6358cf7e84b211a3d039e652afca73454247666f
[linux-2.6.git] / kernel / lockdep.c
1 /*
2  * kernel/lockdep.c
3  *
4  * Runtime locking correctness validator
5  *
6  * Started by Ingo Molnar:
7  *
8  *  Copyright (C) 2006,2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
9  *  Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
10  *
11  * this code maps all the lock dependencies as they occur in a live kernel
12  * and will warn about the following classes of locking bugs:
13  *
14  * - lock inversion scenarios
15  * - circular lock dependencies
16  * - hardirq/softirq safe/unsafe locking bugs
17  *
18  * Bugs are reported even if the current locking scenario does not cause
19  * any deadlock at this point.
20  *
21  * I.e. if anytime in the past two locks were taken in a different order,
22  * even if it happened for another task, even if those were different
23  * locks (but of the same class as this lock), this code will detect it.
24  *
25  * Thanks to Arjan van de Ven for coming up with the initial idea of
26  * mapping lock dependencies runtime.
27  */
28 #define DISABLE_BRANCH_PROFILING
29 #include <linux/mutex.h>
30 #include <linux/sched.h>
31 #include <linux/delay.h>
32 #include <linux/module.h>
33 #include <linux/proc_fs.h>
34 #include <linux/seq_file.h>
35 #include <linux/spinlock.h>
36 #include <linux/kallsyms.h>
37 #include <linux/interrupt.h>
38 #include <linux/stacktrace.h>
39 #include <linux/debug_locks.h>
40 #include <linux/irqflags.h>
41 #include <linux/utsname.h>
42 #include <linux/hash.h>
43 #include <linux/ftrace.h>
44 #include <linux/stringify.h>
45 #include <linux/bitops.h>
46 #include <asm/sections.h>
47
48 #include "lockdep_internals.h"
49
50 #define CREATE_TRACE_POINTS
51 #include <trace/events/lockdep.h>
52
53 #ifdef CONFIG_PROVE_LOCKING
54 int prove_locking = 1;
55 module_param(prove_locking, int, 0644);
56 #else
57 #define prove_locking 0
58 #endif
59
60 #ifdef CONFIG_LOCK_STAT
61 int lock_stat = 1;
62 module_param(lock_stat, int, 0644);
63 #else
64 #define lock_stat 0
65 #endif
66
67 /*
68  * lockdep_lock: protects the lockdep graph, the hashes and the
69  *               class/list/hash allocators.
70  *
71  * This is one of the rare exceptions where it's justified
72  * to use a raw spinlock - we really dont want the spinlock
73  * code to recurse back into the lockdep code...
74  */
75 static raw_spinlock_t lockdep_lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
76
77 static int graph_lock(void)
78 {
79         __raw_spin_lock(&lockdep_lock);
80         /*
81          * Make sure that if another CPU detected a bug while
82          * walking the graph we dont change it (while the other
83          * CPU is busy printing out stuff with the graph lock
84          * dropped already)
85          */
86         if (!debug_locks) {
87                 __raw_spin_unlock(&lockdep_lock);
88                 return 0;
89         }
90         /* prevent any recursions within lockdep from causing deadlocks */
91         current->lockdep_recursion++;
92         return 1;
93 }
94
95 static inline int graph_unlock(void)
96 {
97         if (debug_locks && !__raw_spin_is_locked(&lockdep_lock))
98                 return DEBUG_LOCKS_WARN_ON(1);
99
100         current->lockdep_recursion--;
101         __raw_spin_unlock(&lockdep_lock);
102         return 0;
103 }
104
105 /*
106  * Turn lock debugging off and return with 0 if it was off already,
107  * and also release the graph lock:
108  */
109 static inline int debug_locks_off_graph_unlock(void)
110 {
111         int ret = debug_locks_off();
112
113         __raw_spin_unlock(&lockdep_lock);
114
115         return ret;
116 }
117
118 static int lockdep_initialized;
119
120 unsigned long nr_list_entries;
121 struct lock_list list_entries[MAX_LOCKDEP_ENTRIES];
122
123 /*
124  * All data structures here are protected by the global debug_lock.
125  *
126  * Mutex key structs only get allocated, once during bootup, and never
127  * get freed - this significantly simplifies the debugging code.
128  */
129 unsigned long nr_lock_classes;
130 static struct lock_class lock_classes[MAX_LOCKDEP_KEYS];
131
132 static inline struct lock_class *hlock_class(struct held_lock *hlock)
133 {
134         if (!hlock->class_idx) {
135                 DEBUG_LOCKS_WARN_ON(1);
136                 return NULL;
137         }
138         return lock_classes + hlock->class_idx - 1;
139 }
140
141 #ifdef CONFIG_LOCK_STAT
142 static DEFINE_PER_CPU(struct lock_class_stats[MAX_LOCKDEP_KEYS], lock_stats);
143
144 static int lock_point(unsigned long points[], unsigned long ip)
145 {
146         int i;
147
148         for (i = 0; i < LOCKSTAT_POINTS; i++) {
149                 if (points[i] == 0) {
150                         points[i] = ip;
151                         break;
152                 }
153                 if (points[i] == ip)
154                         break;
155         }
156
157         return i;
158 }
159
160 static void lock_time_inc(struct lock_time *lt, s64 time)
161 {
162         if (time > lt->max)
163                 lt->max = time;
164
165         if (time < lt->min || !lt->min)
166                 lt->min = time;
167
168         lt->total += time;
169         lt->nr++;
170 }
171
172 static inline void lock_time_add(struct lock_time *src, struct lock_time *dst)
173 {
174         dst->min += src->min;
175         dst->max += src->max;
176         dst->total += src->total;
177         dst->nr += src->nr;
178 }
179
180 struct lock_class_stats lock_stats(struct lock_class *class)
181 {
182         struct lock_class_stats stats;
183         int cpu, i;
184
185         memset(&stats, 0, sizeof(struct lock_class_stats));
186         for_each_possible_cpu(cpu) {
187                 struct lock_class_stats *pcs =
188                         &per_cpu(lock_stats, cpu)[class - lock_classes];
189
190                 for (i = 0; i < ARRAY_SIZE(stats.contention_point); i++)
191                         stats.contention_point[i] += pcs->contention_point[i];
192
193                 for (i = 0; i < ARRAY_SIZE(stats.contending_point); i++)
194                         stats.contending_point[i] += pcs->contending_point[i];
195
196                 lock_time_add(&pcs->read_waittime, &stats.read_waittime);
197                 lock_time_add(&pcs->write_waittime, &stats.write_waittime);
198
199                 lock_time_add(&pcs->read_holdtime, &stats.read_holdtime);
200                 lock_time_add(&pcs->write_holdtime, &stats.write_holdtime);
201
202                 for (i = 0; i < ARRAY_SIZE(stats.bounces); i++)
203                         stats.bounces[i] += pcs->bounces[i];
204         }
205
206         return stats;
207 }
208
209 void clear_lock_stats(struct lock_class *class)
210 {
211         int cpu;
212
213         for_each_possible_cpu(cpu) {
214                 struct lock_class_stats *cpu_stats =
215                         &per_cpu(lock_stats, cpu)[class - lock_classes];
216
217                 memset(cpu_stats, 0, sizeof(struct lock_class_stats));
218         }
219         memset(class->contention_point, 0, sizeof(class->contention_point));
220         memset(class->contending_point, 0, sizeof(class->contending_point));
221 }
222
223 static struct lock_class_stats *get_lock_stats(struct lock_class *class)
224 {
225         return &get_cpu_var(lock_stats)[class - lock_classes];
226 }
227
228 static void put_lock_stats(struct lock_class_stats *stats)
229 {
230         put_cpu_var(lock_stats);
231 }
232
233 static void lock_release_holdtime(struct held_lock *hlock)
234 {
235         struct lock_class_stats *stats;
236         s64 holdtime;
237
238         if (!lock_stat)
239                 return;
240
241         holdtime = sched_clock() - hlock->holdtime_stamp;
242
243         stats = get_lock_stats(hlock_class(hlock));
244         if (hlock->read)
245                 lock_time_inc(&stats->read_holdtime, holdtime);
246         else
247                 lock_time_inc(&stats->write_holdtime, holdtime);
248         put_lock_stats(stats);
249 }
250 #else
251 static inline void lock_release_holdtime(struct held_lock *hlock)
252 {
253 }
254 #endif
255
256 /*
257  * We keep a global list of all lock classes. The list only grows,
258  * never shrinks. The list is only accessed with the lockdep
259  * spinlock lock held.
260  */
261 LIST_HEAD(all_lock_classes);
262
263 /*
264  * The lockdep classes are in a hash-table as well, for fast lookup:
265  */
266 #define CLASSHASH_BITS          (MAX_LOCKDEP_KEYS_BITS - 1)
267 #define CLASSHASH_SIZE          (1UL << CLASSHASH_BITS)
268 #define __classhashfn(key)      hash_long((unsigned long)key, CLASSHASH_BITS)
269 #define classhashentry(key)     (classhash_table + __classhashfn((key)))
270
271 static struct list_head classhash_table[CLASSHASH_SIZE];
272
273 /*
274  * We put the lock dependency chains into a hash-table as well, to cache
275  * their existence:
276  */
277 #define CHAINHASH_BITS          (MAX_LOCKDEP_CHAINS_BITS-1)
278 #define CHAINHASH_SIZE          (1UL << CHAINHASH_BITS)
279 #define __chainhashfn(chain)    hash_long(chain, CHAINHASH_BITS)
280 #define chainhashentry(chain)   (chainhash_table + __chainhashfn((chain)))
281
282 static struct list_head chainhash_table[CHAINHASH_SIZE];
283
284 /*
285  * The hash key of the lock dependency chains is a hash itself too:
286  * it's a hash of all locks taken up to that lock, including that lock.
287  * It's a 64-bit hash, because it's important for the keys to be
288  * unique.
289  */
290 #define iterate_chain_key(key1, key2) \
291         (((key1) << MAX_LOCKDEP_KEYS_BITS) ^ \
292         ((key1) >> (64-MAX_LOCKDEP_KEYS_BITS)) ^ \
293         (key2))
294
295 void lockdep_off(void)
296 {
297         current->lockdep_recursion++;
298 }
299 EXPORT_SYMBOL(lockdep_off);
300
301 void lockdep_on(void)
302 {
303         current->lockdep_recursion--;
304 }
305 EXPORT_SYMBOL(lockdep_on);
306
307 /*
308  * Debugging switches:
309  */
310
311 #define VERBOSE                 0
312 #define VERY_VERBOSE            0
313
314 #if VERBOSE
315 # define HARDIRQ_VERBOSE        1
316 # define SOFTIRQ_VERBOSE        1
317 # define RECLAIM_VERBOSE        1
318 #else
319 # define HARDIRQ_VERBOSE        0
320 # define SOFTIRQ_VERBOSE        0
321 # define RECLAIM_VERBOSE        0
322 #endif
323
324 #if VERBOSE || HARDIRQ_VERBOSE || SOFTIRQ_VERBOSE || RECLAIM_VERBOSE
325 /*
326  * Quick filtering for interesting events:
327  */
328 static int class_filter(struct lock_class *class)
329 {
330 #if 0
331         /* Example */
332         if (class->name_version == 1 &&
333                         !strcmp(class->name, "lockname"))
334                 return 1;
335         if (class->name_version == 1 &&
336                         !strcmp(class->name, "&struct->lockfield"))
337                 return 1;
338 #endif
339         /* Filter everything else. 1 would be to allow everything else */
340         return 0;
341 }
342 #endif
343
344 static int verbose(struct lock_class *class)
345 {
346 #if VERBOSE
347         return class_filter(class);
348 #endif
349         return 0;
350 }
351
352 /*
353  * Stack-trace: tightly packed array of stack backtrace
354  * addresses. Protected by the graph_lock.
355  */
356 unsigned long nr_stack_trace_entries;
357 static unsigned long stack_trace[MAX_STACK_TRACE_ENTRIES];
358
359 static int save_trace(struct stack_trace *trace)
360 {
361         trace->nr_entries = 0;
362         trace->max_entries = MAX_STACK_TRACE_ENTRIES - nr_stack_trace_entries;
363         trace->entries = stack_trace + nr_stack_trace_entries;
364
365         trace->skip = 3;
366
367         save_stack_trace(trace);
368
369         trace->max_entries = trace->nr_entries;
370
371         nr_stack_trace_entries += trace->nr_entries;
372
373         if (nr_stack_trace_entries == MAX_STACK_TRACE_ENTRIES) {
374                 if (!debug_locks_off_graph_unlock())
375                         return 0;
376
377                 printk("BUG: MAX_STACK_TRACE_ENTRIES too low!\n");
378                 printk("turning off the locking correctness validator.\n");
379                 dump_stack();
380
381                 return 0;
382         }
383
384         return 1;
385 }
386
387 unsigned int nr_hardirq_chains;
388 unsigned int nr_softirq_chains;
389 unsigned int nr_process_chains;
390 unsigned int max_lockdep_depth;
391 unsigned int max_recursion_depth;
392
393 static unsigned int lockdep_dependency_gen_id;
394
395 static bool lockdep_dependency_visit(struct lock_class *source,
396                                      unsigned int depth)
397 {
398         if (!depth)
399                 lockdep_dependency_gen_id++;
400         if (source->dep_gen_id == lockdep_dependency_gen_id)
401                 return true;
402         source->dep_gen_id = lockdep_dependency_gen_id;
403         return false;
404 }
405
406 #ifdef CONFIG_DEBUG_LOCKDEP
407 /*
408  * We cannot printk in early bootup code. Not even early_printk()
409  * might work. So we mark any initialization errors and printk
410  * about it later on, in lockdep_info().
411  */
412 static int lockdep_init_error;
413 static unsigned long lockdep_init_trace_data[20];
414 static struct stack_trace lockdep_init_trace = {
415         .max_entries = ARRAY_SIZE(lockdep_init_trace_data),
416         .entries = lockdep_init_trace_data,
417 };
418
419 /*
420  * Various lockdep statistics:
421  */
422 atomic_t chain_lookup_hits;
423 atomic_t chain_lookup_misses;
424 atomic_t hardirqs_on_events;
425 atomic_t hardirqs_off_events;
426 atomic_t redundant_hardirqs_on;
427 atomic_t redundant_hardirqs_off;
428 atomic_t softirqs_on_events;
429 atomic_t softirqs_off_events;
430 atomic_t redundant_softirqs_on;
431 atomic_t redundant_softirqs_off;
432 atomic_t nr_unused_locks;
433 atomic_t nr_cyclic_checks;
434 atomic_t nr_cyclic_check_recursions;
435 atomic_t nr_find_usage_forwards_checks;
436 atomic_t nr_find_usage_forwards_recursions;
437 atomic_t nr_find_usage_backwards_checks;
438 atomic_t nr_find_usage_backwards_recursions;
439 #endif
440
441 /*
442  * Locking printouts:
443  */
444
445 #define __USAGE(__STATE)                                                \
446         [LOCK_USED_IN_##__STATE] = "IN-"__stringify(__STATE)"-W",       \
447         [LOCK_ENABLED_##__STATE] = __stringify(__STATE)"-ON-W",         \
448         [LOCK_USED_IN_##__STATE##_READ] = "IN-"__stringify(__STATE)"-R",\
449         [LOCK_ENABLED_##__STATE##_READ] = __stringify(__STATE)"-ON-R",
450
451 static const char *usage_str[] =
452 {
453 #define LOCKDEP_STATE(__STATE) __USAGE(__STATE)
454 #include "lockdep_states.h"
455 #undef LOCKDEP_STATE
456         [LOCK_USED] = "INITIAL USE",
457 };
458
459 const char * __get_key_name(struct lockdep_subclass_key *key, char *str)
460 {
461         return kallsyms_lookup((unsigned long)key, NULL, NULL, NULL, str);
462 }
463
464 static inline unsigned long lock_flag(enum lock_usage_bit bit)
465 {
466         return 1UL << bit;
467 }
468
469 static char get_usage_char(struct lock_class *class, enum lock_usage_bit bit)
470 {
471         char c = '.';
472
473         if (class->usage_mask & lock_flag(bit + 2))
474                 c = '+';
475         if (class->usage_mask & lock_flag(bit)) {
476                 c = '-';
477                 if (class->usage_mask & lock_flag(bit + 2))
478                         c = '?';
479         }
480
481         return c;
482 }
483
484 void get_usage_chars(struct lock_class *class, char usage[LOCK_USAGE_CHARS])
485 {
486         int i = 0;
487
488 #define LOCKDEP_STATE(__STATE)                                          \
489         usage[i++] = get_usage_char(class, LOCK_USED_IN_##__STATE);     \
490         usage[i++] = get_usage_char(class, LOCK_USED_IN_##__STATE##_READ);
491 #include "lockdep_states.h"
492 #undef LOCKDEP_STATE
493
494         usage[i] = '\0';
495 }
496
497 static void print_lock_name(struct lock_class *class)
498 {
499         char str[KSYM_NAME_LEN], usage[LOCK_USAGE_CHARS];
500         const char *name;
501
502         get_usage_chars(class, usage);
503
504         name = class->name;
505         if (!name) {
506                 name = __get_key_name(class->key, str);
507                 printk(" (%s", name);
508         } else {
509                 printk(" (%s", name);
510                 if (class->name_version > 1)
511                         printk("#%d", class->name_version);
512                 if (class->subclass)
513                         printk("/%d", class->subclass);
514         }
515         printk("){%s}", usage);
516 }
517
518 static void print_lockdep_cache(struct lockdep_map *lock)
519 {
520         const char *name;
521         char str[KSYM_NAME_LEN];
522
523         name = lock->name;
524         if (!name)
525                 name = __get_key_name(lock->key->subkeys, str);
526
527         printk("%s", name);
528 }
529
530 static void print_lock(struct held_lock *hlock)
531 {
532         print_lock_name(hlock_class(hlock));
533         printk(", at: ");
534         print_ip_sym(hlock->acquire_ip);
535 }
536
537 static void lockdep_print_held_locks(struct task_struct *curr)
538 {
539         int i, depth = curr->lockdep_depth;
540
541         if (!depth) {
542                 printk("no locks held by %s/%d.\n", curr->comm, task_pid_nr(curr));
543                 return;
544         }
545         printk("%d lock%s held by %s/%d:\n",
546                 depth, depth > 1 ? "s" : "", curr->comm, task_pid_nr(curr));
547
548         for (i = 0; i < depth; i++) {
549                 printk(" #%d: ", i);
550                 print_lock(curr->held_locks + i);
551         }
552 }
553
554 static void print_lock_class_header(struct lock_class *class, int depth)
555 {
556         int bit;
557
558         printk("%*s->", depth, "");
559         print_lock_name(class);
560         printk(" ops: %lu", class->ops);
561         printk(" {\n");
562
563         for (bit = 0; bit < LOCK_USAGE_STATES; bit++) {
564                 if (class->usage_mask & (1 << bit)) {
565                         int len = depth;
566
567                         len += printk("%*s   %s", depth, "", usage_str[bit]);
568                         len += printk(" at:\n");
569                         print_stack_trace(class->usage_traces + bit, len);
570                 }
571         }
572         printk("%*s }\n", depth, "");
573
574         printk("%*s ... key      at: ",depth,"");
575         print_ip_sym((unsigned long)class->key);
576 }
577
578 /*
579  * printk the shortest lock dependencies from @start to @end in reverse order:
580  */
581 static void __used
582 print_shortest_lock_dependencies(struct lock_list *leaf,
583                                 struct lock_list *root)
584 {
585         struct lock_list *entry = leaf;
586         int depth;
587
588         /*compute depth from generated tree by BFS*/
589         depth = get_lock_depth(leaf);
590
591         do {
592                 print_lock_class_header(entry->class, depth);
593                 printk("%*s ... acquired at:\n", depth, "");
594                 print_stack_trace(&entry->trace, 2);
595                 printk("\n");
596
597                 if (depth == 0 && (entry != root)) {
598                         printk("lockdep:%s bad BFS generated tree\n", __func__);
599                         break;
600                 }
601
602                 entry = get_lock_parent(entry);
603                 depth--;
604         } while (entry && (depth >= 0));
605
606         return;
607 }
608 /*
609  * printk all lock dependencies starting at <entry>:
610  */
611 static void __used
612 print_lock_dependencies(struct lock_class *class, int depth)
613 {
614         struct lock_list *entry;
615
616         if (lockdep_dependency_visit(class, depth))
617                 return;
618
619         if (DEBUG_LOCKS_WARN_ON(depth >= 20))
620                 return;
621
622         print_lock_class_header(class, depth);
623
624         list_for_each_entry(entry, &class->locks_after, entry) {
625                 if (DEBUG_LOCKS_WARN_ON(!entry->class))
626                         return;
627
628                 print_lock_dependencies(entry->class, depth + 1);
629
630                 printk("%*s ... acquired at:\n",depth,"");
631                 print_stack_trace(&entry->trace, 2);
632                 printk("\n");
633         }
634 }
635
636 static void print_kernel_version(void)
637 {
638         printk("%s %.*s\n", init_utsname()->release,
639                 (int)strcspn(init_utsname()->version, " "),
640                 init_utsname()->version);
641 }
642
643 static int very_verbose(struct lock_class *class)
644 {
645 #if VERY_VERBOSE
646         return class_filter(class);
647 #endif
648         return 0;
649 }
650
651 /*
652  * Is this the address of a static object:
653  */
654 static int static_obj(void *obj)
655 {
656         unsigned long start = (unsigned long) &_stext,
657                       end   = (unsigned long) &_end,
658                       addr  = (unsigned long) obj;
659 #ifdef CONFIG_SMP
660         int i;
661 #endif
662
663         /*
664          * static variable?
665          */
666         if ((addr >= start) && (addr < end))
667                 return 1;
668
669 #ifdef CONFIG_SMP
670         /*
671          * percpu var?
672          */
673         for_each_possible_cpu(i) {
674                 start = (unsigned long) &__per_cpu_start + per_cpu_offset(i);
675                 end   = (unsigned long) &__per_cpu_start + PERCPU_ENOUGH_ROOM
676                                         + per_cpu_offset(i);
677
678                 if ((addr >= start) && (addr < end))
679                         return 1;
680         }
681 #endif
682
683         /*
684          * module var?
685          */
686         return is_module_address(addr);
687 }
688
689 /*
690  * To make lock name printouts unique, we calculate a unique
691  * class->name_version generation counter:
692  */
693 static int count_matching_names(struct lock_class *new_class)
694 {
695         struct lock_class *class;
696         int count = 0;
697
698         if (!new_class->name)
699                 return 0;
700
701         list_for_each_entry(class, &all_lock_classes, lock_entry) {
702                 if (new_class->key - new_class->subclass == class->key)
703                         return class->name_version;
704                 if (class->name && !strcmp(class->name, new_class->name))
705                         count = max(count, class->name_version);
706         }
707
708         return count + 1;
709 }
710
711 /*
712  * Register a lock's class in the hash-table, if the class is not present
713  * yet. Otherwise we look it up. We cache the result in the lock object
714  * itself, so actual lookup of the hash should be once per lock object.
715  */
716 static inline struct lock_class *
717 look_up_lock_class(struct lockdep_map *lock, unsigned int subclass)
718 {
719         struct lockdep_subclass_key *key;
720         struct list_head *hash_head;
721         struct lock_class *class;
722
723 #ifdef CONFIG_DEBUG_LOCKDEP
724         /*
725          * If the architecture calls into lockdep before initializing
726          * the hashes then we'll warn about it later. (we cannot printk
727          * right now)
728          */
729         if (unlikely(!lockdep_initialized)) {
730                 lockdep_init();
731                 lockdep_init_error = 1;
732                 save_stack_trace(&lockdep_init_trace);
733         }
734 #endif
735
736         /*
737          * Static locks do not have their class-keys yet - for them the key
738          * is the lock object itself:
739          */
740         if (unlikely(!lock->key))
741                 lock->key = (void *)lock;
742
743         /*
744          * NOTE: the class-key must be unique. For dynamic locks, a static
745          * lock_class_key variable is passed in through the mutex_init()
746          * (or spin_lock_init()) call - which acts as the key. For static
747          * locks we use the lock object itself as the key.
748          */
749         BUILD_BUG_ON(sizeof(struct lock_class_key) >
750                         sizeof(struct lockdep_map));
751
752         key = lock->key->subkeys + subclass;
753
754         hash_head = classhashentry(key);
755
756         /*
757          * We can walk the hash lockfree, because the hash only
758          * grows, and we are careful when adding entries to the end:
759          */
760         list_for_each_entry(class, hash_head, hash_entry) {
761                 if (class->key == key) {
762                         WARN_ON_ONCE(class->name != lock->name);
763                         return class;
764                 }
765         }
766
767         return NULL;
768 }
769
770 /*
771  * Register a lock's class in the hash-table, if the class is not present
772  * yet. Otherwise we look it up. We cache the result in the lock object
773  * itself, so actual lookup of the hash should be once per lock object.
774  */
775 static inline struct lock_class *
776 register_lock_class(struct lockdep_map *lock, unsigned int subclass, int force)
777 {
778         struct lockdep_subclass_key *key;
779         struct list_head *hash_head;
780         struct lock_class *class;
781         unsigned long flags;
782
783         class = look_up_lock_class(lock, subclass);
784         if (likely(class))
785                 return class;
786
787         /*
788          * Debug-check: all keys must be persistent!
789          */
790         if (!static_obj(lock->key)) {
791                 debug_locks_off();
792                 printk("INFO: trying to register non-static key.\n");
793                 printk("the code is fine but needs lockdep annotation.\n");
794                 printk("turning off the locking correctness validator.\n");
795                 dump_stack();
796
797                 return NULL;
798         }
799
800         key = lock->key->subkeys + subclass;
801         hash_head = classhashentry(key);
802
803         raw_local_irq_save(flags);
804         if (!graph_lock()) {
805                 raw_local_irq_restore(flags);
806                 return NULL;
807         }
808         /*
809          * We have to do the hash-walk again, to avoid races
810          * with another CPU:
811          */
812         list_for_each_entry(class, hash_head, hash_entry)
813                 if (class->key == key)
814                         goto out_unlock_set;
815         /*
816          * Allocate a new key from the static array, and add it to
817          * the hash:
818          */
819         if (nr_lock_classes >= MAX_LOCKDEP_KEYS) {
820                 if (!debug_locks_off_graph_unlock()) {
821                         raw_local_irq_restore(flags);
822                         return NULL;
823                 }
824                 raw_local_irq_restore(flags);
825
826                 printk("BUG: MAX_LOCKDEP_KEYS too low!\n");
827                 printk("turning off the locking correctness validator.\n");
828                 dump_stack();
829                 return NULL;
830         }
831         class = lock_classes + nr_lock_classes++;
832         debug_atomic_inc(&nr_unused_locks);
833         class->key = key;
834         class->name = lock->name;
835         class->subclass = subclass;
836         INIT_LIST_HEAD(&class->lock_entry);
837         INIT_LIST_HEAD(&class->locks_before);
838         INIT_LIST_HEAD(&class->locks_after);
839         class->name_version = count_matching_names(class);
840         /*
841          * We use RCU's safe list-add method to make
842          * parallel walking of the hash-list safe:
843          */
844         list_add_tail_rcu(&class->hash_entry, hash_head);
845         /*
846          * Add it to the global list of classes:
847          */
848         list_add_tail_rcu(&class->lock_entry, &all_lock_classes);
849
850         if (verbose(class)) {
851                 graph_unlock();
852                 raw_local_irq_restore(flags);
853
854                 printk("\nnew class %p: %s", class->key, class->name);
855                 if (class->name_version > 1)
856                         printk("#%d", class->name_version);
857                 printk("\n");
858                 dump_stack();
859
860                 raw_local_irq_save(flags);
861                 if (!graph_lock()) {
862                         raw_local_irq_restore(flags);
863                         return NULL;
864                 }
865         }
866 out_unlock_set:
867         graph_unlock();
868         raw_local_irq_restore(flags);
869
870         if (!subclass || force)
871                 lock->class_cache = class;
872
873         if (DEBUG_LOCKS_WARN_ON(class->subclass != subclass))
874                 return NULL;
875
876         return class;
877 }
878
879 #ifdef CONFIG_PROVE_LOCKING
880 /*
881  * Allocate a lockdep entry. (assumes the graph_lock held, returns
882  * with NULL on failure)
883  */
884 static struct lock_list *alloc_list_entry(void)
885 {
886         if (nr_list_entries >= MAX_LOCKDEP_ENTRIES) {
887                 if (!debug_locks_off_graph_unlock())
888                         return NULL;
889
890                 printk("BUG: MAX_LOCKDEP_ENTRIES too low!\n");
891                 printk("turning off the locking correctness validator.\n");
892                 dump_stack();
893                 return NULL;
894         }
895         return list_entries + nr_list_entries++;
896 }
897
898 /*
899  * Add a new dependency to the head of the list:
900  */
901 static int add_lock_to_list(struct lock_class *class, struct lock_class *this,
902                             struct list_head *head, unsigned long ip, int distance)
903 {
904         struct lock_list *entry;
905         /*
906          * Lock not present yet - get a new dependency struct and
907          * add it to the list:
908          */
909         entry = alloc_list_entry();
910         if (!entry)
911                 return 0;
912
913         if (!save_trace(&entry->trace))
914                 return 0;
915
916         entry->class = this;
917         entry->distance = distance;
918         /*
919          * Since we never remove from the dependency list, the list can
920          * be walked lockless by other CPUs, it's only allocation
921          * that must be protected by the spinlock. But this also means
922          * we must make new entries visible only once writes to the
923          * entry become visible - hence the RCU op:
924          */
925         list_add_tail_rcu(&entry->entry, head);
926
927         return 1;
928 }
929
930 unsigned long bfs_accessed[BITS_TO_LONGS(MAX_LOCKDEP_ENTRIES)];
931 static struct circular_queue  lock_cq;
932
933 static int __bfs(struct lock_list *source_entry,
934                         void *data,
935                         int (*match)(struct lock_list *entry, void *data),
936                         struct lock_list **target_entry,
937                         int forward)
938 {
939         struct lock_list *entry;
940         struct list_head *head;
941         struct circular_queue *cq = &lock_cq;
942         int ret = 1;
943
944         if (match(source_entry, data)) {
945                 *target_entry = source_entry;
946                 ret = 0;
947                 goto exit;
948         }
949
950         if (forward)
951                 head = &source_entry->class->locks_after;
952         else
953                 head = &source_entry->class->locks_before;
954
955         if (list_empty(head))
956                 goto exit;
957
958         __cq_init(cq);
959         __cq_enqueue(cq, (unsigned long)source_entry);
960
961         while (!__cq_empty(cq)) {
962                 struct lock_list *lock;
963
964                 __cq_dequeue(cq, (unsigned long *)&lock);
965
966                 if (!lock->class) {
967                         ret = -2;
968                         goto exit;
969                 }
970
971                 if (forward)
972                         head = &lock->class->locks_after;
973                 else
974                         head = &lock->class->locks_before;
975
976                 list_for_each_entry(entry, head, entry) {
977                         if (!lock_accessed(entry)) {
978                                 mark_lock_accessed(entry, lock);
979                                 if (match(entry, data)) {
980                                         *target_entry = entry;
981                                         ret = 0;
982                                         goto exit;
983                                 }
984
985                                 if (__cq_enqueue(cq, (unsigned long)entry)) {
986                                         ret = -1;
987                                         goto exit;
988                                 }
989                         }
990                 }
991         }
992 exit:
993         return ret;
994 }
995
996 static inline int __bfs_forwards(struct lock_list *src_entry,
997                         void *data,
998                         int (*match)(struct lock_list *entry, void *data),
999                         struct lock_list **target_entry)
1000 {
1001         return __bfs(src_entry, data, match, target_entry, 1);
1002
1003 }
1004
1005 static inline int __bfs_backwards(struct lock_list *src_entry,
1006                         void *data,
1007                         int (*match)(struct lock_list *entry, void *data),
1008                         struct lock_list **target_entry)
1009 {
1010         return __bfs(src_entry, data, match, target_entry, 0);
1011
1012 }
1013
1014 /*
1015  * Recursive, forwards-direction lock-dependency checking, used for
1016  * both noncyclic checking and for hardirq-unsafe/softirq-unsafe
1017  * checking.
1018  */
1019
1020 /*
1021  * Print a dependency chain entry (this is only done when a deadlock
1022  * has been detected):
1023  */
1024 static noinline int
1025 print_circular_bug_entry(struct lock_list *target, int depth)
1026 {
1027         if (debug_locks_silent)
1028                 return 0;
1029         printk("\n-> #%u", depth);
1030         print_lock_name(target->class);
1031         printk(":\n");
1032         print_stack_trace(&target->trace, 6);
1033
1034         return 0;
1035 }
1036
1037 /*
1038  * When a circular dependency is detected, print the
1039  * header first:
1040  */
1041 static noinline int
1042 print_circular_bug_header(struct lock_list *entry, unsigned int depth,
1043                         struct held_lock *check_src,
1044                         struct held_lock *check_tgt)
1045 {
1046         struct task_struct *curr = current;
1047
1048         if (debug_locks_silent)
1049                 return 0;
1050
1051         printk("\n=======================================================\n");
1052         printk(  "[ INFO: possible circular locking dependency detected ]\n");
1053         print_kernel_version();
1054         printk(  "-------------------------------------------------------\n");
1055         printk("%s/%d is trying to acquire lock:\n",
1056                 curr->comm, task_pid_nr(curr));
1057         print_lock(check_src);
1058         printk("\nbut task is already holding lock:\n");
1059         print_lock(check_tgt);
1060         printk("\nwhich lock already depends on the new lock.\n\n");
1061         printk("\nthe existing dependency chain (in reverse order) is:\n");
1062
1063         print_circular_bug_entry(entry, depth);
1064
1065         return 0;
1066 }
1067
1068 static inline int class_equal(struct lock_list *entry, void *data)
1069 {
1070         return entry->class == data;
1071 }
1072
1073 static noinline int print_circular_bug(struct lock_list *this,
1074                                 struct lock_list *target,
1075                                 struct held_lock *check_src,
1076                                 struct held_lock *check_tgt)
1077 {
1078         struct task_struct *curr = current;
1079         struct lock_list *parent;
1080         int depth;
1081
1082         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1083                 return 0;
1084
1085         if (!save_trace(&this->trace))
1086                 return 0;
1087
1088         depth = get_lock_depth(target);
1089
1090         print_circular_bug_header(target, depth, check_src, check_tgt);
1091
1092         parent = get_lock_parent(target);
1093
1094         while (parent) {
1095                 print_circular_bug_entry(parent, --depth);
1096                 parent = get_lock_parent(parent);
1097         }
1098
1099         printk("\nother info that might help us debug this:\n\n");
1100         lockdep_print_held_locks(curr);
1101
1102         printk("\nstack backtrace:\n");
1103         dump_stack();
1104
1105         return 0;
1106 }
1107
1108 static noinline int print_bfs_bug(int ret)
1109 {
1110         if (!debug_locks_off_graph_unlock())
1111                 return 0;
1112
1113         WARN(1, "lockdep bfs error:%d\n", ret);
1114
1115         return 0;
1116 }
1117
1118 static int noop_count(struct lock_list *entry, void *data)
1119 {
1120         (*(unsigned long *)data)++;
1121         return 0;
1122 }
1123
1124 unsigned long __lockdep_count_forward_deps(struct lock_list *this)
1125 {
1126         unsigned long  count = 0;
1127         struct lock_list *uninitialized_var(target_entry);
1128
1129         __bfs_forwards(this, (void *)&count, noop_count, &target_entry);
1130
1131         return count;
1132 }
1133 unsigned long lockdep_count_forward_deps(struct lock_class *class)
1134 {
1135         unsigned long ret, flags;
1136         struct lock_list this;
1137
1138         this.parent = NULL;
1139         this.class = class;
1140
1141         local_irq_save(flags);
1142         __raw_spin_lock(&lockdep_lock);
1143         ret = __lockdep_count_forward_deps(&this);
1144         __raw_spin_unlock(&lockdep_lock);
1145         local_irq_restore(flags);
1146
1147         return ret;
1148 }
1149
1150 unsigned long __lockdep_count_backward_deps(struct lock_list *this)
1151 {
1152         unsigned long  count = 0;
1153         struct lock_list *uninitialized_var(target_entry);
1154
1155         __bfs_backwards(this, (void *)&count, noop_count, &target_entry);
1156
1157         return count;
1158 }
1159
1160 unsigned long lockdep_count_backward_deps(struct lock_class *class)
1161 {
1162         unsigned long ret, flags;
1163         struct lock_list this;
1164
1165         this.parent = NULL;
1166         this.class = class;
1167
1168         local_irq_save(flags);
1169         __raw_spin_lock(&lockdep_lock);
1170         ret = __lockdep_count_backward_deps(&this);
1171         __raw_spin_unlock(&lockdep_lock);
1172         local_irq_restore(flags);
1173
1174         return ret;
1175 }
1176
1177 /*
1178  * Prove that the dependency graph starting at <entry> can not
1179  * lead to <target>. Print an error and return 0 if it does.
1180  */
1181 static noinline int
1182 check_noncircular(struct lock_list *root, struct lock_class *target,
1183                 struct lock_list **target_entry)
1184 {
1185         int result;
1186
1187         debug_atomic_inc(&nr_cyclic_checks);
1188
1189         result = __bfs_forwards(root, target, class_equal, target_entry);
1190
1191         return result;
1192 }
1193
1194 #if defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING)
1195 /*
1196  * Forwards and backwards subgraph searching, for the purposes of
1197  * proving that two subgraphs can be connected by a new dependency
1198  * without creating any illegal irq-safe -> irq-unsafe lock dependency.
1199  */
1200
1201
1202 #define   BFS_PROCESS_RET(ret)  do { \
1203                                         if (ret < 0) \
1204                                                 return print_bfs_bug(ret); \
1205                                         if (ret == 1) \
1206                                                 return 1; \
1207                                 } while (0)
1208
1209 static inline int usage_match(struct lock_list *entry, void *bit)
1210 {
1211         return entry->class->usage_mask & (1 << (enum lock_usage_bit)bit);
1212 }
1213
1214
1215
1216 /*
1217  * Find a node in the forwards-direction dependency sub-graph starting
1218  * at @root->class that matches @bit.
1219  *
1220  * Return 0 if such a node exists in the subgraph, and put that node
1221  * into *@target_entry.
1222  *
1223  * Return 1 otherwise and keep *@target_entry unchanged.
1224  * Return <0 on error.
1225  */
1226 static int
1227 find_usage_forwards(struct lock_list *root, enum lock_usage_bit bit,
1228                         struct lock_list **target_entry)
1229 {
1230         int result;
1231
1232         debug_atomic_inc(&nr_find_usage_forwards_checks);
1233
1234         result = __bfs_forwards(root, (void *)bit, usage_match, target_entry);
1235
1236         return result;
1237 }
1238
1239 /*
1240  * Find a node in the backwards-direction dependency sub-graph starting
1241  * at @root->class that matches @bit.
1242  *
1243  * Return 0 if such a node exists in the subgraph, and put that node
1244  * into *@target_entry.
1245  *
1246  * Return 1 otherwise and keep *@target_entry unchanged.
1247  * Return <0 on error.
1248  */
1249 static int
1250 find_usage_backwards(struct lock_list *root, enum lock_usage_bit bit,
1251                         struct lock_list **target_entry)
1252 {
1253         int result;
1254
1255         debug_atomic_inc(&nr_find_usage_backwards_checks);
1256
1257         result = __bfs_backwards(root, (void *)bit, usage_match, target_entry);
1258
1259         return result;
1260 }
1261
1262
1263 static int
1264 print_bad_irq_dependency(struct task_struct *curr,
1265                          struct lock_list *prev_root,
1266                          struct lock_list *next_root,
1267                          struct lock_list *backwards_entry,
1268                          struct lock_list *forwards_entry,
1269                          struct held_lock *prev,
1270                          struct held_lock *next,
1271                          enum lock_usage_bit bit1,
1272                          enum lock_usage_bit bit2,
1273                          const char *irqclass)
1274 {
1275         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1276                 return 0;
1277
1278         printk("\n======================================================\n");
1279         printk(  "[ INFO: %s-safe -> %s-unsafe lock order detected ]\n",
1280                 irqclass, irqclass);
1281         print_kernel_version();
1282         printk(  "------------------------------------------------------\n");
1283         printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] is trying to acquire:\n",
1284                 curr->comm, task_pid_nr(curr),
1285                 curr->hardirq_context, hardirq_count() >> HARDIRQ_SHIFT,
1286                 curr->softirq_context, softirq_count() >> SOFTIRQ_SHIFT,
1287                 curr->hardirqs_enabled,
1288                 curr->softirqs_enabled);
1289         print_lock(next);
1290
1291         printk("\nand this task is already holding:\n");
1292         print_lock(prev);
1293         printk("which would create a new lock dependency:\n");
1294         print_lock_name(hlock_class(prev));
1295         printk(" ->");
1296         print_lock_name(hlock_class(next));
1297         printk("\n");
1298
1299         printk("\nbut this new dependency connects a %s-irq-safe lock:\n",
1300                 irqclass);
1301         print_lock_name(backwards_entry->class);
1302         printk("\n... which became %s-irq-safe at:\n", irqclass);
1303
1304         print_stack_trace(backwards_entry->class->usage_traces + bit1, 1);
1305
1306         printk("\nto a %s-irq-unsafe lock:\n", irqclass);
1307         print_lock_name(forwards_entry->class);
1308         printk("\n... which became %s-irq-unsafe at:\n", irqclass);
1309         printk("...");
1310
1311         print_stack_trace(forwards_entry->class->usage_traces + bit2, 1);
1312
1313         printk("\nother info that might help us debug this:\n\n");
1314         lockdep_print_held_locks(curr);
1315
1316         printk("\nthe dependencies between %s-irq-safe lock", irqclass);
1317         printk(" and the holding lock:\n");
1318         if (!save_trace(&prev_root->trace))
1319                 return 0;
1320         print_shortest_lock_dependencies(backwards_entry, prev_root);
1321
1322         printk("\nthe dependencies between the lock to be acquired");
1323         printk(" and %s-irq-unsafe lock:\n", irqclass);
1324         if (!save_trace(&next_root->trace))
1325                 return 0;
1326         print_shortest_lock_dependencies(forwards_entry, next_root);
1327
1328         printk("\nstack backtrace:\n");
1329         dump_stack();
1330
1331         return 0;
1332 }
1333
1334 static int
1335 check_usage(struct task_struct *curr, struct held_lock *prev,
1336             struct held_lock *next, enum lock_usage_bit bit_backwards,
1337             enum lock_usage_bit bit_forwards, const char *irqclass)
1338 {
1339         int ret;
1340         struct lock_list this, that;
1341         struct lock_list *uninitialized_var(target_entry);
1342         struct lock_list *uninitialized_var(target_entry1);
1343
1344         this.parent = NULL;
1345
1346         this.class = hlock_class(prev);
1347         ret = find_usage_backwards(&this, bit_backwards, &target_entry);
1348         BFS_PROCESS_RET(ret);
1349
1350         that.parent = NULL;
1351         that.class = hlock_class(next);
1352         ret = find_usage_forwards(&that, bit_forwards, &target_entry1);
1353         BFS_PROCESS_RET(ret);
1354
1355         return print_bad_irq_dependency(curr, &this, &that,
1356                         target_entry, target_entry1,
1357                         prev, next,
1358                         bit_backwards, bit_forwards, irqclass);
1359 }
1360
1361 static const char *state_names[] = {
1362 #define LOCKDEP_STATE(__STATE) \
1363         __stringify(__STATE),
1364 #include "lockdep_states.h"
1365 #undef LOCKDEP_STATE
1366 };
1367
1368 static const char *state_rnames[] = {
1369 #define LOCKDEP_STATE(__STATE) \
1370         __stringify(__STATE)"-READ",
1371 #include "lockdep_states.h"
1372 #undef LOCKDEP_STATE
1373 };
1374
1375 static inline const char *state_name(enum lock_usage_bit bit)
1376 {
1377         return (bit & 1) ? state_rnames[bit >> 2] : state_names[bit >> 2];
1378 }
1379
1380 static int exclusive_bit(int new_bit)
1381 {
1382         /*
1383          * USED_IN
1384          * USED_IN_READ
1385          * ENABLED
1386          * ENABLED_READ
1387          *
1388          * bit 0 - write/read
1389          * bit 1 - used_in/enabled
1390          * bit 2+  state
1391          */
1392
1393         int state = new_bit & ~3;
1394         int dir = new_bit & 2;
1395
1396         /*
1397          * keep state, bit flip the direction and strip read.
1398          */
1399         return state | (dir ^ 2);
1400 }
1401
1402 static int check_irq_usage(struct task_struct *curr, struct held_lock *prev,
1403                            struct held_lock *next, enum lock_usage_bit bit)
1404 {
1405         /*
1406          * Prove that the new dependency does not connect a hardirq-safe
1407          * lock with a hardirq-unsafe lock - to achieve this we search
1408          * the backwards-subgraph starting at <prev>, and the
1409          * forwards-subgraph starting at <next>:
1410          */
1411         if (!check_usage(curr, prev, next, bit,
1412                            exclusive_bit(bit), state_name(bit)))
1413                 return 0;
1414
1415         bit++; /* _READ */
1416
1417         /*
1418          * Prove that the new dependency does not connect a hardirq-safe-read
1419          * lock with a hardirq-unsafe lock - to achieve this we search
1420          * the backwards-subgraph starting at <prev>, and the
1421          * forwards-subgraph starting at <next>:
1422          */
1423         if (!check_usage(curr, prev, next, bit,
1424                            exclusive_bit(bit), state_name(bit)))
1425                 return 0;
1426
1427         return 1;
1428 }
1429
1430 static int
1431 check_prev_add_irq(struct task_struct *curr, struct held_lock *prev,
1432                 struct held_lock *next)
1433 {
1434 #define LOCKDEP_STATE(__STATE)                                          \
1435         if (!check_irq_usage(curr, prev, next, LOCK_USED_IN_##__STATE)) \
1436                 return 0;
1437 #include "lockdep_states.h"
1438 #undef LOCKDEP_STATE
1439
1440         return 1;
1441 }
1442
1443 static void inc_chains(void)
1444 {
1445         if (current->hardirq_context)
1446                 nr_hardirq_chains++;
1447         else {
1448                 if (current->softirq_context)
1449                         nr_softirq_chains++;
1450                 else
1451                         nr_process_chains++;
1452         }
1453 }
1454
1455 #else
1456
1457 static inline int
1458 check_prev_add_irq(struct task_struct *curr, struct held_lock *prev,
1459                 struct held_lock *next)
1460 {
1461         return 1;
1462 }
1463
1464 static inline void inc_chains(void)
1465 {
1466         nr_process_chains++;
1467 }
1468
1469 #endif
1470
1471 static int
1472 print_deadlock_bug(struct task_struct *curr, struct held_lock *prev,
1473                    struct held_lock *next)
1474 {
1475         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1476                 return 0;
1477
1478         printk("\n=============================================\n");
1479         printk(  "[ INFO: possible recursive locking detected ]\n");
1480         print_kernel_version();
1481         printk(  "---------------------------------------------\n");
1482         printk("%s/%d is trying to acquire lock:\n",
1483                 curr->comm, task_pid_nr(curr));
1484         print_lock(next);
1485         printk("\nbut task is already holding lock:\n");
1486         print_lock(prev);
1487
1488         printk("\nother info that might help us debug this:\n");
1489         lockdep_print_held_locks(curr);
1490
1491         printk("\nstack backtrace:\n");
1492         dump_stack();
1493
1494         return 0;
1495 }
1496
1497 /*
1498  * Check whether we are holding such a class already.
1499  *
1500  * (Note that this has to be done separately, because the graph cannot
1501  * detect such classes of deadlocks.)
1502  *
1503  * Returns: 0 on deadlock detected, 1 on OK, 2 on recursive read
1504  */
1505 static int
1506 check_deadlock(struct task_struct *curr, struct held_lock *next,
1507                struct lockdep_map *next_instance, int read)
1508 {
1509         struct held_lock *prev;
1510         struct held_lock *nest = NULL;
1511         int i;
1512
1513         for (i = 0; i < curr->lockdep_depth; i++) {
1514                 prev = curr->held_locks + i;
1515
1516                 if (prev->instance == next->nest_lock)
1517                         nest = prev;
1518
1519                 if (hlock_class(prev) != hlock_class(next))
1520                         continue;
1521
1522                 /*
1523                  * Allow read-after-read recursion of the same
1524                  * lock class (i.e. read_lock(lock)+read_lock(lock)):
1525                  */
1526                 if ((read == 2) && prev->read)
1527                         return 2;
1528
1529                 /*
1530                  * We're holding the nest_lock, which serializes this lock's
1531                  * nesting behaviour.
1532                  */
1533                 if (nest)
1534                         return 2;
1535
1536                 return print_deadlock_bug(curr, prev, next);
1537         }
1538         return 1;
1539 }
1540
1541 /*
1542  * There was a chain-cache miss, and we are about to add a new dependency
1543  * to a previous lock. We recursively validate the following rules:
1544  *
1545  *  - would the adding of the <prev> -> <next> dependency create a
1546  *    circular dependency in the graph? [== circular deadlock]
1547  *
1548  *  - does the new prev->next dependency connect any hardirq-safe lock
1549  *    (in the full backwards-subgraph starting at <prev>) with any
1550  *    hardirq-unsafe lock (in the full forwards-subgraph starting at
1551  *    <next>)? [== illegal lock inversion with hardirq contexts]
1552  *
1553  *  - does the new prev->next dependency connect any softirq-safe lock
1554  *    (in the full backwards-subgraph starting at <prev>) with any
1555  *    softirq-unsafe lock (in the full forwards-subgraph starting at
1556  *    <next>)? [== illegal lock inversion with softirq contexts]
1557  *
1558  * any of these scenarios could lead to a deadlock.
1559  *
1560  * Then if all the validations pass, we add the forwards and backwards
1561  * dependency.
1562  */
1563 static int
1564 check_prev_add(struct task_struct *curr, struct held_lock *prev,
1565                struct held_lock *next, int distance)
1566 {
1567         struct lock_list *entry;
1568         int ret;
1569         struct lock_list this;
1570         struct lock_list *uninitialized_var(target_entry);
1571
1572         /*
1573          * Prove that the new <prev> -> <next> dependency would not
1574          * create a circular dependency in the graph. (We do this by
1575          * forward-recursing into the graph starting at <next>, and
1576          * checking whether we can reach <prev>.)
1577          *
1578          * We are using global variables to control the recursion, to
1579          * keep the stackframe size of the recursive functions low:
1580          */
1581         this.class = hlock_class(next);
1582         this.parent = NULL;
1583         ret = check_noncircular(&this, hlock_class(prev), &target_entry);
1584         if (unlikely(!ret))
1585                 return print_circular_bug(&this, target_entry, next, prev);
1586         else if (unlikely(ret < 0))
1587                 return print_bfs_bug(ret);
1588
1589         if (!check_prev_add_irq(curr, prev, next))
1590                 return 0;
1591
1592         /*
1593          * For recursive read-locks we do all the dependency checks,
1594          * but we dont store read-triggered dependencies (only
1595          * write-triggered dependencies). This ensures that only the
1596          * write-side dependencies matter, and that if for example a
1597          * write-lock never takes any other locks, then the reads are
1598          * equivalent to a NOP.
1599          */
1600         if (next->read == 2 || prev->read == 2)
1601                 return 1;
1602         /*
1603          * Is the <prev> -> <next> dependency already present?
1604          *
1605          * (this may occur even though this is a new chain: consider
1606          *  e.g. the L1 -> L2 -> L3 -> L4 and the L5 -> L1 -> L2 -> L3
1607          *  chains - the second one will be new, but L1 already has
1608          *  L2 added to its dependency list, due to the first chain.)
1609          */
1610         list_for_each_entry(entry, &hlock_class(prev)->locks_after, entry) {
1611                 if (entry->class == hlock_class(next)) {
1612                         if (distance == 1)
1613                                 entry->distance = 1;
1614                         return 2;
1615                 }
1616         }
1617
1618         /*
1619          * Ok, all validations passed, add the new lock
1620          * to the previous lock's dependency list:
1621          */
1622         ret = add_lock_to_list(hlock_class(prev), hlock_class(next),
1623                                &hlock_class(prev)->locks_after,
1624                                next->acquire_ip, distance);
1625
1626         if (!ret)
1627                 return 0;
1628
1629         ret = add_lock_to_list(hlock_class(next), hlock_class(prev),
1630                                &hlock_class(next)->locks_before,
1631                                next->acquire_ip, distance);
1632         if (!ret)
1633                 return 0;
1634
1635         /*
1636          * Debugging printouts:
1637          */
1638         if (verbose(hlock_class(prev)) || verbose(hlock_class(next))) {
1639                 graph_unlock();
1640                 printk("\n new dependency: ");
1641                 print_lock_name(hlock_class(prev));
1642                 printk(" => ");
1643                 print_lock_name(hlock_class(next));
1644                 printk("\n");
1645                 dump_stack();
1646                 return graph_lock();
1647         }
1648         return 1;
1649 }
1650
1651 /*
1652  * Add the dependency to all directly-previous locks that are 'relevant'.
1653  * The ones that are relevant are (in increasing distance from curr):
1654  * all consecutive trylock entries and the final non-trylock entry - or
1655  * the end of this context's lock-chain - whichever comes first.
1656  */
1657 static int
1658 check_prevs_add(struct task_struct *curr, struct held_lock *next)
1659 {
1660         int depth = curr->lockdep_depth;
1661         struct held_lock *hlock;
1662
1663         /*
1664          * Debugging checks.
1665          *
1666          * Depth must not be zero for a non-head lock:
1667          */
1668         if (!depth)
1669                 goto out_bug;
1670         /*
1671          * At least two relevant locks must exist for this
1672          * to be a head:
1673          */
1674         if (curr->held_locks[depth].irq_context !=
1675                         curr->held_locks[depth-1].irq_context)
1676                 goto out_bug;
1677
1678         for (;;) {
1679                 int distance = curr->lockdep_depth - depth + 1;
1680                 hlock = curr->held_locks + depth-1;
1681                 /*
1682                  * Only non-recursive-read entries get new dependencies
1683                  * added:
1684                  */
1685                 if (hlock->read != 2) {
1686                         if (!check_prev_add(curr, hlock, next, distance))
1687                                 return 0;
1688                         /*
1689                          * Stop after the first non-trylock entry,
1690                          * as non-trylock entries have added their
1691                          * own direct dependencies already, so this
1692                          * lock is connected to them indirectly:
1693                          */
1694                         if (!hlock->trylock)
1695                                 break;
1696                 }
1697                 depth--;
1698                 /*
1699                  * End of lock-stack?
1700                  */
1701                 if (!depth)
1702                         break;
1703                 /*
1704                  * Stop the search if we cross into another context:
1705                  */
1706                 if (curr->held_locks[depth].irq_context !=
1707                                 curr->held_locks[depth-1].irq_context)
1708                         break;
1709         }
1710         return 1;
1711 out_bug:
1712         if (!debug_locks_off_graph_unlock())
1713                 return 0;
1714
1715         WARN_ON(1);
1716
1717         return 0;
1718 }
1719
1720 unsigned long nr_lock_chains;
1721 struct lock_chain lock_chains[MAX_LOCKDEP_CHAINS];
1722 int nr_chain_hlocks;
1723 static u16 chain_hlocks[MAX_LOCKDEP_CHAIN_HLOCKS];
1724
1725 struct lock_class *lock_chain_get_class(struct lock_chain *chain, int i)
1726 {
1727         return lock_classes + chain_hlocks[chain->base + i];
1728 }
1729
1730 /*
1731  * Look up a dependency chain. If the key is not present yet then
1732  * add it and return 1 - in this case the new dependency chain is
1733  * validated. If the key is already hashed, return 0.
1734  * (On return with 1 graph_lock is held.)
1735  */
1736 static inline int lookup_chain_cache(struct task_struct *curr,
1737                                      struct held_lock *hlock,
1738                                      u64 chain_key)
1739 {
1740         struct lock_class *class = hlock_class(hlock);
1741         struct list_head *hash_head = chainhashentry(chain_key);
1742         struct lock_chain *chain;
1743         struct held_lock *hlock_curr, *hlock_next;
1744         int i, j, n, cn;
1745
1746         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
1747                 return 0;
1748         /*
1749          * We can walk it lock-free, because entries only get added
1750          * to the hash:
1751          */
1752         list_for_each_entry(chain, hash_head, entry) {
1753                 if (chain->chain_key == chain_key) {
1754 cache_hit:
1755                         debug_atomic_inc(&chain_lookup_hits);
1756                         if (very_verbose(class))
1757                                 printk("\nhash chain already cached, key: "
1758                                         "%016Lx tail class: [%p] %s\n",
1759                                         (unsigned long long)chain_key,
1760                                         class->key, class->name);
1761                         return 0;
1762                 }
1763         }
1764         if (very_verbose(class))
1765                 printk("\nnew hash chain, key: %016Lx tail class: [%p] %s\n",
1766                         (unsigned long long)chain_key, class->key, class->name);
1767         /*
1768          * Allocate a new chain entry from the static array, and add
1769          * it to the hash:
1770          */
1771         if (!graph_lock())
1772                 return 0;
1773         /*
1774          * We have to walk the chain again locked - to avoid duplicates:
1775          */
1776         list_for_each_entry(chain, hash_head, entry) {
1777                 if (chain->chain_key == chain_key) {
1778                         graph_unlock();
1779                         goto cache_hit;
1780                 }
1781         }
1782         if (unlikely(nr_lock_chains >= MAX_LOCKDEP_CHAINS)) {
1783                 if (!debug_locks_off_graph_unlock())
1784                         return 0;
1785
1786                 printk("BUG: MAX_LOCKDEP_CHAINS too low!\n");
1787                 printk("turning off the locking correctness validator.\n");
1788                 dump_stack();
1789                 return 0;
1790         }
1791         chain = lock_chains + nr_lock_chains++;
1792         chain->chain_key = chain_key;
1793         chain->irq_context = hlock->irq_context;
1794         /* Find the first held_lock of current chain */
1795         hlock_next = hlock;
1796         for (i = curr->lockdep_depth - 1; i >= 0; i--) {
1797                 hlock_curr = curr->held_locks + i;
1798                 if (hlock_curr->irq_context != hlock_next->irq_context)
1799                         break;
1800                 hlock_next = hlock;
1801         }
1802         i++;
1803         chain->depth = curr->lockdep_depth + 1 - i;
1804         cn = nr_chain_hlocks;
1805         while (cn + chain->depth <= MAX_LOCKDEP_CHAIN_HLOCKS) {
1806                 n = cmpxchg(&nr_chain_hlocks, cn, cn + chain->depth);
1807                 if (n == cn)
1808                         break;
1809                 cn = n;
1810         }
1811         if (likely(cn + chain->depth <= MAX_LOCKDEP_CHAIN_HLOCKS)) {
1812                 chain->base = cn;
1813                 for (j = 0; j < chain->depth - 1; j++, i++) {
1814                         int lock_id = curr->held_locks[i].class_idx - 1;
1815                         chain_hlocks[chain->base + j] = lock_id;
1816                 }
1817                 chain_hlocks[chain->base + j] = class - lock_classes;
1818         }
1819         list_add_tail_rcu(&chain->entry, hash_head);
1820         debug_atomic_inc(&chain_lookup_misses);
1821         inc_chains();
1822
1823         return 1;
1824 }
1825
1826 static int validate_chain(struct task_struct *curr, struct lockdep_map *lock,
1827                 struct held_lock *hlock, int chain_head, u64 chain_key)
1828 {
1829         /*
1830          * Trylock needs to maintain the stack of held locks, but it
1831          * does not add new dependencies, because trylock can be done
1832          * in any order.
1833          *
1834          * We look up the chain_key and do the O(N^2) check and update of
1835          * the dependencies only if this is a new dependency chain.
1836          * (If lookup_chain_cache() returns with 1 it acquires
1837          * graph_lock for us)
1838          */
1839         if (!hlock->trylock && (hlock->check == 2) &&
1840             lookup_chain_cache(curr, hlock, chain_key)) {
1841                 /*
1842                  * Check whether last held lock:
1843                  *
1844                  * - is irq-safe, if this lock is irq-unsafe
1845                  * - is softirq-safe, if this lock is hardirq-unsafe
1846                  *
1847                  * And check whether the new lock's dependency graph
1848                  * could lead back to the previous lock.
1849                  *
1850                  * any of these scenarios could lead to a deadlock. If
1851                  * All validations
1852                  */
1853                 int ret = check_deadlock(curr, hlock, lock, hlock->read);
1854
1855                 if (!ret)
1856                         return 0;
1857                 /*
1858                  * Mark recursive read, as we jump over it when
1859                  * building dependencies (just like we jump over
1860                  * trylock entries):
1861                  */
1862                 if (ret == 2)
1863                         hlock->read = 2;
1864                 /*
1865                  * Add dependency only if this lock is not the head
1866                  * of the chain, and if it's not a secondary read-lock:
1867                  */
1868                 if (!chain_head && ret != 2)
1869                         if (!check_prevs_add(curr, hlock))
1870                                 return 0;
1871                 graph_unlock();
1872         } else
1873                 /* after lookup_chain_cache(): */
1874                 if (unlikely(!debug_locks))
1875                         return 0;
1876
1877         return 1;
1878 }
1879 #else
1880 static inline int validate_chain(struct task_struct *curr,
1881                 struct lockdep_map *lock, struct held_lock *hlock,
1882                 int chain_head, u64 chain_key)
1883 {
1884         return 1;
1885 }
1886 #endif
1887
1888 /*
1889  * We are building curr_chain_key incrementally, so double-check
1890  * it from scratch, to make sure that it's done correctly:
1891  */
1892 static void check_chain_key(struct task_struct *curr)
1893 {
1894 #ifdef CONFIG_DEBUG_LOCKDEP
1895         struct held_lock *hlock, *prev_hlock = NULL;
1896         unsigned int i, id;
1897         u64 chain_key = 0;
1898
1899         for (i = 0; i < curr->lockdep_depth; i++) {
1900                 hlock = curr->held_locks + i;
1901                 if (chain_key != hlock->prev_chain_key) {
1902                         debug_locks_off();
1903                         WARN(1, "hm#1, depth: %u [%u], %016Lx != %016Lx\n",
1904                                 curr->lockdep_depth, i,
1905                                 (unsigned long long)chain_key,
1906                                 (unsigned long long)hlock->prev_chain_key);
1907                         return;
1908                 }
1909                 id = hlock->class_idx - 1;
1910                 if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
1911                         return;
1912
1913                 if (prev_hlock && (prev_hlock->irq_context !=
1914                                                         hlock->irq_context))
1915                         chain_key = 0;
1916                 chain_key = iterate_chain_key(chain_key, id);
1917                 prev_hlock = hlock;
1918         }
1919         if (chain_key != curr->curr_chain_key) {
1920                 debug_locks_off();
1921                 WARN(1, "hm#2, depth: %u [%u], %016Lx != %016Lx\n",
1922                         curr->lockdep_depth, i,
1923                         (unsigned long long)chain_key,
1924                         (unsigned long long)curr->curr_chain_key);
1925         }
1926 #endif
1927 }
1928
1929 static int
1930 print_usage_bug(struct task_struct *curr, struct held_lock *this,
1931                 enum lock_usage_bit prev_bit, enum lock_usage_bit new_bit)
1932 {
1933         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1934                 return 0;
1935
1936         printk("\n=================================\n");
1937         printk(  "[ INFO: inconsistent lock state ]\n");
1938         print_kernel_version();
1939         printk(  "---------------------------------\n");
1940
1941         printk("inconsistent {%s} -> {%s} usage.\n",
1942                 usage_str[prev_bit], usage_str[new_bit]);
1943
1944         printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] takes:\n",
1945                 curr->comm, task_pid_nr(curr),
1946                 trace_hardirq_context(curr), hardirq_count() >> HARDIRQ_SHIFT,
1947                 trace_softirq_context(curr), softirq_count() >> SOFTIRQ_SHIFT,
1948                 trace_hardirqs_enabled(curr),
1949                 trace_softirqs_enabled(curr));
1950         print_lock(this);
1951
1952         printk("{%s} state was registered at:\n", usage_str[prev_bit]);
1953         print_stack_trace(hlock_class(this)->usage_traces + prev_bit, 1);
1954
1955         print_irqtrace_events(curr);
1956         printk("\nother info that might help us debug this:\n");
1957         lockdep_print_held_locks(curr);
1958
1959         printk("\nstack backtrace:\n");
1960         dump_stack();
1961
1962         return 0;
1963 }
1964
1965 /*
1966  * Print out an error if an invalid bit is set:
1967  */
1968 static inline int
1969 valid_state(struct task_struct *curr, struct held_lock *this,
1970             enum lock_usage_bit new_bit, enum lock_usage_bit bad_bit)
1971 {
1972         if (unlikely(hlock_class(this)->usage_mask & (1 << bad_bit)))
1973                 return print_usage_bug(curr, this, bad_bit, new_bit);
1974         return 1;
1975 }
1976
1977 static int mark_lock(struct task_struct *curr, struct held_lock *this,
1978                      enum lock_usage_bit new_bit);
1979
1980 #if defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING)
1981
1982 /*
1983  * print irq inversion bug:
1984  */
1985 static int
1986 print_irq_inversion_bug(struct task_struct *curr,
1987                         struct lock_list *root, struct lock_list *other,
1988                         struct held_lock *this, int forwards,
1989                         const char *irqclass)
1990 {
1991         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1992                 return 0;
1993
1994         printk("\n=========================================================\n");
1995         printk(  "[ INFO: possible irq lock inversion dependency detected ]\n");
1996         print_kernel_version();
1997         printk(  "---------------------------------------------------------\n");
1998         printk("%s/%d just changed the state of lock:\n",
1999                 curr->comm, task_pid_nr(curr));
2000         print_lock(this);
2001         if (forwards)
2002                 printk("but this lock took another, %s-unsafe lock in the past:\n", irqclass);
2003         else
2004                 printk("but this lock was taken by another, %s-safe lock in the past:\n", irqclass);
2005         print_lock_name(other->class);
2006         printk("\n\nand interrupts could create inverse lock ordering between them.\n\n");
2007
2008         printk("\nother info that might help us debug this:\n");
2009         lockdep_print_held_locks(curr);
2010
2011         printk("\nthe shortest dependencies between 2nd lock and 1st lock:\n");
2012         if (!save_trace(&root->trace))
2013                 return 0;
2014         print_shortest_lock_dependencies(other, root);
2015
2016         printk("\nstack backtrace:\n");
2017         dump_stack();
2018
2019         return 0;
2020 }
2021
2022 /*
2023  * Prove that in the forwards-direction subgraph starting at <this>
2024  * there is no lock matching <mask>:
2025  */
2026 static int
2027 check_usage_forwards(struct task_struct *curr, struct held_lock *this,
2028                      enum lock_usage_bit bit, const char *irqclass)
2029 {
2030         int ret;
2031         struct lock_list root;
2032         struct lock_list *uninitialized_var(target_entry);
2033
2034         root.parent = NULL;
2035         root.class = hlock_class(this);
2036         ret = find_usage_forwards(&root, bit, &target_entry);
2037         BFS_PROCESS_RET(ret);
2038
2039         return print_irq_inversion_bug(curr, &root, target_entry,
2040                                         this, 1, irqclass);
2041 }
2042
2043 /*
2044  * Prove that in the backwards-direction subgraph starting at <this>
2045  * there is no lock matching <mask>:
2046  */
2047 static int
2048 check_usage_backwards(struct task_struct *curr, struct held_lock *this,
2049                       enum lock_usage_bit bit, const char *irqclass)
2050 {
2051         int ret;
2052         struct lock_list root;
2053         struct lock_list *uninitialized_var(target_entry);
2054
2055         root.parent = NULL;
2056         root.class = hlock_class(this);
2057         ret = find_usage_backwards(&root, bit, &target_entry);
2058         BFS_PROCESS_RET(ret);
2059
2060         return print_irq_inversion_bug(curr, &root, target_entry,
2061                                         this, 1, irqclass);
2062 }
2063
2064 void print_irqtrace_events(struct task_struct *curr)
2065 {
2066         printk("irq event stamp: %u\n", curr->irq_events);
2067         printk("hardirqs last  enabled at (%u): ", curr->hardirq_enable_event);
2068         print_ip_sym(curr->hardirq_enable_ip);
2069         printk("hardirqs last disabled at (%u): ", curr->hardirq_disable_event);
2070         print_ip_sym(curr->hardirq_disable_ip);
2071         printk("softirqs last  enabled at (%u): ", curr->softirq_enable_event);
2072         print_ip_sym(curr->softirq_enable_ip);
2073         printk("softirqs last disabled at (%u): ", curr->softirq_disable_event);
2074         print_ip_sym(curr->softirq_disable_ip);
2075 }
2076
2077 static int HARDIRQ_verbose(struct lock_class *class)
2078 {
2079 #if HARDIRQ_VERBOSE
2080         return class_filter(class);
2081 #endif
2082         return 0;
2083 }
2084
2085 static int SOFTIRQ_verbose(struct lock_class *class)
2086 {
2087 #if SOFTIRQ_VERBOSE
2088         return class_filter(class);
2089 #endif
2090         return 0;
2091 }
2092
2093 static int RECLAIM_FS_verbose(struct lock_class *class)
2094 {
2095 #if RECLAIM_VERBOSE
2096         return class_filter(class);
2097 #endif
2098         return 0;
2099 }
2100
2101 #define STRICT_READ_CHECKS      1
2102
2103 static int (*state_verbose_f[])(struct lock_class *class) = {
2104 #define LOCKDEP_STATE(__STATE) \
2105         __STATE##_verbose,
2106 #include "lockdep_states.h"
2107 #undef LOCKDEP_STATE
2108 };
2109
2110 static inline int state_verbose(enum lock_usage_bit bit,
2111                                 struct lock_class *class)
2112 {
2113         return state_verbose_f[bit >> 2](class);
2114 }
2115
2116 typedef int (*check_usage_f)(struct task_struct *, struct held_lock *,
2117                              enum lock_usage_bit bit, const char *name);
2118
2119 static int
2120 mark_lock_irq(struct task_struct *curr, struct held_lock *this,
2121                 enum lock_usage_bit new_bit)
2122 {
2123         int excl_bit = exclusive_bit(new_bit);
2124         int read = new_bit & 1;
2125         int dir = new_bit & 2;
2126
2127         /*
2128          * mark USED_IN has to look forwards -- to ensure no dependency
2129          * has ENABLED state, which would allow recursion deadlocks.
2130          *
2131          * mark ENABLED has to look backwards -- to ensure no dependee
2132          * has USED_IN state, which, again, would allow  recursion deadlocks.
2133          */
2134         check_usage_f usage = dir ?
2135                 check_usage_backwards : check_usage_forwards;
2136
2137         /*
2138          * Validate that this particular lock does not have conflicting
2139          * usage states.
2140          */
2141         if (!valid_state(curr, this, new_bit, excl_bit))
2142                 return 0;
2143
2144         /*
2145          * Validate that the lock dependencies don't have conflicting usage
2146          * states.
2147          */
2148         if ((!read || !dir || STRICT_READ_CHECKS) &&
2149                         !usage(curr, this, excl_bit, state_name(new_bit & ~1)))
2150                 return 0;
2151
2152         /*
2153          * Check for read in write conflicts
2154          */
2155         if (!read) {
2156                 if (!valid_state(curr, this, new_bit, excl_bit + 1))
2157                         return 0;
2158
2159                 if (STRICT_READ_CHECKS &&
2160                         !usage(curr, this, excl_bit + 1,
2161                                 state_name(new_bit + 1)))
2162                         return 0;
2163         }
2164
2165         if (state_verbose(new_bit, hlock_class(this)))
2166                 return 2;
2167
2168         return 1;
2169 }
2170
2171 enum mark_type {
2172 #define LOCKDEP_STATE(__STATE)  __STATE,
2173 #include "lockdep_states.h"
2174 #undef LOCKDEP_STATE
2175 };
2176
2177 /*
2178  * Mark all held locks with a usage bit:
2179  */
2180 static int
2181 mark_held_locks(struct task_struct *curr, enum mark_type mark)
2182 {
2183         enum lock_usage_bit usage_bit;
2184         struct held_lock *hlock;
2185         int i;
2186
2187         for (i = 0; i < curr->lockdep_depth; i++) {
2188                 hlock = curr->held_locks + i;
2189
2190                 usage_bit = 2 + (mark << 2); /* ENABLED */
2191                 if (hlock->read)
2192                         usage_bit += 1; /* READ */
2193
2194                 BUG_ON(usage_bit >= LOCK_USAGE_STATES);
2195
2196                 if (!mark_lock(curr, hlock, usage_bit))
2197                         return 0;
2198         }
2199
2200         return 1;
2201 }
2202
2203 /*
2204  * Debugging helper: via this flag we know that we are in
2205  * 'early bootup code', and will warn about any invalid irqs-on event:
2206  */
2207 static int early_boot_irqs_enabled;
2208
2209 void early_boot_irqs_off(void)
2210 {
2211         early_boot_irqs_enabled = 0;
2212 }
2213
2214 void early_boot_irqs_on(void)
2215 {
2216         early_boot_irqs_enabled = 1;
2217 }
2218
2219 /*
2220  * Hardirqs will be enabled:
2221  */
2222 void trace_hardirqs_on_caller(unsigned long ip)
2223 {
2224         struct task_struct *curr = current;
2225
2226         time_hardirqs_on(CALLER_ADDR0, ip);
2227
2228         if (unlikely(!debug_locks || current->lockdep_recursion))
2229                 return;
2230
2231         if (DEBUG_LOCKS_WARN_ON(unlikely(!early_boot_irqs_enabled)))
2232                 return;
2233
2234         if (unlikely(curr->hardirqs_enabled)) {
2235                 debug_atomic_inc(&redundant_hardirqs_on);
2236                 return;
2237         }
2238         /* we'll do an OFF -> ON transition: */
2239         curr->hardirqs_enabled = 1;
2240
2241         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2242                 return;
2243         if (DEBUG_LOCKS_WARN_ON(current->hardirq_context))
2244                 return;
2245         /*
2246          * We are going to turn hardirqs on, so set the
2247          * usage bit for all held locks:
2248          */
2249         if (!mark_held_locks(curr, HARDIRQ))
2250                 return;
2251         /*
2252          * If we have softirqs enabled, then set the usage
2253          * bit for all held locks. (disabled hardirqs prevented
2254          * this bit from being set before)
2255          */
2256         if (curr->softirqs_enabled)
2257                 if (!mark_held_locks(curr, SOFTIRQ))
2258                         return;
2259
2260         curr->hardirq_enable_ip = ip;
2261         curr->hardirq_enable_event = ++curr->irq_events;
2262         debug_atomic_inc(&hardirqs_on_events);
2263 }
2264 EXPORT_SYMBOL(trace_hardirqs_on_caller);
2265
2266 void trace_hardirqs_on(void)
2267 {
2268         trace_hardirqs_on_caller(CALLER_ADDR0);
2269 }
2270 EXPORT_SYMBOL(trace_hardirqs_on);
2271
2272 /*
2273  * Hardirqs were disabled:
2274  */
2275 void trace_hardirqs_off_caller(unsigned long ip)
2276 {
2277         struct task_struct *curr = current;
2278
2279         time_hardirqs_off(CALLER_ADDR0, ip);
2280
2281         if (unlikely(!debug_locks || current->lockdep_recursion))
2282                 return;
2283
2284         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2285                 return;
2286
2287         if (curr->hardirqs_enabled) {
2288                 /*
2289                  * We have done an ON -> OFF transition:
2290                  */
2291                 curr->hardirqs_enabled = 0;
2292                 curr->hardirq_disable_ip = ip;
2293                 curr->hardirq_disable_event = ++curr->irq_events;
2294                 debug_atomic_inc(&hardirqs_off_events);
2295         } else
2296                 debug_atomic_inc(&redundant_hardirqs_off);
2297 }
2298 EXPORT_SYMBOL(trace_hardirqs_off_caller);
2299
2300 void trace_hardirqs_off(void)
2301 {
2302         trace_hardirqs_off_caller(CALLER_ADDR0);
2303 }
2304 EXPORT_SYMBOL(trace_hardirqs_off);
2305
2306 /*
2307  * Softirqs will be enabled:
2308  */
2309 void trace_softirqs_on(unsigned long ip)
2310 {
2311         struct task_struct *curr = current;
2312
2313         if (unlikely(!debug_locks))
2314                 return;
2315
2316         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2317                 return;
2318
2319         if (curr->softirqs_enabled) {
2320                 debug_atomic_inc(&redundant_softirqs_on);
2321                 return;
2322         }
2323
2324         /*
2325          * We'll do an OFF -> ON transition:
2326          */
2327         curr->softirqs_enabled = 1;
2328         curr->softirq_enable_ip = ip;
2329         curr->softirq_enable_event = ++curr->irq_events;
2330         debug_atomic_inc(&softirqs_on_events);
2331         /*
2332          * We are going to turn softirqs on, so set the
2333          * usage bit for all held locks, if hardirqs are
2334          * enabled too:
2335          */
2336         if (curr->hardirqs_enabled)
2337                 mark_held_locks(curr, SOFTIRQ);
2338 }
2339
2340 /*
2341  * Softirqs were disabled:
2342  */
2343 void trace_softirqs_off(unsigned long ip)
2344 {
2345         struct task_struct *curr = current;
2346
2347         if (unlikely(!debug_locks))
2348                 return;
2349
2350         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2351                 return;
2352
2353         if (curr->softirqs_enabled) {
2354                 /*
2355                  * We have done an ON -> OFF transition:
2356                  */
2357                 curr->softirqs_enabled = 0;
2358                 curr->softirq_disable_ip = ip;
2359                 curr->softirq_disable_event = ++curr->irq_events;
2360                 debug_atomic_inc(&softirqs_off_events);
2361                 DEBUG_LOCKS_WARN_ON(!softirq_count());
2362         } else
2363                 debug_atomic_inc(&redundant_softirqs_off);
2364 }
2365
2366 static void __lockdep_trace_alloc(gfp_t gfp_mask, unsigned long flags)
2367 {
2368         struct task_struct *curr = current;
2369
2370         if (unlikely(!debug_locks))
2371                 return;
2372
2373         /* no reclaim without waiting on it */
2374         if (!(gfp_mask & __GFP_WAIT))
2375                 return;
2376
2377         /* this guy won't enter reclaim */
2378         if ((curr->flags & PF_MEMALLOC) && !(gfp_mask & __GFP_NOMEMALLOC))
2379                 return;
2380
2381         /* We're only interested __GFP_FS allocations for now */
2382         if (!(gfp_mask & __GFP_FS))
2383                 return;
2384
2385         if (DEBUG_LOCKS_WARN_ON(irqs_disabled_flags(flags)))
2386                 return;
2387
2388         mark_held_locks(curr, RECLAIM_FS);
2389 }
2390
2391 static void check_flags(unsigned long flags);
2392
2393 void lockdep_trace_alloc(gfp_t gfp_mask)
2394 {
2395         unsigned long flags;
2396
2397         if (unlikely(current->lockdep_recursion))
2398                 return;
2399
2400         raw_local_irq_save(flags);
2401         check_flags(flags);
2402         current->lockdep_recursion = 1;
2403         __lockdep_trace_alloc(gfp_mask, flags);
2404         current->lockdep_recursion = 0;
2405         raw_local_irq_restore(flags);
2406 }
2407
2408 static int mark_irqflags(struct task_struct *curr, struct held_lock *hlock)
2409 {
2410         /*
2411          * If non-trylock use in a hardirq or softirq context, then
2412          * mark the lock as used in these contexts:
2413          */
2414         if (!hlock->trylock) {
2415                 if (hlock->read) {
2416                         if (curr->hardirq_context)
2417                                 if (!mark_lock(curr, hlock,
2418                                                 LOCK_USED_IN_HARDIRQ_READ))
2419                                         return 0;
2420                         if (curr->softirq_context)
2421                                 if (!mark_lock(curr, hlock,
2422                                                 LOCK_USED_IN_SOFTIRQ_READ))
2423                                         return 0;
2424                 } else {
2425                         if (curr->hardirq_context)
2426                                 if (!mark_lock(curr, hlock, LOCK_USED_IN_HARDIRQ))
2427                                         return 0;
2428                         if (curr->softirq_context)
2429                                 if (!mark_lock(curr, hlock, LOCK_USED_IN_SOFTIRQ))
2430                                         return 0;
2431                 }
2432         }
2433         if (!hlock->hardirqs_off) {
2434                 if (hlock->read) {
2435                         if (!mark_lock(curr, hlock,
2436                                         LOCK_ENABLED_HARDIRQ_READ))
2437                                 return 0;
2438                         if (curr->softirqs_enabled)
2439                                 if (!mark_lock(curr, hlock,
2440                                                 LOCK_ENABLED_SOFTIRQ_READ))
2441                                         return 0;
2442                 } else {
2443                         if (!mark_lock(curr, hlock,
2444                                         LOCK_ENABLED_HARDIRQ))
2445                                 return 0;
2446                         if (curr->softirqs_enabled)
2447                                 if (!mark_lock(curr, hlock,
2448                                                 LOCK_ENABLED_SOFTIRQ))
2449                                         return 0;
2450                 }
2451         }
2452
2453         /*
2454          * We reuse the irq context infrastructure more broadly as a general
2455          * context checking code. This tests GFP_FS recursion (a lock taken
2456          * during reclaim for a GFP_FS allocation is held over a GFP_FS
2457          * allocation).
2458          */
2459         if (!hlock->trylock && (curr->lockdep_reclaim_gfp & __GFP_FS)) {
2460                 if (hlock->read) {
2461                         if (!mark_lock(curr, hlock, LOCK_USED_IN_RECLAIM_FS_READ))
2462                                         return 0;
2463                 } else {
2464                         if (!mark_lock(curr, hlock, LOCK_USED_IN_RECLAIM_FS))
2465                                         return 0;
2466                 }
2467         }
2468
2469         return 1;
2470 }
2471
2472 static int separate_irq_context(struct task_struct *curr,
2473                 struct held_lock *hlock)
2474 {
2475         unsigned int depth = curr->lockdep_depth;
2476
2477         /*
2478          * Keep track of points where we cross into an interrupt context:
2479          */
2480         hlock->irq_context = 2*(curr->hardirq_context ? 1 : 0) +
2481                                 curr->softirq_context;
2482         if (depth) {
2483                 struct held_lock *prev_hlock;
2484
2485                 prev_hlock = curr->held_locks + depth-1;
2486                 /*
2487                  * If we cross into another context, reset the
2488                  * hash key (this also prevents the checking and the
2489                  * adding of the dependency to 'prev'):
2490                  */
2491                 if (prev_hlock->irq_context != hlock->irq_context)
2492                         return 1;
2493         }
2494         return 0;
2495 }
2496
2497 #else
2498
2499 static inline
2500 int mark_lock_irq(struct task_struct *curr, struct held_lock *this,
2501                 enum lock_usage_bit new_bit)
2502 {
2503         WARN_ON(1);
2504         return 1;
2505 }
2506
2507 static inline int mark_irqflags(struct task_struct *curr,
2508                 struct held_lock *hlock)
2509 {
2510         return 1;
2511 }
2512
2513 static inline int separate_irq_context(struct task_struct *curr,
2514                 struct held_lock *hlock)
2515 {
2516         return 0;
2517 }
2518
2519 void lockdep_trace_alloc(gfp_t gfp_mask)
2520 {
2521 }
2522
2523 #endif
2524
2525 /*
2526  * Mark a lock with a usage bit, and validate the state transition:
2527  */
2528 static int mark_lock(struct task_struct *curr, struct held_lock *this,
2529                              enum lock_usage_bit new_bit)
2530 {
2531         unsigned int new_mask = 1 << new_bit, ret = 1;
2532
2533         /*
2534          * If already set then do not dirty the cacheline,
2535          * nor do any checks:
2536          */
2537         if (likely(hlock_class(this)->usage_mask & new_mask))
2538                 return 1;
2539
2540         if (!graph_lock())
2541                 return 0;
2542         /*
2543          * Make sure we didnt race:
2544          */
2545         if (unlikely(hlock_class(this)->usage_mask & new_mask)) {
2546                 graph_unlock();
2547                 return 1;
2548         }
2549
2550         hlock_class(this)->usage_mask |= new_mask;
2551
2552         if (!save_trace(hlock_class(this)->usage_traces + new_bit))
2553                 return 0;
2554
2555         switch (new_bit) {
2556 #define LOCKDEP_STATE(__STATE)                  \
2557         case LOCK_USED_IN_##__STATE:            \
2558         case LOCK_USED_IN_##__STATE##_READ:     \
2559         case LOCK_ENABLED_##__STATE:            \
2560         case LOCK_ENABLED_##__STATE##_READ:
2561 #include "lockdep_states.h"
2562 #undef LOCKDEP_STATE
2563                 ret = mark_lock_irq(curr, this, new_bit);
2564                 if (!ret)
2565                         return 0;
2566                 break;
2567         case LOCK_USED:
2568                 debug_atomic_dec(&nr_unused_locks);
2569                 break;
2570         default:
2571                 if (!debug_locks_off_graph_unlock())
2572                         return 0;
2573                 WARN_ON(1);
2574                 return 0;
2575         }
2576
2577         graph_unlock();
2578
2579         /*
2580          * We must printk outside of the graph_lock:
2581          */
2582         if (ret == 2) {
2583                 printk("\nmarked lock as {%s}:\n", usage_str[new_bit]);
2584                 print_lock(this);
2585                 print_irqtrace_events(curr);
2586                 dump_stack();
2587         }
2588
2589         return ret;
2590 }
2591
2592 /*
2593  * Initialize a lock instance's lock-class mapping info:
2594  */
2595 void lockdep_init_map(struct lockdep_map *lock, const char *name,
2596                       struct lock_class_key *key, int subclass)
2597 {
2598         lock->class_cache = NULL;
2599 #ifdef CONFIG_LOCK_STAT
2600         lock->cpu = raw_smp_processor_id();
2601 #endif
2602
2603         if (DEBUG_LOCKS_WARN_ON(!name)) {
2604                 lock->name = "NULL";
2605                 return;
2606         }
2607
2608         lock->name = name;
2609
2610         if (DEBUG_LOCKS_WARN_ON(!key))
2611                 return;
2612         /*
2613          * Sanity check, the lock-class key must be persistent:
2614          */
2615         if (!static_obj(key)) {
2616                 printk("BUG: key %p not in .data!\n", key);
2617                 DEBUG_LOCKS_WARN_ON(1);
2618                 return;
2619         }
2620         lock->key = key;
2621
2622         if (unlikely(!debug_locks))
2623                 return;
2624
2625         if (subclass)
2626                 register_lock_class(lock, subclass, 1);
2627 }
2628 EXPORT_SYMBOL_GPL(lockdep_init_map);
2629
2630 /*
2631  * This gets called for every mutex_lock*()/spin_lock*() operation.
2632  * We maintain the dependency maps and validate the locking attempt:
2633  */
2634 static int __lock_acquire(struct lockdep_map *lock, unsigned int subclass,
2635                           int trylock, int read, int check, int hardirqs_off,
2636                           struct lockdep_map *nest_lock, unsigned long ip)
2637 {
2638         struct task_struct *curr = current;
2639         struct lock_class *class = NULL;
2640         struct held_lock *hlock;
2641         unsigned int depth, id;
2642         int chain_head = 0;
2643         u64 chain_key;
2644
2645         if (!prove_locking)
2646                 check = 1;
2647
2648         if (unlikely(!debug_locks))
2649                 return 0;
2650
2651         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2652                 return 0;
2653
2654         if (unlikely(subclass >= MAX_LOCKDEP_SUBCLASSES)) {
2655                 debug_locks_off();
2656                 printk("BUG: MAX_LOCKDEP_SUBCLASSES too low!\n");
2657                 printk("turning off the locking correctness validator.\n");
2658                 dump_stack();
2659                 return 0;
2660         }
2661
2662         if (!subclass)
2663                 class = lock->class_cache;
2664         /*
2665          * Not cached yet or subclass?
2666          */
2667         if (unlikely(!class)) {
2668                 class = register_lock_class(lock, subclass, 0);
2669                 if (!class)
2670                         return 0;
2671         }
2672         debug_atomic_inc((atomic_t *)&class->ops);
2673         if (very_verbose(class)) {
2674                 printk("\nacquire class [%p] %s", class->key, class->name);
2675                 if (class->name_version > 1)
2676                         printk("#%d", class->name_version);
2677                 printk("\n");
2678                 dump_stack();
2679         }
2680
2681         /*
2682          * Add the lock to the list of currently held locks.
2683          * (we dont increase the depth just yet, up until the
2684          * dependency checks are done)
2685          */
2686         depth = curr->lockdep_depth;
2687         if (DEBUG_LOCKS_WARN_ON(depth >= MAX_LOCK_DEPTH))
2688                 return 0;
2689
2690         hlock = curr->held_locks + depth;
2691         if (DEBUG_LOCKS_WARN_ON(!class))
2692                 return 0;
2693         hlock->class_idx = class - lock_classes + 1;
2694         hlock->acquire_ip = ip;
2695         hlock->instance = lock;
2696         hlock->nest_lock = nest_lock;
2697         hlock->trylock = trylock;
2698         hlock->read = read;
2699         hlock->check = check;
2700         hlock->hardirqs_off = !!hardirqs_off;
2701 #ifdef CONFIG_LOCK_STAT
2702         hlock->waittime_stamp = 0;
2703         hlock->holdtime_stamp = sched_clock();
2704 #endif
2705
2706         if (check == 2 && !mark_irqflags(curr, hlock))
2707                 return 0;
2708
2709         /* mark it as used: */
2710         if (!mark_lock(curr, hlock, LOCK_USED))
2711                 return 0;
2712
2713         /*
2714          * Calculate the chain hash: it's the combined hash of all the
2715          * lock keys along the dependency chain. We save the hash value
2716          * at every step so that we can get the current hash easily
2717          * after unlock. The chain hash is then used to cache dependency
2718          * results.
2719          *
2720          * The 'key ID' is what is the most compact key value to drive
2721          * the hash, not class->key.
2722          */
2723         id = class - lock_classes;
2724         if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
2725                 return 0;
2726
2727         chain_key = curr->curr_chain_key;
2728         if (!depth) {
2729                 if (DEBUG_LOCKS_WARN_ON(chain_key != 0))
2730                         return 0;
2731                 chain_head = 1;
2732         }
2733
2734         hlock->prev_chain_key = chain_key;
2735         if (separate_irq_context(curr, hlock)) {
2736                 chain_key = 0;
2737                 chain_head = 1;
2738         }
2739         chain_key = iterate_chain_key(chain_key, id);
2740
2741         if (!validate_chain(curr, lock, hlock, chain_head, chain_key))
2742                 return 0;
2743
2744         curr->curr_chain_key = chain_key;
2745         curr->lockdep_depth++;
2746         check_chain_key(curr);
2747 #ifdef CONFIG_DEBUG_LOCKDEP
2748         if (unlikely(!debug_locks))
2749                 return 0;
2750 #endif
2751         if (unlikely(curr->lockdep_depth >= MAX_LOCK_DEPTH)) {
2752                 debug_locks_off();
2753                 printk("BUG: MAX_LOCK_DEPTH too low!\n");
2754                 printk("turning off the locking correctness validator.\n");
2755                 dump_stack();
2756                 return 0;
2757         }
2758
2759         if (unlikely(curr->lockdep_depth > max_lockdep_depth))
2760                 max_lockdep_depth = curr->lockdep_depth;
2761
2762         return 1;
2763 }
2764
2765 static int
2766 print_unlock_inbalance_bug(struct task_struct *curr, struct lockdep_map *lock,
2767                            unsigned long ip)
2768 {
2769         if (!debug_locks_off())
2770                 return 0;
2771         if (debug_locks_silent)
2772                 return 0;
2773
2774         printk("\n=====================================\n");
2775         printk(  "[ BUG: bad unlock balance detected! ]\n");
2776         printk(  "-------------------------------------\n");
2777         printk("%s/%d is trying to release lock (",
2778                 curr->comm, task_pid_nr(curr));
2779         print_lockdep_cache(lock);
2780         printk(") at:\n");
2781         print_ip_sym(ip);
2782         printk("but there are no more locks to release!\n");
2783         printk("\nother info that might help us debug this:\n");
2784         lockdep_print_held_locks(curr);
2785
2786         printk("\nstack backtrace:\n");
2787         dump_stack();
2788
2789         return 0;
2790 }
2791
2792 /*
2793  * Common debugging checks for both nested and non-nested unlock:
2794  */
2795 static int check_unlock(struct task_struct *curr, struct lockdep_map *lock,
2796                         unsigned long ip)
2797 {
2798         if (unlikely(!debug_locks))
2799                 return 0;
2800         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2801                 return 0;
2802
2803         if (curr->lockdep_depth <= 0)
2804                 return print_unlock_inbalance_bug(curr, lock, ip);
2805
2806         return 1;
2807 }
2808
2809 static int
2810 __lock_set_class(struct lockdep_map *lock, const char *name,
2811                  struct lock_class_key *key, unsigned int subclass,
2812                  unsigned long ip)
2813 {
2814         struct task_struct *curr = current;
2815         struct held_lock *hlock, *prev_hlock;
2816         struct lock_class *class;
2817         unsigned int depth;
2818         int i;
2819
2820         depth = curr->lockdep_depth;
2821         if (DEBUG_LOCKS_WARN_ON(!depth))
2822                 return 0;
2823
2824         prev_hlock = NULL;
2825         for (i = depth-1; i >= 0; i--) {
2826                 hlock = curr->held_locks + i;
2827                 /*
2828                  * We must not cross into another context:
2829                  */
2830                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
2831                         break;
2832                 if (hlock->instance == lock)
2833                         goto found_it;
2834                 prev_hlock = hlock;
2835         }
2836         return print_unlock_inbalance_bug(curr, lock, ip);
2837
2838 found_it:
2839         lockdep_init_map(lock, name, key, 0);
2840         class = register_lock_class(lock, subclass, 0);
2841         hlock->class_idx = class - lock_classes + 1;
2842
2843         curr->lockdep_depth = i;
2844         curr->curr_chain_key = hlock->prev_chain_key;
2845
2846         for (; i < depth; i++) {
2847                 hlock = curr->held_locks + i;
2848                 if (!__lock_acquire(hlock->instance,
2849                         hlock_class(hlock)->subclass, hlock->trylock,
2850                                 hlock->read, hlock->check, hlock->hardirqs_off,
2851                                 hlock->nest_lock, hlock->acquire_ip))
2852                         return 0;
2853         }
2854
2855         if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth))
2856                 return 0;
2857         return 1;
2858 }
2859
2860 /*
2861  * Remove the lock to the list of currently held locks in a
2862  * potentially non-nested (out of order) manner. This is a
2863  * relatively rare operation, as all the unlock APIs default
2864  * to nested mode (which uses lock_release()):
2865  */
2866 static int
2867 lock_release_non_nested(struct task_struct *curr,
2868                         struct lockdep_map *lock, unsigned long ip)
2869 {
2870         struct held_lock *hlock, *prev_hlock;
2871         unsigned int depth;
2872         int i;
2873
2874         /*
2875          * Check whether the lock exists in the current stack
2876          * of held locks:
2877          */
2878         depth = curr->lockdep_depth;
2879         if (DEBUG_LOCKS_WARN_ON(!depth))
2880                 return 0;
2881
2882         prev_hlock = NULL;
2883         for (i = depth-1; i >= 0; i--) {
2884                 hlock = curr->held_locks + i;
2885                 /*
2886                  * We must not cross into another context:
2887                  */
2888                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
2889                         break;
2890                 if (hlock->instance == lock)
2891                         goto found_it;
2892                 prev_hlock = hlock;
2893         }
2894         return print_unlock_inbalance_bug(curr, lock, ip);
2895
2896 found_it:
2897         lock_release_holdtime(hlock);
2898
2899         /*
2900          * We have the right lock to unlock, 'hlock' points to it.
2901          * Now we remove it from the stack, and add back the other
2902          * entries (if any), recalculating the hash along the way:
2903          */
2904         curr->lockdep_depth = i;
2905         curr->curr_chain_key = hlock->prev_chain_key;
2906
2907         for (i++; i < depth; i++) {
2908                 hlock = curr->held_locks + i;
2909                 if (!__lock_acquire(hlock->instance,
2910                         hlock_class(hlock)->subclass, hlock->trylock,
2911                                 hlock->read, hlock->check, hlock->hardirqs_off,
2912                                 hlock->nest_lock, hlock->acquire_ip))
2913                         return 0;
2914         }
2915
2916         if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth - 1))
2917                 return 0;
2918         return 1;
2919 }
2920
2921 /*
2922  * Remove the lock to the list of currently held locks - this gets
2923  * called on mutex_unlock()/spin_unlock*() (or on a failed
2924  * mutex_lock_interruptible()). This is done for unlocks that nest
2925  * perfectly. (i.e. the current top of the lock-stack is unlocked)
2926  */
2927 static int lock_release_nested(struct task_struct *curr,
2928                                struct lockdep_map *lock, unsigned long ip)
2929 {
2930         struct held_lock *hlock;
2931         unsigned int depth;
2932
2933         /*
2934          * Pop off the top of the lock stack:
2935          */
2936         depth = curr->lockdep_depth - 1;
2937         hlock = curr->held_locks + depth;
2938
2939         /*
2940          * Is the unlock non-nested:
2941          */
2942         if (hlock->instance != lock)
2943                 return lock_release_non_nested(curr, lock, ip);
2944         curr->lockdep_depth--;
2945
2946         if (DEBUG_LOCKS_WARN_ON(!depth && (hlock->prev_chain_key != 0)))
2947                 return 0;
2948
2949         curr->curr_chain_key = hlock->prev_chain_key;
2950
2951         lock_release_holdtime(hlock);
2952
2953 #ifdef CONFIG_DEBUG_LOCKDEP
2954         hlock->prev_chain_key = 0;
2955         hlock->class_idx = 0;
2956         hlock->acquire_ip = 0;
2957         hlock->irq_context = 0;
2958 #endif
2959         return 1;
2960 }
2961
2962 /*
2963  * Remove the lock to the list of currently held locks - this gets
2964  * called on mutex_unlock()/spin_unlock*() (or on a failed
2965  * mutex_lock_interruptible()). This is done for unlocks that nest
2966  * perfectly. (i.e. the current top of the lock-stack is unlocked)
2967  */
2968 static void
2969 __lock_release(struct lockdep_map *lock, int nested, unsigned long ip)
2970 {
2971         struct task_struct *curr = current;
2972
2973         if (!check_unlock(curr, lock, ip))
2974                 return;
2975
2976         if (nested) {
2977                 if (!lock_release_nested(curr, lock, ip))
2978                         return;
2979         } else {
2980                 if (!lock_release_non_nested(curr, lock, ip))
2981                         return;
2982         }
2983
2984         check_chain_key(curr);
2985 }
2986
2987 /*
2988  * Check whether we follow the irq-flags state precisely:
2989  */
2990 static void check_flags(unsigned long flags)
2991 {
2992 #if defined(CONFIG_PROVE_LOCKING) && defined(CONFIG_DEBUG_LOCKDEP) && \
2993     defined(CONFIG_TRACE_IRQFLAGS)
2994         if (!debug_locks)
2995                 return;
2996
2997         if (irqs_disabled_flags(flags)) {
2998                 if (DEBUG_LOCKS_WARN_ON(current->hardirqs_enabled)) {
2999                         printk("possible reason: unannotated irqs-off.\n");
3000                 }
3001         } else {
3002                 if (DEBUG_LOCKS_WARN_ON(!current->hardirqs_enabled)) {
3003                         printk("possible reason: unannotated irqs-on.\n");
3004                 }
3005         }
3006
3007         /*
3008          * We dont accurately track softirq state in e.g.
3009          * hardirq contexts (such as on 4KSTACKS), so only
3010          * check if not in hardirq contexts:
3011          */
3012         if (!hardirq_count()) {
3013                 if (softirq_count())
3014                         DEBUG_LOCKS_WARN_ON(current->softirqs_enabled);
3015                 else
3016                         DEBUG_LOCKS_WARN_ON(!current->softirqs_enabled);
3017         }
3018
3019         if (!debug_locks)
3020                 print_irqtrace_events(current);
3021 #endif
3022 }
3023
3024 void lock_set_class(struct lockdep_map *lock, const char *name,
3025                     struct lock_class_key *key, unsigned int subclass,
3026                     unsigned long ip)
3027 {
3028         unsigned long flags;
3029
3030         if (unlikely(current->lockdep_recursion))
3031                 return;
3032
3033         raw_local_irq_save(flags);
3034         current->lockdep_recursion = 1;
3035         check_flags(flags);
3036         if (__lock_set_class(lock, name, key, subclass, ip))
3037                 check_chain_key(current);
3038         current->lockdep_recursion = 0;
3039         raw_local_irq_restore(flags);
3040 }
3041 EXPORT_SYMBOL_GPL(lock_set_class);
3042
3043 /*
3044  * We are not always called with irqs disabled - do that here,
3045  * and also avoid lockdep recursion:
3046  */
3047 void lock_acquire(struct lockdep_map *lock, unsigned int subclass,
3048                           int trylock, int read, int check,
3049                           struct lockdep_map *nest_lock, unsigned long ip)
3050 {
3051         unsigned long flags;
3052
3053         trace_lock_acquire(lock, subclass, trylock, read, check, nest_lock, ip);
3054
3055         if (unlikely(current->lockdep_recursion))
3056                 return;
3057
3058         raw_local_irq_save(flags);
3059         check_flags(flags);
3060
3061         current->lockdep_recursion = 1;
3062         __lock_acquire(lock, subclass, trylock, read, check,
3063                        irqs_disabled_flags(flags), nest_lock, ip);
3064         current->lockdep_recursion = 0;
3065         raw_local_irq_restore(flags);
3066 }
3067 EXPORT_SYMBOL_GPL(lock_acquire);
3068
3069 void lock_release(struct lockdep_map *lock, int nested,
3070                           unsigned long ip)
3071 {
3072         unsigned long flags;
3073
3074         trace_lock_release(lock, nested, ip);
3075
3076         if (unlikely(current->lockdep_recursion))
3077                 return;
3078
3079         raw_local_irq_save(flags);
3080         check_flags(flags);
3081         current->lockdep_recursion = 1;
3082         __lock_release(lock, nested, ip);
3083         current->lockdep_recursion = 0;
3084         raw_local_irq_restore(flags);
3085 }
3086 EXPORT_SYMBOL_GPL(lock_release);
3087
3088 void lockdep_set_current_reclaim_state(gfp_t gfp_mask)
3089 {
3090         current->lockdep_reclaim_gfp = gfp_mask;
3091 }
3092
3093 void lockdep_clear_current_reclaim_state(void)
3094 {
3095         current->lockdep_reclaim_gfp = 0;
3096 }
3097
3098 #ifdef CONFIG_LOCK_STAT
3099 static int
3100 print_lock_contention_bug(struct task_struct *curr, struct lockdep_map *lock,
3101                            unsigned long ip)
3102 {
3103         if (!debug_locks_off())
3104                 return 0;
3105         if (debug_locks_silent)
3106                 return 0;
3107
3108         printk("\n=================================\n");
3109         printk(  "[ BUG: bad contention detected! ]\n");
3110         printk(  "---------------------------------\n");
3111         printk("%s/%d is trying to contend lock (",
3112                 curr->comm, task_pid_nr(curr));
3113         print_lockdep_cache(lock);
3114         printk(") at:\n");
3115         print_ip_sym(ip);
3116         printk("but there are no locks held!\n");
3117         printk("\nother info that might help us debug this:\n");
3118         lockdep_print_held_locks(curr);
3119
3120         printk("\nstack backtrace:\n");
3121         dump_stack();
3122
3123         return 0;
3124 }
3125
3126 static void
3127 __lock_contended(struct lockdep_map *lock, unsigned long ip)
3128 {
3129         struct task_struct *curr = current;
3130         struct held_lock *hlock, *prev_hlock;
3131         struct lock_class_stats *stats;
3132         unsigned int depth;
3133         int i, contention_point, contending_point;
3134
3135         depth = curr->lockdep_depth;
3136         if (DEBUG_LOCKS_WARN_ON(!depth))
3137                 return;
3138
3139         prev_hlock = NULL;
3140         for (i = depth-1; i >= 0; i--) {
3141                 hlock = curr->held_locks + i;
3142                 /*
3143                  * We must not cross into another context:
3144                  */
3145                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3146                         break;
3147                 if (hlock->instance == lock)
3148                         goto found_it;
3149                 prev_hlock = hlock;
3150         }
3151         print_lock_contention_bug(curr, lock, ip);
3152         return;
3153
3154 found_it:
3155         hlock->waittime_stamp = sched_clock();
3156
3157         contention_point = lock_point(hlock_class(hlock)->contention_point, ip);
3158         contending_point = lock_point(hlock_class(hlock)->contending_point,
3159                                       lock->ip);
3160
3161         stats = get_lock_stats(hlock_class(hlock));
3162         if (contention_point < LOCKSTAT_POINTS)
3163                 stats->contention_point[contention_point]++;
3164         if (contending_point < LOCKSTAT_POINTS)
3165                 stats->contending_point[contending_point]++;
3166         if (lock->cpu != smp_processor_id())
3167                 stats->bounces[bounce_contended + !!hlock->read]++;
3168         put_lock_stats(stats);
3169 }
3170
3171 static void
3172 __lock_acquired(struct lockdep_map *lock, unsigned long ip)
3173 {
3174         struct task_struct *curr = current;
3175         struct held_lock *hlock, *prev_hlock;
3176         struct lock_class_stats *stats;
3177         unsigned int depth;
3178         u64 now;
3179         s64 waittime = 0;
3180         int i, cpu;
3181
3182         depth = curr->lockdep_depth;
3183         if (DEBUG_LOCKS_WARN_ON(!depth))
3184                 return;
3185
3186         prev_hlock = NULL;
3187         for (i = depth-1; i >= 0; i--) {
3188                 hlock = curr->held_locks + i;
3189                 /*
3190                  * We must not cross into another context:
3191                  */
3192                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3193                         break;
3194                 if (hlock->instance == lock)
3195                         goto found_it;
3196                 prev_hlock = hlock;
3197         }
3198         print_lock_contention_bug(curr, lock, _RET_IP_);
3199         return;
3200
3201 found_it:
3202         cpu = smp_processor_id();
3203         if (hlock->waittime_stamp) {
3204                 now = sched_clock();
3205                 waittime = now - hlock->waittime_stamp;
3206                 hlock->holdtime_stamp = now;
3207         }
3208
3209         trace_lock_acquired(lock, ip, waittime);
3210
3211         stats = get_lock_stats(hlock_class(hlock));
3212         if (waittime) {
3213                 if (hlock->read)
3214                         lock_time_inc(&stats->read_waittime, waittime);
3215                 else
3216                         lock_time_inc(&stats->write_waittime, waittime);
3217         }
3218         if (lock->cpu != cpu)
3219                 stats->bounces[bounce_acquired + !!hlock->read]++;
3220         put_lock_stats(stats);
3221
3222         lock->cpu = cpu;
3223         lock->ip = ip;
3224 }
3225
3226 void lock_contended(struct lockdep_map *lock, unsigned long ip)
3227 {
3228         unsigned long flags;
3229
3230         trace_lock_contended(lock, ip);
3231
3232         if (unlikely(!lock_stat))
3233                 return;
3234
3235         if (unlikely(current->lockdep_recursion))
3236                 return;
3237
3238         raw_local_irq_save(flags);
3239         check_flags(flags);
3240         current->lockdep_recursion = 1;
3241         __lock_contended(lock, ip);
3242         current->lockdep_recursion = 0;
3243         raw_local_irq_restore(flags);
3244 }
3245 EXPORT_SYMBOL_GPL(lock_contended);
3246
3247 void lock_acquired(struct lockdep_map *lock, unsigned long ip)
3248 {
3249         unsigned long flags;
3250
3251         if (unlikely(!lock_stat))
3252                 return;
3253
3254         if (unlikely(current->lockdep_recursion))
3255                 return;
3256
3257         raw_local_irq_save(flags);
3258         check_flags(flags);
3259         current->lockdep_recursion = 1;
3260         __lock_acquired(lock, ip);
3261         current->lockdep_recursion = 0;
3262         raw_local_irq_restore(flags);
3263 }
3264 EXPORT_SYMBOL_GPL(lock_acquired);
3265 #endif
3266
3267 /*
3268  * Used by the testsuite, sanitize the validator state
3269  * after a simulated failure:
3270  */
3271
3272 void lockdep_reset(void)
3273 {
3274         unsigned long flags;
3275         int i;
3276
3277         raw_local_irq_save(flags);
3278         current->curr_chain_key = 0;
3279         current->lockdep_depth = 0;
3280         current->lockdep_recursion = 0;
3281         memset(current->held_locks, 0, MAX_LOCK_DEPTH*sizeof(struct held_lock));
3282         nr_hardirq_chains = 0;
3283         nr_softirq_chains = 0;
3284         nr_process_chains = 0;
3285         debug_locks = 1;
3286         for (i = 0; i < CHAINHASH_SIZE; i++)
3287                 INIT_LIST_HEAD(chainhash_table + i);
3288         raw_local_irq_restore(flags);
3289 }
3290
3291 static void zap_class(struct lock_class *class)
3292 {
3293         int i;
3294
3295         /*
3296          * Remove all dependencies this lock is
3297          * involved in:
3298          */
3299         for (i = 0; i < nr_list_entries; i++) {
3300                 if (list_entries[i].class == class)
3301                         list_del_rcu(&list_entries[i].entry);
3302         }
3303         /*
3304          * Unhash the class and remove it from the all_lock_classes list:
3305          */
3306         list_del_rcu(&class->hash_entry);
3307         list_del_rcu(&class->lock_entry);
3308
3309         class->key = NULL;
3310 }
3311
3312 static inline int within(const void *addr, void *start, unsigned long size)
3313 {
3314         return addr >= start && addr < start + size;
3315 }
3316
3317 void lockdep_free_key_range(void *start, unsigned long size)
3318 {
3319         struct lock_class *class, *next;
3320         struct list_head *head;
3321         unsigned long flags;
3322         int i;
3323         int locked;
3324
3325         raw_local_irq_save(flags);
3326         locked = graph_lock();
3327
3328         /*
3329          * Unhash all classes that were created by this module:
3330          */
3331         for (i = 0; i < CLASSHASH_SIZE; i++) {
3332                 head = classhash_table + i;
3333                 if (list_empty(head))
3334                         continue;
3335                 list_for_each_entry_safe(class, next, head, hash_entry) {
3336                         if (within(class->key, start, size))
3337                                 zap_class(class);
3338                         else if (within(class->name, start, size))
3339                                 zap_class(class);
3340                 }
3341         }
3342
3343         if (locked)
3344                 graph_unlock();
3345         raw_local_irq_restore(flags);
3346 }
3347
3348 void lockdep_reset_lock(struct lockdep_map *lock)
3349 {
3350         struct lock_class *class, *next;
3351         struct list_head *head;
3352         unsigned long flags;
3353         int i, j;
3354         int locked;
3355
3356         raw_local_irq_save(flags);
3357
3358         /*
3359          * Remove all classes this lock might have:
3360          */
3361         for (j = 0; j < MAX_LOCKDEP_SUBCLASSES; j++) {
3362                 /*
3363                  * If the class exists we look it up and zap it:
3364                  */
3365                 class = look_up_lock_class(lock, j);
3366                 if (class)
3367                         zap_class(class);
3368         }
3369         /*
3370          * Debug check: in the end all mapped classes should
3371          * be gone.
3372          */
3373         locked = graph_lock();
3374         for (i = 0; i < CLASSHASH_SIZE; i++) {
3375                 head = classhash_table + i;
3376                 if (list_empty(head))
3377                         continue;
3378                 list_for_each_entry_safe(class, next, head, hash_entry) {
3379                         if (unlikely(class == lock->class_cache)) {
3380                                 if (debug_locks_off_graph_unlock())
3381                                         WARN_ON(1);
3382                                 goto out_restore;
3383                         }
3384                 }
3385         }
3386         if (locked)
3387                 graph_unlock();
3388
3389 out_restore:
3390         raw_local_irq_restore(flags);
3391 }
3392
3393 void lockdep_init(void)
3394 {
3395         int i;
3396
3397         /*
3398          * Some architectures have their own start_kernel()
3399          * code which calls lockdep_init(), while we also
3400          * call lockdep_init() from the start_kernel() itself,
3401          * and we want to initialize the hashes only once:
3402          */
3403         if (lockdep_initialized)
3404                 return;
3405
3406         for (i = 0; i < CLASSHASH_SIZE; i++)
3407                 INIT_LIST_HEAD(classhash_table + i);
3408
3409         for (i = 0; i < CHAINHASH_SIZE; i++)
3410                 INIT_LIST_HEAD(chainhash_table + i);
3411
3412         lockdep_initialized = 1;
3413 }
3414
3415 void __init lockdep_info(void)
3416 {
3417         printk("Lock dependency validator: Copyright (c) 2006 Red Hat, Inc., Ingo Molnar\n");
3418
3419         printk("... MAX_LOCKDEP_SUBCLASSES:  %lu\n", MAX_LOCKDEP_SUBCLASSES);
3420         printk("... MAX_LOCK_DEPTH:          %lu\n", MAX_LOCK_DEPTH);
3421         printk("... MAX_LOCKDEP_KEYS:        %lu\n", MAX_LOCKDEP_KEYS);
3422         printk("... CLASSHASH_SIZE:          %lu\n", CLASSHASH_SIZE);
3423         printk("... MAX_LOCKDEP_ENTRIES:     %lu\n", MAX_LOCKDEP_ENTRIES);
3424         printk("... MAX_LOCKDEP_CHAINS:      %lu\n", MAX_LOCKDEP_CHAINS);
3425         printk("... CHAINHASH_SIZE:          %lu\n", CHAINHASH_SIZE);
3426
3427         printk(" memory used by lock dependency info: %lu kB\n",
3428                 (sizeof(struct lock_class) * MAX_LOCKDEP_KEYS +
3429                 sizeof(struct list_head) * CLASSHASH_SIZE +
3430                 sizeof(struct lock_list) * MAX_LOCKDEP_ENTRIES +
3431                 sizeof(struct lock_chain) * MAX_LOCKDEP_CHAINS +
3432                 sizeof(struct list_head) * CHAINHASH_SIZE) / 1024
3433 #ifdef CONFIG_PROVE_LOCKING
3434                 + sizeof(struct circular_queue) + sizeof(bfs_accessed)
3435 #endif
3436                 );
3437
3438         printk(" per task-struct memory footprint: %lu bytes\n",
3439                 sizeof(struct held_lock) * MAX_LOCK_DEPTH);
3440
3441 #ifdef CONFIG_DEBUG_LOCKDEP
3442         if (lockdep_init_error) {
3443                 printk("WARNING: lockdep init error! Arch code didn't call lockdep_init() early enough?\n");
3444                 printk("Call stack leading to lockdep invocation was:\n");
3445                 print_stack_trace(&lockdep_init_trace, 0);
3446         }
3447 #endif
3448 }
3449
3450 static void
3451 print_freed_lock_bug(struct task_struct *curr, const void *mem_from,
3452                      const void *mem_to, struct held_lock *hlock)
3453 {
3454         if (!debug_locks_off())
3455                 return;
3456         if (debug_locks_silent)
3457                 return;
3458
3459         printk("\n=========================\n");
3460         printk(  "[ BUG: held lock freed! ]\n");
3461         printk(  "-------------------------\n");
3462         printk("%s/%d is freeing memory %p-%p, with a lock still held there!\n",
3463                 curr->comm, task_pid_nr(curr), mem_from, mem_to-1);
3464         print_lock(hlock);
3465         lockdep_print_held_locks(curr);
3466
3467         printk("\nstack backtrace:\n");
3468         dump_stack();
3469 }
3470
3471 static inline int not_in_range(const void* mem_from, unsigned long mem_len,
3472                                 const void* lock_from, unsigned long lock_len)
3473 {
3474         return lock_from + lock_len <= mem_from ||
3475                 mem_from + mem_len <= lock_from;
3476 }
3477
3478 /*
3479  * Called when kernel memory is freed (or unmapped), or if a lock
3480  * is destroyed or reinitialized - this code checks whether there is
3481  * any held lock in the memory range of <from> to <to>:
3482  */
3483 void debug_check_no_locks_freed(const void *mem_from, unsigned long mem_len)
3484 {
3485         struct task_struct *curr = current;
3486         struct held_lock *hlock;
3487         unsigned long flags;
3488         int i;
3489
3490         if (unlikely(!debug_locks))
3491                 return;
3492
3493         local_irq_save(flags);
3494         for (i = 0; i < curr->lockdep_depth; i++) {
3495                 hlock = curr->held_locks + i;
3496
3497                 if (not_in_range(mem_from, mem_len, hlock->instance,
3498                                         sizeof(*hlock->instance)))
3499                         continue;
3500
3501                 print_freed_lock_bug(curr, mem_from, mem_from + mem_len, hlock);
3502                 break;
3503         }
3504         local_irq_restore(flags);
3505 }
3506 EXPORT_SYMBOL_GPL(debug_check_no_locks_freed);
3507
3508 static void print_held_locks_bug(struct task_struct *curr)
3509 {
3510         if (!debug_locks_off())
3511                 return;
3512         if (debug_locks_silent)
3513                 return;
3514
3515         printk("\n=====================================\n");
3516         printk(  "[ BUG: lock held at task exit time! ]\n");
3517         printk(  "-------------------------------------\n");
3518         printk("%s/%d is exiting with locks still held!\n",
3519                 curr->comm, task_pid_nr(curr));
3520         lockdep_print_held_locks(curr);
3521
3522         printk("\nstack backtrace:\n");
3523         dump_stack();
3524 }
3525
3526 void debug_check_no_locks_held(struct task_struct *task)
3527 {
3528         if (unlikely(task->lockdep_depth > 0))
3529                 print_held_locks_bug(task);
3530 }
3531
3532 void debug_show_all_locks(void)
3533 {
3534         struct task_struct *g, *p;
3535         int count = 10;
3536         int unlock = 1;
3537
3538         if (unlikely(!debug_locks)) {
3539                 printk("INFO: lockdep is turned off.\n");
3540                 return;
3541         }
3542         printk("\nShowing all locks held in the system:\n");
3543
3544         /*
3545          * Here we try to get the tasklist_lock as hard as possible,
3546          * if not successful after 2 seconds we ignore it (but keep
3547          * trying). This is to enable a debug printout even if a
3548          * tasklist_lock-holding task deadlocks or crashes.
3549          */
3550 retry:
3551         if (!read_trylock(&tasklist_lock)) {
3552                 if (count == 10)
3553                         printk("hm, tasklist_lock locked, retrying... ");
3554                 if (count) {
3555                         count--;
3556                         printk(" #%d", 10-count);
3557                         mdelay(200);
3558                         goto retry;
3559                 }
3560                 printk(" ignoring it.\n");
3561                 unlock = 0;
3562         } else {
3563                 if (count != 10)
3564                         printk(KERN_CONT " locked it.\n");
3565         }
3566
3567         do_each_thread(g, p) {
3568                 /*
3569                  * It's not reliable to print a task's held locks
3570                  * if it's not sleeping (or if it's not the current
3571                  * task):
3572                  */
3573                 if (p->state == TASK_RUNNING && p != current)
3574                         continue;
3575                 if (p->lockdep_depth)
3576                         lockdep_print_held_locks(p);
3577                 if (!unlock)
3578                         if (read_trylock(&tasklist_lock))
3579                                 unlock = 1;
3580         } while_each_thread(g, p);
3581
3582         printk("\n");
3583         printk("=============================================\n\n");
3584
3585         if (unlock)
3586                 read_unlock(&tasklist_lock);
3587 }
3588 EXPORT_SYMBOL_GPL(debug_show_all_locks);
3589
3590 /*
3591  * Careful: only use this function if you are sure that
3592  * the task cannot run in parallel!
3593  */
3594 void __debug_show_held_locks(struct task_struct *task)
3595 {
3596         if (unlikely(!debug_locks)) {
3597                 printk("INFO: lockdep is turned off.\n");
3598                 return;
3599         }
3600         lockdep_print_held_locks(task);
3601 }
3602 EXPORT_SYMBOL_GPL(__debug_show_held_locks);
3603
3604 void debug_show_held_locks(struct task_struct *task)
3605 {
3606                 __debug_show_held_locks(task);
3607 }
3608 EXPORT_SYMBOL_GPL(debug_show_held_locks);
3609
3610 void lockdep_sys_exit(void)
3611 {
3612         struct task_struct *curr = current;
3613
3614         if (unlikely(curr->lockdep_depth)) {
3615                 if (!debug_locks_off())
3616                         return;
3617                 printk("\n================================================\n");
3618                 printk(  "[ BUG: lock held when returning to user space! ]\n");
3619                 printk(  "------------------------------------------------\n");
3620                 printk("%s/%d is leaving the kernel with locks still held!\n",
3621                                 curr->comm, curr->pid);
3622                 lockdep_print_held_locks(curr);
3623         }
3624 }