9a1e2bcc4b8d2062ad72d491f19ef49e7e18434d
[sfrench/cifs-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                 return NULL;
796         }
797         class = lock_classes + nr_lock_classes++;
798         debug_atomic_inc(&nr_unused_locks);
799         class->key = key;
800         class->name = lock->name;
801         class->subclass = subclass;
802         INIT_LIST_HEAD(&class->lock_entry);
803         INIT_LIST_HEAD(&class->locks_before);
804         INIT_LIST_HEAD(&class->locks_after);
805         class->name_version = count_matching_names(class);
806         /*
807          * We use RCU's safe list-add method to make
808          * parallel walking of the hash-list safe:
809          */
810         list_add_tail_rcu(&class->hash_entry, hash_head);
811         /*
812          * Add it to the global list of classes:
813          */
814         list_add_tail_rcu(&class->lock_entry, &all_lock_classes);
815
816         if (verbose(class)) {
817                 graph_unlock();
818                 raw_local_irq_restore(flags);
819
820                 printk("\nnew class %p: %s", class->key, class->name);
821                 if (class->name_version > 1)
822                         printk("#%d", class->name_version);
823                 printk("\n");
824                 dump_stack();
825
826                 raw_local_irq_save(flags);
827                 if (!graph_lock()) {
828                         raw_local_irq_restore(flags);
829                         return NULL;
830                 }
831         }
832 out_unlock_set:
833         graph_unlock();
834         raw_local_irq_restore(flags);
835
836         if (!subclass || force)
837                 lock->class_cache = class;
838
839         if (DEBUG_LOCKS_WARN_ON(class->subclass != subclass))
840                 return NULL;
841
842         return class;
843 }
844
845 #ifdef CONFIG_PROVE_LOCKING
846 /*
847  * Allocate a lockdep entry. (assumes the graph_lock held, returns
848  * with NULL on failure)
849  */
850 static struct lock_list *alloc_list_entry(void)
851 {
852         if (nr_list_entries >= MAX_LOCKDEP_ENTRIES) {
853                 if (!debug_locks_off_graph_unlock())
854                         return NULL;
855
856                 printk("BUG: MAX_LOCKDEP_ENTRIES too low!\n");
857                 printk("turning off the locking correctness validator.\n");
858                 return NULL;
859         }
860         return list_entries + nr_list_entries++;
861 }
862
863 /*
864  * Add a new dependency to the head of the list:
865  */
866 static int add_lock_to_list(struct lock_class *class, struct lock_class *this,
867                             struct list_head *head, unsigned long ip, int distance)
868 {
869         struct lock_list *entry;
870         /*
871          * Lock not present yet - get a new dependency struct and
872          * add it to the list:
873          */
874         entry = alloc_list_entry();
875         if (!entry)
876                 return 0;
877
878         if (!save_trace(&entry->trace))
879                 return 0;
880
881         entry->class = this;
882         entry->distance = distance;
883         /*
884          * Since we never remove from the dependency list, the list can
885          * be walked lockless by other CPUs, it's only allocation
886          * that must be protected by the spinlock. But this also means
887          * we must make new entries visible only once writes to the
888          * entry become visible - hence the RCU op:
889          */
890         list_add_tail_rcu(&entry->entry, head);
891
892         return 1;
893 }
894
895 /*
896  * Recursive, forwards-direction lock-dependency checking, used for
897  * both noncyclic checking and for hardirq-unsafe/softirq-unsafe
898  * checking.
899  *
900  * (to keep the stackframe of the recursive functions small we
901  *  use these global variables, and we also mark various helper
902  *  functions as noinline.)
903  */
904 static struct held_lock *check_source, *check_target;
905
906 /*
907  * Print a dependency chain entry (this is only done when a deadlock
908  * has been detected):
909  */
910 static noinline int
911 print_circular_bug_entry(struct lock_list *target, unsigned int depth)
912 {
913         if (debug_locks_silent)
914                 return 0;
915         printk("\n-> #%u", depth);
916         print_lock_name(target->class);
917         printk(":\n");
918         print_stack_trace(&target->trace, 6);
919
920         return 0;
921 }
922
923 /*
924  * When a circular dependency is detected, print the
925  * header first:
926  */
927 static noinline int
928 print_circular_bug_header(struct lock_list *entry, unsigned int depth)
929 {
930         struct task_struct *curr = current;
931
932         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
933                 return 0;
934
935         printk("\n=======================================================\n");
936         printk(  "[ INFO: possible circular locking dependency detected ]\n");
937         print_kernel_version();
938         printk(  "-------------------------------------------------------\n");
939         printk("%s/%d is trying to acquire lock:\n",
940                 curr->comm, task_pid_nr(curr));
941         print_lock(check_source);
942         printk("\nbut task is already holding lock:\n");
943         print_lock(check_target);
944         printk("\nwhich lock already depends on the new lock.\n\n");
945         printk("\nthe existing dependency chain (in reverse order) is:\n");
946
947         print_circular_bug_entry(entry, depth);
948
949         return 0;
950 }
951
952 static noinline int print_circular_bug_tail(void)
953 {
954         struct task_struct *curr = current;
955         struct lock_list this;
956
957         if (debug_locks_silent)
958                 return 0;
959
960         this.class = hlock_class(check_source);
961         if (!save_trace(&this.trace))
962                 return 0;
963
964         print_circular_bug_entry(&this, 0);
965
966         printk("\nother info that might help us debug this:\n\n");
967         lockdep_print_held_locks(curr);
968
969         printk("\nstack backtrace:\n");
970         dump_stack();
971
972         return 0;
973 }
974
975 #define RECURSION_LIMIT 40
976
977 static int noinline print_infinite_recursion_bug(void)
978 {
979         if (!debug_locks_off_graph_unlock())
980                 return 0;
981
982         WARN_ON(1);
983
984         return 0;
985 }
986
987 unsigned long __lockdep_count_forward_deps(struct lock_class *class,
988                                            unsigned int depth)
989 {
990         struct lock_list *entry;
991         unsigned long ret = 1;
992
993         if (lockdep_dependency_visit(class, depth))
994                 return 0;
995
996         /*
997          * Recurse this class's dependency list:
998          */
999         list_for_each_entry(entry, &class->locks_after, entry)
1000                 ret += __lockdep_count_forward_deps(entry->class, depth + 1);
1001
1002         return ret;
1003 }
1004
1005 unsigned long lockdep_count_forward_deps(struct lock_class *class)
1006 {
1007         unsigned long ret, flags;
1008
1009         local_irq_save(flags);
1010         __raw_spin_lock(&lockdep_lock);
1011         ret = __lockdep_count_forward_deps(class, 0);
1012         __raw_spin_unlock(&lockdep_lock);
1013         local_irq_restore(flags);
1014
1015         return ret;
1016 }
1017
1018 unsigned long __lockdep_count_backward_deps(struct lock_class *class,
1019                                             unsigned int depth)
1020 {
1021         struct lock_list *entry;
1022         unsigned long ret = 1;
1023
1024         if (lockdep_dependency_visit(class, depth))
1025                 return 0;
1026         /*
1027          * Recurse this class's dependency list:
1028          */
1029         list_for_each_entry(entry, &class->locks_before, entry)
1030                 ret += __lockdep_count_backward_deps(entry->class, depth + 1);
1031
1032         return ret;
1033 }
1034
1035 unsigned long lockdep_count_backward_deps(struct lock_class *class)
1036 {
1037         unsigned long ret, flags;
1038
1039         local_irq_save(flags);
1040         __raw_spin_lock(&lockdep_lock);
1041         ret = __lockdep_count_backward_deps(class, 0);
1042         __raw_spin_unlock(&lockdep_lock);
1043         local_irq_restore(flags);
1044
1045         return ret;
1046 }
1047
1048 /*
1049  * Prove that the dependency graph starting at <entry> can not
1050  * lead to <target>. Print an error and return 0 if it does.
1051  */
1052 static noinline int
1053 check_noncircular(struct lock_class *source, unsigned int depth)
1054 {
1055         struct lock_list *entry;
1056
1057         if (lockdep_dependency_visit(source, depth))
1058                 return 1;
1059
1060         debug_atomic_inc(&nr_cyclic_check_recursions);
1061         if (depth > max_recursion_depth)
1062                 max_recursion_depth = depth;
1063         if (depth >= RECURSION_LIMIT)
1064                 return print_infinite_recursion_bug();
1065         /*
1066          * Check this lock's dependency list:
1067          */
1068         list_for_each_entry(entry, &source->locks_after, entry) {
1069                 if (entry->class == hlock_class(check_target))
1070                         return print_circular_bug_header(entry, depth+1);
1071                 debug_atomic_inc(&nr_cyclic_checks);
1072                 if (!check_noncircular(entry->class, depth+1))
1073                         return print_circular_bug_entry(entry, depth+1);
1074         }
1075         return 1;
1076 }
1077
1078 #if defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING)
1079 /*
1080  * Forwards and backwards subgraph searching, for the purposes of
1081  * proving that two subgraphs can be connected by a new dependency
1082  * without creating any illegal irq-safe -> irq-unsafe lock dependency.
1083  */
1084 static enum lock_usage_bit find_usage_bit;
1085 static struct lock_class *forwards_match, *backwards_match;
1086
1087 /*
1088  * Find a node in the forwards-direction dependency sub-graph starting
1089  * at <source> that matches <find_usage_bit>.
1090  *
1091  * Return 2 if such a node exists in the subgraph, and put that node
1092  * into <forwards_match>.
1093  *
1094  * Return 1 otherwise and keep <forwards_match> unchanged.
1095  * Return 0 on error.
1096  */
1097 static noinline int
1098 find_usage_forwards(struct lock_class *source, unsigned int depth)
1099 {
1100         struct lock_list *entry;
1101         int ret;
1102
1103         if (lockdep_dependency_visit(source, depth))
1104                 return 1;
1105
1106         if (depth > max_recursion_depth)
1107                 max_recursion_depth = depth;
1108         if (depth >= RECURSION_LIMIT)
1109                 return print_infinite_recursion_bug();
1110
1111         debug_atomic_inc(&nr_find_usage_forwards_checks);
1112         if (source->usage_mask & (1 << find_usage_bit)) {
1113                 forwards_match = source;
1114                 return 2;
1115         }
1116
1117         /*
1118          * Check this lock's dependency list:
1119          */
1120         list_for_each_entry(entry, &source->locks_after, entry) {
1121                 debug_atomic_inc(&nr_find_usage_forwards_recursions);
1122                 ret = find_usage_forwards(entry->class, depth+1);
1123                 if (ret == 2 || ret == 0)
1124                         return ret;
1125         }
1126         return 1;
1127 }
1128
1129 /*
1130  * Find a node in the backwards-direction dependency sub-graph starting
1131  * at <source> that matches <find_usage_bit>.
1132  *
1133  * Return 2 if such a node exists in the subgraph, and put that node
1134  * into <backwards_match>.
1135  *
1136  * Return 1 otherwise and keep <backwards_match> unchanged.
1137  * Return 0 on error.
1138  */
1139 static noinline int
1140 find_usage_backwards(struct lock_class *source, unsigned int depth)
1141 {
1142         struct lock_list *entry;
1143         int ret;
1144
1145         if (lockdep_dependency_visit(source, depth))
1146                 return 1;
1147
1148         if (!__raw_spin_is_locked(&lockdep_lock))
1149                 return DEBUG_LOCKS_WARN_ON(1);
1150
1151         if (depth > max_recursion_depth)
1152                 max_recursion_depth = depth;
1153         if (depth >= RECURSION_LIMIT)
1154                 return print_infinite_recursion_bug();
1155
1156         debug_atomic_inc(&nr_find_usage_backwards_checks);
1157         if (source->usage_mask & (1 << find_usage_bit)) {
1158                 backwards_match = source;
1159                 return 2;
1160         }
1161
1162         if (!source && debug_locks_off_graph_unlock()) {
1163                 WARN_ON(1);
1164                 return 0;
1165         }
1166
1167         /*
1168          * Check this lock's dependency list:
1169          */
1170         list_for_each_entry(entry, &source->locks_before, entry) {
1171                 debug_atomic_inc(&nr_find_usage_backwards_recursions);
1172                 ret = find_usage_backwards(entry->class, depth+1);
1173                 if (ret == 2 || ret == 0)
1174                         return ret;
1175         }
1176         return 1;
1177 }
1178
1179 static int
1180 print_bad_irq_dependency(struct task_struct *curr,
1181                          struct held_lock *prev,
1182                          struct held_lock *next,
1183                          enum lock_usage_bit bit1,
1184                          enum lock_usage_bit bit2,
1185                          const char *irqclass)
1186 {
1187         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1188                 return 0;
1189
1190         printk("\n======================================================\n");
1191         printk(  "[ INFO: %s-safe -> %s-unsafe lock order detected ]\n",
1192                 irqclass, irqclass);
1193         print_kernel_version();
1194         printk(  "------------------------------------------------------\n");
1195         printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] is trying to acquire:\n",
1196                 curr->comm, task_pid_nr(curr),
1197                 curr->hardirq_context, hardirq_count() >> HARDIRQ_SHIFT,
1198                 curr->softirq_context, softirq_count() >> SOFTIRQ_SHIFT,
1199                 curr->hardirqs_enabled,
1200                 curr->softirqs_enabled);
1201         print_lock(next);
1202
1203         printk("\nand this task is already holding:\n");
1204         print_lock(prev);
1205         printk("which would create a new lock dependency:\n");
1206         print_lock_name(hlock_class(prev));
1207         printk(" ->");
1208         print_lock_name(hlock_class(next));
1209         printk("\n");
1210
1211         printk("\nbut this new dependency connects a %s-irq-safe lock:\n",
1212                 irqclass);
1213         print_lock_name(backwards_match);
1214         printk("\n... which became %s-irq-safe at:\n", irqclass);
1215
1216         print_stack_trace(backwards_match->usage_traces + bit1, 1);
1217
1218         printk("\nto a %s-irq-unsafe lock:\n", irqclass);
1219         print_lock_name(forwards_match);
1220         printk("\n... which became %s-irq-unsafe at:\n", irqclass);
1221         printk("...");
1222
1223         print_stack_trace(forwards_match->usage_traces + bit2, 1);
1224
1225         printk("\nother info that might help us debug this:\n\n");
1226         lockdep_print_held_locks(curr);
1227
1228         printk("\nthe %s-irq-safe lock's dependencies:\n", irqclass);
1229         print_lock_dependencies(backwards_match, 0);
1230
1231         printk("\nthe %s-irq-unsafe lock's dependencies:\n", irqclass);
1232         print_lock_dependencies(forwards_match, 0);
1233
1234         printk("\nstack backtrace:\n");
1235         dump_stack();
1236
1237         return 0;
1238 }
1239
1240 static int
1241 check_usage(struct task_struct *curr, struct held_lock *prev,
1242             struct held_lock *next, enum lock_usage_bit bit_backwards,
1243             enum lock_usage_bit bit_forwards, const char *irqclass)
1244 {
1245         int ret;
1246
1247         find_usage_bit = bit_backwards;
1248         /* fills in <backwards_match> */
1249         ret = find_usage_backwards(hlock_class(prev), 0);
1250         if (!ret || ret == 1)
1251                 return ret;
1252
1253         find_usage_bit = bit_forwards;
1254         ret = find_usage_forwards(hlock_class(next), 0);
1255         if (!ret || ret == 1)
1256                 return ret;
1257         /* ret == 2 */
1258         return print_bad_irq_dependency(curr, prev, next,
1259                         bit_backwards, bit_forwards, irqclass);
1260 }
1261
1262 static const char *state_names[] = {
1263 #define LOCKDEP_STATE(__STATE) \
1264         __stringify(__STATE),
1265 #include "lockdep_states.h"
1266 #undef LOCKDEP_STATE
1267 };
1268
1269 static const char *state_rnames[] = {
1270 #define LOCKDEP_STATE(__STATE) \
1271         __stringify(__STATE)"-READ",
1272 #include "lockdep_states.h"
1273 #undef LOCKDEP_STATE
1274 };
1275
1276 static inline const char *state_name(enum lock_usage_bit bit)
1277 {
1278         return (bit & 1) ? state_rnames[bit >> 2] : state_names[bit >> 2];
1279 }
1280
1281 static int exclusive_bit(int new_bit)
1282 {
1283         /*
1284          * USED_IN
1285          * USED_IN_READ
1286          * ENABLED
1287          * ENABLED_READ
1288          *
1289          * bit 0 - write/read
1290          * bit 1 - used_in/enabled
1291          * bit 2+  state
1292          */
1293
1294         int state = new_bit & ~3;
1295         int dir = new_bit & 2;
1296
1297         /*
1298          * keep state, bit flip the direction and strip read.
1299          */
1300         return state | (dir ^ 2);
1301 }
1302
1303 static int check_irq_usage(struct task_struct *curr, struct held_lock *prev,
1304                            struct held_lock *next, enum lock_usage_bit bit)
1305 {
1306         /*
1307          * Prove that the new dependency does not connect a hardirq-safe
1308          * lock with a hardirq-unsafe lock - to achieve this we search
1309          * the backwards-subgraph starting at <prev>, and the
1310          * forwards-subgraph starting at <next>:
1311          */
1312         if (!check_usage(curr, prev, next, bit,
1313                            exclusive_bit(bit), state_name(bit)))
1314                 return 0;
1315
1316         bit++; /* _READ */
1317
1318         /*
1319          * Prove that the new dependency does not connect a hardirq-safe-read
1320          * lock with a hardirq-unsafe lock - to achieve this we search
1321          * the backwards-subgraph starting at <prev>, and the
1322          * forwards-subgraph starting at <next>:
1323          */
1324         if (!check_usage(curr, prev, next, bit,
1325                            exclusive_bit(bit), state_name(bit)))
1326                 return 0;
1327
1328         return 1;
1329 }
1330
1331 static int
1332 check_prev_add_irq(struct task_struct *curr, struct held_lock *prev,
1333                 struct held_lock *next)
1334 {
1335 #define LOCKDEP_STATE(__STATE)                                          \
1336         if (!check_irq_usage(curr, prev, next, LOCK_USED_IN_##__STATE)) \
1337                 return 0;
1338 #include "lockdep_states.h"
1339 #undef LOCKDEP_STATE
1340
1341         return 1;
1342 }
1343
1344 static void inc_chains(void)
1345 {
1346         if (current->hardirq_context)
1347                 nr_hardirq_chains++;
1348         else {
1349                 if (current->softirq_context)
1350                         nr_softirq_chains++;
1351                 else
1352                         nr_process_chains++;
1353         }
1354 }
1355
1356 #else
1357
1358 static inline int
1359 check_prev_add_irq(struct task_struct *curr, struct held_lock *prev,
1360                 struct held_lock *next)
1361 {
1362         return 1;
1363 }
1364
1365 static inline void inc_chains(void)
1366 {
1367         nr_process_chains++;
1368 }
1369
1370 #endif
1371
1372 static int
1373 print_deadlock_bug(struct task_struct *curr, struct held_lock *prev,
1374                    struct held_lock *next)
1375 {
1376         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1377                 return 0;
1378
1379         printk("\n=============================================\n");
1380         printk(  "[ INFO: possible recursive locking detected ]\n");
1381         print_kernel_version();
1382         printk(  "---------------------------------------------\n");
1383         printk("%s/%d is trying to acquire lock:\n",
1384                 curr->comm, task_pid_nr(curr));
1385         print_lock(next);
1386         printk("\nbut task is already holding lock:\n");
1387         print_lock(prev);
1388
1389         printk("\nother info that might help us debug this:\n");
1390         lockdep_print_held_locks(curr);
1391
1392         printk("\nstack backtrace:\n");
1393         dump_stack();
1394
1395         return 0;
1396 }
1397
1398 /*
1399  * Check whether we are holding such a class already.
1400  *
1401  * (Note that this has to be done separately, because the graph cannot
1402  * detect such classes of deadlocks.)
1403  *
1404  * Returns: 0 on deadlock detected, 1 on OK, 2 on recursive read
1405  */
1406 static int
1407 check_deadlock(struct task_struct *curr, struct held_lock *next,
1408                struct lockdep_map *next_instance, int read)
1409 {
1410         struct held_lock *prev;
1411         struct held_lock *nest = NULL;
1412         int i;
1413
1414         for (i = 0; i < curr->lockdep_depth; i++) {
1415                 prev = curr->held_locks + i;
1416
1417                 if (prev->instance == next->nest_lock)
1418                         nest = prev;
1419
1420                 if (hlock_class(prev) != hlock_class(next))
1421                         continue;
1422
1423                 /*
1424                  * Allow read-after-read recursion of the same
1425                  * lock class (i.e. read_lock(lock)+read_lock(lock)):
1426                  */
1427                 if ((read == 2) && prev->read)
1428                         return 2;
1429
1430                 /*
1431                  * We're holding the nest_lock, which serializes this lock's
1432                  * nesting behaviour.
1433                  */
1434                 if (nest)
1435                         return 2;
1436
1437                 return print_deadlock_bug(curr, prev, next);
1438         }
1439         return 1;
1440 }
1441
1442 /*
1443  * There was a chain-cache miss, and we are about to add a new dependency
1444  * to a previous lock. We recursively validate the following rules:
1445  *
1446  *  - would the adding of the <prev> -> <next> dependency create a
1447  *    circular dependency in the graph? [== circular deadlock]
1448  *
1449  *  - does the new prev->next dependency connect any hardirq-safe lock
1450  *    (in the full backwards-subgraph starting at <prev>) with any
1451  *    hardirq-unsafe lock (in the full forwards-subgraph starting at
1452  *    <next>)? [== illegal lock inversion with hardirq contexts]
1453  *
1454  *  - does the new prev->next dependency connect any softirq-safe lock
1455  *    (in the full backwards-subgraph starting at <prev>) with any
1456  *    softirq-unsafe lock (in the full forwards-subgraph starting at
1457  *    <next>)? [== illegal lock inversion with softirq contexts]
1458  *
1459  * any of these scenarios could lead to a deadlock.
1460  *
1461  * Then if all the validations pass, we add the forwards and backwards
1462  * dependency.
1463  */
1464 static int
1465 check_prev_add(struct task_struct *curr, struct held_lock *prev,
1466                struct held_lock *next, int distance)
1467 {
1468         struct lock_list *entry;
1469         int ret;
1470
1471         /*
1472          * Prove that the new <prev> -> <next> dependency would not
1473          * create a circular dependency in the graph. (We do this by
1474          * forward-recursing into the graph starting at <next>, and
1475          * checking whether we can reach <prev>.)
1476          *
1477          * We are using global variables to control the recursion, to
1478          * keep the stackframe size of the recursive functions low:
1479          */
1480         check_source = next;
1481         check_target = prev;
1482         if (!(check_noncircular(hlock_class(next), 0)))
1483                 return print_circular_bug_tail();
1484
1485         if (!check_prev_add_irq(curr, prev, next))
1486                 return 0;
1487
1488         /*
1489          * For recursive read-locks we do all the dependency checks,
1490          * but we dont store read-triggered dependencies (only
1491          * write-triggered dependencies). This ensures that only the
1492          * write-side dependencies matter, and that if for example a
1493          * write-lock never takes any other locks, then the reads are
1494          * equivalent to a NOP.
1495          */
1496         if (next->read == 2 || prev->read == 2)
1497                 return 1;
1498         /*
1499          * Is the <prev> -> <next> dependency already present?
1500          *
1501          * (this may occur even though this is a new chain: consider
1502          *  e.g. the L1 -> L2 -> L3 -> L4 and the L5 -> L1 -> L2 -> L3
1503          *  chains - the second one will be new, but L1 already has
1504          *  L2 added to its dependency list, due to the first chain.)
1505          */
1506         list_for_each_entry(entry, &hlock_class(prev)->locks_after, entry) {
1507                 if (entry->class == hlock_class(next)) {
1508                         if (distance == 1)
1509                                 entry->distance = 1;
1510                         return 2;
1511                 }
1512         }
1513
1514         /*
1515          * Ok, all validations passed, add the new lock
1516          * to the previous lock's dependency list:
1517          */
1518         ret = add_lock_to_list(hlock_class(prev), hlock_class(next),
1519                                &hlock_class(prev)->locks_after,
1520                                next->acquire_ip, distance);
1521
1522         if (!ret)
1523                 return 0;
1524
1525         ret = add_lock_to_list(hlock_class(next), hlock_class(prev),
1526                                &hlock_class(next)->locks_before,
1527                                next->acquire_ip, distance);
1528         if (!ret)
1529                 return 0;
1530
1531         /*
1532          * Debugging printouts:
1533          */
1534         if (verbose(hlock_class(prev)) || verbose(hlock_class(next))) {
1535                 graph_unlock();
1536                 printk("\n new dependency: ");
1537                 print_lock_name(hlock_class(prev));
1538                 printk(" => ");
1539                 print_lock_name(hlock_class(next));
1540                 printk("\n");
1541                 dump_stack();
1542                 return graph_lock();
1543         }
1544         return 1;
1545 }
1546
1547 /*
1548  * Add the dependency to all directly-previous locks that are 'relevant'.
1549  * The ones that are relevant are (in increasing distance from curr):
1550  * all consecutive trylock entries and the final non-trylock entry - or
1551  * the end of this context's lock-chain - whichever comes first.
1552  */
1553 static int
1554 check_prevs_add(struct task_struct *curr, struct held_lock *next)
1555 {
1556         int depth = curr->lockdep_depth;
1557         struct held_lock *hlock;
1558
1559         /*
1560          * Debugging checks.
1561          *
1562          * Depth must not be zero for a non-head lock:
1563          */
1564         if (!depth)
1565                 goto out_bug;
1566         /*
1567          * At least two relevant locks must exist for this
1568          * to be a head:
1569          */
1570         if (curr->held_locks[depth].irq_context !=
1571                         curr->held_locks[depth-1].irq_context)
1572                 goto out_bug;
1573
1574         for (;;) {
1575                 int distance = curr->lockdep_depth - depth + 1;
1576                 hlock = curr->held_locks + depth-1;
1577                 /*
1578                  * Only non-recursive-read entries get new dependencies
1579                  * added:
1580                  */
1581                 if (hlock->read != 2) {
1582                         if (!check_prev_add(curr, hlock, next, distance))
1583                                 return 0;
1584                         /*
1585                          * Stop after the first non-trylock entry,
1586                          * as non-trylock entries have added their
1587                          * own direct dependencies already, so this
1588                          * lock is connected to them indirectly:
1589                          */
1590                         if (!hlock->trylock)
1591                                 break;
1592                 }
1593                 depth--;
1594                 /*
1595                  * End of lock-stack?
1596                  */
1597                 if (!depth)
1598                         break;
1599                 /*
1600                  * Stop the search if we cross into another context:
1601                  */
1602                 if (curr->held_locks[depth].irq_context !=
1603                                 curr->held_locks[depth-1].irq_context)
1604                         break;
1605         }
1606         return 1;
1607 out_bug:
1608         if (!debug_locks_off_graph_unlock())
1609                 return 0;
1610
1611         WARN_ON(1);
1612
1613         return 0;
1614 }
1615
1616 unsigned long nr_lock_chains;
1617 struct lock_chain lock_chains[MAX_LOCKDEP_CHAINS];
1618 int nr_chain_hlocks;
1619 static u16 chain_hlocks[MAX_LOCKDEP_CHAIN_HLOCKS];
1620
1621 struct lock_class *lock_chain_get_class(struct lock_chain *chain, int i)
1622 {
1623         return lock_classes + chain_hlocks[chain->base + i];
1624 }
1625
1626 /*
1627  * Look up a dependency chain. If the key is not present yet then
1628  * add it and return 1 - in this case the new dependency chain is
1629  * validated. If the key is already hashed, return 0.
1630  * (On return with 1 graph_lock is held.)
1631  */
1632 static inline int lookup_chain_cache(struct task_struct *curr,
1633                                      struct held_lock *hlock,
1634                                      u64 chain_key)
1635 {
1636         struct lock_class *class = hlock_class(hlock);
1637         struct list_head *hash_head = chainhashentry(chain_key);
1638         struct lock_chain *chain;
1639         struct held_lock *hlock_curr, *hlock_next;
1640         int i, j, n, cn;
1641
1642         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
1643                 return 0;
1644         /*
1645          * We can walk it lock-free, because entries only get added
1646          * to the hash:
1647          */
1648         list_for_each_entry(chain, hash_head, entry) {
1649                 if (chain->chain_key == chain_key) {
1650 cache_hit:
1651                         debug_atomic_inc(&chain_lookup_hits);
1652                         if (very_verbose(class))
1653                                 printk("\nhash chain already cached, key: "
1654                                         "%016Lx tail class: [%p] %s\n",
1655                                         (unsigned long long)chain_key,
1656                                         class->key, class->name);
1657                         return 0;
1658                 }
1659         }
1660         if (very_verbose(class))
1661                 printk("\nnew hash chain, key: %016Lx tail class: [%p] %s\n",
1662                         (unsigned long long)chain_key, class->key, class->name);
1663         /*
1664          * Allocate a new chain entry from the static array, and add
1665          * it to the hash:
1666          */
1667         if (!graph_lock())
1668                 return 0;
1669         /*
1670          * We have to walk the chain again locked - to avoid duplicates:
1671          */
1672         list_for_each_entry(chain, hash_head, entry) {
1673                 if (chain->chain_key == chain_key) {
1674                         graph_unlock();
1675                         goto cache_hit;
1676                 }
1677         }
1678         if (unlikely(nr_lock_chains >= MAX_LOCKDEP_CHAINS)) {
1679                 if (!debug_locks_off_graph_unlock())
1680                         return 0;
1681
1682                 printk("BUG: MAX_LOCKDEP_CHAINS too low!\n");
1683                 printk("turning off the locking correctness validator.\n");
1684                 return 0;
1685         }
1686         chain = lock_chains + nr_lock_chains++;
1687         chain->chain_key = chain_key;
1688         chain->irq_context = hlock->irq_context;
1689         /* Find the first held_lock of current chain */
1690         hlock_next = hlock;
1691         for (i = curr->lockdep_depth - 1; i >= 0; i--) {
1692                 hlock_curr = curr->held_locks + i;
1693                 if (hlock_curr->irq_context != hlock_next->irq_context)
1694                         break;
1695                 hlock_next = hlock;
1696         }
1697         i++;
1698         chain->depth = curr->lockdep_depth + 1 - i;
1699         cn = nr_chain_hlocks;
1700         while (cn + chain->depth <= MAX_LOCKDEP_CHAIN_HLOCKS) {
1701                 n = cmpxchg(&nr_chain_hlocks, cn, cn + chain->depth);
1702                 if (n == cn)
1703                         break;
1704                 cn = n;
1705         }
1706         if (likely(cn + chain->depth <= MAX_LOCKDEP_CHAIN_HLOCKS)) {
1707                 chain->base = cn;
1708                 for (j = 0; j < chain->depth - 1; j++, i++) {
1709                         int lock_id = curr->held_locks[i].class_idx - 1;
1710                         chain_hlocks[chain->base + j] = lock_id;
1711                 }
1712                 chain_hlocks[chain->base + j] = class - lock_classes;
1713         }
1714         list_add_tail_rcu(&chain->entry, hash_head);
1715         debug_atomic_inc(&chain_lookup_misses);
1716         inc_chains();
1717
1718         return 1;
1719 }
1720
1721 static int validate_chain(struct task_struct *curr, struct lockdep_map *lock,
1722                 struct held_lock *hlock, int chain_head, u64 chain_key)
1723 {
1724         /*
1725          * Trylock needs to maintain the stack of held locks, but it
1726          * does not add new dependencies, because trylock can be done
1727          * in any order.
1728          *
1729          * We look up the chain_key and do the O(N^2) check and update of
1730          * the dependencies only if this is a new dependency chain.
1731          * (If lookup_chain_cache() returns with 1 it acquires
1732          * graph_lock for us)
1733          */
1734         if (!hlock->trylock && (hlock->check == 2) &&
1735             lookup_chain_cache(curr, hlock, chain_key)) {
1736                 /*
1737                  * Check whether last held lock:
1738                  *
1739                  * - is irq-safe, if this lock is irq-unsafe
1740                  * - is softirq-safe, if this lock is hardirq-unsafe
1741                  *
1742                  * And check whether the new lock's dependency graph
1743                  * could lead back to the previous lock.
1744                  *
1745                  * any of these scenarios could lead to a deadlock. If
1746                  * All validations
1747                  */
1748                 int ret = check_deadlock(curr, hlock, lock, hlock->read);
1749
1750                 if (!ret)
1751                         return 0;
1752                 /*
1753                  * Mark recursive read, as we jump over it when
1754                  * building dependencies (just like we jump over
1755                  * trylock entries):
1756                  */
1757                 if (ret == 2)
1758                         hlock->read = 2;
1759                 /*
1760                  * Add dependency only if this lock is not the head
1761                  * of the chain, and if it's not a secondary read-lock:
1762                  */
1763                 if (!chain_head && ret != 2)
1764                         if (!check_prevs_add(curr, hlock))
1765                                 return 0;
1766                 graph_unlock();
1767         } else
1768                 /* after lookup_chain_cache(): */
1769                 if (unlikely(!debug_locks))
1770                         return 0;
1771
1772         return 1;
1773 }
1774 #else
1775 static inline int validate_chain(struct task_struct *curr,
1776                 struct lockdep_map *lock, struct held_lock *hlock,
1777                 int chain_head, u64 chain_key)
1778 {
1779         return 1;
1780 }
1781 #endif
1782
1783 /*
1784  * We are building curr_chain_key incrementally, so double-check
1785  * it from scratch, to make sure that it's done correctly:
1786  */
1787 static void check_chain_key(struct task_struct *curr)
1788 {
1789 #ifdef CONFIG_DEBUG_LOCKDEP
1790         struct held_lock *hlock, *prev_hlock = NULL;
1791         unsigned int i, id;
1792         u64 chain_key = 0;
1793
1794         for (i = 0; i < curr->lockdep_depth; i++) {
1795                 hlock = curr->held_locks + i;
1796                 if (chain_key != hlock->prev_chain_key) {
1797                         debug_locks_off();
1798                         WARN(1, "hm#1, depth: %u [%u], %016Lx != %016Lx\n",
1799                                 curr->lockdep_depth, i,
1800                                 (unsigned long long)chain_key,
1801                                 (unsigned long long)hlock->prev_chain_key);
1802                         return;
1803                 }
1804                 id = hlock->class_idx - 1;
1805                 if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
1806                         return;
1807
1808                 if (prev_hlock && (prev_hlock->irq_context !=
1809                                                         hlock->irq_context))
1810                         chain_key = 0;
1811                 chain_key = iterate_chain_key(chain_key, id);
1812                 prev_hlock = hlock;
1813         }
1814         if (chain_key != curr->curr_chain_key) {
1815                 debug_locks_off();
1816                 WARN(1, "hm#2, depth: %u [%u], %016Lx != %016Lx\n",
1817                         curr->lockdep_depth, i,
1818                         (unsigned long long)chain_key,
1819                         (unsigned long long)curr->curr_chain_key);
1820         }
1821 #endif
1822 }
1823
1824 static int
1825 print_usage_bug(struct task_struct *curr, struct held_lock *this,
1826                 enum lock_usage_bit prev_bit, enum lock_usage_bit new_bit)
1827 {
1828         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1829                 return 0;
1830
1831         printk("\n=================================\n");
1832         printk(  "[ INFO: inconsistent lock state ]\n");
1833         print_kernel_version();
1834         printk(  "---------------------------------\n");
1835
1836         printk("inconsistent {%s} -> {%s} usage.\n",
1837                 usage_str[prev_bit], usage_str[new_bit]);
1838
1839         printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] takes:\n",
1840                 curr->comm, task_pid_nr(curr),
1841                 trace_hardirq_context(curr), hardirq_count() >> HARDIRQ_SHIFT,
1842                 trace_softirq_context(curr), softirq_count() >> SOFTIRQ_SHIFT,
1843                 trace_hardirqs_enabled(curr),
1844                 trace_softirqs_enabled(curr));
1845         print_lock(this);
1846
1847         printk("{%s} state was registered at:\n", usage_str[prev_bit]);
1848         print_stack_trace(hlock_class(this)->usage_traces + prev_bit, 1);
1849
1850         print_irqtrace_events(curr);
1851         printk("\nother info that might help us debug this:\n");
1852         lockdep_print_held_locks(curr);
1853
1854         printk("\nstack backtrace:\n");
1855         dump_stack();
1856
1857         return 0;
1858 }
1859
1860 /*
1861  * Print out an error if an invalid bit is set:
1862  */
1863 static inline int
1864 valid_state(struct task_struct *curr, struct held_lock *this,
1865             enum lock_usage_bit new_bit, enum lock_usage_bit bad_bit)
1866 {
1867         if (unlikely(hlock_class(this)->usage_mask & (1 << bad_bit)))
1868                 return print_usage_bug(curr, this, bad_bit, new_bit);
1869         return 1;
1870 }
1871
1872 static int mark_lock(struct task_struct *curr, struct held_lock *this,
1873                      enum lock_usage_bit new_bit);
1874
1875 #if defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING)
1876
1877 /*
1878  * print irq inversion bug:
1879  */
1880 static int
1881 print_irq_inversion_bug(struct task_struct *curr, struct lock_class *other,
1882                         struct held_lock *this, int forwards,
1883                         const char *irqclass)
1884 {
1885         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1886                 return 0;
1887
1888         printk("\n=========================================================\n");
1889         printk(  "[ INFO: possible irq lock inversion dependency detected ]\n");
1890         print_kernel_version();
1891         printk(  "---------------------------------------------------------\n");
1892         printk("%s/%d just changed the state of lock:\n",
1893                 curr->comm, task_pid_nr(curr));
1894         print_lock(this);
1895         if (forwards)
1896                 printk("but this lock took another, %s-unsafe lock in the past:\n", irqclass);
1897         else
1898                 printk("but this lock was taken by another, %s-safe lock in the past:\n", irqclass);
1899         print_lock_name(other);
1900         printk("\n\nand interrupts could create inverse lock ordering between them.\n\n");
1901
1902         printk("\nother info that might help us debug this:\n");
1903         lockdep_print_held_locks(curr);
1904
1905         printk("\nthe first lock's dependencies:\n");
1906         print_lock_dependencies(hlock_class(this), 0);
1907
1908         printk("\nthe second lock's dependencies:\n");
1909         print_lock_dependencies(other, 0);
1910
1911         printk("\nstack backtrace:\n");
1912         dump_stack();
1913
1914         return 0;
1915 }
1916
1917 /*
1918  * Prove that in the forwards-direction subgraph starting at <this>
1919  * there is no lock matching <mask>:
1920  */
1921 static int
1922 check_usage_forwards(struct task_struct *curr, struct held_lock *this,
1923                      enum lock_usage_bit bit, const char *irqclass)
1924 {
1925         int ret;
1926
1927         find_usage_bit = bit;
1928         /* fills in <forwards_match> */
1929         ret = find_usage_forwards(hlock_class(this), 0);
1930         if (!ret || ret == 1)
1931                 return ret;
1932
1933         return print_irq_inversion_bug(curr, forwards_match, this, 1, irqclass);
1934 }
1935
1936 /*
1937  * Prove that in the backwards-direction subgraph starting at <this>
1938  * there is no lock matching <mask>:
1939  */
1940 static int
1941 check_usage_backwards(struct task_struct *curr, struct held_lock *this,
1942                       enum lock_usage_bit bit, const char *irqclass)
1943 {
1944         int ret;
1945
1946         find_usage_bit = bit;
1947         /* fills in <backwards_match> */
1948         ret = find_usage_backwards(hlock_class(this), 0);
1949         if (!ret || ret == 1)
1950                 return ret;
1951
1952         return print_irq_inversion_bug(curr, backwards_match, this, 0, irqclass);
1953 }
1954
1955 void print_irqtrace_events(struct task_struct *curr)
1956 {
1957         printk("irq event stamp: %u\n", curr->irq_events);
1958         printk("hardirqs last  enabled at (%u): ", curr->hardirq_enable_event);
1959         print_ip_sym(curr->hardirq_enable_ip);
1960         printk("hardirqs last disabled at (%u): ", curr->hardirq_disable_event);
1961         print_ip_sym(curr->hardirq_disable_ip);
1962         printk("softirqs last  enabled at (%u): ", curr->softirq_enable_event);
1963         print_ip_sym(curr->softirq_enable_ip);
1964         printk("softirqs last disabled at (%u): ", curr->softirq_disable_event);
1965         print_ip_sym(curr->softirq_disable_ip);
1966 }
1967
1968 static int HARDIRQ_verbose(struct lock_class *class)
1969 {
1970 #if HARDIRQ_VERBOSE
1971         return class_filter(class);
1972 #endif
1973         return 0;
1974 }
1975
1976 static int SOFTIRQ_verbose(struct lock_class *class)
1977 {
1978 #if SOFTIRQ_VERBOSE
1979         return class_filter(class);
1980 #endif
1981         return 0;
1982 }
1983
1984 static int RECLAIM_FS_verbose(struct lock_class *class)
1985 {
1986 #if RECLAIM_VERBOSE
1987         return class_filter(class);
1988 #endif
1989         return 0;
1990 }
1991
1992 #define STRICT_READ_CHECKS      1
1993
1994 static int (*state_verbose_f[])(struct lock_class *class) = {
1995 #define LOCKDEP_STATE(__STATE) \
1996         __STATE##_verbose,
1997 #include "lockdep_states.h"
1998 #undef LOCKDEP_STATE
1999 };
2000
2001 static inline int state_verbose(enum lock_usage_bit bit,
2002                                 struct lock_class *class)
2003 {
2004         return state_verbose_f[bit >> 2](class);
2005 }
2006
2007 typedef int (*check_usage_f)(struct task_struct *, struct held_lock *,
2008                              enum lock_usage_bit bit, const char *name);
2009
2010 static int
2011 mark_lock_irq(struct task_struct *curr, struct held_lock *this,
2012                 enum lock_usage_bit new_bit)
2013 {
2014         int excl_bit = exclusive_bit(new_bit);
2015         int read = new_bit & 1;
2016         int dir = new_bit & 2;
2017
2018         /*
2019          * mark USED_IN has to look forwards -- to ensure no dependency
2020          * has ENABLED state, which would allow recursion deadlocks.
2021          *
2022          * mark ENABLED has to look backwards -- to ensure no dependee
2023          * has USED_IN state, which, again, would allow  recursion deadlocks.
2024          */
2025         check_usage_f usage = dir ?
2026                 check_usage_backwards : check_usage_forwards;
2027
2028         /*
2029          * Validate that this particular lock does not have conflicting
2030          * usage states.
2031          */
2032         if (!valid_state(curr, this, new_bit, excl_bit))
2033                 return 0;
2034
2035         /*
2036          * Validate that the lock dependencies don't have conflicting usage
2037          * states.
2038          */
2039         if ((!read || !dir || STRICT_READ_CHECKS) &&
2040                         !usage(curr, this, excl_bit, state_name(new_bit & ~1)))
2041                 return 0;
2042
2043         /*
2044          * Check for read in write conflicts
2045          */
2046         if (!read) {
2047                 if (!valid_state(curr, this, new_bit, excl_bit + 1))
2048                         return 0;
2049
2050                 if (STRICT_READ_CHECKS &&
2051                         !usage(curr, this, excl_bit + 1,
2052                                 state_name(new_bit + 1)))
2053                         return 0;
2054         }
2055
2056         if (state_verbose(new_bit, hlock_class(this)))
2057                 return 2;
2058
2059         return 1;
2060 }
2061
2062 enum mark_type {
2063 #define LOCKDEP_STATE(__STATE)  __STATE,
2064 #include "lockdep_states.h"
2065 #undef LOCKDEP_STATE
2066 };
2067
2068 /*
2069  * Mark all held locks with a usage bit:
2070  */
2071 static int
2072 mark_held_locks(struct task_struct *curr, enum mark_type mark)
2073 {
2074         enum lock_usage_bit usage_bit;
2075         struct held_lock *hlock;
2076         int i;
2077
2078         for (i = 0; i < curr->lockdep_depth; i++) {
2079                 hlock = curr->held_locks + i;
2080
2081                 usage_bit = 2 + (mark << 2); /* ENABLED */
2082                 if (hlock->read)
2083                         usage_bit += 1; /* READ */
2084
2085                 BUG_ON(usage_bit >= LOCK_USAGE_STATES);
2086
2087                 if (!mark_lock(curr, hlock, usage_bit))
2088                         return 0;
2089         }
2090
2091         return 1;
2092 }
2093
2094 /*
2095  * Debugging helper: via this flag we know that we are in
2096  * 'early bootup code', and will warn about any invalid irqs-on event:
2097  */
2098 static int early_boot_irqs_enabled;
2099
2100 void early_boot_irqs_off(void)
2101 {
2102         early_boot_irqs_enabled = 0;
2103 }
2104
2105 void early_boot_irqs_on(void)
2106 {
2107         early_boot_irqs_enabled = 1;
2108 }
2109
2110 /*
2111  * Hardirqs will be enabled:
2112  */
2113 void trace_hardirqs_on_caller(unsigned long ip)
2114 {
2115         struct task_struct *curr = current;
2116
2117         time_hardirqs_on(CALLER_ADDR0, ip);
2118
2119         if (unlikely(!debug_locks || current->lockdep_recursion))
2120                 return;
2121
2122         if (DEBUG_LOCKS_WARN_ON(unlikely(!early_boot_irqs_enabled)))
2123                 return;
2124
2125         if (unlikely(curr->hardirqs_enabled)) {
2126                 debug_atomic_inc(&redundant_hardirqs_on);
2127                 return;
2128         }
2129         /* we'll do an OFF -> ON transition: */
2130         curr->hardirqs_enabled = 1;
2131
2132         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2133                 return;
2134         if (DEBUG_LOCKS_WARN_ON(current->hardirq_context))
2135                 return;
2136         /*
2137          * We are going to turn hardirqs on, so set the
2138          * usage bit for all held locks:
2139          */
2140         if (!mark_held_locks(curr, HARDIRQ))
2141                 return;
2142         /*
2143          * If we have softirqs enabled, then set the usage
2144          * bit for all held locks. (disabled hardirqs prevented
2145          * this bit from being set before)
2146          */
2147         if (curr->softirqs_enabled)
2148                 if (!mark_held_locks(curr, SOFTIRQ))
2149                         return;
2150
2151         curr->hardirq_enable_ip = ip;
2152         curr->hardirq_enable_event = ++curr->irq_events;
2153         debug_atomic_inc(&hardirqs_on_events);
2154 }
2155 EXPORT_SYMBOL(trace_hardirqs_on_caller);
2156
2157 void trace_hardirqs_on(void)
2158 {
2159         trace_hardirqs_on_caller(CALLER_ADDR0);
2160 }
2161 EXPORT_SYMBOL(trace_hardirqs_on);
2162
2163 /*
2164  * Hardirqs were disabled:
2165  */
2166 void trace_hardirqs_off_caller(unsigned long ip)
2167 {
2168         struct task_struct *curr = current;
2169
2170         time_hardirqs_off(CALLER_ADDR0, ip);
2171
2172         if (unlikely(!debug_locks || current->lockdep_recursion))
2173                 return;
2174
2175         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2176                 return;
2177
2178         if (curr->hardirqs_enabled) {
2179                 /*
2180                  * We have done an ON -> OFF transition:
2181                  */
2182                 curr->hardirqs_enabled = 0;
2183                 curr->hardirq_disable_ip = ip;
2184                 curr->hardirq_disable_event = ++curr->irq_events;
2185                 debug_atomic_inc(&hardirqs_off_events);
2186         } else
2187                 debug_atomic_inc(&redundant_hardirqs_off);
2188 }
2189 EXPORT_SYMBOL(trace_hardirqs_off_caller);
2190
2191 void trace_hardirqs_off(void)
2192 {
2193         trace_hardirqs_off_caller(CALLER_ADDR0);
2194 }
2195 EXPORT_SYMBOL(trace_hardirqs_off);
2196
2197 /*
2198  * Softirqs will be enabled:
2199  */
2200 void trace_softirqs_on(unsigned long ip)
2201 {
2202         struct task_struct *curr = current;
2203
2204         if (unlikely(!debug_locks))
2205                 return;
2206
2207         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2208                 return;
2209
2210         if (curr->softirqs_enabled) {
2211                 debug_atomic_inc(&redundant_softirqs_on);
2212                 return;
2213         }
2214
2215         /*
2216          * We'll do an OFF -> ON transition:
2217          */
2218         curr->softirqs_enabled = 1;
2219         curr->softirq_enable_ip = ip;
2220         curr->softirq_enable_event = ++curr->irq_events;
2221         debug_atomic_inc(&softirqs_on_events);
2222         /*
2223          * We are going to turn softirqs on, so set the
2224          * usage bit for all held locks, if hardirqs are
2225          * enabled too:
2226          */
2227         if (curr->hardirqs_enabled)
2228                 mark_held_locks(curr, SOFTIRQ);
2229 }
2230
2231 /*
2232  * Softirqs were disabled:
2233  */
2234 void trace_softirqs_off(unsigned long ip)
2235 {
2236         struct task_struct *curr = current;
2237
2238         if (unlikely(!debug_locks))
2239                 return;
2240
2241         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2242                 return;
2243
2244         if (curr->softirqs_enabled) {
2245                 /*
2246                  * We have done an ON -> OFF transition:
2247                  */
2248                 curr->softirqs_enabled = 0;
2249                 curr->softirq_disable_ip = ip;
2250                 curr->softirq_disable_event = ++curr->irq_events;
2251                 debug_atomic_inc(&softirqs_off_events);
2252                 DEBUG_LOCKS_WARN_ON(!softirq_count());
2253         } else
2254                 debug_atomic_inc(&redundant_softirqs_off);
2255 }
2256
2257 void lockdep_trace_alloc(gfp_t gfp_mask)
2258 {
2259         struct task_struct *curr = current;
2260
2261         if (unlikely(!debug_locks))
2262                 return;
2263
2264         /* no reclaim without waiting on it */
2265         if (!(gfp_mask & __GFP_WAIT))
2266                 return;
2267
2268         /* this guy won't enter reclaim */
2269         if ((curr->flags & PF_MEMALLOC) && !(gfp_mask & __GFP_NOMEMALLOC))
2270                 return;
2271
2272         /* We're only interested __GFP_FS allocations for now */
2273         if (!(gfp_mask & __GFP_FS))
2274                 return;
2275
2276         if (DEBUG_LOCKS_WARN_ON(irqs_disabled()))
2277                 return;
2278
2279         mark_held_locks(curr, RECLAIM_FS);
2280 }
2281
2282 static int mark_irqflags(struct task_struct *curr, struct held_lock *hlock)
2283 {
2284         /*
2285          * If non-trylock use in a hardirq or softirq context, then
2286          * mark the lock as used in these contexts:
2287          */
2288         if (!hlock->trylock) {
2289                 if (hlock->read) {
2290                         if (curr->hardirq_context)
2291                                 if (!mark_lock(curr, hlock,
2292                                                 LOCK_USED_IN_HARDIRQ_READ))
2293                                         return 0;
2294                         if (curr->softirq_context)
2295                                 if (!mark_lock(curr, hlock,
2296                                                 LOCK_USED_IN_SOFTIRQ_READ))
2297                                         return 0;
2298                 } else {
2299                         if (curr->hardirq_context)
2300                                 if (!mark_lock(curr, hlock, LOCK_USED_IN_HARDIRQ))
2301                                         return 0;
2302                         if (curr->softirq_context)
2303                                 if (!mark_lock(curr, hlock, LOCK_USED_IN_SOFTIRQ))
2304                                         return 0;
2305                 }
2306         }
2307         if (!hlock->hardirqs_off) {
2308                 if (hlock->read) {
2309                         if (!mark_lock(curr, hlock,
2310                                         LOCK_ENABLED_HARDIRQ_READ))
2311                                 return 0;
2312                         if (curr->softirqs_enabled)
2313                                 if (!mark_lock(curr, hlock,
2314                                                 LOCK_ENABLED_SOFTIRQ_READ))
2315                                         return 0;
2316                 } else {
2317                         if (!mark_lock(curr, hlock,
2318                                         LOCK_ENABLED_HARDIRQ))
2319                                 return 0;
2320                         if (curr->softirqs_enabled)
2321                                 if (!mark_lock(curr, hlock,
2322                                                 LOCK_ENABLED_SOFTIRQ))
2323                                         return 0;
2324                 }
2325         }
2326
2327         /*
2328          * We reuse the irq context infrastructure more broadly as a general
2329          * context checking code. This tests GFP_FS recursion (a lock taken
2330          * during reclaim for a GFP_FS allocation is held over a GFP_FS
2331          * allocation).
2332          */
2333         if (!hlock->trylock && (curr->lockdep_reclaim_gfp & __GFP_FS)) {
2334                 if (hlock->read) {
2335                         if (!mark_lock(curr, hlock, LOCK_USED_IN_RECLAIM_FS_READ))
2336                                         return 0;
2337                 } else {
2338                         if (!mark_lock(curr, hlock, LOCK_USED_IN_RECLAIM_FS))
2339                                         return 0;
2340                 }
2341         }
2342
2343         return 1;
2344 }
2345
2346 static int separate_irq_context(struct task_struct *curr,
2347                 struct held_lock *hlock)
2348 {
2349         unsigned int depth = curr->lockdep_depth;
2350
2351         /*
2352          * Keep track of points where we cross into an interrupt context:
2353          */
2354         hlock->irq_context = 2*(curr->hardirq_context ? 1 : 0) +
2355                                 curr->softirq_context;
2356         if (depth) {
2357                 struct held_lock *prev_hlock;
2358
2359                 prev_hlock = curr->held_locks + depth-1;
2360                 /*
2361                  * If we cross into another context, reset the
2362                  * hash key (this also prevents the checking and the
2363                  * adding of the dependency to 'prev'):
2364                  */
2365                 if (prev_hlock->irq_context != hlock->irq_context)
2366                         return 1;
2367         }
2368         return 0;
2369 }
2370
2371 #else
2372
2373 static inline
2374 int mark_lock_irq(struct task_struct *curr, struct held_lock *this,
2375                 enum lock_usage_bit new_bit)
2376 {
2377         WARN_ON(1);
2378         return 1;
2379 }
2380
2381 static inline int mark_irqflags(struct task_struct *curr,
2382                 struct held_lock *hlock)
2383 {
2384         return 1;
2385 }
2386
2387 static inline int separate_irq_context(struct task_struct *curr,
2388                 struct held_lock *hlock)
2389 {
2390         return 0;
2391 }
2392
2393 void lockdep_trace_alloc(gfp_t gfp_mask)
2394 {
2395 }
2396
2397 #endif
2398
2399 /*
2400  * Mark a lock with a usage bit, and validate the state transition:
2401  */
2402 static int mark_lock(struct task_struct *curr, struct held_lock *this,
2403                              enum lock_usage_bit new_bit)
2404 {
2405         unsigned int new_mask = 1 << new_bit, ret = 1;
2406
2407         /*
2408          * If already set then do not dirty the cacheline,
2409          * nor do any checks:
2410          */
2411         if (likely(hlock_class(this)->usage_mask & new_mask))
2412                 return 1;
2413
2414         if (!graph_lock())
2415                 return 0;
2416         /*
2417          * Make sure we didnt race:
2418          */
2419         if (unlikely(hlock_class(this)->usage_mask & new_mask)) {
2420                 graph_unlock();
2421                 return 1;
2422         }
2423
2424         hlock_class(this)->usage_mask |= new_mask;
2425
2426         if (!save_trace(hlock_class(this)->usage_traces + new_bit))
2427                 return 0;
2428
2429         switch (new_bit) {
2430 #define LOCKDEP_STATE(__STATE)                  \
2431         case LOCK_USED_IN_##__STATE:            \
2432         case LOCK_USED_IN_##__STATE##_READ:     \
2433         case LOCK_ENABLED_##__STATE:            \
2434         case LOCK_ENABLED_##__STATE##_READ:
2435 #include "lockdep_states.h"
2436 #undef LOCKDEP_STATE
2437                 ret = mark_lock_irq(curr, this, new_bit);
2438                 if (!ret)
2439                         return 0;
2440                 break;
2441         case LOCK_USED:
2442                 debug_atomic_dec(&nr_unused_locks);
2443                 break;
2444         default:
2445                 if (!debug_locks_off_graph_unlock())
2446                         return 0;
2447                 WARN_ON(1);
2448                 return 0;
2449         }
2450
2451         graph_unlock();
2452
2453         /*
2454          * We must printk outside of the graph_lock:
2455          */
2456         if (ret == 2) {
2457                 printk("\nmarked lock as {%s}:\n", usage_str[new_bit]);
2458                 print_lock(this);
2459                 print_irqtrace_events(curr);
2460                 dump_stack();
2461         }
2462
2463         return ret;
2464 }
2465
2466 /*
2467  * Initialize a lock instance's lock-class mapping info:
2468  */
2469 void lockdep_init_map(struct lockdep_map *lock, const char *name,
2470                       struct lock_class_key *key, int subclass)
2471 {
2472         if (unlikely(!debug_locks))
2473                 return;
2474
2475         if (DEBUG_LOCKS_WARN_ON(!key))
2476                 return;
2477         if (DEBUG_LOCKS_WARN_ON(!name))
2478                 return;
2479         /*
2480          * Sanity check, the lock-class key must be persistent:
2481          */
2482         if (!static_obj(key)) {
2483                 printk("BUG: key %p not in .data!\n", key);
2484                 DEBUG_LOCKS_WARN_ON(1);
2485                 return;
2486         }
2487         lock->name = name;
2488         lock->key = key;
2489         lock->class_cache = NULL;
2490 #ifdef CONFIG_LOCK_STAT
2491         lock->cpu = raw_smp_processor_id();
2492 #endif
2493         if (subclass)
2494                 register_lock_class(lock, subclass, 1);
2495 }
2496 EXPORT_SYMBOL_GPL(lockdep_init_map);
2497
2498 /*
2499  * This gets called for every mutex_lock*()/spin_lock*() operation.
2500  * We maintain the dependency maps and validate the locking attempt:
2501  */
2502 static int __lock_acquire(struct lockdep_map *lock, unsigned int subclass,
2503                           int trylock, int read, int check, int hardirqs_off,
2504                           struct lockdep_map *nest_lock, unsigned long ip)
2505 {
2506         struct task_struct *curr = current;
2507         struct lock_class *class = NULL;
2508         struct held_lock *hlock;
2509         unsigned int depth, id;
2510         int chain_head = 0;
2511         u64 chain_key;
2512
2513         if (!prove_locking)
2514                 check = 1;
2515
2516         if (unlikely(!debug_locks))
2517                 return 0;
2518
2519         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2520                 return 0;
2521
2522         if (unlikely(subclass >= MAX_LOCKDEP_SUBCLASSES)) {
2523                 debug_locks_off();
2524                 printk("BUG: MAX_LOCKDEP_SUBCLASSES too low!\n");
2525                 printk("turning off the locking correctness validator.\n");
2526                 return 0;
2527         }
2528
2529         if (!subclass)
2530                 class = lock->class_cache;
2531         /*
2532          * Not cached yet or subclass?
2533          */
2534         if (unlikely(!class)) {
2535                 class = register_lock_class(lock, subclass, 0);
2536                 if (!class)
2537                         return 0;
2538         }
2539         debug_atomic_inc((atomic_t *)&class->ops);
2540         if (very_verbose(class)) {
2541                 printk("\nacquire class [%p] %s", class->key, class->name);
2542                 if (class->name_version > 1)
2543                         printk("#%d", class->name_version);
2544                 printk("\n");
2545                 dump_stack();
2546         }
2547
2548         /*
2549          * Add the lock to the list of currently held locks.
2550          * (we dont increase the depth just yet, up until the
2551          * dependency checks are done)
2552          */
2553         depth = curr->lockdep_depth;
2554         if (DEBUG_LOCKS_WARN_ON(depth >= MAX_LOCK_DEPTH))
2555                 return 0;
2556
2557         hlock = curr->held_locks + depth;
2558         if (DEBUG_LOCKS_WARN_ON(!class))
2559                 return 0;
2560         hlock->class_idx = class - lock_classes + 1;
2561         hlock->acquire_ip = ip;
2562         hlock->instance = lock;
2563         hlock->nest_lock = nest_lock;
2564         hlock->trylock = trylock;
2565         hlock->read = read;
2566         hlock->check = check;
2567         hlock->hardirqs_off = !!hardirqs_off;
2568 #ifdef CONFIG_LOCK_STAT
2569         hlock->waittime_stamp = 0;
2570         hlock->holdtime_stamp = sched_clock();
2571 #endif
2572
2573         if (check == 2 && !mark_irqflags(curr, hlock))
2574                 return 0;
2575
2576         /* mark it as used: */
2577         if (!mark_lock(curr, hlock, LOCK_USED))
2578                 return 0;
2579
2580         /*
2581          * Calculate the chain hash: it's the combined hash of all the
2582          * lock keys along the dependency chain. We save the hash value
2583          * at every step so that we can get the current hash easily
2584          * after unlock. The chain hash is then used to cache dependency
2585          * results.
2586          *
2587          * The 'key ID' is what is the most compact key value to drive
2588          * the hash, not class->key.
2589          */
2590         id = class - lock_classes;
2591         if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
2592                 return 0;
2593
2594         chain_key = curr->curr_chain_key;
2595         if (!depth) {
2596                 if (DEBUG_LOCKS_WARN_ON(chain_key != 0))
2597                         return 0;
2598                 chain_head = 1;
2599         }
2600
2601         hlock->prev_chain_key = chain_key;
2602         if (separate_irq_context(curr, hlock)) {
2603                 chain_key = 0;
2604                 chain_head = 1;
2605         }
2606         chain_key = iterate_chain_key(chain_key, id);
2607
2608         if (!validate_chain(curr, lock, hlock, chain_head, chain_key))
2609                 return 0;
2610
2611         curr->curr_chain_key = chain_key;
2612         curr->lockdep_depth++;
2613         check_chain_key(curr);
2614 #ifdef CONFIG_DEBUG_LOCKDEP
2615         if (unlikely(!debug_locks))
2616                 return 0;
2617 #endif
2618         if (unlikely(curr->lockdep_depth >= MAX_LOCK_DEPTH)) {
2619                 debug_locks_off();
2620                 printk("BUG: MAX_LOCK_DEPTH too low!\n");
2621                 printk("turning off the locking correctness validator.\n");
2622                 return 0;
2623         }
2624
2625         if (unlikely(curr->lockdep_depth > max_lockdep_depth))
2626                 max_lockdep_depth = curr->lockdep_depth;
2627
2628         return 1;
2629 }
2630
2631 static int
2632 print_unlock_inbalance_bug(struct task_struct *curr, struct lockdep_map *lock,
2633                            unsigned long ip)
2634 {
2635         if (!debug_locks_off())
2636                 return 0;
2637         if (debug_locks_silent)
2638                 return 0;
2639
2640         printk("\n=====================================\n");
2641         printk(  "[ BUG: bad unlock balance detected! ]\n");
2642         printk(  "-------------------------------------\n");
2643         printk("%s/%d is trying to release lock (",
2644                 curr->comm, task_pid_nr(curr));
2645         print_lockdep_cache(lock);
2646         printk(") at:\n");
2647         print_ip_sym(ip);
2648         printk("but there are no more locks to release!\n");
2649         printk("\nother info that might help us debug this:\n");
2650         lockdep_print_held_locks(curr);
2651
2652         printk("\nstack backtrace:\n");
2653         dump_stack();
2654
2655         return 0;
2656 }
2657
2658 /*
2659  * Common debugging checks for both nested and non-nested unlock:
2660  */
2661 static int check_unlock(struct task_struct *curr, struct lockdep_map *lock,
2662                         unsigned long ip)
2663 {
2664         if (unlikely(!debug_locks))
2665                 return 0;
2666         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2667                 return 0;
2668
2669         if (curr->lockdep_depth <= 0)
2670                 return print_unlock_inbalance_bug(curr, lock, ip);
2671
2672         return 1;
2673 }
2674
2675 static int
2676 __lock_set_class(struct lockdep_map *lock, const char *name,
2677                  struct lock_class_key *key, unsigned int subclass,
2678                  unsigned long ip)
2679 {
2680         struct task_struct *curr = current;
2681         struct held_lock *hlock, *prev_hlock;
2682         struct lock_class *class;
2683         unsigned int depth;
2684         int i;
2685
2686         depth = curr->lockdep_depth;
2687         if (DEBUG_LOCKS_WARN_ON(!depth))
2688                 return 0;
2689
2690         prev_hlock = NULL;
2691         for (i = depth-1; i >= 0; i--) {
2692                 hlock = curr->held_locks + i;
2693                 /*
2694                  * We must not cross into another context:
2695                  */
2696                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
2697                         break;
2698                 if (hlock->instance == lock)
2699                         goto found_it;
2700                 prev_hlock = hlock;
2701         }
2702         return print_unlock_inbalance_bug(curr, lock, ip);
2703
2704 found_it:
2705         lockdep_init_map(lock, name, key, 0);
2706         class = register_lock_class(lock, subclass, 0);
2707         hlock->class_idx = class - lock_classes + 1;
2708
2709         curr->lockdep_depth = i;
2710         curr->curr_chain_key = hlock->prev_chain_key;
2711
2712         for (; i < depth; i++) {
2713                 hlock = curr->held_locks + i;
2714                 if (!__lock_acquire(hlock->instance,
2715                         hlock_class(hlock)->subclass, hlock->trylock,
2716                                 hlock->read, hlock->check, hlock->hardirqs_off,
2717                                 hlock->nest_lock, hlock->acquire_ip))
2718                         return 0;
2719         }
2720
2721         if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth))
2722                 return 0;
2723         return 1;
2724 }
2725
2726 /*
2727  * Remove the lock to the list of currently held locks in a
2728  * potentially non-nested (out of order) manner. This is a
2729  * relatively rare operation, as all the unlock APIs default
2730  * to nested mode (which uses lock_release()):
2731  */
2732 static int
2733 lock_release_non_nested(struct task_struct *curr,
2734                         struct lockdep_map *lock, unsigned long ip)
2735 {
2736         struct held_lock *hlock, *prev_hlock;
2737         unsigned int depth;
2738         int i;
2739
2740         /*
2741          * Check whether the lock exists in the current stack
2742          * of held locks:
2743          */
2744         depth = curr->lockdep_depth;
2745         if (DEBUG_LOCKS_WARN_ON(!depth))
2746                 return 0;
2747
2748         prev_hlock = NULL;
2749         for (i = depth-1; i >= 0; i--) {
2750                 hlock = curr->held_locks + i;
2751                 /*
2752                  * We must not cross into another context:
2753                  */
2754                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
2755                         break;
2756                 if (hlock->instance == lock)
2757                         goto found_it;
2758                 prev_hlock = hlock;
2759         }
2760         return print_unlock_inbalance_bug(curr, lock, ip);
2761
2762 found_it:
2763         lock_release_holdtime(hlock);
2764
2765         /*
2766          * We have the right lock to unlock, 'hlock' points to it.
2767          * Now we remove it from the stack, and add back the other
2768          * entries (if any), recalculating the hash along the way:
2769          */
2770         curr->lockdep_depth = i;
2771         curr->curr_chain_key = hlock->prev_chain_key;
2772
2773         for (i++; i < depth; i++) {
2774                 hlock = curr->held_locks + i;
2775                 if (!__lock_acquire(hlock->instance,
2776                         hlock_class(hlock)->subclass, hlock->trylock,
2777                                 hlock->read, hlock->check, hlock->hardirqs_off,
2778                                 hlock->nest_lock, hlock->acquire_ip))
2779                         return 0;
2780         }
2781
2782         if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth - 1))
2783                 return 0;
2784         return 1;
2785 }
2786
2787 /*
2788  * Remove the lock to the list of currently held locks - this gets
2789  * called on mutex_unlock()/spin_unlock*() (or on a failed
2790  * mutex_lock_interruptible()). This is done for unlocks that nest
2791  * perfectly. (i.e. the current top of the lock-stack is unlocked)
2792  */
2793 static int lock_release_nested(struct task_struct *curr,
2794                                struct lockdep_map *lock, unsigned long ip)
2795 {
2796         struct held_lock *hlock;
2797         unsigned int depth;
2798
2799         /*
2800          * Pop off the top of the lock stack:
2801          */
2802         depth = curr->lockdep_depth - 1;
2803         hlock = curr->held_locks + depth;
2804
2805         /*
2806          * Is the unlock non-nested:
2807          */
2808         if (hlock->instance != lock)
2809                 return lock_release_non_nested(curr, lock, ip);
2810         curr->lockdep_depth--;
2811
2812         if (DEBUG_LOCKS_WARN_ON(!depth && (hlock->prev_chain_key != 0)))
2813                 return 0;
2814
2815         curr->curr_chain_key = hlock->prev_chain_key;
2816
2817         lock_release_holdtime(hlock);
2818
2819 #ifdef CONFIG_DEBUG_LOCKDEP
2820         hlock->prev_chain_key = 0;
2821         hlock->class_idx = 0;
2822         hlock->acquire_ip = 0;
2823         hlock->irq_context = 0;
2824 #endif
2825         return 1;
2826 }
2827
2828 /*
2829  * Remove the lock to the list of currently held locks - this gets
2830  * called on mutex_unlock()/spin_unlock*() (or on a failed
2831  * mutex_lock_interruptible()). This is done for unlocks that nest
2832  * perfectly. (i.e. the current top of the lock-stack is unlocked)
2833  */
2834 static void
2835 __lock_release(struct lockdep_map *lock, int nested, unsigned long ip)
2836 {
2837         struct task_struct *curr = current;
2838
2839         if (!check_unlock(curr, lock, ip))
2840                 return;
2841
2842         if (nested) {
2843                 if (!lock_release_nested(curr, lock, ip))
2844                         return;
2845         } else {
2846                 if (!lock_release_non_nested(curr, lock, ip))
2847                         return;
2848         }
2849
2850         check_chain_key(curr);
2851 }
2852
2853 /*
2854  * Check whether we follow the irq-flags state precisely:
2855  */
2856 static void check_flags(unsigned long flags)
2857 {
2858 #if defined(CONFIG_PROVE_LOCKING) && defined(CONFIG_DEBUG_LOCKDEP) && \
2859     defined(CONFIG_TRACE_IRQFLAGS)
2860         if (!debug_locks)
2861                 return;
2862
2863         if (irqs_disabled_flags(flags)) {
2864                 if (DEBUG_LOCKS_WARN_ON(current->hardirqs_enabled)) {
2865                         printk("possible reason: unannotated irqs-off.\n");
2866                 }
2867         } else {
2868                 if (DEBUG_LOCKS_WARN_ON(!current->hardirqs_enabled)) {
2869                         printk("possible reason: unannotated irqs-on.\n");
2870                 }
2871         }
2872
2873         /*
2874          * We dont accurately track softirq state in e.g.
2875          * hardirq contexts (such as on 4KSTACKS), so only
2876          * check if not in hardirq contexts:
2877          */
2878         if (!hardirq_count()) {
2879                 if (softirq_count())
2880                         DEBUG_LOCKS_WARN_ON(current->softirqs_enabled);
2881                 else
2882                         DEBUG_LOCKS_WARN_ON(!current->softirqs_enabled);
2883         }
2884
2885         if (!debug_locks)
2886                 print_irqtrace_events(current);
2887 #endif
2888 }
2889
2890 void lock_set_class(struct lockdep_map *lock, const char *name,
2891                     struct lock_class_key *key, unsigned int subclass,
2892                     unsigned long ip)
2893 {
2894         unsigned long flags;
2895
2896         if (unlikely(current->lockdep_recursion))
2897                 return;
2898
2899         raw_local_irq_save(flags);
2900         current->lockdep_recursion = 1;
2901         check_flags(flags);
2902         if (__lock_set_class(lock, name, key, subclass, ip))
2903                 check_chain_key(current);
2904         current->lockdep_recursion = 0;
2905         raw_local_irq_restore(flags);
2906 }
2907 EXPORT_SYMBOL_GPL(lock_set_class);
2908
2909 /*
2910  * We are not always called with irqs disabled - do that here,
2911  * and also avoid lockdep recursion:
2912  */
2913 void lock_acquire(struct lockdep_map *lock, unsigned int subclass,
2914                           int trylock, int read, int check,
2915                           struct lockdep_map *nest_lock, unsigned long ip)
2916 {
2917         unsigned long flags;
2918
2919         if (unlikely(current->lockdep_recursion))
2920                 return;
2921
2922         raw_local_irq_save(flags);
2923         check_flags(flags);
2924
2925         current->lockdep_recursion = 1;
2926         __lock_acquire(lock, subclass, trylock, read, check,
2927                        irqs_disabled_flags(flags), nest_lock, ip);
2928         current->lockdep_recursion = 0;
2929         raw_local_irq_restore(flags);
2930 }
2931 EXPORT_SYMBOL_GPL(lock_acquire);
2932
2933 void lock_release(struct lockdep_map *lock, int nested,
2934                           unsigned long ip)
2935 {
2936         unsigned long flags;
2937
2938         if (unlikely(current->lockdep_recursion))
2939                 return;
2940
2941         raw_local_irq_save(flags);
2942         check_flags(flags);
2943         current->lockdep_recursion = 1;
2944         __lock_release(lock, nested, ip);
2945         current->lockdep_recursion = 0;
2946         raw_local_irq_restore(flags);
2947 }
2948 EXPORT_SYMBOL_GPL(lock_release);
2949
2950 void lockdep_set_current_reclaim_state(gfp_t gfp_mask)
2951 {
2952         current->lockdep_reclaim_gfp = gfp_mask;
2953 }
2954
2955 void lockdep_clear_current_reclaim_state(void)
2956 {
2957         current->lockdep_reclaim_gfp = 0;
2958 }
2959
2960 #ifdef CONFIG_LOCK_STAT
2961 static int
2962 print_lock_contention_bug(struct task_struct *curr, struct lockdep_map *lock,
2963                            unsigned long ip)
2964 {
2965         if (!debug_locks_off())
2966                 return 0;
2967         if (debug_locks_silent)
2968                 return 0;
2969
2970         printk("\n=================================\n");
2971         printk(  "[ BUG: bad contention detected! ]\n");
2972         printk(  "---------------------------------\n");
2973         printk("%s/%d is trying to contend lock (",
2974                 curr->comm, task_pid_nr(curr));
2975         print_lockdep_cache(lock);
2976         printk(") at:\n");
2977         print_ip_sym(ip);
2978         printk("but there are no locks held!\n");
2979         printk("\nother info that might help us debug this:\n");
2980         lockdep_print_held_locks(curr);
2981
2982         printk("\nstack backtrace:\n");
2983         dump_stack();
2984
2985         return 0;
2986 }
2987
2988 static void
2989 __lock_contended(struct lockdep_map *lock, unsigned long ip)
2990 {
2991         struct task_struct *curr = current;
2992         struct held_lock *hlock, *prev_hlock;
2993         struct lock_class_stats *stats;
2994         unsigned int depth;
2995         int i, contention_point, contending_point;
2996
2997         depth = curr->lockdep_depth;
2998         if (DEBUG_LOCKS_WARN_ON(!depth))
2999                 return;
3000
3001         prev_hlock = NULL;
3002         for (i = depth-1; i >= 0; i--) {
3003                 hlock = curr->held_locks + i;
3004                 /*
3005                  * We must not cross into another context:
3006                  */
3007                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3008                         break;
3009                 if (hlock->instance == lock)
3010                         goto found_it;
3011                 prev_hlock = hlock;
3012         }
3013         print_lock_contention_bug(curr, lock, ip);
3014         return;
3015
3016 found_it:
3017         hlock->waittime_stamp = sched_clock();
3018
3019         contention_point = lock_point(hlock_class(hlock)->contention_point, ip);
3020         contending_point = lock_point(hlock_class(hlock)->contending_point,
3021                                       lock->ip);
3022
3023         stats = get_lock_stats(hlock_class(hlock));
3024         if (contention_point < LOCKSTAT_POINTS)
3025                 stats->contention_point[contention_point]++;
3026         if (contending_point < LOCKSTAT_POINTS)
3027                 stats->contending_point[contending_point]++;
3028         if (lock->cpu != smp_processor_id())
3029                 stats->bounces[bounce_contended + !!hlock->read]++;
3030         put_lock_stats(stats);
3031 }
3032
3033 static void
3034 __lock_acquired(struct lockdep_map *lock, unsigned long ip)
3035 {
3036         struct task_struct *curr = current;
3037         struct held_lock *hlock, *prev_hlock;
3038         struct lock_class_stats *stats;
3039         unsigned int depth;
3040         u64 now;
3041         s64 waittime = 0;
3042         int i, cpu;
3043
3044         depth = curr->lockdep_depth;
3045         if (DEBUG_LOCKS_WARN_ON(!depth))
3046                 return;
3047
3048         prev_hlock = NULL;
3049         for (i = depth-1; i >= 0; i--) {
3050                 hlock = curr->held_locks + i;
3051                 /*
3052                  * We must not cross into another context:
3053                  */
3054                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3055                         break;
3056                 if (hlock->instance == lock)
3057                         goto found_it;
3058                 prev_hlock = hlock;
3059         }
3060         print_lock_contention_bug(curr, lock, _RET_IP_);
3061         return;
3062
3063 found_it:
3064         cpu = smp_processor_id();
3065         if (hlock->waittime_stamp) {
3066                 now = sched_clock();
3067                 waittime = now - hlock->waittime_stamp;
3068                 hlock->holdtime_stamp = now;
3069         }
3070
3071         stats = get_lock_stats(hlock_class(hlock));
3072         if (waittime) {
3073                 if (hlock->read)
3074                         lock_time_inc(&stats->read_waittime, waittime);
3075                 else
3076                         lock_time_inc(&stats->write_waittime, waittime);
3077         }
3078         if (lock->cpu != cpu)
3079                 stats->bounces[bounce_acquired + !!hlock->read]++;
3080         put_lock_stats(stats);
3081
3082         lock->cpu = cpu;
3083         lock->ip = ip;
3084 }
3085
3086 void lock_contended(struct lockdep_map *lock, unsigned long ip)
3087 {
3088         unsigned long flags;
3089
3090         if (unlikely(!lock_stat))
3091                 return;
3092
3093         if (unlikely(current->lockdep_recursion))
3094                 return;
3095
3096         raw_local_irq_save(flags);
3097         check_flags(flags);
3098         current->lockdep_recursion = 1;
3099         __lock_contended(lock, ip);
3100         current->lockdep_recursion = 0;
3101         raw_local_irq_restore(flags);
3102 }
3103 EXPORT_SYMBOL_GPL(lock_contended);
3104
3105 void lock_acquired(struct lockdep_map *lock, unsigned long ip)
3106 {
3107         unsigned long flags;
3108
3109         if (unlikely(!lock_stat))
3110                 return;
3111
3112         if (unlikely(current->lockdep_recursion))
3113                 return;
3114
3115         raw_local_irq_save(flags);
3116         check_flags(flags);
3117         current->lockdep_recursion = 1;
3118         __lock_acquired(lock, ip);
3119         current->lockdep_recursion = 0;
3120         raw_local_irq_restore(flags);
3121 }
3122 EXPORT_SYMBOL_GPL(lock_acquired);
3123 #endif
3124
3125 /*
3126  * Used by the testsuite, sanitize the validator state
3127  * after a simulated failure:
3128  */
3129
3130 void lockdep_reset(void)
3131 {
3132         unsigned long flags;
3133         int i;
3134
3135         raw_local_irq_save(flags);
3136         current->curr_chain_key = 0;
3137         current->lockdep_depth = 0;
3138         current->lockdep_recursion = 0;
3139         memset(current->held_locks, 0, MAX_LOCK_DEPTH*sizeof(struct held_lock));
3140         nr_hardirq_chains = 0;
3141         nr_softirq_chains = 0;
3142         nr_process_chains = 0;
3143         debug_locks = 1;
3144         for (i = 0; i < CHAINHASH_SIZE; i++)
3145                 INIT_LIST_HEAD(chainhash_table + i);
3146         raw_local_irq_restore(flags);
3147 }
3148
3149 static void zap_class(struct lock_class *class)
3150 {
3151         int i;
3152
3153         /*
3154          * Remove all dependencies this lock is
3155          * involved in:
3156          */
3157         for (i = 0; i < nr_list_entries; i++) {
3158                 if (list_entries[i].class == class)
3159                         list_del_rcu(&list_entries[i].entry);
3160         }
3161         /*
3162          * Unhash the class and remove it from the all_lock_classes list:
3163          */
3164         list_del_rcu(&class->hash_entry);
3165         list_del_rcu(&class->lock_entry);
3166
3167         class->key = NULL;
3168 }
3169
3170 static inline int within(const void *addr, void *start, unsigned long size)
3171 {
3172         return addr >= start && addr < start + size;
3173 }
3174
3175 void lockdep_free_key_range(void *start, unsigned long size)
3176 {
3177         struct lock_class *class, *next;
3178         struct list_head *head;
3179         unsigned long flags;
3180         int i;
3181         int locked;
3182
3183         raw_local_irq_save(flags);
3184         locked = graph_lock();
3185
3186         /*
3187          * Unhash all classes that were created by this module:
3188          */
3189         for (i = 0; i < CLASSHASH_SIZE; i++) {
3190                 head = classhash_table + i;
3191                 if (list_empty(head))
3192                         continue;
3193                 list_for_each_entry_safe(class, next, head, hash_entry) {
3194                         if (within(class->key, start, size))
3195                                 zap_class(class);
3196                         else if (within(class->name, start, size))
3197                                 zap_class(class);
3198                 }
3199         }
3200
3201         if (locked)
3202                 graph_unlock();
3203         raw_local_irq_restore(flags);
3204 }
3205
3206 void lockdep_reset_lock(struct lockdep_map *lock)
3207 {
3208         struct lock_class *class, *next;
3209         struct list_head *head;
3210         unsigned long flags;
3211         int i, j;
3212         int locked;
3213
3214         raw_local_irq_save(flags);
3215
3216         /*
3217          * Remove all classes this lock might have:
3218          */
3219         for (j = 0; j < MAX_LOCKDEP_SUBCLASSES; j++) {
3220                 /*
3221                  * If the class exists we look it up and zap it:
3222                  */
3223                 class = look_up_lock_class(lock, j);
3224                 if (class)
3225                         zap_class(class);
3226         }
3227         /*
3228          * Debug check: in the end all mapped classes should
3229          * be gone.
3230          */
3231         locked = graph_lock();
3232         for (i = 0; i < CLASSHASH_SIZE; i++) {
3233                 head = classhash_table + i;
3234                 if (list_empty(head))
3235                         continue;
3236                 list_for_each_entry_safe(class, next, head, hash_entry) {
3237                         if (unlikely(class == lock->class_cache)) {
3238                                 if (debug_locks_off_graph_unlock())
3239                                         WARN_ON(1);
3240                                 goto out_restore;
3241                         }
3242                 }
3243         }
3244         if (locked)
3245                 graph_unlock();
3246
3247 out_restore:
3248         raw_local_irq_restore(flags);
3249 }
3250
3251 void lockdep_init(void)
3252 {
3253         int i;
3254
3255         /*
3256          * Some architectures have their own start_kernel()
3257          * code which calls lockdep_init(), while we also
3258          * call lockdep_init() from the start_kernel() itself,
3259          * and we want to initialize the hashes only once:
3260          */
3261         if (lockdep_initialized)
3262                 return;
3263
3264         for (i = 0; i < CLASSHASH_SIZE; i++)
3265                 INIT_LIST_HEAD(classhash_table + i);
3266
3267         for (i = 0; i < CHAINHASH_SIZE; i++)
3268                 INIT_LIST_HEAD(chainhash_table + i);
3269
3270         lockdep_initialized = 1;
3271 }
3272
3273 void __init lockdep_info(void)
3274 {
3275         printk("Lock dependency validator: Copyright (c) 2006 Red Hat, Inc., Ingo Molnar\n");
3276
3277         printk("... MAX_LOCKDEP_SUBCLASSES:  %lu\n", MAX_LOCKDEP_SUBCLASSES);
3278         printk("... MAX_LOCK_DEPTH:          %lu\n", MAX_LOCK_DEPTH);
3279         printk("... MAX_LOCKDEP_KEYS:        %lu\n", MAX_LOCKDEP_KEYS);
3280         printk("... CLASSHASH_SIZE:          %lu\n", CLASSHASH_SIZE);
3281         printk("... MAX_LOCKDEP_ENTRIES:     %lu\n", MAX_LOCKDEP_ENTRIES);
3282         printk("... MAX_LOCKDEP_CHAINS:      %lu\n", MAX_LOCKDEP_CHAINS);
3283         printk("... CHAINHASH_SIZE:          %lu\n", CHAINHASH_SIZE);
3284
3285         printk(" memory used by lock dependency info: %lu kB\n",
3286                 (sizeof(struct lock_class) * MAX_LOCKDEP_KEYS +
3287                 sizeof(struct list_head) * CLASSHASH_SIZE +
3288                 sizeof(struct lock_list) * MAX_LOCKDEP_ENTRIES +
3289                 sizeof(struct lock_chain) * MAX_LOCKDEP_CHAINS +
3290                 sizeof(struct list_head) * CHAINHASH_SIZE) / 1024);
3291
3292         printk(" per task-struct memory footprint: %lu bytes\n",
3293                 sizeof(struct held_lock) * MAX_LOCK_DEPTH);
3294
3295 #ifdef CONFIG_DEBUG_LOCKDEP
3296         if (lockdep_init_error) {
3297                 printk("WARNING: lockdep init error! Arch code didn't call lockdep_init() early enough?\n");
3298                 printk("Call stack leading to lockdep invocation was:\n");
3299                 print_stack_trace(&lockdep_init_trace, 0);
3300         }
3301 #endif
3302 }
3303
3304 static void
3305 print_freed_lock_bug(struct task_struct *curr, const void *mem_from,
3306                      const void *mem_to, struct held_lock *hlock)
3307 {
3308         if (!debug_locks_off())
3309                 return;
3310         if (debug_locks_silent)
3311                 return;
3312
3313         printk("\n=========================\n");
3314         printk(  "[ BUG: held lock freed! ]\n");
3315         printk(  "-------------------------\n");
3316         printk("%s/%d is freeing memory %p-%p, with a lock still held there!\n",
3317                 curr->comm, task_pid_nr(curr), mem_from, mem_to-1);
3318         print_lock(hlock);
3319         lockdep_print_held_locks(curr);
3320
3321         printk("\nstack backtrace:\n");
3322         dump_stack();
3323 }
3324
3325 static inline int not_in_range(const void* mem_from, unsigned long mem_len,
3326                                 const void* lock_from, unsigned long lock_len)
3327 {
3328         return lock_from + lock_len <= mem_from ||
3329                 mem_from + mem_len <= lock_from;
3330 }
3331
3332 /*
3333  * Called when kernel memory is freed (or unmapped), or if a lock
3334  * is destroyed or reinitialized - this code checks whether there is
3335  * any held lock in the memory range of <from> to <to>:
3336  */
3337 void debug_check_no_locks_freed(const void *mem_from, unsigned long mem_len)
3338 {
3339         struct task_struct *curr = current;
3340         struct held_lock *hlock;
3341         unsigned long flags;
3342         int i;
3343
3344         if (unlikely(!debug_locks))
3345                 return;
3346
3347         local_irq_save(flags);
3348         for (i = 0; i < curr->lockdep_depth; i++) {
3349                 hlock = curr->held_locks + i;
3350
3351                 if (not_in_range(mem_from, mem_len, hlock->instance,
3352                                         sizeof(*hlock->instance)))
3353                         continue;
3354
3355                 print_freed_lock_bug(curr, mem_from, mem_from + mem_len, hlock);
3356                 break;
3357         }
3358         local_irq_restore(flags);
3359 }
3360 EXPORT_SYMBOL_GPL(debug_check_no_locks_freed);
3361
3362 static void print_held_locks_bug(struct task_struct *curr)
3363 {
3364         if (!debug_locks_off())
3365                 return;
3366         if (debug_locks_silent)
3367                 return;
3368
3369         printk("\n=====================================\n");
3370         printk(  "[ BUG: lock held at task exit time! ]\n");
3371         printk(  "-------------------------------------\n");
3372         printk("%s/%d is exiting with locks still held!\n",
3373                 curr->comm, task_pid_nr(curr));
3374         lockdep_print_held_locks(curr);
3375
3376         printk("\nstack backtrace:\n");
3377         dump_stack();
3378 }
3379
3380 void debug_check_no_locks_held(struct task_struct *task)
3381 {
3382         if (unlikely(task->lockdep_depth > 0))
3383                 print_held_locks_bug(task);
3384 }
3385
3386 void debug_show_all_locks(void)
3387 {
3388         struct task_struct *g, *p;
3389         int count = 10;
3390         int unlock = 1;
3391
3392         if (unlikely(!debug_locks)) {
3393                 printk("INFO: lockdep is turned off.\n");
3394                 return;
3395         }
3396         printk("\nShowing all locks held in the system:\n");
3397
3398         /*
3399          * Here we try to get the tasklist_lock as hard as possible,
3400          * if not successful after 2 seconds we ignore it (but keep
3401          * trying). This is to enable a debug printout even if a
3402          * tasklist_lock-holding task deadlocks or crashes.
3403          */
3404 retry:
3405         if (!read_trylock(&tasklist_lock)) {
3406                 if (count == 10)
3407                         printk("hm, tasklist_lock locked, retrying... ");
3408                 if (count) {
3409                         count--;
3410                         printk(" #%d", 10-count);
3411                         mdelay(200);
3412                         goto retry;
3413                 }
3414                 printk(" ignoring it.\n");
3415                 unlock = 0;
3416         } else {
3417                 if (count != 10)
3418                         printk(KERN_CONT " locked it.\n");
3419         }
3420
3421         do_each_thread(g, p) {
3422                 /*
3423                  * It's not reliable to print a task's held locks
3424                  * if it's not sleeping (or if it's not the current
3425                  * task):
3426                  */
3427                 if (p->state == TASK_RUNNING && p != current)
3428                         continue;
3429                 if (p->lockdep_depth)
3430                         lockdep_print_held_locks(p);
3431                 if (!unlock)
3432                         if (read_trylock(&tasklist_lock))
3433                                 unlock = 1;
3434         } while_each_thread(g, p);
3435
3436         printk("\n");
3437         printk("=============================================\n\n");
3438
3439         if (unlock)
3440                 read_unlock(&tasklist_lock);
3441 }
3442 EXPORT_SYMBOL_GPL(debug_show_all_locks);
3443
3444 /*
3445  * Careful: only use this function if you are sure that
3446  * the task cannot run in parallel!
3447  */
3448 void __debug_show_held_locks(struct task_struct *task)
3449 {
3450         if (unlikely(!debug_locks)) {
3451                 printk("INFO: lockdep is turned off.\n");
3452                 return;
3453         }
3454         lockdep_print_held_locks(task);
3455 }
3456 EXPORT_SYMBOL_GPL(__debug_show_held_locks);
3457
3458 void debug_show_held_locks(struct task_struct *task)
3459 {
3460                 __debug_show_held_locks(task);