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