                 return class;
753
754         /*
755          * Debug-check: all keys must be persistent!
756          */
757         if (!static_obj(lock->key)) {
758                 debug_locks_off();
759                 printk("INFO: trying to register non-static key.\n");
760                 printk("the code is fine but needs lockdep annotation.\n");
761                 printk("turning off the locking correctness validator.\n");
762                 dump_stack();
763
764                 return NULL;
765         }
766
767         key = lock->key->subkeys + subclass;
768         hash_head = classhashentry(key);
769
770         raw_local_irq_save(flags);
771         if (!graph_lock()) {
772                 raw_local_irq_restore(flags);
773                 return NULL;
774         }
775         /*
776          * We have to do the hash-walk again, to avoid races
777          * with another CPU:
778          */
779         list_for_each_entry(class, hash_head, hash_entry)
780                 if (class->key == key)
781                         goto out_unlock_set;
782         /*
783          * Allocate a new key from the static array, and add it to
784          * the hash:
785          */
786         if (nr_lock_classes >= MAX_LOCKDEP_KEYS) {
787                 if (!debug_locks_off_graph_unlock()) {
788                         raw_local_irq_restore(flags);
789                         return NULL;
790                 }
791                 raw_local_irq_restore(flags);
792
793                 printk("BUG: MAX_LOCKDEP_KEYS too low!\n");
794                 printk("turning off the locking correctness validator.\n");
795                 dump_stack();
796                 return NULL;
797         }
798         class = lock_classes + nr_lock_classes++;
799         debug_atomic_inc(&nr_unused_locks);
800         class->key = key;
801         class->name = lock->name;
802         class->subclass = subclass;
803         INIT_LIST_HEAD(&class->lock_entry);
804         INIT_LIST_HEAD(&class->locks_before);
805         INIT_LIST_HEAD(&class->locks_after);
806         class->name_version = count_matching_names(class);
807         /*
808          * We use RCU's safe list-add method to make
809          * parallel walking of the hash-list safe:
810          */
811         list_add_tail_rcu(&class->hash_entry, hash_head);
812         /*
813          * Add it to the global list of classes:
814          */
815         list_add_tail_rcu(&class->lock_entry, &all_lock_classes);
816
817         if (verbose(class)) {
818                 graph_unlock();
819                 raw_local_irq_restore(flags);
820
821                 printk("\nnew class %p: %s", class->key, class->name);
822                 if (class->name_version > 1)
823                         printk("#%d", class->name_version);
824                 printk("\n");
825                 dump_stack();
826
827                 raw_local_irq_save(flags);
828                 if (!graph_lock()) {
829                         raw_local_irq_restore(flags);
830                         return NULL;
831                 }
832         }
833 out_unlock_set:
834         graph_unlock();
835         raw_local_irq_restore(flags);
836
837         if (!subclass || force)
838                 lock->class_cache = class;
839
840         if (DEBUG_LOCKS_WARN_ON(class->subclass != subclass))
841                 return NULL;
842
843         return class;
844 }
845
846 #ifdef CONFIG_PROVE_LOCKING
847 /*
848  * Allocate a lockdep entry. (assumes the graph_lock held, returns
849  * with NULL on failure)
850  */
851 static struct lock_list *alloc_list_entry(void)
852 {
853         if (nr_list_entries >= MAX_LOCKDEP_ENTRIES) {
854                 if (!debug_locks_off_graph_unlock())
855                         return NULL;
856
857                 printk("BUG: MAX_LOCKDEP_ENTRIES too low!\n");
858                 printk("turning off the locking correctness validator.\n");
859                 dump_stack();
860                 return NULL;
861         }
862         return list_entries + nr_list_entries++;
863 }
864
865 /*
866  * Add a new dependency to the head of the list:
867  */
868 static int add_lock_to_list(struct lock_class *class, struct lock_class *this,
869                             struct list_head *head, unsigned long ip, int distance)
870 {
871         struct lock_list *entry;
872         /*
873          * Lock not present yet - get a new dependency struct and
874          * add it to the list:
875          */
876         entry = alloc_list_entry();
877         if (!entry)
878                 return 0;
879
880         if (!save_trace(&entry->trace))
881                 return 0;
882
883         entry->class = this;
884         entry->distance = distance;
885         /*
886          * Since we never remove from the dependency list, the list can
887          * be walked lockless by other CPUs, it's only allocation
888          * that must be protected by the spinlock. But this also means
889          * we must make new entries visible only once writes to the
890          * entry become visible - hence the RCU op:
891          */
892         list_add_tail_rcu(&entry->entry, head);
893
894         return 1;
895 }
896
897 /*
898  * Recursive, forwards-direction lock-dependency checking, used for
899  * both noncyclic checking and for hardirq-unsafe/softirq-unsafe
900  * checking.
901  *
902  * (to keep the stackframe of the recursive functions small we
903  *  use these global variables, and we also mark various helper
904  *  functions as noinline.)
905  */
906 static struct held_lock *check_source, *check_target;
907
908 /*
909  * Print a dependency chain entry (this is only done when a deadlock
910  * has been detected):
911  */
912 static noinline int
913 print_circular_bug_entry(struct lock_list *target, unsigned int depth)
914 {
915         if (debug_locks_silent)
916                 return 0;
917         printk("\n-> #%u", depth);
918         print_lock_name(target->class);
919         printk(":\n");
920         print_stack_trace(&target->trace, 6);
921
922         return 0;
923 }
924
925 /*
926  * When a circular dependency is detected, print the
927  * header first:
928  */
929 static noinline int
930 print_circular_bug_header(struct lock_list *entry, unsigned int depth)
931 {
932         struct task_struct *curr = current;
933
934         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
935                 return 0;
936
937         printk("\n=======================================================\n");
938         printk(  "[ INFO: possible circular locking dependency detected ]\n");
939         print_kernel_version();
940         printk(  "-------------------------------------------------------\n");
941         printk("%s/%d is trying to acquire lock:\n",
942                 curr->comm, task_pid_nr(curr));
943         print_lock(check_source);
944         printk("\nbut task is already holding lock:\n");
945         print_lock(check_target);
946         printk("\nwhich lock already depends on the new lock.\n\n");
947         printk("\nthe existing dependency chain (in reverse order) is:\n");
948
949         print_circular_bug_entry(entry, depth);
950
951         return 0;
952 }
953
954 static noinline int print_circular_bug_tail(void)
955 {
956         struct task_struct *curr = current;
957         struct lock_list this;
958
959         if (debug_locks_silent)
960                 return 0;
961
962         this.class = hlock_class(check_source);
963         if (!save_trace(&this.trace))
964                 return 0;
965
966         print_circular_bug_entry(&this, 0);
967
968         printk("\nother info that might help us debug this:\n\n");
969         lockdep_print_held_locks(curr);
970
971         printk("\nstack backtrace:\n");
972         dump_stack();
973
974         return 0;
975 }
976
977 #define RECURSION_LIMIT 40
978
979 static int noinline print_infinite_recursion_bug(void)
980 {
981         if (!debug_locks_off_graph_unlock())
982                 return 0;
983
984         WARN_ON(1);
985
986         return 0;
987 }
988
989 unsigned long __lockdep_count_forward_deps(struct lock_class *class,
990                                            unsigned int depth)
991 {
992         struct lock_list *entry;
993         unsigned long ret = 1;
994
995         if (lockdep_dependency_visit(class, depth))
996                 return 0;
997
998         /*
999          * Recurse this class's dependency list:
1000          */
1001         list_for_each_entry(entry, &class->locks_after, entry)
1002                 ret += __lockdep_count_forward_deps(entry->class, depth + 1);
1003
1004         return ret;
1005 }
1006
1007 unsigned long lockdep_count_forward_deps(struct lock_class *class)
1008 {
1009         unsigned long ret, flags;
1010
1011         local_irq_save(flags);
1012         __raw_spin_lock(&lockdep_lock);
1013         ret = __lockdep_count_forward_deps(class, 0);
1014         __raw_spin_unlock(&lockdep_lock);
1015         local_irq_restore(flags);
1016
1017         return ret;
1018 }
1019
1020 unsigned long __lockdep_count_backward_deps(struct lock_class *class,
1021                                             unsigned int depth)
1022 {
1023         struct lock_list *entry;
1024         unsigned long ret = 1;
1025
1026         if (lockdep_dependency_visit(class, depth))
1027                 return 0;
1028         /*
1029          * Recurse this class's dependency list:
1030          */
1031         list_for_each_entry(entry, &class->locks_before, entry)
1032                 ret += __lockdep_count_backward_deps(entry->class, depth + 1);
1033
1034         return ret;
1035 }
1036
1037 unsigned long lockdep_count_backward_deps(struct lock_class *class)
1038 {
1039         unsigned long ret, flags;
1040
1041         local_irq_save(flags);
1042         __raw_spin_lock(&lockdep_lock);
1043         ret = __lockdep_count_backward_deps(class, 0);
1044         __raw_spin_unlock(&lockdep_lock);
1045         local_irq_restore(flags);
1046
1047         return ret;
1048 }
1049
1050 /*
1051  * Prove that the dependency graph starting at <entry> can not
1052  * lead to <target>. Print an error and return 0 if it does.
1053  */
1054 static noinline int
1055 check_noncircular(struct lock_class *source, unsigned int depth)
1056 {
1057         struct lock_list *entry;
1058
1059         if (lockdep_dependency_visit(source, depth))
1060                 return 1;
1061
1062         debug_atomic_inc(&nr_cyclic_check_recursions);
1063         if (depth > max_recursion_depth)
1064                 max_recursion_depth = depth;
1065         if (depth >= RECURSION_LIMIT)
1066                 return print_infinite_recursion_bug();
1067         /*
1068          * Check this lock's dependency list:
1069          */
1070         list_for_each_entry(entry, &source->locks_after, entry) {
1071                 if (entry->class == hlock_class(check_target))
1072                         return print_circular_bug_header(entry, depth+1);
1073                 debug_atomic_inc(&nr_cyclic_checks);
1074                 if (!check_noncircular(entry->class, depth+1))
1075                         return print_circular_bug_entry(entry, depth+1);
1076         }
1077         return 1;
1078 }
1079
1080 #if defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING)
1081 /*
1082  * Forwards and backwards subgraph searching, for the purposes of
1083  * proving that two subgraphs can be connected by a new dependency
1084  * without creating any illegal irq-safe -> irq-unsafe lock dependency.
1085  */
1086 static enum lock_usage_bit find_usage_bit;
1087 static struct lock_class *forwards_match, *backwards_match;
1088
1089 /*
1090  * Find a node in the forwards-direction dependency sub-graph starting
1091  * at <source> that matches <find_usage_bit>.
1092  *
1093  * Return 2 if such a node exists in the subgraph, and put that node
1094  * into <forwards_match>.
1095  *
1096  * Return 1 otherwise and keep <forwards_match> unchanged.
1097  * Return 0 on error.
1098  */
1099 static noinline int
1100 find_usage_forwards(struct lock_class *source, unsigned int depth)
1101 {
1102         struct lock_list *entry;
1103         int ret;
1104
1105         if (lockdep_dependency_visit(source, depth))
1106                 return 1;
1107
1108         if (depth > max_recursion_depth)
1109                 max_recursion_depth = depth;
1110         if (depth >= RECURSION_LIMIT)
1111                 return print_infinite_recursion_bug();
1112
1113         debug_atomic_inc(&nr_find_usage_forwards_checks);
1114         if (source->usage_mask & (1 << find_usage_bit)) {
1115                 forwards_match = source;
1116                 return 2;
1117         }
1118
1119         /*
1120          * Check this lock's dependency list:
1121          */
1122         list_for_each_entry(entry, &source->locks_after, entry) {
1123                 debug_atomic_inc(&nr_find_usage_forwards_recursions);
1124                 ret = find_usage_forwards(entry->class, depth+1);
1125                 if (ret == 2 || ret == 0)
1126                         return ret;
1127         }
1128         return 1;
1129 }
1130
1131 /*
1132  * Find a node in the backwards-direction dependency sub-graph starting
1133  * at <source> that matches <find_usage_bit>.
1134  *
1135  * Return 2 if such a node exists in the subgraph, and put that node
1136  * into <backwards_match>.
1137  *
1138  * Return 1 otherwise and keep <backwards_match> unchanged.
1139  * Return 0 on error.
1140  */
1141 static noinline int
1142 find_usage_backwards(struct lock_class *source, unsigned int depth)
1143 {
1144         struct lock_list *entry;
1145         int ret;
1146
1147         if (lockdep_dependency_visit(source, depth))
1148                 return 1;
1149
1150         if (!__raw_spin_is_locked(&lockdep_lock))
1151                 return DEBUG_LOCKS_WARN_ON(1);
1152
1153         if (depth > max_recursion_depth)
1154                 max_recursion_depth = depth;
1155         if (depth >= RECURSION_LIMIT)
1156                 return print_infinite_recursion_bug();
1157
1158         debug_atomic_inc(&nr_find_usage_backwards_checks);
1159         if (source->usage_mask & (1 << find_usage_bit)) {
1160                 backwards_match = source;
1161                 return 2;
1162         }
1163
1164         if (!source && debug_locks_off_graph_unlock()) {
1165                 WARN_ON(1);
1166                 return 0;
1167         }
1168
1169         /*
1170          * Check this lock's dependency list:
1171          */
1172         list_for_each_entry(entry, &source->locks_before, entry) {
1173                 debug_atomic_inc(&nr_find_usage_backwards_recursions);
1174                 ret = find_usage_backwards(entry->class, depth+1);
1175                 if (ret == 2 || ret == 0)
1176                         return ret;
1177         }
1178         return 1;
1179 }
1180
1181 static int
1182 print_bad_irq_dependency(struct task_struct *curr,
1183                          struct held_lock *prev,
1184                          struct held_lock *next,
1185                          enum lock_usage_bit bit1,
1186                          enum lock_usage_bit bit2,
1187                          const char *irqclass)
1188 {
1189         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1190                 return 0;
1191
1192         printk("\n======================================================\n");
1193         printk(  "[ INFO: %s-safe -> %s-unsafe lock order detected ]\n",
1194                 irqclass, irqclass);
1195         print_kernel_version();
1196         printk(  "------------------------------------------------------\n");
1197         printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] is trying to acquire:\n",
1198                 curr->comm, task_pid_nr(curr),
1199                 curr->hardirq_context, hardirq_count() >> HARDIRQ_SHIFT,
1200                 curr->softirq_context, softirq_count() >> SOFTIRQ_SHIFT,
1201                 curr->hardirqs_enabled,
1202                 curr->softirqs_enabled);
1203         print_lock(next);
1204
1205         printk("\nand this task is already holding:\n");
1206         print_lock(prev);
1207         printk("which would create a new lock dependency:\n");
1208         print_lock_name(hlock_class(prev));
1209         printk(" ->");
1210         print_lock_name(hlock_class(next));
1211         printk("\n");
1212
1213         printk("\nbut this new dependency connects a %s-irq-safe lock:\n",
1214                 irqclass);
1215         print_lock_name(backwards_match);
1216         printk("\n... which became %s-irq-safe at:\n", irqclass);
1217
1218         print_stack_trace(backwards_match->usage_traces + bit1, 1);
1219
1220         printk("\nto a %s-irq-unsafe lock:\n", irqclass);
1221         print_lock_name(forwards_match);
1222         printk("\n... which became %s-irq-unsafe at:\n", irqclass);
1223         printk("...");
1224
1225         print_stack_trace(forwards_match->usage_traces + bit2, 1);
1226
1227         printk("\nother info that might help us debug this:\n\n");
1228         lockdep_print_held_locks(curr);
1229
1230         printk("\nthe %s-irq-safe lock's dependencies:\n", irqclass);
1231         print_lock_dependencies(backwards_match, 0);
1232
1233         printk("\nthe %s-irq-unsafe lock's dependencies:\n", irqclass);
1234         print_lock_dependencies(forwards_match, 0);
1235
1236         printk("\nstack backtrace:\n");
1237         dump_stack();
1238
1239         return 0;
1240 }
1241
1242 static int
1243 check_usage(struct task_struct *curr, struct held_lock *prev,
1244             struct held_lock *next, enum lock_usage_bit bit_backwards,
1245             enum lock_usage_bit bit_forwards, const char *irqclass)
1246 {
1247         int ret;
1248
1249         find_usage_bit = bit_backwards;
1250         /* fills in <backwards_match> */
1251         ret = find_usage_backwards(hlock_class(prev), 0);
1252         if (!ret || ret == 1)
1253                 return ret;
1254
1255         find_usage_bit = bit_forwards;
1256         ret = find_usage_forwards(hlock_class(next), 0);
1257         if (!ret || ret == 1)
1258                 return ret;
1259         /* ret == 2 */
1260         return print_bad_irq_dependency(curr, prev, next,
1261                         bit_backwards, bit_forwards, irqclass);
1262 }
1263
1264 static const char *state_names[] = {
1265 #define LOCKDEP_STATE(__STATE) \
1266         __stringify(__STATE),
1267 #include "lockdep_states.h"
1268 #undef LOCKDEP_STATE
1269 };
1270
1271 static const char *state_rnames[] = {
1272 #define LOCKDEP_STATE(__STATE) \
1273         __stringify(__STATE)"-READ",
1274 #include "lockdep_states.h"
1275 #undef LOCKDEP_STATE
1276 };
1277
1278 static inline const char *state_name(enum lock_usage_bit bit)
1279 {
1280         return (bit & 1) ? state_rnames[bit >> 2] : state_names[bit >> 2];
1281 }
1282
1283 static int exclusive_bit(int new_bit)
1284 {
1285         /*
1286          * USED_IN
1287          * USED_IN_READ
1288          * ENABLED
1289          * ENABLED_READ
1290          *
1291          * bit 0 - write/read
1292          * bit 1 - used_in/enabled
1293          * bit 2+  state
1294          */
1295
1296         int state = new_bit & ~3;
1297         int dir = new_bit & 2;
1298
1299         /*
1300          * keep state, bit flip the direction and strip read.
1301          */
1302         return state | (dir ^ 2);
1303 }
1304
1305 static int check_irq_usage(struct task_struct *curr, struct held_lock *prev,
1306                            struct held_lock *next, enum lock_usage_bit bit)
1307 {
1308         /*
1309          * Prove that the new dependency does not connect a hardirq-safe
1310          * lock with a hardirq-unsafe lock - to achieve this we search
1311          * the backwards-subgraph starting at <prev>, and the
1312          * forwards-subgraph starting at <next>:
1313          */
1314         if (!check_usage(curr, prev, next, bit,
1315                            exclusive_bit(bit), state_name(bit)))
1316                 return 0;
1317
1318         bit++; /* _READ */
1319
1320         /*
1321          * Prove that the new dependency does not connect a hardirq-safe-read
1322          * lock with a hardirq-unsafe lock - to achieve this we search
1323          * the backwards-subgraph starting at <prev>, and the
1324          * forwards-subgraph starting at <next>:
1325          */
1326         if (!check_usage(curr, prev, next, bit,
1327                            exclusive_bit(bit), state_name(bit)))
1328                 return 0;
1329
1330         return 1;
1331 }
1332
1333 static int
1334 check_prev_add_irq(struct task_struct *curr, struct held_lock *prev,
1335                 struct held_lock *next)
1336 {
1337 #define LOCKDEP_STATE(__STATE)                                          \
1338         if (!check_irq_usage(curr, prev, next, LOCK_USED_IN_##__STATE)) \
1339                 return 0;
1340 #include "lockdep_states.h"
1341 #undef LOCKDEP_STATE
1342
1343         return 1;
1344 }
1345
1346 static void inc_chains(void)
1347 {
1348         if (current->hardirq_context)
1349                 nr_hardirq_chains++;
1350         else {
1351                 if (current->softirq_context)
1352                         nr_softirq_chains++;
1353                 else
1354                         nr_process_chains++;
1355         }
1356 }
1357
1358 #else
1359
1360 static inline int
1361 check_prev_add_irq(struct task_struct *curr, struct held_lock *prev,
1362                 struct held_lock *next)
1363 {
1364         return 1;
1365 }
1366
1367 static inline void inc_chains(void)
1368 {
1369         nr_process_chains++;
1370 }
1371
1372 #endif
1373
1374 static int
1375 print_deadlock_bug(struct task_struct *curr, struct held_lock *prev,
1376                    struct held_lock *next)
1377 {
1378         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1379                 return 0;
1380
1381         printk("\n=============================================\n");
1382         printk(  "[ INFO: possible recursive locking detected ]\n");
1383         print_kernel_version();
1384         printk(  "---------------------------------------------\n");
1385         printk("%s/%d is trying to acquire lock:\n",
1386                 curr->comm, task_pid_nr(curr));
1387         print_lock(next);
1388         printk("\nbut task is already holding lock:\n");
1389         print_lock(prev);
1390
1391         printk("\nother info that might help us debug this:\n");
1392         lockdep_print_held_locks(curr);
1393
1394         printk("\nstack backtrace:\n");
1395         dump_stack();
1396
1397         return 0;
1398 }
1399
1400 /*
1401  * Check whether we are holding such a class already.
1402  *
1403  * (Note that this has to be done separately, because the graph cannot
1404  * detect such classes of deadlocks.)
1405  *
1406  * Returns: 0 on deadlock detected, 1 on OK, 2 on recursive read
1407  */
1408 static int
1409 check_deadlock(struct task_struct *curr, struct held_lock *next,
1410                struct lockdep_map *next_instance, int read)
1411 {
1412         struct held_lock *prev;
1413         struct held_lock *nest = NULL;
1414         int i;
1415
1416         for (i = 0; i < curr->lockdep_depth; i++) {
1417                 prev = curr->held_locks + i;
1418
1419                 if (prev->instance == next->nest_lock)
1420                         nest = prev;
1421
1422                 if (hlock_class(prev) != hlock_class(next))
1423                         continue;
1424
1425                 /*
1426                  * Allow read-after-read recursion of the same
1427                  * lock class (i.e. read_lock(lock)+read_lock(lock)):
1428                  */
1429                 if ((read == 2) && prev->read)
1430                         return 2;
1431
1432                 /*
1433                  * We're holding the nest_lock, which serializes this lock's
1434                  * nesting behaviour.
1435                  */
1436                 if (nest)
1437                         return 2;
1438
1439                 return print_deadlock_bug(curr, prev, next);
1440         }
1441         return 1;
1442 }
1443
1444 /*
1445  * There was a chain-cache miss, and we are about to add a new dependency
1446  * to a previous lock. We recursively validate the following rules:
1447  *
1448  *  - would the adding of the <prev> -> <next> dependency create a
1449  *    circular dependency in the graph? [== circular deadlock]
1450  *
1451  *  - does the new prev->next dependency connect any hardirq-safe lock
1452  *    (in the full backwards-subgraph starting at <prev>) with any
1453  *    hardirq-unsafe lock (in the full forwards-subgraph starting at
1454  *    <next>)? [== illegal lock inversion with hardirq contexts]
1455  *
1456  *  - does the new prev->next dependency connect any softirq-safe lock
1457  *    (in the full backwards-subgraph starting at <prev>) with any
1458  *    softirq-unsafe lock (in the full forwards-subgraph starting at
1459  *    <next>)? [== illegal lock inversion with softirq contexts]
1460  *
1461  * any of these scenarios could lead to a deadlock.
1462  *
1463  * Then if all the validations pass, we add the forwards and backwards
1464  * dependency.
1465  */
1466 static int
1467 check_prev_add(struct task_struct *curr, struct held_lock *prev,
1468                struct held_lock *next, int distance)
1469 {
1470         struct lock_list *entry;
1471         int ret;
1472
1473         /*
1474          * Prove that the new <prev> -> <next> dependency would not
1475          * create a circular dependency in the graph. (We do this by
1476          * forward-recursing into the graph starting at <next>, and
1477          * checking whether we can reach <prev>.)
1478          *
1479          * We are using global variables to control the recursion, to
1480          * keep the stackframe size of the recursive functions low:
1481          */
1482         check_source = next;
1483         check_target = prev;
1484         if (!(check_noncircular(hlock_class(next), 0)))
1485                 return print_circular_bug_tail();
1486
1487         if (!check_prev_add_irq(curr, prev, next))
1488                 return 0;
1489
1490         /*
1491          * For recursive read-locks we do all the dependency checks,
1492          * but we dont store read-triggered dependencies (only
1493          * write-triggered dependencies). This ensures that only the
1494          * write-side dependencies matter, and that if for example a
1495          * write-lock never takes any other locks, then the reads are
1496          * equivalent to a NOP.
1497          */
1498         if (next->read == 2 || prev->read == 2)
1499                 return 1;
1500         /*
1501          * Is the <prev> -> <next> dependency already present?
1502          *
1503          * (this may occur even though this is a new chain: consider
1504          *  e.g. the L1 -> L2 -> L3 -> L4 and the L5 -> L1 -> L2 -> L3
1505          *  chains - the second one will be new, but L1 already has
1506          *  L2 added to its dependency list, due to the first chain.)
1507          */
1508         list_for_each_entry(entry, &hlock_class(prev)->locks_after, entry) {
1509                 if (entry->class == hlock_class(next)) {
1510                         if (distance == 1)
1511                                 entry->distance = 1;
1512                         return 2;
1513                 }
1514         }
1515
1516         /*
1517          * Ok, all validations passed, add the new lock
1518          * to the previous lock's dependency list:
1519          */
1520         ret = add_lock_to_list(hlock_class(prev), hlock_class(next),
1521                                &hlock_class(prev)->locks_after,
1522                                next->acquire_ip, distance);
1523
1524         if (!ret)
1525                 return 0;
1526
1527         ret = add_lock_to_list(hlock_class(next), hlock_class(prev),
1528                                &hlock_class(next)->locks_before,
1529                                next->acquire_ip, distance);
1530         if (!ret)
1531                 return 0;
1532
1533         /*
1534          * Debugging printouts:
1535          */
1536         if (verbose(hlock_class(prev)) || verbose(hlock_class(next))) {
1537                 graph_unlock();
1538                 printk("\n new dependency: ");
1539                 print_lock_name(hlock_class(prev));
1540                 printk(" => ");
1541                 print_lock_name(hlock_class(next));
1542                 printk("\n");
1543                 dump_stack();
1544                 return graph_lock();
1545         }
1546         return 1;
1547 }
1548
1549 /*
1550  * Add the dependency to all directly-previous locks that are 'relevant'.
1551  * The ones that are relevant are (in increasing distance from curr):
1552  * all consecutive trylock entries and the final non-trylock entry - or
1553  * the end of this context's lock-chain - whichever comes first.
1554  */
1555 static int
1556 check_prevs_add(struct task_struct *curr, struct held_lock *next)
1557 {
1558         int depth = curr->lockdep_depth;
1559         struct held_lock *hlock;
1560
1561         /*
1562          * Debugging checks.
1563          *
1564          * Depth must not be zero for a non-head lock:
1565          */
1566         if (!depth)
1567                 goto out_bug;
1568         /*
1569          * At least two relevant locks must exist for this
1570          * to be a head:
1571          */
1572         if (curr->held_locks[depth].irq_context !=
1573                         curr->held_locks[depth-1].irq_context)
1574                 goto out_bug;
1575
1576         for (;;) {
1577                 int distance = curr->lockdep_depth - depth + 1;
1578                 hlock = curr->held_locks + depth-1;
1579                 /*
1580                  * Only non-recursive-read entries get new dependencies
1581                  * added:
1582                  */
1583                 if (hlock->read != 2) {
1584                         if (!check_prev_add(curr, hlock, next, distance))
1585                                 return 0;
1586                         /*
1587                          * Stop after the first non-trylock entry,
1588                          * as non-trylock entries have added their
1589                          * own direct dependencies already, so this
1590                          * lock is connected to them indirectly:
1591                          */
1592                         if (!hlock->trylock)
1593                                 break;
1594                 }
1595                 depth--;
1596                 /*
1597                  * End of lock-stack?
1598                  */
1599                 if (!depth)
1600                         break;
1601                 /*
1602                  * Stop the search if we cross into another context:
1603                  */
1604                 if (curr->held_locks[depth].irq_context !=
1605                                 curr->held_locks[depth-1].irq_context)
1606                         break;
1607         }
1608         return 1;
1609 out_bug:
1610         if (!debug_locks_off_graph_unlock())
1611                 return 0;
1612
1613         WARN_ON(1);
1614
1615         return 0;
1616 }
1617
1618 unsigned long nr_lock_chains;
1619 struct lock_chain lock_chains[MAX_LOCKDEP_CHAINS];
1620 int nr_chain_hlocks;
1621 static u16 chain_hlocks[MAX_LOCKDEP_CHAIN_HLOCKS];
1622
1623 struct lock_class *lock_chain_get_class(struct lock_chain *chain, int i)
1624 {
1625         return lock_classes + chain_hlocks[chain->base + i];
1626 }
1627
1628 /*
1629  * Look up a dependency chain. If the key is not present yet then
1630  * add it and return 1 - in this case the new dependency chain is
1631  * validated. If the key is already hashed, return 0.
1632  * (On return with 1 graph_lock is held.)
1633  */
1634 static inline int lookup_chain_cache(struct task_struct *curr,
1635                                      struct held_lock *hlock,
1636                                      u64 chain_key)
1637 {
1638         struct lock_class *class = hlock_class(hlock);
1639         struct list_head *hash_head = chainhashentry(chain_key);
1640         struct lock_chain *chain;
1641         struct held_lock *hlock_curr, *hlock_next;
1642         int i, j, n, cn;
1643
1644         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
1645                 return 0;
1646         /*
1647          * We can walk it lock-free, because entries only get added
1648          * to the hash:
1649          */
1650         list_for_each_entry(chain, hash_head, entry) {
1651                 if (chain->chain_key == chain_key) {
1652 cache_hit:
1653                         debug_atomic_inc(&chain_lookup_hits);
1654                         if (very_verbose(class))
1655                                 printk("\nhash chain already cached, key: "
1656                                         "%016Lx tail class: [%p] %s\n",
1657                                         (unsigned long long)chain_key,
1658                                         class->key, class->name);
1659                         return 0;
1660                 }
1661         }
1662         if (very_verbose(class))
1663                 printk("\nnew hash chain, key: %016Lx tail class: [%p] %s\n",
1664                         (unsigned long long)chain_key, class->key, class->name);
1665         /*
1666          * Allocate a new chain entry from the static array, and add
1667          * it to the hash:
1668          */
1669         if (!graph_lock())
1670                 return 0;
1671         /*
1672          * We have to walk the chain again locked - to avoid duplicates:
1673          */
1674         list_for_each_entry(chain, hash_head, entry) {
1675                 if (chain->chain_key == chain_key) {
1676                         graph_unlock();
1677                         goto cache_hit;
1678                 }
1679         }
1680         if (unlikely(nr_lock_chains >= MAX_LOCKDEP_CHAINS)) {
1681                 if (!debug_locks_off_graph_unlock())
1682                         return 0;
1683
1684                 printk("BUG: MAX_LOCKDEP_CHAINS too low!\n");
1685                 printk("turning off the locking correctness validator.\n");
1686                 dump_stack();
1687                 return 0;
1688         }
1689         chain = lock_chains + nr_lock_chains++;
1690         chain->chain_key = chain_key;
1691         chain->irq_context = hlock->irq_context;
1692         /* Find the first held_lock of current chain */
1693         hlock_next = hlock;
1694         for (i = curr->lockdep_depth - 1; i >= 0; i--) {
1695                 hlock_curr = curr->held_locks + i;
1696                 if (hlock_curr->irq_context != hlock_next->irq_context)
1697                         break;
1698                 hlock_next = hlock;
1699         }
1700         i++;
1701         chain->depth = curr->lockdep_depth + 1 - i;
1702         cn = nr_chain_hlocks;
1703         while (cn + chain->depth <= MAX_LOCKDEP_CHAIN_HLOCKS) {
1704                 n = cmpxchg(&nr_chain_hlocks, cn, cn + chain->depth);
1705                 if (n == cn)
1706                         break;
1707                 cn = n;
1708         }
1709         if (likely(cn + chain->depth <= MAX_LOCKDEP_CHAIN_HLOCKS)) {
1710                 chain->base = cn;
1711                 for (j = 0; j < chain->depth - 1; j++, i++) {
1712                         int lock_id = curr->held_locks[i].class_idx - 1;
1713                         chain_hlocks[chain->base + j] = lock_id;
1714                 }
1715                 chain_hlocks[chain->base + j] = class - lock_classes;
1716         }
1717         list_add_tail_rcu(&chain->entry, hash_head);
1718         debug_atomic_inc(&chain_lookup_misses);
1719         inc_chains();
1720
1721         return 1;
1722 }
1723
1724 static int validate_chain(struct task_struct *curr, struct lockdep_map *lock,
1725                 struct held_lock *hlock, int chain_head, u64 chain_key)
1726 {
1727         /*
1728          * Trylock needs to maintain the stack of held locks, but it
1729          * does not add new dependencies, because trylock can be done
1730          * in any order.
1731          *
1732          * We look up the chain_key and do the O(N^2) check and update of
1733          * the dependencies only if this is a new dependency chain.
1734          * (If lookup_chain_cache() returns with 1 it acquires
1735          * graph_lock for us)
1736          */
1737         if (!hlock->trylock && (hlock->check == 2) &&
1738             lookup_chain_cache(curr, hlock, chain_key)) {
1739                 /*
1740                  * Check whether last held lock:
1741                  *
1742                  * - is irq-safe, if this lock is irq-unsafe
1743                  * - is softirq-safe, if this lock is hardirq-unsafe
1744                  *
1745                  * And check whether the new lock's dependency graph
1746                  * could lead back to the previous lock.
1747                  *
1748                  * any of these scenarios could lead to a deadlock. If
1749                  * All validations
1750                  */
1751                 int ret = check_deadlock(curr, hlock, lock, hlock->read);
1752
1753                 if (!ret)
1754                         return 0;
1755                 /*
1756                  * Mark recursive read, as we jump over it when
1757                  * building dependencies (just like we jump over
1758                  * trylock entries):
1759                  */
1760                 if (ret == 2)
1761                         hlock->read = 2;
1762                 /*
1763                  * Add dependency only if this lock is not the head
1764                  * of the chain, and if it's not a secondary read-lock:
1765                  */
1766                 if (!chain_head && ret != 2)
1767                         if (!check_prevs_add(curr, hlock))
1768                                 return 0;
1769                 graph_unlock();
1770         } else
1771                 /* after lookup_chain_cache(): */
1772                 if (unlikely(!debug_locks))
1773                         return 0;
1774
1775         return 1;
1776 }
1777 #else
1778 static inline int validate_chain(struct task_struct *curr,
1779                 struct lockdep_map *lock, struct held_lock *hlock,
1780                 int chain_head, u64 chain_key)
1781 {
1782         return 1;
1783 }
1784 #endif
1785
1786 /*
1787  * We are building curr_chain_key incrementally, so double-check
1788  * it from scratch, to make sure that it's done correctly:
1789  */
1790 static void check_chain_key(struct task_struct *curr)
1791 {
1792 #ifdef CONFIG_DEBUG_LOCKDEP
1793         struct held_lock *hlock, *prev_hlock = NULL;
1794         unsigned int i, id;
1795         u64 chain_key = 0;
1796
1797         for (i = 0; i < curr->lockdep_depth; i++) {
1798                 hlock = curr->held_locks + i;
1799                 if (chain_key != hlock->prev_chain_key) {
1800                         debug_locks_off();
1801                         WARN(1, "hm#1, depth: %u [%u], %016Lx != %016Lx\n",
1802                                 curr->lockdep_depth, i,
1803                                 (unsigned long long)chain_key,
1804                                 (unsigned long long)hlock->prev_chain_key);
1805                         return;
1806                 }
1807                 id = hlock->class_idx - 1;
1808                 if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
1809                         return;
1810
1811                 if (prev_hlock && (prev_hlock->irq_context !=
1812                                                         hlock->irq_context))
1813                         chain_key = 0;
1814                 chain_key = iterate_chain_key(chain_key, id);
1815                 prev_hlock = hlock;
1816         }
1817         if (chain_key != curr->curr_chain_key) {
1818                 debug_locks_off();
1819                 WARN(1, "hm#2, depth: %u [%u], %016Lx != %016Lx\n",
1820                         curr->lockdep_depth, i,
1821                         (unsigned long long)chain_key,
1822                         (unsigned long long)curr->curr_chain_key);
1823         }
1824 #endif
1825 }
1826
1827 static int
1828 print_usage_bug(struct task_struct *curr, struct held_lock *this,
1829                 enum lock_usage_bit prev_bit, enum lock_usage_bit new_bit)
1830 {
1831         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1832                 return 0;
1833
1834         printk("\n=================================\n");
1835         printk(  "[ INFO: inconsistent lock state ]\n");
1836         print_kernel_version();
1837         printk(  "---------------------------------\n");
1838
1839         printk("inconsistent {%s} -> {%s} usage.\n",
1840                 usage_str[prev_bit], usage_str[new_bit]);
1841
1842         printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] takes:\n",
1843                 curr->comm, task_pid_nr(curr),
1844                 trace_hardirq_context(curr), hardirq_count() >> HARDIRQ_SHIFT,
1845                 trace_softirq_context(curr), softirq_count() >> SOFTIRQ_SHIFT,
1846                 trace_hardirqs_enabled(curr),
1847                 trace_softirqs_enabled(curr));
1848         print_lock(this);
1849
1850         printk("{%s} state was registered at:\n", usage_str[prev_bit]);
1851         print_stack_trace(hlock_class(this)->usage_traces + prev_bit, 1);
1852
1853         print_irqtrace_events(curr);
1854         printk("\nother info that might help us debug this:\n");
1855         lockdep_print_held_locks(curr);
1856
1857         printk("\nstack backtrace:\n");
1858         dump_stack();
1859
1860         return 0;
1861 }
1862
1863 /*
1864  * Print out an error if an invalid bit is set:
1865  */
1866 static inline int
1867 valid_state(struct task_struct *curr, struct held_lock *this,
1868             enum lock_usage_bit new_bit, enum lock_usage_bit bad_bit)
1869 {
1870         if (unlikely(hlock_class(this)->usage_mask & (1 << bad_bit)))
1871                 return print_usage_bug(curr, this, bad_bit, new_bit);
1872         return 1;
1873 }
1874
1875 static int mark_lock(struct task_struct *curr, struct held_lock *this,
1876                      enum lock_usage_bit new_bit);
1877
1878 #if defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING)
1879
1880 /*
1881  * print irq inversion bug:
1882  */
1883 static int
1884 print_irq_inversion_bug(struct task_struct *curr, struct lock_class *other,
1885                         struct held_lock *this, int forwards,
1886                         const char *irqclass)
1887 {
1888         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1889                 return 0;
1890
1891         printk("\n=========================================================\n");
1892         printk(  "[ INFO: possible irq lock inversion dependency detected ]\n");
1893         print_kernel_version();
1894         printk(  "---------------------------------------------------------\n");
1895         printk("%s/%d just changed the state of lock:\n",
1896                 curr->comm, task_pid_nr(curr));
1897         print_lock(this);
1898         if (forwards)
1899                 printk("but this lock took another, %s-unsafe lock in the past:\n", irqclass);
1900         else
1901                 printk("but this lock was taken by another, %s-safe lock in the past:\n", irqclass);
1902         print_lock_name(other);
1903         printk("\n\nand interrupts could create inverse lock ordering between them.\n\n");
1904
1905         printk("\nother info that might help us debug this:\n");
1906         lockdep_print_held_locks(curr);
1907
1908         printk("\nthe first lock's dependencies:\n");
1909         print_lock_dependencies(hlock_class(this), 0);
1910
1911         printk("\nthe second lock's dependencies:\n");
1912         print_lock_dependencies(other, 0);
1913
1914         printk("\nstack backtrace:\n");
1915         dump_stack();
1916
1917         return 0;
1918 }
1919
1920 /*
1921  * Prove that in the forwards-direction subgraph starting at <this>
1922  * there is no lock matching <mask>:
1923  */
1924 static int
1925 check_usage_forwards(struct task_struct *curr, struct held_lock *this,
1926                      enum lock_usage_bit bit, const char *irqclass)
1927 {
1928         int ret;
1929
1930         find_usage_bit = bit;
1931         /* fills in <forwards_match> */
1932         ret = find_usage_forwards(hlock_class(this), 0);
1933         if (!ret || ret == 1)
1934                 return ret;
1935
1936         return print_irq_inversion_bug(curr, forwards_match, this, 1, irqclass);
1937 }
1938
1939 /*
1940  * Prove that in the backwards-direction subgraph starting at <this>
1941  * there is no lock matching <mask>:
1942  */
1943 static int
1944 check_usage_backwards(struct task_struct *curr, struct held_lock *this,
1945                       enum lock_usage_bit bit, const char *irqclass)
1946 {
1947         int ret;
1948
1949         find_usage_bit = bit;
1950         /* fills in <backwards_match> */
1951         ret = find_usage_backwards(hlock_class(this), 0);
1952         if (!ret || ret == 1)
1953                 return ret;
1954
1955         return print_irq_inversion_bug(curr, backwards_match, this, 0, irqclass);
1956 }
1957
1958 void print_irqtrace_events(struct task_struct *curr)
1959 {
1960         printk("irq event stamp: %u\n", curr->irq_events);
1961         printk("hardirqs last  enabled at (%u): ", curr->hardirq_enable_event);
1962         print_ip_sym(curr->hardirq_enable_ip);
1963         printk("hardirqs last disabled at (%u): ", curr->hardirq_disable_event);
1964         print_ip_sym(curr->hardirq_disable_ip);
1965         printk("softirqs last  enabled at (%u): ", curr->softirq_enable_event);
1966         print_ip_sym(curr->softirq_enable_ip);
1967         printk("softirqs last disabled at (%u): ", curr->softirq_disable_event);
1968         print_ip_sym(curr->softirq_disable_ip);
1969 }
1970
1971 static int HARDIRQ_verbose(struct lock_class *class)
1972 {
1973 #if HARDIRQ_VERBOSE
1974         return class_filter(class);
1975 #endif
1976         return 0;
1977 }
1978
1979 static int SOFTIRQ_verbose(struct lock_class *class)
1980 {
1981 #if SOFTIRQ_VERBOSE
1982         return class_filter(class);
1983 #endif
1984         return 0;
1985 }
1986
1987 static int RECLAIM_FS_verbose(struct lock_class *class)
1988 {
1989 #if RECLAIM_VERBOSE
1990         return class_filter(class);
1991 #endif
1992         return 0;
1993 }
1994
1995 #define STRICT_READ_CHECKS      1
1996
1997 static int (*state_verbose_f[])(struct lock_class *class) = {
1998 #define LOCKDEP_STATE(__STATE) \
1999         __STATE##_verbose,
2000 #include "lockdep_states.h"
2001 #undef LOCKDEP_STATE
2002 };
2003
2004 static inline int state_verbose(enum lock_usage_bit bit,
2005                                 struct lock_class *class)
2006 {
2007         return state_verbose_f[bit >> 2](class);
2008 }
2009
2010 typedef int (*check_usage_f)(struct task_struct *, struct held_lock *,
2011                              enum lock_usage_bit bit, const char *name);
2012
2013 static int
2014 mark_lock_irq(struct task_struct *curr, struct held_lock *this,
2015                 enum lock_usage_bit new_bit)
2016 {
2017         int excl_bit = exclusive_bit(new_bit);
2018         int read = new_bit & 1;
2019         int dir = new_bit & 2;
2020
2021         /*
2022          * mark USED_IN has to look forwards -- to ensure no dependency
2023          * has ENABLED state, which would allow recursion deadlocks.
2024          *
2025          * mark ENABLED has to look backwards -- to ensure no dependee
2026          * has USED_IN state, which, again, would allow  recursion deadlocks.
2027          */
2028         check_usage_f usage = dir ?
2029                 check_usage_backwards : check_usage_forwards;
2030
2031         /*
2032          * Validate that this particular lock does not have conflicting
2033          * usage states.
2034          */
2035         if (!valid_state(curr, this, new_bit, excl_bit))
2036                 return 0;
2037
2038         /*
2039          * Validate that the lock dependencies don't have conflicting usage
2040          * states.
2041          */
2042         if ((!read || !dir || STRICT_READ_CHECKS) &&
2043                         !usage(curr, this, excl_bit, state_name(new_bit & ~1)))
2044                 return 0;
2045
2046         /*
2047          * Check for read in write conflicts
2048          */
2049         if (!read) {
2050                 if (!valid_state(curr, this, new_bit, excl_bit + 1))
2051                         return 0;
2052
2053                 if (STRICT_READ_CHECKS &&
2054                         !usage(curr, this, excl_bit + 1,
2055                                 state_name(new_bit + 1)))
2056                         return 0;
2057         }
2058
2059         if (state_verbose(new_bit, hlock_class(this)))
2060                 return 2;
2061
2062         return 1;
2063 }
2064
2065 enum mark_type {
2066 #define LOCKDEP_STATE(__STATE)  __STATE,
2067 #include "lockdep_states.h"
2068 #undef LOCKDEP_STATE
2069 };
2070
2071 /*
2072  * Mark all held locks with a usage bit:
2073  */
2074 static int
2075 mark_held_locks(struct task_struct *curr, enum mark_type mark)
2076 {
2077         enum lock_usage_bit usage_bit;
2078         struct held_lock *hlock;
2079         int i;
2080
2081         for (i = 0; i < curr->lockdep_depth; i++) {
2082                 hlock = curr->held_locks + i;
2083
2084                 usage_bit = 2 + (mark << 2); /* ENABLED */
2085                 if (hlock->read)
2086                         usage_bit += 1; /* READ */
2087
2088                 BUG_ON(usage_bit >= LOCK_USAGE_STATES);
2089
2090                 if (!mark_lock(curr, hlock, usage_bit))
2091                         return 0;
2092         }
2093
2094         return 1;
2095 }
2096
2097 /*
2098  * Debugging helper: via this flag we know that we are in
2099  * 'early bootup code', and will warn about any invalid irqs-on event:
2100  */
2101 static int early_boot_irqs_enabled;
2102
2103 void early_boot_irqs_off(void)
2104 {
2105         early_boot_irqs_enabled = 0;
2106 }
2107
2108 void early_boot_irqs_on(void)
2109 {
2110         early_boot_irqs_enabled = 1;
2111 }
2112
2113 /*
2114  * Hardirqs will be enabled:
2115  */
2116 void trace_hardirqs_on_caller(unsigned long ip)
2117 {
2118         struct task_struct *curr = current;
2119
2120         time_hardirqs_on(CALLER_ADDR0, ip);
2121
2122         if (unlikely(!debug_locks || current->lockdep_recursion))
2123                 return;
2124
2125         if (DEBUG_LOCKS_WARN_ON(unlikely(!early_boot_irqs_enabled)))
2126                 return;
2127
2128         if (unlikely(curr->hardirqs_enabled)) {
2129                 debug_atomic_inc(&redundant_hardirqs_on);
2130                 return;
2131         }
2132         /* we'll do an OFF -> ON transition: */
2133         curr->hardirqs_enabled = 1;
2134
2135         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2136                 return;
2137         if (DEBUG_LOCKS_WARN_ON(current->hardirq_context))
2138                 return;
2139         /*
2140          * We are going to turn hardirqs on, so set the
2141          * usage bit for all held locks:
2142          */
2143         if (!mark_held_locks(curr, HARDIRQ))
2144                 return;
2145         /*
2146          * If we have softirqs enabled, then set the usage
2147          * bit for all held locks. (disabled hardirqs prevented
2148          * this bit from being set before)
2149          */
2150         if (curr->softirqs_enabled)
2151                 if (!mark_held_locks(curr, SOFTIRQ))
2152                         return;
2153
2154         curr->hardirq_enable_ip = ip;
2155         curr->hardirq_enable_event = ++curr->irq_events;
2156         debug_atomic_inc(&hardirqs_on_events);
2157 }
2158 EXPORT_SYMBOL(trace_hardirqs_on_caller);
2159
2160 void trace_hardirqs_on(void)
2161 {
2162         trace_hardirqs_on_caller(CALLER_ADDR0);
2163 }
2164 EXPORT_SYMBOL(trace_hardirqs_on);
2165
2166 /*
2167  * Hardirqs were disabled:
2168  */
2169 void trace_hardirqs_off_caller(unsigned long ip)
2170 {
2171         struct task_struct *curr = current;
2172
2173         time_hardirqs_off(CALLER_ADDR0, ip);
2174
2175         if (unlikely(!debug_locks || current->lockdep_recursion))
2176                 return;
2177
2178         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2179                 return;
2180
2181         if (curr->hardirqs_enabled) {
2182                 /*
2183                  * We have done an ON -> OFF transition:
2184                  */
2185                 curr->hardirqs_enabled = 0;
2186                 curr->hardirq_disable_ip = ip;
2187                 curr->hardirq_disable_event = ++curr->irq_events;
2188                 debug_atomic_inc(&hardirqs_off_events);
2189         } else
2190                 debug_atomic_inc(&redundant_hardirqs_off);
2191 }
2192 EXPORT_SYMBOL(trace_hardirqs_off_caller);
2193
2194 void trace_hardirqs_off(void)
2195 {
2196         trace_hardirqs_off_caller(CALLER_ADDR0);
2197 }
2198 EXPORT_SYMBOL(trace_hardirqs_off);
2199
2200 /*
2201  * Softirqs will be enabled:
2202  */
2203 void trace_softirqs_on(unsigned long ip)
2204 {
2205         struct task_struct *curr = current;
2206
2207         if (unlikely(!debug_locks))
2208                 return;
2209
2210         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2211                 return;
2212
2213         if (curr->softirqs_enabled) {
2214                 debug_atomic_inc(&redundant_softirqs_on);
2215                 return;
2216         }
2217
2218         /*
2219          * We'll do an OFF -> ON transition:
2220          */
2221         curr->softirqs_enabled = 1;
2222         curr->softirq_enable_ip = ip;
2223         curr->softirq_enable_event = ++curr->irq_events;
2224         debug_atomic_inc(&softirqs_on_events);
2225         /*
2226          * We are going to turn softirqs on, so set the
2227          * usage bit for all held locks, if hardirqs are
2228          * enabled too:
2229          */
2230         if (curr->hardirqs_enabled)
2231                 mark_held_locks(curr, SOFTIRQ);
2232 }
2233
2234 /*
2235  * Softirqs were disabled:
2236  */
2237 void trace_softirqs_off(unsigned long ip)
2238 {
2239         struct task_struct *curr = current;
2240
2241         if (unlikely(!debug_locks))
2242                 return;
2243
2244         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2245                 return;
2246
2247         if (curr->softirqs_enabled) {
2248                 /*
2249                  * We have done an ON -> OFF transition:
2250                  */
2251                 curr->softirqs_enabled = 0;
2252                 curr->softirq_disable_ip = ip;
2253                 curr->softirq_disable_event = ++curr->irq_events;
2254                 debug_atomic_inc(&softirqs_off_events);
2255                 DEBUG_LOCKS_WARN_ON(!softirq_count());
2256         } else
2257                 debug_atomic_inc(&redundant_softirqs_off);
2258 }
2259
2260 static void __lockdep_trace_alloc(gfp_t gfp_mask, unsigned long flags)
2261 {
2262         struct task_struct *curr = current;
2263
2264         if (unlikely(!debug_locks))
2265                 return;
2266
2267         /* no reclaim without waiting on it */
2268         if (!(gfp_mask & __GFP_WAIT))
2269                 return;
2270
2271         /* this guy won't enter reclaim */
2272         if ((curr->flags & PF_MEMALLOC) && !(gfp_mask & __GFP_NOMEMALLOC))
2273                 return;
2274
2275         /* We're only interested __GFP_FS allocations for now */
2276         if (!(gfp_mask & __GFP_FS))
2277                 return;
2278
2279         if (DEBUG_LOCKS_WARN_ON(irqs_disabled_flags(flags)))
2280                 return;
2281
2282         mark_held_locks(curr, RECLAIM_FS);
2283 }
2284
2285 static void check_flags(unsigned long flags);
2286
2287 void lockdep_trace_alloc(gfp_t gfp_mask)
2288 {
2289         unsigned long flags;
2290
2291         if (unlikely(current->lockdep_recursion))
2292                 return;
2293
2294         raw_local_irq_save(flags);
2295         check_flags(flags);
2296         current->lockdep_recursion = 1;
2297         __lockdep_trace_alloc(gfp_mask, flags);
2298         current->lockdep_recursion = 0;
2299         raw_local_irq_restore(flags);
2300 }
2301
2302 static int mark_irqflags(struct task_struct *curr, struct held_lock *hlock)
2303 {
2304         /*
2305          * If non-trylock use in a hardirq or softirq context, then
2306          * mark the lock as used in these contexts:
2307          */
2308         if (!hlock->trylock) {
2309                 if (hlock->read) {
2310                         if (curr->hardirq_context)
2311                                 if (!mark_lock(curr, hlock,
2312                                                 LOCK_USED_IN_HARDIRQ_READ))
2313                                         return 0;
2314                         if (curr->softirq_context)
2315                                 if (!mark_lock(curr, hlock,
2316                                                 LOCK_USED_IN_SOFTIRQ_READ))
2317                                         return 0;
2318                 } else {
2319                         if (curr->hardirq_context)
2320                                 if (!mark_lock(curr, hlock, LOCK_USED_IN_HARDIRQ))
2321                                         return 0;
2322                         if (curr->softirq_context)
2323                                 if (!mark_lock(curr, hlock, LOCK_USED_IN_SOFTIRQ))
2324                                         return 0;
2325                 }
2326         }
2327         if (!hlock->hardirqs_off) {
2328                 if (hlock->read) {
2329                         if (!mark_lock(curr, hlock,
2330                                         LOCK_ENABLED_HARDIRQ_READ))
2331                                 return 0;
2332                         if (curr->softirqs_enabled)
2333                                 if (!mark_lock(curr, hlock,
2334                                                 LOCK_ENABLED_SOFTIRQ_READ))
2335                                         return 0;
2336                 } else {
2337                         if (!mark_lock(curr, hlock,
2338                                         LOCK_ENABLED_HARDIRQ))
2339                                 return 0;
2340                         if (curr->softirqs_enabled)
2341                                 if (!mark_lock(curr, hlock,
2342                                                 LOCK_ENABLED_SOFTIRQ))
2343                                         return 0;
2344                 }
2345         }
2346
2347         /*
2348          * We reuse the irq context infrastructure more broadly as a general
2349          * context checking code. This tests GFP_FS recursion (a lock taken
2350          * during reclaim for a GFP_FS allocation is held over a GFP_FS
2351          * allocation).
2352          */
2353         if (!hlock->trylock && (curr->lockdep_reclaim_gfp & __GFP_FS)) {
2354                 if (hlock->read) {
2355                         if (!mark_lock(curr, hlock, LOCK_USED_IN_RECLAIM_FS_READ))
2356                                         return 0;
2357                 } else {
2358                         if (!mark_lock(curr, hlock, LOCK_USED_IN_RECLAIM_FS))
2359                                         return 0;
2360                 }
2361         }
2362
2363         return 1;
2364 }
2365
2366 static int separate_irq_context(struct task_struct *curr,
2367                 struct held_lock *hlock)
2368 {
2369         unsigned int depth = curr->lockdep_depth;
2370
2371         /*
2372          * Keep track of points where we cross into an interrupt context:
2373          */
2374         hlock->irq_context = 2*(curr->hardirq_context ? 1 : 0) +
2375                                 curr->softirq_context;
2376         if (depth) {
2377                 struct held_lock *prev_hlock;
2378
2379                 prev_hlock = curr->held_locks + depth-1;
2380                 /*
2381                  * If we cross into another context, reset the
2382                  * hash key (this also prevents the checking and the
2383                  * adding of the dependency to 'prev'):
2384                  */
2385                 if (prev_hlock->irq_context != hlock->irq_context)
2386                         return 1;
2387         }
2388         return 0;
2389 }
2390
2391 #else
2392
2393 static inline
2394 int mark_lock_irq(struct task_struct *curr, struct held_lock *this,
2395                 enum lock_usage_bit new_bit)
2396 {
2397         WARN_ON(1);
2398         return 1;
2399 }
2400
2401 static inline int mark_irqflags(struct task_struct *curr,
2402                 struct held_lock *hlock)
2403 {
2404         return 1;
2405 }
2406
2407 static inline int separate_irq_context(struct task_struct *curr,
2408                 struct held_lock *hlock)
2409 {
2410         return 0;
2411 }
2412
2413 void lockdep_trace_alloc(gfp_t gfp_mask)
2414 {
2415 }
2416
2417 #endif
2418
2419 /*
2420  * Mark a lock with a usage bit, and validate the state transition:
2421  */
2422 static int mark_lock(struct task_struct *curr, struct held_lock *this,
2423                              enum lock_usage_bit new_bit)
2424 {
2425         unsigned int new_mask = 1 << new_bit, ret = 1;
2426
2427         /*
2428          * If already set then do not dirty the cacheline,
2429          * nor do any checks:
2430          */
2431         if (likely(hlock_class(this)->usage_mask & new_mask))
2432                 return 1;
2433
2434         if (!graph_lock())
2435                 return 0;
2436         /*
2437          * Make sure we didnt race:
2438          */
2439         if (unlikely(hlock_class(this)->usage_mask & new_mask)) {
2440                 graph_unlock();
2441                 return 1;
2442         }
2443
2444         hlock_class(this)->usage_mask |= new_mask;
2445
2446         if (!save_trace(hlock_class(this)->usage_traces + new_bit))
2447                 return 0;
2448
2449         switch (new_bit) {
2450 #define LOCKDEP_STATE(__STATE)                  \
2451         case LOCK_USED_IN_##__STATE:            \
2452         case LOCK_USED_IN_##__STATE##_READ:     \
2453         case LOCK_ENABLED_##__STATE:            \
2454         case LOCK_ENABLED_##__STATE##_READ:
2455 #include "lockdep_states.h"
2456 #undef LOCKDEP_STATE
2457                 ret = mark_lock_irq(curr, this, new_bit);
2458                 if (!ret)
2459                         return 0;
2460                 break;
2461         case LOCK_USED:
2462                 debug_atomic_dec(&nr_unused_locks);
2463                 break;
2464         default:
2465                 if (!debug_locks_off_graph_unlock())
2466                         return 0;
2467                 WARN_ON(1);
2468                 return 0;
2469         }
2470
2471         graph_unlock();
2472
2473         /*
2474          * We must printk outside of the graph_lock:
2475          */
2476         if (ret == 2) {
2477                 printk("\nmarked lock as {%s}:\n", usage_str[new_bit]);
2478                 print_lock(this);
2479                 print_irqtrace_events(curr);
2480                 dump_stack();
2481         }
2482
2483         return ret;
2484 }
2485
2486 /*
2487  * Initialize a lock instance's lock-class mapping info:
2488  */
2489 void lockdep_init_map(struct lockdep_map *lock, const char *name,
2490                       struct lock_class_key *key, int subclass)
2491 {
2492         if (unlikely(!debug_locks))
2493                 return;
2494
2495         if (DEBUG_LOCKS_WARN_ON(!key))
2496                 return;
2497         if (DEBUG_LOCKS_WARN_ON(!name))
2498                 return;
2499         /*
2500          * Sanity check, the lock-class key must be persistent:
2501          */
2502         if (!static_obj(key)) {
2503                 printk("BUG: key %p not in .data!\n", key);
2504                 DEBUG_LOCKS_WARN_ON(1);
2505                 return;
2506         }
2507         lock->name = name;
2508         lock->key = key;
2509         lock->class_cache = NULL;
2510 #ifdef CONFIG_LOCK_STAT
2511         lock->cpu = raw_smp_processor_id();
2512 #endif
2513         if (subclass)
2514                 register_lock_class(lock, subclass, 1);
2515 }
2516 EXPORT_SYMBOL_GPL(lockdep_init_map);
2517
2518 /*
2519  * This gets called for every mutex_lock*()/spin_lock*() operation.
2520  * We maintain the dependency maps and validate the locking attempt:
2521  */
2522 static int __lock_acquire(struct lockdep_map *lock, unsigned int subclass,
2523                           int trylock, int read, int check, int hardirqs_off,
2524                           struct lockdep_map *nest_lock, unsigned long ip)
2525 {
2526         struct task_struct *curr = current;
2527         struct lock_class *class = NULL;
2528         struct held_lock *hlock;
2529         unsigned int depth, id;
2530         int chain_head = 0;
2531         u64 chain_key;
2532
2533         if (!prove_locking)
2534                 check = 1;
2535
2536         if (unlikely(!debug_locks))
2537                 return 0;
2538
2539         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2540                 return 0;
2541
2542         if (unlikely(subclass >= MAX_LOCKDEP_SUBCLASSES)) {
2543                 debug_locks_off();
2544                 printk("BUG: MAX_LOCKDEP_SUBCLASSES too low!\n");
2545                 printk("turning off the locking correctness validator.\n");
2546                 dump_stack();
2547                 return 0;
2548         }
2549
2550         if (!subclass)
2551                 class = lock->class_cache;
2552         /*
2553          * Not cached yet or subclass?
2554          */
2555         if (unlikely(!class)) {
2556                 class = register_lock_class(lock, subclass, 0);
2557                 if (!class)
2558                         return 0;
2559         }
2560         debug_atomic_inc((atomic_t *)&class->ops);
2561         if (very_verbose(class)) {
2562                 printk("\nacquire class [%p] %s", class->key, class->name);
2563                 if (class->name_version > 1)
2564                         printk("#%d", class->name_version);
2565                 printk("\n");
2566                 dump_stack();
2567         }
2568
2569         /*
2570          * Add the lock to the list of currently held locks.
2571          * (we dont increase the depth just yet, up until the
2572          * dependency checks are done)
2573          */
2574         depth = curr->lockdep_depth;
2575         if (DEBUG_LOCKS_WARN_ON(depth >= MAX_LOCK_DEPTH))
2576                 return 0;
2577
2578         hlock = curr->held_locks + depth;
2579         if (DEBUG_LOCKS_WARN_ON(!class))
2580                 return 0;
2581         hlock->class_idx = class - lock_classes + 1;
2582         hlock->acquire_ip = ip;
2583         hlock->instance = lock;
2584         hlock->nest_lock = nest_lock;
2585         hlock->trylock = trylock;
2586         hlock->read = read;
2587         hlock->check = check;
2588         hlock->hardirqs_off = !!hardirqs_off;
2589 #ifdef CONFIG_LOCK_STAT
2590         hlock->waittime_stamp = 0;
2591         hlock->holdtime_stamp = sched_clock();
2592 #endif
2593
2594         if (check == 2 && !mark_irqflags(curr, hlock))
2595                 return 0;
2596
2597         /* mark it as used: */
2598         if (!mark_lock(curr, hlock, LOCK_USED))
2599                 return 0;
2600
2601         /*
2602          * Calculate the chain hash: it's the combined hash of all the
2603          * lock keys along the dependency chain. We save the hash value
2604          * at every step so that we can get the current hash easily
2605          * after unlock. The chain hash is then used to cache dependency
2606          * results.
2607          *
2608          * The 'key ID' is what is the most compact key value to drive
2609          * the hash, not class->key.
2610          */
2611         id = class - lock_classes;
2612         if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
2613                 return 0;
2614
2615         chain_key = curr->curr_chain_key;
2616         if (!depth) {
2617                 if (DEBUG_LOCKS_WARN_ON(chain_key != 0))
2618                         return 0;
2619                 chain_head = 1;
2620         }
2621
2622         hlock->prev_chain_key = chain_key;
2623         if (separate_irq_context(curr, hlock)) {
2624                 chain_key = 0;
2625                 chain_head = 1;
2626         }
2627         chain_key = iterate_chain_key(chain_key, id);
2628
2629         if (!validate_chain(curr, lock, hlock, chain_head, chain_key))
2630                 return 0;
2631
2632         curr->curr_chain_key = chain_key;
2633         curr->lockdep_depth++;
2634         check_chain_key(curr);
2635 #ifdef CONFIG_DEBUG_LOCKDEP
2636         if (unlikely(!debug_locks))
2637                 return 0;
2638 #endif
2639         if (unlikely(curr->lockdep_depth >= MAX_LOCK_DEPTH)) {
2640                 debug_locks_off();
2641                 printk("BUG: MAX_LOCK_DEPTH too low!\n");
2642                 printk("turning off the locking correctness validator.\n");
2643                 dump_stack();
2644                 return 0;
2645         }
2646
2647         if (unlikely(curr->lockdep_depth > max_lockdep_depth))
2648                 max_lockdep_depth = curr->lockdep_depth;
2649
2650         return 1;
2651 }
2652
2653 static int
2654 print_unlock_inbalance_bug(struct task_struct *curr, struct lockdep_map *lock,
2655                            unsigned long ip)
2656 {
2657         if (!debug_locks_off())
2658                 return 0;
2659         if (debug_locks_silent)
2660                 return 0;
2661
2662         printk("\n=====================================\n");
2663         printk(  "[ BUG: bad unlock balance detected! ]\n");
2664         printk(  "-------------------------------------\n");
2665         printk("%s/%d is trying to release lock (",
2666                 curr->comm, task_pid_nr(curr));
2667         print_lockdep_cache(lock);
2668         printk(") at:\n");
2669         print_ip_sym(ip);
2670         printk("but there are no more locks to release!\n");
2671         printk("\nother info that might help us debug this:\n");
2672         lockdep_print_held_locks(curr);
2673
2674         printk("\nstack backtrace:\n");
2675         dump_stack();
2676
2677         return 0;
2678 }
2679
2680 /*
2681  * Common debugging checks for both nested and non-nested unlock:
2682  */
2683 static int check_unlock(struct task_struct *curr, struct lockdep_map *lock,
2684                         unsigned long ip)
2685 {
2686         if (unlikely(!debug_locks))
2687                 return 0;
2688         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2689                 return 0;
2690
2691         if (curr->lockdep_depth <= 0)
2692                 return print_unlock_inbalance_bug(curr, lock, ip);
2693
2694         return 1;
2695 }
2696
2697 static int
2698 __lock_set_class(struct lockdep_map *lock, const char *name,
2699                  struct lock_class_key *key, unsigned int subclass,
2700                  unsigned long ip)
2701 {
2702         struct task_struct *curr = current;
2703         struct held_lock *hlock, *prev_hlock;
2704         struct lock_class *class;
2705         unsigned int depth;
2706         int i;
2707
2708         depth = curr->lockdep_depth;
2709         if (DEBUG_LOCKS_WARN_ON(!depth))
2710                 return 0;
2711
2712         prev_hlock = NULL;
2713         for (i = depth-1; i >= 0; i--) {
2714                 hlock = curr->held_locks + i;
2715                 /*
2716                  * We must not cross into another context:
2717                  */
2718                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
2719                         break;
2720                 if (hlock->instance == lock)
2721                         goto found_it;
2722                 prev_hlock = hlock;
2723         }
2724         return print_unlock_inbalance_bug(curr, lock, ip);
2725
2726 found_it:
2727         lockdep_init_map(lock, name, key, 0);
2728         class = register_lock_class(lock, subclass, 0);
2729         hlock->class_idx = class - lock_classes + 1;
2730
2731         curr->lockdep_depth = i;
2732         curr->curr_chain_key = hlock->prev_chain_key;
2733
2734         for (; i < depth; i++) {
2735                 hlock = curr->held_locks + i;
2736                 if (!__lock_acquire(hlock->instance,
2737                         hlock_class(hlock)->subclass, hlock->trylock,
2738                                 hlock->read, hlock->check, hlock->hardirqs_off,
2739                                 hlock->nest_lock, hlock->acquire_ip))
2740                         return 0;
2741         }
2742
2743         if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth))
2744                 return 0;
2745         return 1;
2746 }
2747
2748 /*
2749  * Remove the lock to the list of currently held locks in a
2750  * potentially non-nested (out of order) manner. This is a
2751  * relatively rare operation, as all the unlock APIs default
2752  * to nested mode (which uses lock_release()):
2753  */
2754 static int
2755 lock_release_non_nested(struct task_struct *curr,
2756                         struct lockdep_map *lock, unsigned long ip)
2757 {
2758         struct held_lock *hlock, *prev_hlock;
2759         unsigned int depth;
2760         int i;
2761
2762         /*
2763          * Check whether the lock exists in the current stack
2764          * of held locks:
2765          */
2766         depth = curr->lockdep_depth;
2767         if (DEBUG_LOCKS_WARN_ON(!depth))
2768                 return 0;
2769
2770         prev_hlock = NULL;
2771         for (i = depth-1; i >= 0; i--) {
2772                 hlock = curr->held_locks + i;
2773                 /*
2774                  * We must not cross into another context:
2775                  */
2776                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
2777                         break;
2778                 if (hlock->instance == lock)
2779                         goto found_it;
2780                 prev_hlock = hlock;
2781         }
2782         return print_unlock_inbalance_bug(curr, lock, ip);
2783
2784 found_it:
2785         lock_release_holdtime(hlock);
2786
2787         /*
2788          * We have the right lock to unlock, 'hlock' points to it.
2789          * Now we remove it from the stack, and add back the other
2790          * entries (if any), recalculating the hash along the way:
2791          */
2792         curr->lockdep_depth = i;
2793         curr->curr_chain_key = hlock->prev_chain_key;
2794
2795         for (i++; i < depth; i++) {
2796                 hlock = curr->held_locks + i;
2797                 if (!__lock_acquire(hlock->instance,
2798                         hlock_class(hlock)->subclass, hlock->trylock,
2799                                 hlock->read, hlock->check, hlock->hardirqs_off,
2800                                 hlock->nest_lock, hlock->acquire_ip))
2801                         return 0;
2802         }
2803
2804         if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth - 1))
2805                 return 0;
2806         return 1;
2807 }
2808
2809 /*
2810  * Remove the lock to the list of currently held locks - this gets
2811  * called on mutex_unlock()/spin_unlock*() (or on a failed
2812  * mutex_lock_interruptible()). This is done for unlocks that nest
2813  * perfectly. (i.e. the current top of the lock-stack is unlocked)
2814  */
2815 static int lock_release_nested(struct task_struct *curr,
2816                                struct lockdep_map *lock, unsigned long ip)
2817 {
2818         struct held_lock *hlock;
2819         unsigned int depth;
2820
2821         /*
2822          * Pop off the top of the lock stack:
2823          */
2824         depth = curr->lockdep_depth - 1;
2825         hlock = curr->held_locks + depth;
2826
2827         /*
2828          * Is the unlock non-nested:
2829          */
2830         if (hlock->instance != lock)
2831                 return lock_release_non_nested(curr, lock, ip);
2832         curr->lockdep_depth--;
2833
2834         if (DEBUG_LOCKS_WARN_ON(!depth && (hlock->prev_chain_key != 0)))
2835                 return 0;
2836
2837         curr->curr_chain_key = hlock->prev_chain_key;
2838
2839         lock_release_holdtime(hlock);
2840
2841 #ifdef CONFIG_DEBUG_LOCKDEP
2842         hlock->prev_chain_key = 0;
2843         hlock->class_idx = 0;
2844         hlock->acquire_ip = 0;
2845         hlock->irq_context = 0;
2846 #endif
2847         return 1;
2848 }
2849
2850 /*
2851  * Remove the lock to the list of currently held locks - this gets
2852  * called on mutex_unlock()/spin_unlock*() (or on a failed
2853  * mutex_lock_interruptible()). This is done for unlocks that nest
2854  * perfectly. (i.e. the current top of the lock-stack is unlocked)
2855  */
2856 static void
2857 __lock_release(struct lockdep_map *lock, int nested, unsigned long ip)
2858 {
2859         struct task_struct *curr = current;
2860
2861         if (!check_unlock(curr, lock, ip))
2862                 return;
2863
2864         if (nested) {
2865                 if (!lock_release_nested(curr, lock, ip))
2866                         return;
2867         } else {
2868                 if (!lock_release_non_nested(curr, lock, ip))
2869                         return;
2870         }
2871
2872         check_chain_key(curr);
2873 }
2874
2875 /*
2876  * Check whether we follow the irq-flags state precisely:
2877  */
2878 static void check_flags(unsigned long flags)
2879 {
2880 #if defined(CONFIG_PROVE_LOCKING) && defined(CONFIG_DEBUG_LOCKDEP) && \
2881     defined(CONFIG_TRACE_IRQFLAGS)
2882         if (!debug_locks)
2883                 return;
2884
2885         if (irqs_disabled_flags(flags)) {
2886                 if (DEBUG_LOCKS_WARN_ON(current->hardirqs_enabled)) {
2887                         printk("possible reason: unannotated irqs-off.\n");
2888                 }
2889         } else {
2890                 if (DEBUG_LOCKS_WARN_ON(!current->hardirqs_enabled)) {
2891                         printk("possible reason: unannotated irqs-on.\n");
2892                 }
2893         }
2894
2895         /*
2896          * We dont accurately track softirq state in e.g.
2897          * hardirq contexts (such as on 4KSTACKS), so only
2898          * check if not in hardirq contexts:
2899          */
2900         if (!hardirq_count()) {
2901                 if (softirq_count())
2902                         DEBUG_LOCKS_WARN_ON(current->softirqs_enabled);
2903                 else
2904                         DEBUG_LOCKS_WARN_ON(!current->softirqs_enabled);
2905         }
2906
2907         if (!debug_locks)
2908                 print_irqtrace_events(current);
2909 #endif
2910 }
2911
2912 void lock_set_class(struct lockdep_map *lock, const char *name,
2913                     struct lock_class_key *key, unsigned int subclass,
2914                     unsigned long ip)
2915 {
2916         unsigned long flags;
2917
2918         if (unlikely(current->lockdep_recursion))
2919                 return;
2920
2921         raw_local_irq_save(flags);
2922         current->lockdep_recursion = 1;
2923         check_flags(flags);
2924         if (__lock_set_class(lock, name, key, subclass, ip))
2925                 check_chain_key(current);
2926         current->lockdep_recursion = 0;
2927         raw_local_irq_restore(flags);
2928 }
2929 EXPORT_SYMBOL_GPL(lock_set_class);
2930
2931 /*
2932  * We are not always called with irqs disabled - do that here,
2933  * and also avoid lockdep recursion:
2934  */
2935 void lock_acquire(struct lockdep_map *lock, unsigned int subclass,
2936                           int trylock, int read, int check,
2937                           struct lockdep_map *nest_lock, unsigned long ip)
2938 {
2939         unsigned long flags;
2940
2941         if (unlikely(current->lockdep_recursion))
2942                 return;
2943
2944         raw_local_irq_save(flags);
2945         check_flags(flags);
2946
2947         current->lockdep_recursion = 1;
2948         __lock_acquire(lock, subclass, trylock, read, check,
2949                        irqs_disabled_flags(flags), nest_lock, ip);
2950         current->lockdep_recursion = 0;
2951         raw_local_irq_restore(flags);
2952 }
2953 EXPORT_SYMBOL_GPL(lock_acquire);
2954
2955 void lock_release(struct lockdep_map *lock, int nested,
2956                           unsigned long ip)
2957 {
2958         unsigned long flags;
2959
2960         if (unlikely(current->lockdep_recursion))
2961                 return;
2962
2963         raw_local_irq_save(flags);
2964         check_flags(flags);
2965         current->lockdep_recursion = 1;
2966         __lock_release(lock, nested, ip);
2967         current->lockdep_recursion = 0;
2968         raw_local_irq_restore(flags);
2969 }
2970 EXPORT_SYMBOL_GPL(lock_release);
2971
2972 void lockdep_set_current_reclaim_state(gfp_t gfp_mask)
2973 {
2974         current->lockdep_reclaim_gfp = gfp_mask;
2975 }
2976
2977 void lockdep_clear_current_reclaim_state(void)
2978 {
2979         current->lockdep_reclaim_gfp = 0;
2980 }
2981
2982 #ifdef CONFIG_LOCK_STAT
2983 static int
2984 print_lock_contention_bug(struct task_struct *curr, struct lockdep_map *lock,
2985                            unsigned long ip)
2986 {
2987         if (!debug_locks_off())
2988                 return 0;
2989         if (debug_locks_silent)
2990                 return 0;
2991
2992         printk("\n=================================\n");
2993         printk(  "[ BUG: bad contention detected! ]\n");
2994         printk(  "---------------------------------\n");
2995         printk("%s/%d is trying to contend lock (",
2996                 curr->comm, task_pid_nr(curr));
2997         print_lockdep_cache(lock);
2998         printk(") at:\n");
2999         print_ip_sym(ip);
3000         printk("but there are no locks held!\n");
3001         printk("\nother info that might help us debug this:\n");
3002         lockdep_print_held_locks(curr);
3003
3004         printk("\nstack backtrace:\n");
3005         dump_stack();
3006
3007         return 0;
3008 }
3009
3010 static void
3011 __lock_contended(struct lockdep_map *lock, unsigned long ip)
3012 {
3013         struct task_struct *curr = current;
3014         struct held_lock *hlock, *prev_hlock;
3015         struct lock_class_stats *stats;
3016         unsigned int depth;
3017         int i, contention_point, contending_point;
3018
3019         depth = curr->lockdep_depth;
3020         if (DEBUG_LOCKS_WARN_ON(!depth))
3021                 return;
3022
3023         prev_hlock = NULL;
3024         for (i = depth-1; i >= 0; i--) {
3025                 hlock = curr->held_locks + i;
3026                 /*
3027                  * We must not cross into another context:
3028                  */
3029                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3030                         break;
3031                 if (hlock->instance == lock)
3032                         goto found_it;
3033                 prev_hlock = hlock;
3034         }
3035         print_lock_contention_bug(curr, lock, ip);
3036         return;
3037
3038 found_it:
3039         hlock->waittime_stamp = sched_clock();
3040
3041         contention_point = lock_point(hlock_class(hlock)->contention_point, ip);
3042         contending_point = lock_point(hlock_class(hlock)->contending_point,
3043                                       lock->ip);
3044
3045         stats = get_lock_stats(hlock_class(hlock));
3046         if (contention_point < LOCKSTAT_POINTS)
3047                 stats->contention_point[contention_point]++;
3048         if (contending_point < LOCKSTAT_POINTS)
3049                 stats->contending_point[contending_point]++;
3050         if (lock->cpu != smp_processor_id())
3051                 stats->bounces[bounce_contended + !!hlock->read]++;
3052         put_lock_stats(stats);
3053 }
3054
3055 static void
3056 __lock_acquired(struct lockdep_map *lock, unsigned long ip)
3057 {
3058         struct task_struct *curr = current;
3059         struct held_lock *hlock, *prev_hlock;
3060         struct lock_class_stats *stats;
3061         unsigned int depth;
3062         u64 now;
3063         s64 waittime = 0;
3064         int i, cpu;
3065
3066         depth = curr->lockdep_depth;
3067         if (DEBUG_LOCKS_WARN_ON(!depth))
3068                 return;
3069
3070         prev_hlock = NULL;
3071         for (i = depth-1; i >= 0; i--) {
3072                 hlock = curr->held_locks + i;
3073                 /*
3074                  * We must not cross into another context:
3075                  */
3076                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3077                         break;
3078                 if (hlock->instance == lock)
3079                         goto found_it;
3080                 prev_hlock = hlock;
3081         }
3082         print_lock_contention_bug(curr, lock, _RET_IP_);
3083         return;
3084
3085 found_it:
3086         cpu = smp_processor_id();
3087         if (hlock->waittime_stamp) {
3088                 now = sched_clock();
3089                 waittime = now - hlock->waittime_stamp;
3090                 hlock->holdtime_stamp = now;
3091         }
3092
3093         stats = get_lock_stats(hlock_class(hlock));
3094         if (waittime) {
3095                 if (hlock->read)
3096                         lock_time_inc(&stats->read_waittime, waittime);
3097                 else
3098                         lock_time_inc(&stats->write_waittime, waittime);
3099         }
3100         if (lock->cpu != cpu)
3101                 stats->bounces[bounce_acquired + !!hlock->read]++;
3102         put_lock_stats(stats);
3103
3104         lock->cpu = cpu;
3105         lock->ip = ip;
3106 }
3107
3108 void lock_contended(struct lockdep_map *lock, unsigned long ip)
3109 {
3110         unsigned long flags;
3111
3112         if (unlikely(!lock_stat))
3113                 return;
3114
3115         if (unlikely(current->lockdep_recursion))
3116                 return;
3117
3118         raw_local_irq_save(flags);
3119         check_flags(flags);
3120         current->lockdep_recursion = 1;
3121         __lock_contended(lock, ip);
3122         current->lockdep_recursion = 0;
3123         raw_local_irq_restore(flags);
3124 }
3125 EXPORT_SYMBOL_GPL(lock_contended);
3126
3127 void lock_acquired(struct lockdep_map *lock, unsigned long ip)
3128 {
3129         unsigned long flags;
3130
3131         if (unlikely(!lock_stat))
3132                 return;
3133
3134         if (unlikely(current->lockdep_recursion))
3135                 return;
3136
3137         raw_local_irq_save(flags);
3138         check_flags(flags);
3139         current->lockdep_recursion = 1;
3140         __lock_acquired(lock, ip);
3141         current->lockdep_recursion = 0;
3142         raw_local_irq_restore(flags);
3143 }
3144 EXPORT_SYMBOL_GPL(lock_acquired);
3145 #endif
3146
3147 /*
3148  * Used by the testsuite, sanitize the validator state
3149  * after a simulated failure:
3150  */
3151
3152 void lockdep_reset(void)
3153 {
3154         unsigned long flags;
3155         int i;
3156
3157         raw_local_irq_save(flags);
3158         current->curr_chain_key = 0;
3159         current->lockdep_depth = 0;
3160         current->lockdep_recursion = 0;
3161         memset(current->held_locks, 0, MAX_LOCK_DEPTH*sizeof(struct held_lock));
3162         nr_hardirq_chains = 0;
3163         nr_softirq_chains = 0;
3164         nr_process_chains = 0;
3165         debug_locks = 1;
3166         for (i = 0; i < CHAINHASH_SIZE; i++)
3167                 INIT_LIST_HEAD(chainhash_table + i);
3168         raw_local_irq_restore(flags);
3169 }
3170
3171 static void zap_class(struct lock_class *class)
3172 {
3173         int i;
3174
3175         /*
3176          * Remove all dependencies this lock is
3177          * involved in:
3178          */
3179         for (i = 0; i < nr_list_entries; i++) {
3180                 if (list_entries[i].class == class)
3181                         list_del_rcu(&list_entries[i].entry);
3182         }
3183         /*
3184          * Unhash the class and remove it from the all_lock_classes list:
3185          */
3186         list_del_rcu(&class->hash_entry);
3187         list_del_rcu(&class->lock_entry);
3188
3189         class->key = NULL;
3190 }
3191
3192 static inline int within(const void *addr, void *start, unsigned long size)
3193 {
3194         return addr >= start && addr < start + size;
3195 }
3196
3197 void lockdep_free_key_range(void *start, unsigned long size)
3198 {
3199         struct lock_class *class, *next;
3200         struct list_head *head;
3201         unsigned long flags;
3202         int i;
3203         int locked;
3204
3205         raw_local_irq_save(flags);
3206         locked = graph_lock();
3207
3208         /*
3209          * Unhash all classes that were created by this module:
3210          */
3211         for (i = 0; i < CLASSHASH_SIZE; i++) {
3212                 head = classhash_table + i;
3213                 if (list_empty(head))
3214                         continue;
3215                 list_for_each_entry_safe(class, next, head, hash_entry) {
3216                         if (within(class->key, start, size))
3217                                 zap_class(class);
3218                         else if (within(class->name, start, size))
3219                                 zap_class(class);
3220                 }
3221         }
3222
3223         if (locked)
3224                 graph_unlock();
3225         raw_local_irq_restore(flags);
3226 }
3227
3228 void lockdep_reset_lock(struct lockdep_map *lock)
3229 {
3230         struct lock_class *class, *next;
3231         struct list_head *head;
3232         unsigned long flags;
3233         int i, j;
3234         int locked;
3235
3236         raw_local_irq_save(flags);
3237
3238         /*
3239          * Remove all classes this lock might have:
3240          */
3241         for (j = 0; j < MAX_LOCKDEP_SUBCLASSES; j++) {
3242                 /*
3243                  * If the class exists we look it up and zap it:
3244                  */
3245                 class = look_up_lock_class(lock, j);
3246                 if (class)
3247                         zap_class(class);
3248         }
3249         /*
3250          * Debug check: in the end all mapped classes should
3251          * be gone.
3252          */
3253         locked = graph_lock();
3254         for (i = 0; i < CLASSHASH_SIZE; i++) {
3255                 head = classhash_table + i;
3256                 if (list_empty(head))
3257                         continue;
3258                 list_for_each_entry_safe(class, next, head, hash_entry) {
3259                         if (unlikely(class == lock->class_cache)) {
3260                                 if (debug_locks_off_graph_unlock())
3261                                         WARN_ON(1);
3262                                 goto out_restore;
3263                         }
3264                 }
3265         }
3266         if (locked)
3267                 graph_unlock();
3268
3269 out_restore:
3270         raw_local_irq_restore(flags);
3271 }
3272
3273 void lockdep_init(void)
3274 {
3275         int i;
3276
3277         /*
3278          * Some architectures have their own start_kernel()
3279          * code which calls lockdep_init(), while we also
3280          * call lockdep_init() from the start_kernel() itself,
3281          * and we want to initialize the hashes only once:
3282          */
3283         if (lockdep_initialized)
3284                 return;
3285
3286         for (i = 0; i < CLASSHASH_SIZE; i++)
3287                 INIT_LIST_HEAD(classhash_table + i);
3288
3289         for (i = 0; i < CHAINHASH_SIZE; i++)
3290                 INIT_LIST_HEAD(chainhash_table + i);
3291
3292         lockdep_initialized = 1;
3293 }
3294
3295 void __init lockdep_info(void)
3296 {
3297         printk("Lock dependency validator: Copyright (c) 2006 Red Hat, Inc., Ingo Molnar\n");
3298
3299         printk("... MAX_LOCKDEP_SUBCLASSES:  %lu\n", MAX_LOCKDEP_SUBCLASSES);
3300         printk("... MAX_LOCK_DEPTH:          %lu\n", MAX_LOCK_DEPTH);
3301         printk("... MAX_LOCKDEP_KEYS:        %lu\n", MAX_LOCKDEP_KEYS);
3302         printk("... CLASSHASH_SIZE:          %lu\n", CLASSHASH_SIZE);
3303         printk("... MAX_LOCKDEP_ENTRIES:     %lu\n", MAX_LOCKDEP_ENTRIES);
3304         printk("... MAX_LOCKDEP_CHAINS:      %lu\n", MAX_LOCKDEP_CHAINS);
3305         printk("... CHAINHASH_SIZE:          %lu\n", CHAINHASH_SIZE);
3306
3307         printk(" memory used by lock dependency info: %lu kB\n",
3308                 (sizeof(struct lock_class) * MAX_LOCKDEP_KEYS +
3309                 sizeof(struct list_head) * CLASSHASH_SIZE +
3310                 sizeof(struct lock_list) * MAX_LOCKDEP_ENTRIES +
3311                 sizeof(struct lock_chain) * MAX_LOCKDEP_CHAINS +
3312                 sizeof(struct list_head) * CHAINHASH_SIZE) / 1024);
3313
3314         printk(" per task-struct memory footprint: %lu bytes\n",
3315                 sizeof(struct held_lock) * MAX_LOCK_DEPTH);
3316
3317 #ifdef CONFIG_DEBUG_LOCKDEP
3318         if (lockdep_init_error) {
3319                 printk("WARNING: lockdep init error! Arch code didn't call lockdep_init() early enough?\n");
3320                 printk("Call stack leading to lockdep invocation was:\n");
3321                 print_stack_trace(&lockdep_init_trace, 0);
3322         }
3323 #endif
3324 }
3325
3326 static void
3327 print_freed_lock_bug(struct task_struct *curr, const void *mem_from,
3328                      const void *mem_to, struct held_lock *hlock)
3329 {
3330         if (!debug_locks_off())
3331                 return;
3332         if (debug_locks_silent)
3333                 return;
3334
3335         printk("\n=========================\n");
3336         printk(  "[ BUG: held lock freed! ]\n");
3337         printk(  "-------------------------\n");
3338         printk("%s/%d is freeing memory %p-%p, with a lock still held there!\n",
3339                 curr->comm, task_pid_nr(curr), mem_from, mem_to-1);
3340         print_lock(hlock);
3341         lockdep_print_held_locks(curr);
3342
3343         printk("\nstack backtrace:\n");
3344         dump_stack();
3345 }
3346
3347 static inline int not_in_range(const void* mem_from, unsigned long mem_len,
3348                                 const void* lock_from, unsigned long lock_len)
3349 {
3350         return lock_from + lock_len <= mem_from ||
3351                 mem_from + mem_len <= lock_from;
3352 }
3353
3354 /*
3355  * Called when kernel memory is freed (or unmapped), or if a lock
3356  * is destroyed or reinitialized - this code checks whether there is
3357  * any held lock in the memory range of <from> to <to>:
3358  */
3359 void debug_check_no_locks_freed(const void *mem_from, unsigned long mem_len)
3360 {
3361         struct task_struct *curr = current;
3362         struct held_lock *hlock;
3363         unsigned long flags;
3364         int i;
3365
3366         if (unlikely(!debug_locks))
3367                 return;
3368
3369         local_irq_save(flags);
3370         for (i = 0; i < curr->lockdep_depth; i++) {
3371                 hlock = curr->held_locks + i;
3372
3373                 if (not_in_range(mem_from, mem_len, hlock->instance,
3374                                         sizeof(*hlock->instance)))
3375                         continue;
3376
3377                 print_freed_lock_bug(curr, mem_from, mem_from + mem_len, hlock);
3378                 break;
3379         }
3380         local_irq_restore(flags);
3381 }
3382 EXPORT_SYMBOL_GPL(debug_check_no_locks_freed);
3383
3384 static void print_held_locks_bug(struct task_struct *curr)
3385 {
3386         if (!debug_locks_off())
3387                 return;
3388         if (debug_locks_silent)
3389                 return;
3390
3391         printk("\n=====================================\n");
3392         printk(  "[ BUG: lock held at task exit time! ]\n");
3393         printk(  "-------------------------------------\n");
3394         printk("%s/%d is exiting with locks still held!\n",
3395                 curr->comm, task_pid_nr(curr));
3396         lockdep_print_held_locks(curr);
3397
3398         printk("\nstack backtrace:\n");
3399         dump_stack();
3400 }
3401
3402 void debug_check_no_locks_held(struct task_struct *task)
3403 {
3404         if (unlikely(task->lockdep_depth > 0))
3405                 print_held_locks_bug(task);
3406 }
3407
3408 void debug_show_all_locks(void)
3409 {
3410         struct task_struct *g, *p;
3411         int count = 10;
3412         int unlock = 1;
3413
3414         if (unlikely(!debug_locks)) {
3415                 printk("INFO: lockdep is turned off.\n");
3416                 return;
3417         }
3418         printk("\nShowing all locks held in the system:\n");
3419
3420         /*
3421          * Here we try to get the tasklist_lock as hard as possible,
3422          * if not successful after 2 seconds we ignore it (but keep
3423          * trying). This is to enable a debug printout even if a
3424          * tasklist_lock-holding task deadlocks or crashes.
3425          */
3426 retry:
3427         if (!read_trylock(&tasklist_lock)) {
3428                 if (count == 10)
3429                         printk("hm, tasklist_lock locked, retrying... ");
3430                 if (count) {
3431                         count--;
3432                         printk(" #%d", 10-count);
3433                         mdelay(200);
3434                         goto retry;
3435                 }
3436                 printk(" ignoring it.\n");
3437                 unlock = 0;
3438         } else {
3439                 if (count != 10)
3440                         printk(KERN_CONT " locked it.\n");
3441         }
3442
3443         do_each_thread(g, p) {
3444                 /*
3445                  * It's not reliable to print a task's held locks
3446                  * if it's not sleeping (or if it's not the current
3447                  * task):
3448                  */
3449                 if (p->state == TASK_RUNNING && p != current)
3450                         continue;
3451                 if (p->lockdep_depth)
3452                         lockdep_print_held_locks(p);
3453                 if (!unlock)
3454                         if (read_trylock(&tasklist_lock))
3455                                 unlock = 1;
3456         } while_each_thread(g, p);
3457
3458         printk("\n");
3459         printk("=============================================\n\n");
3460
3461         if (unlock)
3462                 read_unlock(&tasklist_lock);
3463 }
3464 EXPORT_SYMBOL_GPL(debug_show_all_locks);
3465
3466 /*
3467  * Careful: only use this function if you are sure that
3468  * the task cannot run in parallel!
3469  */
3470 void __debug_show_held_locks(struct task_struct *task)
3471 {
3472         if (unlikely(!debug_locks)) {
3473                 printk("INFO: lockdep is turned off.\n");
3474                 return;
3475         }
3476         lockdep_print_held_locks(task);
3477 }
3478 EXPORT_SYMBOL_GPL(__debug_show_held_locks);
3479
3480 void debug_show_held_locks(struct task_struct *task)
3481 {
3482                 __debug_show_held_locks(task);
3483 }
3484 EXPORT_SYMBOL_GPL(debug_show_held_locks);
3485
3486 void lockdep_sys_exit(void)
3487 {
3488         struct task_struct *curr = current;
3489
3490         if (unlikely(curr->lockdep_depth)) {
3491                 if (!debug_locks_off())
3492                         return;
3493                 printk("\n================================================\n");
3494                 printk(  "[ BUG: lock held when returning to user space! ]\n");
3495                 printk(  "------------------------------------------------\n");
3496                 printk("%s/%d is leaving the kernel with locks still held!\n",
3497                                 curr->comm, curr->pid);
3498                 lockdep_print_held_locks(curr);
3499         }
3500 }