Merge tag 'perf-core-for-mingo' of git://git.kernel.org/pub/scm/linux/kernel/git...
[linux-2.6-microblaze.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 <linux/gfp.h>
47 #include <linux/kmemcheck.h>
48
49 #include <asm/sections.h>
50
51 #include "lockdep_internals.h"
52
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/lock.h>
55
56 #ifdef CONFIG_PROVE_LOCKING
57 int prove_locking = 1;
58 module_param(prove_locking, int, 0644);
59 #else
60 #define prove_locking 0
61 #endif
62
63 #ifdef CONFIG_LOCK_STAT
64 int lock_stat = 1;
65 module_param(lock_stat, int, 0644);
66 #else
67 #define lock_stat 0
68 #endif
69
70 /*
71  * lockdep_lock: protects the lockdep graph, the hashes and the
72  *               class/list/hash allocators.
73  *
74  * This is one of the rare exceptions where it's justified
75  * to use a raw spinlock - we really dont want the spinlock
76  * code to recurse back into the lockdep code...
77  */
78 static arch_spinlock_t lockdep_lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
79
80 static int graph_lock(void)
81 {
82         arch_spin_lock(&lockdep_lock);
83         /*
84          * Make sure that if another CPU detected a bug while
85          * walking the graph we dont change it (while the other
86          * CPU is busy printing out stuff with the graph lock
87          * dropped already)
88          */
89         if (!debug_locks) {
90                 arch_spin_unlock(&lockdep_lock);
91                 return 0;
92         }
93         /* prevent any recursions within lockdep from causing deadlocks */
94         current->lockdep_recursion++;
95         return 1;
96 }
97
98 static inline int graph_unlock(void)
99 {
100         if (debug_locks && !arch_spin_is_locked(&lockdep_lock)) {
101                 /*
102                  * The lockdep graph lock isn't locked while we expect it to
103                  * be, we're confused now, bye!
104                  */
105                 return DEBUG_LOCKS_WARN_ON(1);
106         }
107
108         current->lockdep_recursion--;
109         arch_spin_unlock(&lockdep_lock);
110         return 0;
111 }
112
113 /*
114  * Turn lock debugging off and return with 0 if it was off already,
115  * and also release the graph lock:
116  */
117 static inline int debug_locks_off_graph_unlock(void)
118 {
119         int ret = debug_locks_off();
120
121         arch_spin_unlock(&lockdep_lock);
122
123         return ret;
124 }
125
126 static int lockdep_initialized;
127
128 unsigned long nr_list_entries;
129 static struct lock_list list_entries[MAX_LOCKDEP_ENTRIES];
130
131 /*
132  * All data structures here are protected by the global debug_lock.
133  *
134  * Mutex key structs only get allocated, once during bootup, and never
135  * get freed - this significantly simplifies the debugging code.
136  */
137 unsigned long nr_lock_classes;
138 static struct lock_class lock_classes[MAX_LOCKDEP_KEYS];
139
140 static inline struct lock_class *hlock_class(struct held_lock *hlock)
141 {
142         if (!hlock->class_idx) {
143                 /*
144                  * Someone passed in garbage, we give up.
145                  */
146                 DEBUG_LOCKS_WARN_ON(1);
147                 return NULL;
148         }
149         return lock_classes + hlock->class_idx - 1;
150 }
151
152 #ifdef CONFIG_LOCK_STAT
153 static DEFINE_PER_CPU(struct lock_class_stats[MAX_LOCKDEP_KEYS],
154                       cpu_lock_stats);
155
156 static inline u64 lockstat_clock(void)
157 {
158         return local_clock();
159 }
160
161 static int lock_point(unsigned long points[], unsigned long ip)
162 {
163         int i;
164
165         for (i = 0; i < LOCKSTAT_POINTS; i++) {
166                 if (points[i] == 0) {
167                         points[i] = ip;
168                         break;
169                 }
170                 if (points[i] == ip)
171                         break;
172         }
173
174         return i;
175 }
176
177 static void lock_time_inc(struct lock_time *lt, u64 time)
178 {
179         if (time > lt->max)
180                 lt->max = time;
181
182         if (time < lt->min || !lt->nr)
183                 lt->min = time;
184
185         lt->total += time;
186         lt->nr++;
187 }
188
189 static inline void lock_time_add(struct lock_time *src, struct lock_time *dst)
190 {
191         if (!src->nr)
192                 return;
193
194         if (src->max > dst->max)
195                 dst->max = src->max;
196
197         if (src->min < dst->min || !dst->nr)
198                 dst->min = src->min;
199
200         dst->total += src->total;
201         dst->nr += src->nr;
202 }
203
204 struct lock_class_stats lock_stats(struct lock_class *class)
205 {
206         struct lock_class_stats stats;
207         int cpu, i;
208
209         memset(&stats, 0, sizeof(struct lock_class_stats));
210         for_each_possible_cpu(cpu) {
211                 struct lock_class_stats *pcs =
212                         &per_cpu(cpu_lock_stats, cpu)[class - lock_classes];
213
214                 for (i = 0; i < ARRAY_SIZE(stats.contention_point); i++)
215                         stats.contention_point[i] += pcs->contention_point[i];
216
217                 for (i = 0; i < ARRAY_SIZE(stats.contending_point); i++)
218                         stats.contending_point[i] += pcs->contending_point[i];
219
220                 lock_time_add(&pcs->read_waittime, &stats.read_waittime);
221                 lock_time_add(&pcs->write_waittime, &stats.write_waittime);
222
223                 lock_time_add(&pcs->read_holdtime, &stats.read_holdtime);
224                 lock_time_add(&pcs->write_holdtime, &stats.write_holdtime);
225
226                 for (i = 0; i < ARRAY_SIZE(stats.bounces); i++)
227                         stats.bounces[i] += pcs->bounces[i];
228         }
229
230         return stats;
231 }
232
233 void clear_lock_stats(struct lock_class *class)
234 {
235         int cpu;
236
237         for_each_possible_cpu(cpu) {
238                 struct lock_class_stats *cpu_stats =
239                         &per_cpu(cpu_lock_stats, cpu)[class - lock_classes];
240
241                 memset(cpu_stats, 0, sizeof(struct lock_class_stats));
242         }
243         memset(class->contention_point, 0, sizeof(class->contention_point));
244         memset(class->contending_point, 0, sizeof(class->contending_point));
245 }
246
247 static struct lock_class_stats *get_lock_stats(struct lock_class *class)
248 {
249         return &get_cpu_var(cpu_lock_stats)[class - lock_classes];
250 }
251
252 static void put_lock_stats(struct lock_class_stats *stats)
253 {
254         put_cpu_var(cpu_lock_stats);
255 }
256
257 static void lock_release_holdtime(struct held_lock *hlock)
258 {
259         struct lock_class_stats *stats;
260         u64 holdtime;
261
262         if (!lock_stat)
263                 return;
264
265         holdtime = lockstat_clock() - hlock->holdtime_stamp;
266
267         stats = get_lock_stats(hlock_class(hlock));
268         if (hlock->read)
269                 lock_time_inc(&stats->read_holdtime, holdtime);
270         else
271                 lock_time_inc(&stats->write_holdtime, holdtime);
272         put_lock_stats(stats);
273 }
274 #else
275 static inline void lock_release_holdtime(struct held_lock *hlock)
276 {
277 }
278 #endif
279
280 /*
281  * We keep a global list of all lock classes. The list only grows,
282  * never shrinks. The list is only accessed with the lockdep
283  * spinlock lock held.
284  */
285 LIST_HEAD(all_lock_classes);
286
287 /*
288  * The lockdep classes are in a hash-table as well, for fast lookup:
289  */
290 #define CLASSHASH_BITS          (MAX_LOCKDEP_KEYS_BITS - 1)
291 #define CLASSHASH_SIZE          (1UL << CLASSHASH_BITS)
292 #define __classhashfn(key)      hash_long((unsigned long)key, CLASSHASH_BITS)
293 #define classhashentry(key)     (classhash_table + __classhashfn((key)))
294
295 static struct list_head classhash_table[CLASSHASH_SIZE];
296
297 /*
298  * We put the lock dependency chains into a hash-table as well, to cache
299  * their existence:
300  */
301 #define CHAINHASH_BITS          (MAX_LOCKDEP_CHAINS_BITS-1)
302 #define CHAINHASH_SIZE          (1UL << CHAINHASH_BITS)
303 #define __chainhashfn(chain)    hash_long(chain, CHAINHASH_BITS)
304 #define chainhashentry(chain)   (chainhash_table + __chainhashfn((chain)))
305
306 static struct list_head chainhash_table[CHAINHASH_SIZE];
307
308 /*
309  * The hash key of the lock dependency chains is a hash itself too:
310  * it's a hash of all locks taken up to that lock, including that lock.
311  * It's a 64-bit hash, because it's important for the keys to be
312  * unique.
313  */
314 #define iterate_chain_key(key1, key2) \
315         (((key1) << MAX_LOCKDEP_KEYS_BITS) ^ \
316         ((key1) >> (64-MAX_LOCKDEP_KEYS_BITS)) ^ \
317         (key2))
318
319 void lockdep_off(void)
320 {
321         current->lockdep_recursion++;
322 }
323 EXPORT_SYMBOL(lockdep_off);
324
325 void lockdep_on(void)
326 {
327         current->lockdep_recursion--;
328 }
329 EXPORT_SYMBOL(lockdep_on);
330
331 /*
332  * Debugging switches:
333  */
334
335 #define VERBOSE                 0
336 #define VERY_VERBOSE            0
337
338 #if VERBOSE
339 # define HARDIRQ_VERBOSE        1
340 # define SOFTIRQ_VERBOSE        1
341 # define RECLAIM_VERBOSE        1
342 #else
343 # define HARDIRQ_VERBOSE        0
344 # define SOFTIRQ_VERBOSE        0
345 # define RECLAIM_VERBOSE        0
346 #endif
347
348 #if VERBOSE || HARDIRQ_VERBOSE || SOFTIRQ_VERBOSE || RECLAIM_VERBOSE
349 /*
350  * Quick filtering for interesting events:
351  */
352 static int class_filter(struct lock_class *class)
353 {
354 #if 0
355         /* Example */
356         if (class->name_version == 1 &&
357                         !strcmp(class->name, "lockname"))
358                 return 1;
359         if (class->name_version == 1 &&
360                         !strcmp(class->name, "&struct->lockfield"))
361                 return 1;
362 #endif
363         /* Filter everything else. 1 would be to allow everything else */
364         return 0;
365 }
366 #endif
367
368 static int verbose(struct lock_class *class)
369 {
370 #if VERBOSE
371         return class_filter(class);
372 #endif
373         return 0;
374 }
375
376 /*
377  * Stack-trace: tightly packed array of stack backtrace
378  * addresses. Protected by the graph_lock.
379  */
380 unsigned long nr_stack_trace_entries;
381 static unsigned long stack_trace[MAX_STACK_TRACE_ENTRIES];
382
383 static void print_lockdep_off(const char *bug_msg)
384 {
385         printk(KERN_DEBUG "%s\n", bug_msg);
386         printk(KERN_DEBUG "turning off the locking correctness validator.\n");
387         printk(KERN_DEBUG "Please attach the output of /proc/lock_stat to the bug report\n");
388 }
389
390 static int save_trace(struct stack_trace *trace)
391 {
392         trace->nr_entries = 0;
393         trace->max_entries = MAX_STACK_TRACE_ENTRIES - nr_stack_trace_entries;
394         trace->entries = stack_trace + nr_stack_trace_entries;
395
396         trace->skip = 3;
397
398         save_stack_trace(trace);
399
400         /*
401          * Some daft arches put -1 at the end to indicate its a full trace.
402          *
403          * <rant> this is buggy anyway, since it takes a whole extra entry so a
404          * complete trace that maxes out the entries provided will be reported
405          * as incomplete, friggin useless </rant>
406          */
407         if (trace->nr_entries != 0 &&
408             trace->entries[trace->nr_entries-1] == ULONG_MAX)
409                 trace->nr_entries--;
410
411         trace->max_entries = trace->nr_entries;
412
413         nr_stack_trace_entries += trace->nr_entries;
414
415         if (nr_stack_trace_entries >= MAX_STACK_TRACE_ENTRIES-1) {
416                 if (!debug_locks_off_graph_unlock())
417                         return 0;
418
419                 print_lockdep_off("BUG: MAX_STACK_TRACE_ENTRIES too low!");
420                 dump_stack();
421
422                 return 0;
423         }
424
425         return 1;
426 }
427
428 unsigned int nr_hardirq_chains;
429 unsigned int nr_softirq_chains;
430 unsigned int nr_process_chains;
431 unsigned int max_lockdep_depth;
432
433 #ifdef CONFIG_DEBUG_LOCKDEP
434 /*
435  * We cannot printk in early bootup code. Not even early_printk()
436  * might work. So we mark any initialization errors and printk
437  * about it later on, in lockdep_info().
438  */
439 static int lockdep_init_error;
440 static const char *lock_init_error;
441 static unsigned long lockdep_init_trace_data[20];
442 static struct stack_trace lockdep_init_trace = {
443         .max_entries = ARRAY_SIZE(lockdep_init_trace_data),
444         .entries = lockdep_init_trace_data,
445 };
446
447 /*
448  * Various lockdep statistics:
449  */
450 DEFINE_PER_CPU(struct lockdep_stats, lockdep_stats);
451 #endif
452
453 /*
454  * Locking printouts:
455  */
456
457 #define __USAGE(__STATE)                                                \
458         [LOCK_USED_IN_##__STATE] = "IN-"__stringify(__STATE)"-W",       \
459         [LOCK_ENABLED_##__STATE] = __stringify(__STATE)"-ON-W",         \
460         [LOCK_USED_IN_##__STATE##_READ] = "IN-"__stringify(__STATE)"-R",\
461         [LOCK_ENABLED_##__STATE##_READ] = __stringify(__STATE)"-ON-R",
462
463 static const char *usage_str[] =
464 {
465 #define LOCKDEP_STATE(__STATE) __USAGE(__STATE)
466 #include "lockdep_states.h"
467 #undef LOCKDEP_STATE
468         [LOCK_USED] = "INITIAL USE",
469 };
470
471 const char * __get_key_name(struct lockdep_subclass_key *key, char *str)
472 {
473         return kallsyms_lookup((unsigned long)key, NULL, NULL, NULL, str);
474 }
475
476 static inline unsigned long lock_flag(enum lock_usage_bit bit)
477 {
478         return 1UL << bit;
479 }
480
481 static char get_usage_char(struct lock_class *class, enum lock_usage_bit bit)
482 {
483         char c = '.';
484
485         if (class->usage_mask & lock_flag(bit + 2))
486                 c = '+';
487         if (class->usage_mask & lock_flag(bit)) {
488                 c = '-';
489                 if (class->usage_mask & lock_flag(bit + 2))
490                         c = '?';
491         }
492
493         return c;
494 }
495
496 void get_usage_chars(struct lock_class *class, char usage[LOCK_USAGE_CHARS])
497 {
498         int i = 0;
499
500 #define LOCKDEP_STATE(__STATE)                                          \
501         usage[i++] = get_usage_char(class, LOCK_USED_IN_##__STATE);     \
502         usage[i++] = get_usage_char(class, LOCK_USED_IN_##__STATE##_READ);
503 #include "lockdep_states.h"
504 #undef LOCKDEP_STATE
505
506         usage[i] = '\0';
507 }
508
509 static void __print_lock_name(struct lock_class *class)
510 {
511         char str[KSYM_NAME_LEN];
512         const char *name;
513
514         name = class->name;
515         if (!name) {
516                 name = __get_key_name(class->key, str);
517                 printk("%s", name);
518         } else {
519                 printk("%s", name);
520                 if (class->name_version > 1)
521                         printk("#%d", class->name_version);
522                 if (class->subclass)
523                         printk("/%d", class->subclass);
524         }
525 }
526
527 static void print_lock_name(struct lock_class *class)
528 {
529         char usage[LOCK_USAGE_CHARS];
530
531         get_usage_chars(class, usage);
532
533         printk(" (");
534         __print_lock_name(class);
535         printk("){%s}", usage);
536 }
537
538 static void print_lockdep_cache(struct lockdep_map *lock)
539 {
540         const char *name;
541         char str[KSYM_NAME_LEN];
542
543         name = lock->name;
544         if (!name)
545                 name = __get_key_name(lock->key->subkeys, str);
546
547         printk("%s", name);
548 }
549
550 static void print_lock(struct held_lock *hlock)
551 {
552         print_lock_name(hlock_class(hlock));
553         printk(", at: ");
554         print_ip_sym(hlock->acquire_ip);
555 }
556
557 static void lockdep_print_held_locks(struct task_struct *curr)
558 {
559         int i, depth = curr->lockdep_depth;
560
561         if (!depth) {
562                 printk("no locks held by %s/%d.\n", curr->comm, task_pid_nr(curr));
563                 return;
564         }
565         printk("%d lock%s held by %s/%d:\n",
566                 depth, depth > 1 ? "s" : "", curr->comm, task_pid_nr(curr));
567
568         for (i = 0; i < depth; i++) {
569                 printk(" #%d: ", i);
570                 print_lock(curr->held_locks + i);
571         }
572 }
573
574 static void print_kernel_ident(void)
575 {
576         printk("%s %.*s %s\n", init_utsname()->release,
577                 (int)strcspn(init_utsname()->version, " "),
578                 init_utsname()->version,
579                 print_tainted());
580 }
581
582 static int very_verbose(struct lock_class *class)
583 {
584 #if VERY_VERBOSE
585         return class_filter(class);
586 #endif
587         return 0;
588 }
589
590 /*
591  * Is this the address of a static object:
592  */
593 static int static_obj(void *obj)
594 {
595         unsigned long start = (unsigned long) &_stext,
596                       end   = (unsigned long) &_end,
597                       addr  = (unsigned long) obj;
598
599         /*
600          * static variable?
601          */
602         if ((addr >= start) && (addr < end))
603                 return 1;
604
605         if (arch_is_kernel_data(addr))
606                 return 1;
607
608         /*
609          * in-kernel percpu var?
610          */
611         if (is_kernel_percpu_address(addr))
612                 return 1;
613
614         /*
615          * module static or percpu var?
616          */
617         return is_module_address(addr) || is_module_percpu_address(addr);
618 }
619
620 /*
621  * To make lock name printouts unique, we calculate a unique
622  * class->name_version generation counter:
623  */
624 static int count_matching_names(struct lock_class *new_class)
625 {
626         struct lock_class *class;
627         int count = 0;
628
629         if (!new_class->name)
630                 return 0;
631
632         list_for_each_entry(class, &all_lock_classes, lock_entry) {
633                 if (new_class->key - new_class->subclass == class->key)
634                         return class->name_version;
635                 if (class->name && !strcmp(class->name, new_class->name))
636                         count = max(count, class->name_version);
637         }
638
639         return count + 1;
640 }
641
642 /*
643  * Register a lock's class in the hash-table, if the class is not present
644  * yet. Otherwise we look it up. We cache the result in the lock object
645  * itself, so actual lookup of the hash should be once per lock object.
646  */
647 static inline struct lock_class *
648 look_up_lock_class(struct lockdep_map *lock, unsigned int subclass)
649 {
650         struct lockdep_subclass_key *key;
651         struct list_head *hash_head;
652         struct lock_class *class;
653
654 #ifdef CONFIG_DEBUG_LOCKDEP
655         /*
656          * If the architecture calls into lockdep before initializing
657          * the hashes then we'll warn about it later. (we cannot printk
658          * right now)
659          */
660         if (unlikely(!lockdep_initialized)) {
661                 lockdep_init();
662                 lockdep_init_error = 1;
663                 lock_init_error = lock->name;
664                 save_stack_trace(&lockdep_init_trace);
665         }
666 #endif
667
668         if (unlikely(subclass >= MAX_LOCKDEP_SUBCLASSES)) {
669                 debug_locks_off();
670                 printk(KERN_ERR
671                         "BUG: looking up invalid subclass: %u\n", subclass);
672                 printk(KERN_ERR
673                         "turning off the locking correctness validator.\n");
674                 dump_stack();
675                 return NULL;
676         }
677
678         /*
679          * Static locks do not have their class-keys yet - for them the key
680          * is the lock object itself:
681          */
682         if (unlikely(!lock->key))
683                 lock->key = (void *)lock;
684
685         /*
686          * NOTE: the class-key must be unique. For dynamic locks, a static
687          * lock_class_key variable is passed in through the mutex_init()
688          * (or spin_lock_init()) call - which acts as the key. For static
689          * locks we use the lock object itself as the key.
690          */
691         BUILD_BUG_ON(sizeof(struct lock_class_key) >
692                         sizeof(struct lockdep_map));
693
694         key = lock->key->subkeys + subclass;
695
696         hash_head = classhashentry(key);
697
698         /*
699          * We can walk the hash lockfree, because the hash only
700          * grows, and we are careful when adding entries to the end:
701          */
702         list_for_each_entry(class, hash_head, hash_entry) {
703                 if (class->key == key) {
704                         /*
705                          * Huh! same key, different name? Did someone trample
706                          * on some memory? We're most confused.
707                          */
708                         WARN_ON_ONCE(class->name != lock->name);
709                         return class;
710                 }
711         }
712
713         return NULL;
714 }
715
716 /*
717  * Register a lock's class in the hash-table, if the class is not present
718  * yet. Otherwise we look it up. We cache the result in the lock object
719  * itself, so actual lookup of the hash should be once per lock object.
720  */
721 static inline struct lock_class *
722 register_lock_class(struct lockdep_map *lock, unsigned int subclass, int force)
723 {
724         struct lockdep_subclass_key *key;
725         struct list_head *hash_head;
726         struct lock_class *class;
727         unsigned long flags;
728
729         class = look_up_lock_class(lock, subclass);
730         if (likely(class))
731                 goto out_set_class_cache;
732
733         /*
734          * Debug-check: all keys must be persistent!
735          */
736         if (!static_obj(lock->key)) {
737                 debug_locks_off();
738                 printk("INFO: trying to register non-static key.\n");
739                 printk("the code is fine but needs lockdep annotation.\n");
740                 printk("turning off the locking correctness validator.\n");
741                 dump_stack();
742
743                 return NULL;
744         }
745
746         key = lock->key->subkeys + subclass;
747         hash_head = classhashentry(key);
748
749         raw_local_irq_save(flags);
750         if (!graph_lock()) {
751                 raw_local_irq_restore(flags);
752                 return NULL;
753         }
754         /*
755          * We have to do the hash-walk again, to avoid races
756          * with another CPU:
757          */
758         list_for_each_entry(class, hash_head, hash_entry)
759                 if (class->key == key)
760                         goto out_unlock_set;
761         /*
762          * Allocate a new key from the static array, and add it to
763          * the hash:
764          */
765         if (nr_lock_classes >= MAX_LOCKDEP_KEYS) {
766                 if (!debug_locks_off_graph_unlock()) {
767                         raw_local_irq_restore(flags);
768                         return NULL;
769                 }
770                 raw_local_irq_restore(flags);
771
772                 print_lockdep_off("BUG: MAX_LOCKDEP_KEYS too low!");
773                 dump_stack();
774                 return NULL;
775         }
776         class = lock_classes + nr_lock_classes++;
777         debug_atomic_inc(nr_unused_locks);
778         class->key = key;
779         class->name = lock->name;
780         class->subclass = subclass;
781         INIT_LIST_HEAD(&class->lock_entry);
782         INIT_LIST_HEAD(&class->locks_before);
783         INIT_LIST_HEAD(&class->locks_after);
784         class->name_version = count_matching_names(class);
785         /*
786          * We use RCU's safe list-add method to make
787          * parallel walking of the hash-list safe:
788          */
789         list_add_tail_rcu(&class->hash_entry, hash_head);
790         /*
791          * Add it to the global list of classes:
792          */
793         list_add_tail_rcu(&class->lock_entry, &all_lock_classes);
794
795         if (verbose(class)) {
796                 graph_unlock();
797                 raw_local_irq_restore(flags);
798
799                 printk("\nnew class %p: %s", class->key, class->name);
800                 if (class->name_version > 1)
801                         printk("#%d", class->name_version);
802                 printk("\n");
803                 dump_stack();
804
805                 raw_local_irq_save(flags);
806                 if (!graph_lock()) {
807                         raw_local_irq_restore(flags);
808                         return NULL;
809                 }
810         }
811 out_unlock_set:
812         graph_unlock();
813         raw_local_irq_restore(flags);
814
815 out_set_class_cache:
816         if (!subclass || force)
817                 lock->class_cache[0] = class;
818         else if (subclass < NR_LOCKDEP_CACHING_CLASSES)
819                 lock->class_cache[subclass] = class;
820
821         /*
822          * Hash collision, did we smoke some? We found a class with a matching
823          * hash but the subclass -- which is hashed in -- didn't match.
824          */
825         if (DEBUG_LOCKS_WARN_ON(class->subclass != subclass))
826                 return NULL;
827
828         return class;
829 }
830
831 #ifdef CONFIG_PROVE_LOCKING
832 /*
833  * Allocate a lockdep entry. (assumes the graph_lock held, returns
834  * with NULL on failure)
835  */
836 static struct lock_list *alloc_list_entry(void)
837 {
838         if (nr_list_entries >= MAX_LOCKDEP_ENTRIES) {
839                 if (!debug_locks_off_graph_unlock())
840                         return NULL;
841
842                 print_lockdep_off("BUG: MAX_LOCKDEP_ENTRIES too low!");
843                 dump_stack();
844                 return NULL;
845         }
846         return list_entries + nr_list_entries++;
847 }
848
849 /*
850  * Add a new dependency to the head of the list:
851  */
852 static int add_lock_to_list(struct lock_class *class, struct lock_class *this,
853                             struct list_head *head, unsigned long ip,
854                             int distance, struct stack_trace *trace)
855 {
856         struct lock_list *entry;
857         /*
858          * Lock not present yet - get a new dependency struct and
859          * add it to the list:
860          */
861         entry = alloc_list_entry();
862         if (!entry)
863                 return 0;
864
865         entry->class = this;
866         entry->distance = distance;
867         entry->trace = *trace;
868         /*
869          * Since we never remove from the dependency list, the list can
870          * be walked lockless by other CPUs, it's only allocation
871          * that must be protected by the spinlock. But this also means
872          * we must make new entries visible only once writes to the
873          * entry become visible - hence the RCU op:
874          */
875         list_add_tail_rcu(&entry->entry, head);
876
877         return 1;
878 }
879
880 /*
881  * For good efficiency of modular, we use power of 2
882  */
883 #define MAX_CIRCULAR_QUEUE_SIZE         4096UL
884 #define CQ_MASK                         (MAX_CIRCULAR_QUEUE_SIZE-1)
885
886 /*
887  * The circular_queue and helpers is used to implement the
888  * breadth-first search(BFS)algorithem, by which we can build
889  * the shortest path from the next lock to be acquired to the
890  * previous held lock if there is a circular between them.
891  */
892 struct circular_queue {
893         unsigned long element[MAX_CIRCULAR_QUEUE_SIZE];
894         unsigned int  front, rear;
895 };
896
897 static struct circular_queue lock_cq;
898
899 unsigned int max_bfs_queue_depth;
900
901 static unsigned int lockdep_dependency_gen_id;
902
903 static inline void __cq_init(struct circular_queue *cq)
904 {
905         cq->front = cq->rear = 0;
906         lockdep_dependency_gen_id++;
907 }
908
909 static inline int __cq_empty(struct circular_queue *cq)
910 {
911         return (cq->front == cq->rear);
912 }
913
914 static inline int __cq_full(struct circular_queue *cq)
915 {
916         return ((cq->rear + 1) & CQ_MASK) == cq->front;
917 }
918
919 static inline int __cq_enqueue(struct circular_queue *cq, unsigned long elem)
920 {
921         if (__cq_full(cq))
922                 return -1;
923
924         cq->element[cq->rear] = elem;
925         cq->rear = (cq->rear + 1) & CQ_MASK;
926         return 0;
927 }
928
929 static inline int __cq_dequeue(struct circular_queue *cq, unsigned long *elem)
930 {
931         if (__cq_empty(cq))
932                 return -1;
933
934         *elem = cq->element[cq->front];
935         cq->front = (cq->front + 1) & CQ_MASK;
936         return 0;
937 }
938
939 static inline unsigned int  __cq_get_elem_count(struct circular_queue *cq)
940 {
941         return (cq->rear - cq->front) & CQ_MASK;
942 }
943
944 static inline void mark_lock_accessed(struct lock_list *lock,
945                                         struct lock_list *parent)
946 {
947         unsigned long nr;
948
949         nr = lock - list_entries;
950         WARN_ON(nr >= nr_list_entries); /* Out-of-bounds, input fail */
951         lock->parent = parent;
952         lock->class->dep_gen_id = lockdep_dependency_gen_id;
953 }
954
955 static inline unsigned long lock_accessed(struct lock_list *lock)
956 {
957         unsigned long nr;
958
959         nr = lock - list_entries;
960         WARN_ON(nr >= nr_list_entries); /* Out-of-bounds, input fail */
961         return lock->class->dep_gen_id == lockdep_dependency_gen_id;
962 }
963
964 static inline struct lock_list *get_lock_parent(struct lock_list *child)
965 {
966         return child->parent;
967 }
968
969 static inline int get_lock_depth(struct lock_list *child)
970 {
971         int depth = 0;
972         struct lock_list *parent;
973
974         while ((parent = get_lock_parent(child))) {
975                 child = parent;
976                 depth++;
977         }
978         return depth;
979 }
980
981 static int __bfs(struct lock_list *source_entry,
982                  void *data,
983                  int (*match)(struct lock_list *entry, void *data),
984                  struct lock_list **target_entry,
985                  int forward)
986 {
987         struct lock_list *entry;
988         struct list_head *head;
989         struct circular_queue *cq = &lock_cq;
990         int ret = 1;
991
992         if (match(source_entry, data)) {
993                 *target_entry = source_entry;
994                 ret = 0;
995                 goto exit;
996         }
997
998         if (forward)
999                 head = &source_entry->class->locks_after;
1000         else
1001                 head = &source_entry->class->locks_before;
1002
1003         if (list_empty(head))
1004                 goto exit;
1005
1006         __cq_init(cq);
1007         __cq_enqueue(cq, (unsigned long)source_entry);
1008
1009         while (!__cq_empty(cq)) {
1010                 struct lock_list *lock;
1011
1012                 __cq_dequeue(cq, (unsigned long *)&lock);
1013
1014                 if (!lock->class) {
1015                         ret = -2;
1016                         goto exit;
1017                 }
1018
1019                 if (forward)
1020                         head = &lock->class->locks_after;
1021                 else
1022                         head = &lock->class->locks_before;
1023
1024                 list_for_each_entry(entry, head, entry) {
1025                         if (!lock_accessed(entry)) {
1026                                 unsigned int cq_depth;
1027                                 mark_lock_accessed(entry, lock);
1028                                 if (match(entry, data)) {
1029                                         *target_entry = entry;
1030                                         ret = 0;
1031                                         goto exit;
1032                                 }
1033
1034                                 if (__cq_enqueue(cq, (unsigned long)entry)) {
1035                                         ret = -1;
1036                                         goto exit;
1037                                 }
1038                                 cq_depth = __cq_get_elem_count(cq);
1039                                 if (max_bfs_queue_depth < cq_depth)
1040                                         max_bfs_queue_depth = cq_depth;
1041                         }
1042                 }
1043         }
1044 exit:
1045         return ret;
1046 }
1047
1048 static inline int __bfs_forwards(struct lock_list *src_entry,
1049                         void *data,
1050                         int (*match)(struct lock_list *entry, void *data),
1051                         struct lock_list **target_entry)
1052 {
1053         return __bfs(src_entry, data, match, target_entry, 1);
1054
1055 }
1056
1057 static inline int __bfs_backwards(struct lock_list *src_entry,
1058                         void *data,
1059                         int (*match)(struct lock_list *entry, void *data),
1060                         struct lock_list **target_entry)
1061 {
1062         return __bfs(src_entry, data, match, target_entry, 0);
1063
1064 }
1065
1066 /*
1067  * Recursive, forwards-direction lock-dependency checking, used for
1068  * both noncyclic checking and for hardirq-unsafe/softirq-unsafe
1069  * checking.
1070  */
1071
1072 /*
1073  * Print a dependency chain entry (this is only done when a deadlock
1074  * has been detected):
1075  */
1076 static noinline int
1077 print_circular_bug_entry(struct lock_list *target, int depth)
1078 {
1079         if (debug_locks_silent)
1080                 return 0;
1081         printk("\n-> #%u", depth);
1082         print_lock_name(target->class);
1083         printk(":\n");
1084         print_stack_trace(&target->trace, 6);
1085
1086         return 0;
1087 }
1088
1089 static void
1090 print_circular_lock_scenario(struct held_lock *src,
1091                              struct held_lock *tgt,
1092                              struct lock_list *prt)
1093 {
1094         struct lock_class *source = hlock_class(src);
1095         struct lock_class *target = hlock_class(tgt);
1096         struct lock_class *parent = prt->class;
1097
1098         /*
1099          * A direct locking problem where unsafe_class lock is taken
1100          * directly by safe_class lock, then all we need to show
1101          * is the deadlock scenario, as it is obvious that the
1102          * unsafe lock is taken under the safe lock.
1103          *
1104          * But if there is a chain instead, where the safe lock takes
1105          * an intermediate lock (middle_class) where this lock is
1106          * not the same as the safe lock, then the lock chain is
1107          * used to describe the problem. Otherwise we would need
1108          * to show a different CPU case for each link in the chain
1109          * from the safe_class lock to the unsafe_class lock.
1110          */
1111         if (parent != source) {
1112                 printk("Chain exists of:\n  ");
1113                 __print_lock_name(source);
1114                 printk(" --> ");
1115                 __print_lock_name(parent);
1116                 printk(" --> ");
1117                 __print_lock_name(target);
1118                 printk("\n\n");
1119         }
1120
1121         printk(" Possible unsafe locking scenario:\n\n");
1122         printk("       CPU0                    CPU1\n");
1123         printk("       ----                    ----\n");
1124         printk("  lock(");
1125         __print_lock_name(target);
1126         printk(");\n");
1127         printk("                               lock(");
1128         __print_lock_name(parent);
1129         printk(");\n");
1130         printk("                               lock(");
1131         __print_lock_name(target);
1132         printk(");\n");
1133         printk("  lock(");
1134         __print_lock_name(source);
1135         printk(");\n");
1136         printk("\n *** DEADLOCK ***\n\n");
1137 }
1138
1139 /*
1140  * When a circular dependency is detected, print the
1141  * header first:
1142  */
1143 static noinline int
1144 print_circular_bug_header(struct lock_list *entry, unsigned int depth,
1145                         struct held_lock *check_src,
1146                         struct held_lock *check_tgt)
1147 {
1148         struct task_struct *curr = current;
1149
1150         if (debug_locks_silent)
1151                 return 0;
1152
1153         printk("\n");
1154         printk("======================================================\n");
1155         printk("[ INFO: possible circular locking dependency detected ]\n");
1156         print_kernel_ident();
1157         printk("-------------------------------------------------------\n");
1158         printk("%s/%d is trying to acquire lock:\n",
1159                 curr->comm, task_pid_nr(curr));
1160         print_lock(check_src);
1161         printk("\nbut task is already holding lock:\n");
1162         print_lock(check_tgt);
1163         printk("\nwhich lock already depends on the new lock.\n\n");
1164         printk("\nthe existing dependency chain (in reverse order) is:\n");
1165
1166         print_circular_bug_entry(entry, depth);
1167
1168         return 0;
1169 }
1170
1171 static inline int class_equal(struct lock_list *entry, void *data)
1172 {
1173         return entry->class == data;
1174 }
1175
1176 static noinline int print_circular_bug(struct lock_list *this,
1177                                 struct lock_list *target,
1178                                 struct held_lock *check_src,
1179                                 struct held_lock *check_tgt)
1180 {
1181         struct task_struct *curr = current;
1182         struct lock_list *parent;
1183         struct lock_list *first_parent;
1184         int depth;
1185
1186         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1187                 return 0;
1188
1189         if (!save_trace(&this->trace))
1190                 return 0;
1191
1192         depth = get_lock_depth(target);
1193
1194         print_circular_bug_header(target, depth, check_src, check_tgt);
1195
1196         parent = get_lock_parent(target);
1197         first_parent = parent;
1198
1199         while (parent) {
1200                 print_circular_bug_entry(parent, --depth);
1201                 parent = get_lock_parent(parent);
1202         }
1203
1204         printk("\nother info that might help us debug this:\n\n");
1205         print_circular_lock_scenario(check_src, check_tgt,
1206                                      first_parent);
1207
1208         lockdep_print_held_locks(curr);
1209
1210         printk("\nstack backtrace:\n");
1211         dump_stack();
1212
1213         return 0;
1214 }
1215
1216 static noinline int print_bfs_bug(int ret)
1217 {
1218         if (!debug_locks_off_graph_unlock())
1219                 return 0;
1220
1221         /*
1222          * Breadth-first-search failed, graph got corrupted?
1223          */
1224         WARN(1, "lockdep bfs error:%d\n", ret);
1225
1226         return 0;
1227 }
1228
1229 static int noop_count(struct lock_list *entry, void *data)
1230 {
1231         (*(unsigned long *)data)++;
1232         return 0;
1233 }
1234
1235 unsigned long __lockdep_count_forward_deps(struct lock_list *this)
1236 {
1237         unsigned long  count = 0;
1238         struct lock_list *uninitialized_var(target_entry);
1239
1240         __bfs_forwards(this, (void *)&count, noop_count, &target_entry);
1241
1242         return count;
1243 }
1244 unsigned long lockdep_count_forward_deps(struct lock_class *class)
1245 {
1246         unsigned long ret, flags;
1247         struct lock_list this;
1248
1249         this.parent = NULL;
1250         this.class = class;
1251
1252         local_irq_save(flags);
1253         arch_spin_lock(&lockdep_lock);
1254         ret = __lockdep_count_forward_deps(&this);
1255         arch_spin_unlock(&lockdep_lock);
1256         local_irq_restore(flags);
1257
1258         return ret;
1259 }
1260
1261 unsigned long __lockdep_count_backward_deps(struct lock_list *this)
1262 {
1263         unsigned long  count = 0;
1264         struct lock_list *uninitialized_var(target_entry);
1265
1266         __bfs_backwards(this, (void *)&count, noop_count, &target_entry);
1267
1268         return count;
1269 }
1270
1271 unsigned long lockdep_count_backward_deps(struct lock_class *class)
1272 {
1273         unsigned long ret, flags;
1274         struct lock_list this;
1275
1276         this.parent = NULL;
1277         this.class = class;
1278
1279         local_irq_save(flags);
1280         arch_spin_lock(&lockdep_lock);
1281         ret = __lockdep_count_backward_deps(&this);
1282         arch_spin_unlock(&lockdep_lock);
1283         local_irq_restore(flags);
1284
1285         return ret;
1286 }
1287
1288 /*
1289  * Prove that the dependency graph starting at <entry> can not
1290  * lead to <target>. Print an error and return 0 if it does.
1291  */
1292 static noinline int
1293 check_noncircular(struct lock_list *root, struct lock_class *target,
1294                 struct lock_list **target_entry)
1295 {
1296         int result;
1297
1298         debug_atomic_inc(nr_cyclic_checks);
1299
1300         result = __bfs_forwards(root, target, class_equal, target_entry);
1301
1302         return result;
1303 }
1304
1305 #if defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING)
1306 /*
1307  * Forwards and backwards subgraph searching, for the purposes of
1308  * proving that two subgraphs can be connected by a new dependency
1309  * without creating any illegal irq-safe -> irq-unsafe lock dependency.
1310  */
1311
1312 static inline int usage_match(struct lock_list *entry, void *bit)
1313 {
1314         return entry->class->usage_mask & (1 << (enum lock_usage_bit)bit);
1315 }
1316
1317
1318
1319 /*
1320  * Find a node in the forwards-direction dependency sub-graph starting
1321  * at @root->class that matches @bit.
1322  *
1323  * Return 0 if such a node exists in the subgraph, and put that node
1324  * into *@target_entry.
1325  *
1326  * Return 1 otherwise and keep *@target_entry unchanged.
1327  * Return <0 on error.
1328  */
1329 static int
1330 find_usage_forwards(struct lock_list *root, enum lock_usage_bit bit,
1331                         struct lock_list **target_entry)
1332 {
1333         int result;
1334
1335         debug_atomic_inc(nr_find_usage_forwards_checks);
1336
1337         result = __bfs_forwards(root, (void *)bit, usage_match, target_entry);
1338
1339         return result;
1340 }
1341
1342 /*
1343  * Find a node in the backwards-direction dependency sub-graph starting
1344  * at @root->class that matches @bit.
1345  *
1346  * Return 0 if such a node exists in the subgraph, and put that node
1347  * into *@target_entry.
1348  *
1349  * Return 1 otherwise and keep *@target_entry unchanged.
1350  * Return <0 on error.
1351  */
1352 static int
1353 find_usage_backwards(struct lock_list *root, enum lock_usage_bit bit,
1354                         struct lock_list **target_entry)
1355 {
1356         int result;
1357
1358         debug_atomic_inc(nr_find_usage_backwards_checks);
1359
1360         result = __bfs_backwards(root, (void *)bit, usage_match, target_entry);
1361
1362         return result;
1363 }
1364
1365 static void print_lock_class_header(struct lock_class *class, int depth)
1366 {
1367         int bit;
1368
1369         printk("%*s->", depth, "");
1370         print_lock_name(class);
1371         printk(" ops: %lu", class->ops);
1372         printk(" {\n");
1373
1374         for (bit = 0; bit < LOCK_USAGE_STATES; bit++) {
1375                 if (class->usage_mask & (1 << bit)) {
1376                         int len = depth;
1377
1378                         len += printk("%*s   %s", depth, "", usage_str[bit]);
1379                         len += printk(" at:\n");
1380                         print_stack_trace(class->usage_traces + bit, len);
1381                 }
1382         }
1383         printk("%*s }\n", depth, "");
1384
1385         printk("%*s ... key      at: ",depth,"");
1386         print_ip_sym((unsigned long)class->key);
1387 }
1388
1389 /*
1390  * printk the shortest lock dependencies from @start to @end in reverse order:
1391  */
1392 static void __used
1393 print_shortest_lock_dependencies(struct lock_list *leaf,
1394                                 struct lock_list *root)
1395 {
1396         struct lock_list *entry = leaf;
1397         int depth;
1398
1399         /*compute depth from generated tree by BFS*/
1400         depth = get_lock_depth(leaf);
1401
1402         do {
1403                 print_lock_class_header(entry->class, depth);
1404                 printk("%*s ... acquired at:\n", depth, "");
1405                 print_stack_trace(&entry->trace, 2);
1406                 printk("\n");
1407
1408                 if (depth == 0 && (entry != root)) {
1409                         printk("lockdep:%s bad path found in chain graph\n", __func__);
1410                         break;
1411                 }
1412
1413                 entry = get_lock_parent(entry);
1414                 depth--;
1415         } while (entry && (depth >= 0));
1416
1417         return;
1418 }
1419
1420 static void
1421 print_irq_lock_scenario(struct lock_list *safe_entry,
1422                         struct lock_list *unsafe_entry,
1423                         struct lock_class *prev_class,
1424                         struct lock_class *next_class)
1425 {
1426         struct lock_class *safe_class = safe_entry->class;
1427         struct lock_class *unsafe_class = unsafe_entry->class;
1428         struct lock_class *middle_class = prev_class;
1429
1430         if (middle_class == safe_class)
1431                 middle_class = next_class;
1432
1433         /*
1434          * A direct locking problem where unsafe_class lock is taken
1435          * directly by safe_class lock, then all we need to show
1436          * is the deadlock scenario, as it is obvious that the
1437          * unsafe lock is taken under the safe lock.
1438          *
1439          * But if there is a chain instead, where the safe lock takes
1440          * an intermediate lock (middle_class) where this lock is
1441          * not the same as the safe lock, then the lock chain is
1442          * used to describe the problem. Otherwise we would need
1443          * to show a different CPU case for each link in the chain
1444          * from the safe_class lock to the unsafe_class lock.
1445          */
1446         if (middle_class != unsafe_class) {
1447                 printk("Chain exists of:\n  ");
1448                 __print_lock_name(safe_class);
1449                 printk(" --> ");
1450                 __print_lock_name(middle_class);
1451                 printk(" --> ");
1452                 __print_lock_name(unsafe_class);
1453                 printk("\n\n");
1454         }
1455
1456         printk(" Possible interrupt unsafe locking scenario:\n\n");
1457         printk("       CPU0                    CPU1\n");
1458         printk("       ----                    ----\n");
1459         printk("  lock(");
1460         __print_lock_name(unsafe_class);
1461         printk(");\n");
1462         printk("                               local_irq_disable();\n");
1463         printk("                               lock(");
1464         __print_lock_name(safe_class);
1465         printk(");\n");
1466         printk("                               lock(");
1467         __print_lock_name(middle_class);
1468         printk(");\n");
1469         printk("  <Interrupt>\n");
1470         printk("    lock(");
1471         __print_lock_name(safe_class);
1472         printk(");\n");
1473         printk("\n *** DEADLOCK ***\n\n");
1474 }
1475
1476 static int
1477 print_bad_irq_dependency(struct task_struct *curr,
1478                          struct lock_list *prev_root,
1479                          struct lock_list *next_root,
1480                          struct lock_list *backwards_entry,
1481                          struct lock_list *forwards_entry,
1482                          struct held_lock *prev,
1483                          struct held_lock *next,
1484                          enum lock_usage_bit bit1,
1485                          enum lock_usage_bit bit2,
1486                          const char *irqclass)
1487 {
1488         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1489                 return 0;
1490
1491         printk("\n");
1492         printk("======================================================\n");
1493         printk("[ INFO: %s-safe -> %s-unsafe lock order detected ]\n",
1494                 irqclass, irqclass);
1495         print_kernel_ident();
1496         printk("------------------------------------------------------\n");
1497         printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] is trying to acquire:\n",
1498                 curr->comm, task_pid_nr(curr),
1499                 curr->hardirq_context, hardirq_count() >> HARDIRQ_SHIFT,
1500                 curr->softirq_context, softirq_count() >> SOFTIRQ_SHIFT,
1501                 curr->hardirqs_enabled,
1502                 curr->softirqs_enabled);
1503         print_lock(next);
1504
1505         printk("\nand this task is already holding:\n");
1506         print_lock(prev);
1507         printk("which would create a new lock dependency:\n");
1508         print_lock_name(hlock_class(prev));
1509         printk(" ->");
1510         print_lock_name(hlock_class(next));
1511         printk("\n");
1512
1513         printk("\nbut this new dependency connects a %s-irq-safe lock:\n",
1514                 irqclass);
1515         print_lock_name(backwards_entry->class);
1516         printk("\n... which became %s-irq-safe at:\n", irqclass);
1517
1518         print_stack_trace(backwards_entry->class->usage_traces + bit1, 1);
1519
1520         printk("\nto a %s-irq-unsafe lock:\n", irqclass);
1521         print_lock_name(forwards_entry->class);
1522         printk("\n... which became %s-irq-unsafe at:\n", irqclass);
1523         printk("...");
1524
1525         print_stack_trace(forwards_entry->class->usage_traces + bit2, 1);
1526
1527         printk("\nother info that might help us debug this:\n\n");
1528         print_irq_lock_scenario(backwards_entry, forwards_entry,
1529                                 hlock_class(prev), hlock_class(next));
1530
1531         lockdep_print_held_locks(curr);
1532
1533         printk("\nthe dependencies between %s-irq-safe lock", irqclass);
1534         printk(" and the holding lock:\n");
1535         if (!save_trace(&prev_root->trace))
1536                 return 0;
1537         print_shortest_lock_dependencies(backwards_entry, prev_root);
1538
1539         printk("\nthe dependencies between the lock to be acquired");
1540         printk(" and %s-irq-unsafe lock:\n", irqclass);
1541         if (!save_trace(&next_root->trace))
1542                 return 0;
1543         print_shortest_lock_dependencies(forwards_entry, next_root);
1544
1545         printk("\nstack backtrace:\n");
1546         dump_stack();
1547
1548         return 0;
1549 }
1550
1551 static int
1552 check_usage(struct task_struct *curr, struct held_lock *prev,
1553             struct held_lock *next, enum lock_usage_bit bit_backwards,
1554             enum lock_usage_bit bit_forwards, const char *irqclass)
1555 {
1556         int ret;
1557         struct lock_list this, that;
1558         struct lock_list *uninitialized_var(target_entry);
1559         struct lock_list *uninitialized_var(target_entry1);
1560
1561         this.parent = NULL;
1562
1563         this.class = hlock_class(prev);
1564         ret = find_usage_backwards(&this, bit_backwards, &target_entry);
1565         if (ret < 0)
1566                 return print_bfs_bug(ret);
1567         if (ret == 1)
1568                 return ret;
1569
1570         that.parent = NULL;
1571         that.class = hlock_class(next);
1572         ret = find_usage_forwards(&that, bit_forwards, &target_entry1);
1573         if (ret < 0)
1574                 return print_bfs_bug(ret);
1575         if (ret == 1)
1576                 return ret;
1577
1578         return print_bad_irq_dependency(curr, &this, &that,
1579                         target_entry, target_entry1,
1580                         prev, next,
1581                         bit_backwards, bit_forwards, irqclass);
1582 }
1583
1584 static const char *state_names[] = {
1585 #define LOCKDEP_STATE(__STATE) \
1586         __stringify(__STATE),
1587 #include "lockdep_states.h"
1588 #undef LOCKDEP_STATE
1589 };
1590
1591 static const char *state_rnames[] = {
1592 #define LOCKDEP_STATE(__STATE) \
1593         __stringify(__STATE)"-READ",
1594 #include "lockdep_states.h"
1595 #undef LOCKDEP_STATE
1596 };
1597
1598 static inline const char *state_name(enum lock_usage_bit bit)
1599 {
1600         return (bit & 1) ? state_rnames[bit >> 2] : state_names[bit >> 2];
1601 }
1602
1603 static int exclusive_bit(int new_bit)
1604 {
1605         /*
1606          * USED_IN
1607          * USED_IN_READ
1608          * ENABLED
1609          * ENABLED_READ
1610          *
1611          * bit 0 - write/read
1612          * bit 1 - used_in/enabled
1613          * bit 2+  state
1614          */
1615
1616         int state = new_bit & ~3;
1617         int dir = new_bit & 2;
1618
1619         /*
1620          * keep state, bit flip the direction and strip read.
1621          */
1622         return state | (dir ^ 2);
1623 }
1624
1625 static int check_irq_usage(struct task_struct *curr, struct held_lock *prev,
1626                            struct held_lock *next, enum lock_usage_bit bit)
1627 {
1628         /*
1629          * Prove that the new dependency does not connect a hardirq-safe
1630          * lock with a hardirq-unsafe lock - to achieve this we search
1631          * the backwards-subgraph starting at <prev>, and the
1632          * forwards-subgraph starting at <next>:
1633          */
1634         if (!check_usage(curr, prev, next, bit,
1635                            exclusive_bit(bit), state_name(bit)))
1636                 return 0;
1637
1638         bit++; /* _READ */
1639
1640         /*
1641          * Prove that the new dependency does not connect a hardirq-safe-read
1642          * lock with a hardirq-unsafe lock - to achieve this we search
1643          * the backwards-subgraph starting at <prev>, and the
1644          * forwards-subgraph starting at <next>:
1645          */
1646         if (!check_usage(curr, prev, next, bit,
1647                            exclusive_bit(bit), state_name(bit)))
1648                 return 0;
1649
1650         return 1;
1651 }
1652
1653 static int
1654 check_prev_add_irq(struct task_struct *curr, struct held_lock *prev,
1655                 struct held_lock *next)
1656 {
1657 #define LOCKDEP_STATE(__STATE)                                          \
1658         if (!check_irq_usage(curr, prev, next, LOCK_USED_IN_##__STATE)) \
1659                 return 0;
1660 #include "lockdep_states.h"
1661 #undef LOCKDEP_STATE
1662
1663         return 1;
1664 }
1665
1666 static void inc_chains(void)
1667 {
1668         if (current->hardirq_context)
1669                 nr_hardirq_chains++;
1670         else {
1671                 if (current->softirq_context)
1672                         nr_softirq_chains++;
1673                 else
1674                         nr_process_chains++;
1675         }
1676 }
1677
1678 #else
1679
1680 static inline int
1681 check_prev_add_irq(struct task_struct *curr, struct held_lock *prev,
1682                 struct held_lock *next)
1683 {
1684         return 1;
1685 }
1686
1687 static inline void inc_chains(void)
1688 {
1689         nr_process_chains++;
1690 }
1691
1692 #endif
1693
1694 static void
1695 print_deadlock_scenario(struct held_lock *nxt,
1696                              struct held_lock *prv)
1697 {
1698         struct lock_class *next = hlock_class(nxt);
1699         struct lock_class *prev = hlock_class(prv);
1700
1701         printk(" Possible unsafe locking scenario:\n\n");
1702         printk("       CPU0\n");
1703         printk("       ----\n");
1704         printk("  lock(");
1705         __print_lock_name(prev);
1706         printk(");\n");
1707         printk("  lock(");
1708         __print_lock_name(next);
1709         printk(");\n");
1710         printk("\n *** DEADLOCK ***\n\n");
1711         printk(" May be due to missing lock nesting notation\n\n");
1712 }
1713
1714 static int
1715 print_deadlock_bug(struct task_struct *curr, struct held_lock *prev,
1716                    struct held_lock *next)
1717 {
1718         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
1719                 return 0;
1720
1721         printk("\n");
1722         printk("=============================================\n");
1723         printk("[ INFO: possible recursive locking detected ]\n");
1724         print_kernel_ident();
1725         printk("---------------------------------------------\n");
1726         printk("%s/%d is trying to acquire lock:\n",
1727                 curr->comm, task_pid_nr(curr));
1728         print_lock(next);
1729         printk("\nbut task is already holding lock:\n");
1730         print_lock(prev);
1731
1732         printk("\nother info that might help us debug this:\n");
1733         print_deadlock_scenario(next, prev);
1734         lockdep_print_held_locks(curr);
1735
1736         printk("\nstack backtrace:\n");
1737         dump_stack();
1738
1739         return 0;
1740 }
1741
1742 /*
1743  * Check whether we are holding such a class already.
1744  *
1745  * (Note that this has to be done separately, because the graph cannot
1746  * detect such classes of deadlocks.)
1747  *
1748  * Returns: 0 on deadlock detected, 1 on OK, 2 on recursive read
1749  */
1750 static int
1751 check_deadlock(struct task_struct *curr, struct held_lock *next,
1752                struct lockdep_map *next_instance, int read)
1753 {
1754         struct held_lock *prev;
1755         struct held_lock *nest = NULL;
1756         int i;
1757
1758         for (i = 0; i < curr->lockdep_depth; i++) {
1759                 prev = curr->held_locks + i;
1760
1761                 if (prev->instance == next->nest_lock)
1762                         nest = prev;
1763
1764                 if (hlock_class(prev) != hlock_class(next))
1765                         continue;
1766
1767                 /*
1768                  * Allow read-after-read recursion of the same
1769                  * lock class (i.e. read_lock(lock)+read_lock(lock)):
1770                  */
1771                 if ((read == 2) && prev->read)
1772                         return 2;
1773
1774                 /*
1775                  * We're holding the nest_lock, which serializes this lock's
1776                  * nesting behaviour.
1777                  */
1778                 if (nest)
1779                         return 2;
1780
1781                 return print_deadlock_bug(curr, prev, next);
1782         }
1783         return 1;
1784 }
1785
1786 /*
1787  * There was a chain-cache miss, and we are about to add a new dependency
1788  * to a previous lock. We recursively validate the following rules:
1789  *
1790  *  - would the adding of the <prev> -> <next> dependency create a
1791  *    circular dependency in the graph? [== circular deadlock]
1792  *
1793  *  - does the new prev->next dependency connect any hardirq-safe lock
1794  *    (in the full backwards-subgraph starting at <prev>) with any
1795  *    hardirq-unsafe lock (in the full forwards-subgraph starting at
1796  *    <next>)? [== illegal lock inversion with hardirq contexts]
1797  *
1798  *  - does the new prev->next dependency connect any softirq-safe lock
1799  *    (in the full backwards-subgraph starting at <prev>) with any
1800  *    softirq-unsafe lock (in the full forwards-subgraph starting at
1801  *    <next>)? [== illegal lock inversion with softirq contexts]
1802  *
1803  * any of these scenarios could lead to a deadlock.
1804  *
1805  * Then if all the validations pass, we add the forwards and backwards
1806  * dependency.
1807  */
1808 static int
1809 check_prev_add(struct task_struct *curr, struct held_lock *prev,
1810                struct held_lock *next, int distance, int trylock_loop)
1811 {
1812         struct lock_list *entry;
1813         int ret;
1814         struct lock_list this;
1815         struct lock_list *uninitialized_var(target_entry);
1816         /*
1817          * Static variable, serialized by the graph_lock().
1818          *
1819          * We use this static variable to save the stack trace in case
1820          * we call into this function multiple times due to encountering
1821          * trylocks in the held lock stack.
1822          */
1823         static struct stack_trace trace;
1824
1825         /*
1826          * Prove that the new <prev> -> <next> dependency would not
1827          * create a circular dependency in the graph. (We do this by
1828          * forward-recursing into the graph starting at <next>, and
1829          * checking whether we can reach <prev>.)
1830          *
1831          * We are using global variables to control the recursion, to
1832          * keep the stackframe size of the recursive functions low:
1833          */
1834         this.class = hlock_class(next);
1835         this.parent = NULL;
1836         ret = check_noncircular(&this, hlock_class(prev), &target_entry);
1837         if (unlikely(!ret))
1838                 return print_circular_bug(&this, target_entry, next, prev);
1839         else if (unlikely(ret < 0))
1840                 return print_bfs_bug(ret);
1841
1842         if (!check_prev_add_irq(curr, prev, next))
1843                 return 0;
1844
1845         /*
1846          * For recursive read-locks we do all the dependency checks,
1847          * but we dont store read-triggered dependencies (only
1848          * write-triggered dependencies). This ensures that only the
1849          * write-side dependencies matter, and that if for example a
1850          * write-lock never takes any other locks, then the reads are
1851          * equivalent to a NOP.
1852          */
1853         if (next->read == 2 || prev->read == 2)
1854                 return 1;
1855         /*
1856          * Is the <prev> -> <next> dependency already present?
1857          *
1858          * (this may occur even though this is a new chain: consider
1859          *  e.g. the L1 -> L2 -> L3 -> L4 and the L5 -> L1 -> L2 -> L3
1860          *  chains - the second one will be new, but L1 already has
1861          *  L2 added to its dependency list, due to the first chain.)
1862          */
1863         list_for_each_entry(entry, &hlock_class(prev)->locks_after, entry) {
1864                 if (entry->class == hlock_class(next)) {
1865                         if (distance == 1)
1866                                 entry->distance = 1;
1867                         return 2;
1868                 }
1869         }
1870
1871         if (!trylock_loop && !save_trace(&trace))
1872                 return 0;
1873
1874         /*
1875          * Ok, all validations passed, add the new lock
1876          * to the previous lock's dependency list:
1877          */
1878         ret = add_lock_to_list(hlock_class(prev), hlock_class(next),
1879                                &hlock_class(prev)->locks_after,
1880                                next->acquire_ip, distance, &trace);
1881
1882         if (!ret)
1883                 return 0;
1884
1885         ret = add_lock_to_list(hlock_class(next), hlock_class(prev),
1886                                &hlock_class(next)->locks_before,
1887                                next->acquire_ip, distance, &trace);
1888         if (!ret)
1889                 return 0;
1890
1891         /*
1892          * Debugging printouts:
1893          */
1894         if (verbose(hlock_class(prev)) || verbose(hlock_class(next))) {
1895                 graph_unlock();
1896                 printk("\n new dependency: ");
1897                 print_lock_name(hlock_class(prev));
1898                 printk(" => ");
1899                 print_lock_name(hlock_class(next));
1900                 printk("\n");
1901                 dump_stack();
1902                 return graph_lock();
1903         }
1904         return 1;
1905 }
1906
1907 /*
1908  * Add the dependency to all directly-previous locks that are 'relevant'.
1909  * The ones that are relevant are (in increasing distance from curr):
1910  * all consecutive trylock entries and the final non-trylock entry - or
1911  * the end of this context's lock-chain - whichever comes first.
1912  */
1913 static int
1914 check_prevs_add(struct task_struct *curr, struct held_lock *next)
1915 {
1916         int depth = curr->lockdep_depth;
1917         int trylock_loop = 0;
1918         struct held_lock *hlock;
1919
1920         /*
1921          * Debugging checks.
1922          *
1923          * Depth must not be zero for a non-head lock:
1924          */
1925         if (!depth)
1926                 goto out_bug;
1927         /*
1928          * At least two relevant locks must exist for this
1929          * to be a head:
1930          */
1931         if (curr->held_locks[depth].irq_context !=
1932                         curr->held_locks[depth-1].irq_context)
1933                 goto out_bug;
1934
1935         for (;;) {
1936                 int distance = curr->lockdep_depth - depth + 1;
1937                 hlock = curr->held_locks + depth-1;
1938                 /*
1939                  * Only non-recursive-read entries get new dependencies
1940                  * added:
1941                  */
1942                 if (hlock->read != 2) {
1943                         if (!check_prev_add(curr, hlock, next,
1944                                                 distance, trylock_loop))
1945                                 return 0;
1946                         /*
1947                          * Stop after the first non-trylock entry,
1948                          * as non-trylock entries have added their
1949                          * own direct dependencies already, so this
1950                          * lock is connected to them indirectly:
1951                          */
1952                         if (!hlock->trylock)
1953                                 break;
1954                 }
1955                 depth--;
1956                 /*
1957                  * End of lock-stack?
1958                  */
1959                 if (!depth)
1960                         break;
1961                 /*
1962                  * Stop the search if we cross into another context:
1963                  */
1964                 if (curr->held_locks[depth].irq_context !=
1965                                 curr->held_locks[depth-1].irq_context)
1966                         break;
1967                 trylock_loop = 1;
1968         }
1969         return 1;
1970 out_bug:
1971         if (!debug_locks_off_graph_unlock())
1972                 return 0;
1973
1974         /*
1975          * Clearly we all shouldn't be here, but since we made it we
1976          * can reliable say we messed up our state. See the above two
1977          * gotos for reasons why we could possibly end up here.
1978          */
1979         WARN_ON(1);
1980
1981         return 0;
1982 }
1983
1984 unsigned long nr_lock_chains;
1985 struct lock_chain lock_chains[MAX_LOCKDEP_CHAINS];
1986 int nr_chain_hlocks;
1987 static u16 chain_hlocks[MAX_LOCKDEP_CHAIN_HLOCKS];
1988
1989 struct lock_class *lock_chain_get_class(struct lock_chain *chain, int i)
1990 {
1991         return lock_classes + chain_hlocks[chain->base + i];
1992 }
1993
1994 /*
1995  * Look up a dependency chain. If the key is not present yet then
1996  * add it and return 1 - in this case the new dependency chain is
1997  * validated. If the key is already hashed, return 0.
1998  * (On return with 1 graph_lock is held.)
1999  */
2000 static inline int lookup_chain_cache(struct task_struct *curr,
2001                                      struct held_lock *hlock,
2002                                      u64 chain_key)
2003 {
2004         struct lock_class *class = hlock_class(hlock);
2005         struct list_head *hash_head = chainhashentry(chain_key);
2006         struct lock_chain *chain;
2007         struct held_lock *hlock_curr;
2008         int i, j;
2009
2010         /*
2011          * We might need to take the graph lock, ensure we've got IRQs
2012          * disabled to make this an IRQ-safe lock.. for recursion reasons
2013          * lockdep won't complain about its own locking errors.
2014          */
2015         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2016                 return 0;
2017         /*
2018          * We can walk it lock-free, because entries only get added
2019          * to the hash:
2020          */
2021         list_for_each_entry(chain, hash_head, entry) {
2022                 if (chain->chain_key == chain_key) {
2023 cache_hit:
2024                         debug_atomic_inc(chain_lookup_hits);
2025                         if (very_verbose(class))
2026                                 printk("\nhash chain already cached, key: "
2027                                         "%016Lx tail class: [%p] %s\n",
2028                                         (unsigned long long)chain_key,
2029                                         class->key, class->name);
2030                         return 0;
2031                 }
2032         }
2033         if (very_verbose(class))
2034                 printk("\nnew hash chain, key: %016Lx tail class: [%p] %s\n",
2035                         (unsigned long long)chain_key, class->key, class->name);
2036         /*
2037          * Allocate a new chain entry from the static array, and add
2038          * it to the hash:
2039          */
2040         if (!graph_lock())
2041                 return 0;
2042         /*
2043          * We have to walk the chain again locked - to avoid duplicates:
2044          */
2045         list_for_each_entry(chain, hash_head, entry) {
2046                 if (chain->chain_key == chain_key) {
2047                         graph_unlock();
2048                         goto cache_hit;
2049                 }
2050         }
2051         if (unlikely(nr_lock_chains >= MAX_LOCKDEP_CHAINS)) {
2052                 if (!debug_locks_off_graph_unlock())
2053                         return 0;
2054
2055                 print_lockdep_off("BUG: MAX_LOCKDEP_CHAINS too low!");
2056                 dump_stack();
2057                 return 0;
2058         }
2059         chain = lock_chains + nr_lock_chains++;
2060         chain->chain_key = chain_key;
2061         chain->irq_context = hlock->irq_context;
2062         /* Find the first held_lock of current chain */
2063         for (i = curr->lockdep_depth - 1; i >= 0; i--) {
2064                 hlock_curr = curr->held_locks + i;
2065                 if (hlock_curr->irq_context != hlock->irq_context)
2066                         break;
2067         }
2068         i++;
2069         chain->depth = curr->lockdep_depth + 1 - i;
2070         if (likely(nr_chain_hlocks + chain->depth <= MAX_LOCKDEP_CHAIN_HLOCKS)) {
2071                 chain->base = nr_chain_hlocks;
2072                 nr_chain_hlocks += chain->depth;
2073                 for (j = 0; j < chain->depth - 1; j++, i++) {
2074                         int lock_id = curr->held_locks[i].class_idx - 1;
2075                         chain_hlocks[chain->base + j] = lock_id;
2076                 }
2077                 chain_hlocks[chain->base + j] = class - lock_classes;
2078         }
2079         list_add_tail_rcu(&chain->entry, hash_head);
2080         debug_atomic_inc(chain_lookup_misses);
2081         inc_chains();
2082
2083         return 1;
2084 }
2085
2086 static int validate_chain(struct task_struct *curr, struct lockdep_map *lock,
2087                 struct held_lock *hlock, int chain_head, u64 chain_key)
2088 {
2089         /*
2090          * Trylock needs to maintain the stack of held locks, but it
2091          * does not add new dependencies, because trylock can be done
2092          * in any order.
2093          *
2094          * We look up the chain_key and do the O(N^2) check and update of
2095          * the dependencies only if this is a new dependency chain.
2096          * (If lookup_chain_cache() returns with 1 it acquires
2097          * graph_lock for us)
2098          */
2099         if (!hlock->trylock && (hlock->check == 2) &&
2100             lookup_chain_cache(curr, hlock, chain_key)) {
2101                 /*
2102                  * Check whether last held lock:
2103                  *
2104                  * - is irq-safe, if this lock is irq-unsafe
2105                  * - is softirq-safe, if this lock is hardirq-unsafe
2106                  *
2107                  * And check whether the new lock's dependency graph
2108                  * could lead back to the previous lock.
2109                  *
2110                  * any of these scenarios could lead to a deadlock. If
2111                  * All validations
2112                  */
2113                 int ret = check_deadlock(curr, hlock, lock, hlock->read);
2114
2115                 if (!ret)
2116                         return 0;
2117                 /*
2118                  * Mark recursive read, as we jump over it when
2119                  * building dependencies (just like we jump over
2120                  * trylock entries):
2121                  */
2122                 if (ret == 2)
2123                         hlock->read = 2;
2124                 /*
2125                  * Add dependency only if this lock is not the head
2126                  * of the chain, and if it's not a secondary read-lock:
2127                  */
2128                 if (!chain_head && ret != 2)
2129                         if (!check_prevs_add(curr, hlock))
2130                                 return 0;
2131                 graph_unlock();
2132         } else
2133                 /* after lookup_chain_cache(): */
2134                 if (unlikely(!debug_locks))
2135                         return 0;
2136
2137         return 1;
2138 }
2139 #else
2140 static inline int validate_chain(struct task_struct *curr,
2141                 struct lockdep_map *lock, struct held_lock *hlock,
2142                 int chain_head, u64 chain_key)
2143 {
2144         return 1;
2145 }
2146 #endif
2147
2148 /*
2149  * We are building curr_chain_key incrementally, so double-check
2150  * it from scratch, to make sure that it's done correctly:
2151  */
2152 static void check_chain_key(struct task_struct *curr)
2153 {
2154 #ifdef CONFIG_DEBUG_LOCKDEP
2155         struct held_lock *hlock, *prev_hlock = NULL;
2156         unsigned int i, id;
2157         u64 chain_key = 0;
2158
2159         for (i = 0; i < curr->lockdep_depth; i++) {
2160                 hlock = curr->held_locks + i;
2161                 if (chain_key != hlock->prev_chain_key) {
2162                         debug_locks_off();
2163                         /*
2164                          * We got mighty confused, our chain keys don't match
2165                          * with what we expect, someone trample on our task state?
2166                          */
2167                         WARN(1, "hm#1, depth: %u [%u], %016Lx != %016Lx\n",
2168                                 curr->lockdep_depth, i,
2169                                 (unsigned long long)chain_key,
2170                                 (unsigned long long)hlock->prev_chain_key);
2171                         return;
2172                 }
2173                 id = hlock->class_idx - 1;
2174                 /*
2175                  * Whoops ran out of static storage again?
2176                  */
2177                 if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
2178                         return;
2179
2180                 if (prev_hlock && (prev_hlock->irq_context !=
2181                                                         hlock->irq_context))
2182                         chain_key = 0;
2183                 chain_key = iterate_chain_key(chain_key, id);
2184                 prev_hlock = hlock;
2185         }
2186         if (chain_key != curr->curr_chain_key) {
2187                 debug_locks_off();
2188                 /*
2189                  * More smoking hash instead of calculating it, damn see these
2190                  * numbers float.. I bet that a pink elephant stepped on my memory.
2191                  */
2192                 WARN(1, "hm#2, depth: %u [%u], %016Lx != %016Lx\n",
2193                         curr->lockdep_depth, i,
2194                         (unsigned long long)chain_key,
2195                         (unsigned long long)curr->curr_chain_key);
2196         }
2197 #endif
2198 }
2199
2200 static void
2201 print_usage_bug_scenario(struct held_lock *lock)
2202 {
2203         struct lock_class *class = hlock_class(lock);
2204
2205         printk(" Possible unsafe locking scenario:\n\n");
2206         printk("       CPU0\n");
2207         printk("       ----\n");
2208         printk("  lock(");
2209         __print_lock_name(class);
2210         printk(");\n");
2211         printk("  <Interrupt>\n");
2212         printk("    lock(");
2213         __print_lock_name(class);
2214         printk(");\n");
2215         printk("\n *** DEADLOCK ***\n\n");
2216 }
2217
2218 static int
2219 print_usage_bug(struct task_struct *curr, struct held_lock *this,
2220                 enum lock_usage_bit prev_bit, enum lock_usage_bit new_bit)
2221 {
2222         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
2223                 return 0;
2224
2225         printk("\n");
2226         printk("=================================\n");
2227         printk("[ INFO: inconsistent lock state ]\n");
2228         print_kernel_ident();
2229         printk("---------------------------------\n");
2230
2231         printk("inconsistent {%s} -> {%s} usage.\n",
2232                 usage_str[prev_bit], usage_str[new_bit]);
2233
2234         printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] takes:\n",
2235                 curr->comm, task_pid_nr(curr),
2236                 trace_hardirq_context(curr), hardirq_count() >> HARDIRQ_SHIFT,
2237                 trace_softirq_context(curr), softirq_count() >> SOFTIRQ_SHIFT,
2238                 trace_hardirqs_enabled(curr),
2239                 trace_softirqs_enabled(curr));
2240         print_lock(this);
2241
2242         printk("{%s} state was registered at:\n", usage_str[prev_bit]);
2243         print_stack_trace(hlock_class(this)->usage_traces + prev_bit, 1);
2244
2245         print_irqtrace_events(curr);
2246         printk("\nother info that might help us debug this:\n");
2247         print_usage_bug_scenario(this);
2248
2249         lockdep_print_held_locks(curr);
2250
2251         printk("\nstack backtrace:\n");
2252         dump_stack();
2253
2254         return 0;
2255 }
2256
2257 /*
2258  * Print out an error if an invalid bit is set:
2259  */
2260 static inline int
2261 valid_state(struct task_struct *curr, struct held_lock *this,
2262             enum lock_usage_bit new_bit, enum lock_usage_bit bad_bit)
2263 {
2264         if (unlikely(hlock_class(this)->usage_mask & (1 << bad_bit)))
2265                 return print_usage_bug(curr, this, bad_bit, new_bit);
2266         return 1;
2267 }
2268
2269 static int mark_lock(struct task_struct *curr, struct held_lock *this,
2270                      enum lock_usage_bit new_bit);
2271
2272 #if defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING)
2273
2274 /*
2275  * print irq inversion bug:
2276  */
2277 static int
2278 print_irq_inversion_bug(struct task_struct *curr,
2279                         struct lock_list *root, struct lock_list *other,
2280                         struct held_lock *this, int forwards,
2281                         const char *irqclass)
2282 {
2283         struct lock_list *entry = other;
2284         struct lock_list *middle = NULL;
2285         int depth;
2286
2287         if (!debug_locks_off_graph_unlock() || debug_locks_silent)
2288                 return 0;
2289
2290         printk("\n");
2291         printk("=========================================================\n");
2292         printk("[ INFO: possible irq lock inversion dependency detected ]\n");
2293         print_kernel_ident();
2294         printk("---------------------------------------------------------\n");
2295         printk("%s/%d just changed the state of lock:\n",
2296                 curr->comm, task_pid_nr(curr));
2297         print_lock(this);
2298         if (forwards)
2299                 printk("but this lock took another, %s-unsafe lock in the past:\n", irqclass);
2300         else
2301                 printk("but this lock was taken by another, %s-safe lock in the past:\n", irqclass);
2302         print_lock_name(other->class);
2303         printk("\n\nand interrupts could create inverse lock ordering between them.\n\n");
2304
2305         printk("\nother info that might help us debug this:\n");
2306
2307         /* Find a middle lock (if one exists) */
2308         depth = get_lock_depth(other);
2309         do {
2310                 if (depth == 0 && (entry != root)) {
2311                         printk("lockdep:%s bad path found in chain graph\n", __func__);
2312                         break;
2313                 }
2314                 middle = entry;
2315                 entry = get_lock_parent(entry);
2316                 depth--;
2317         } while (entry && entry != root && (depth >= 0));
2318         if (forwards)
2319                 print_irq_lock_scenario(root, other,
2320                         middle ? middle->class : root->class, other->class);
2321         else
2322                 print_irq_lock_scenario(other, root,
2323                         middle ? middle->class : other->class, root->class);
2324
2325         lockdep_print_held_locks(curr);
2326
2327         printk("\nthe shortest dependencies between 2nd lock and 1st lock:\n");
2328         if (!save_trace(&root->trace))
2329                 return 0;
2330         print_shortest_lock_dependencies(other, root);
2331
2332         printk("\nstack backtrace:\n");
2333         dump_stack();
2334
2335         return 0;
2336 }
2337
2338 /*
2339  * Prove that in the forwards-direction subgraph starting at <this>
2340  * there is no lock matching <mask>:
2341  */
2342 static int
2343 check_usage_forwards(struct task_struct *curr, struct held_lock *this,
2344                      enum lock_usage_bit bit, const char *irqclass)
2345 {
2346         int ret;
2347         struct lock_list root;
2348         struct lock_list *uninitialized_var(target_entry);
2349
2350         root.parent = NULL;
2351         root.class = hlock_class(this);
2352         ret = find_usage_forwards(&root, bit, &target_entry);
2353         if (ret < 0)
2354                 return print_bfs_bug(ret);
2355         if (ret == 1)
2356                 return ret;
2357
2358         return print_irq_inversion_bug(curr, &root, target_entry,
2359                                         this, 1, irqclass);
2360 }
2361
2362 /*
2363  * Prove that in the backwards-direction subgraph starting at <this>
2364  * there is no lock matching <mask>:
2365  */
2366 static int
2367 check_usage_backwards(struct task_struct *curr, struct held_lock *this,
2368                       enum lock_usage_bit bit, const char *irqclass)
2369 {
2370         int ret;
2371         struct lock_list root;
2372         struct lock_list *uninitialized_var(target_entry);
2373
2374         root.parent = NULL;
2375         root.class = hlock_class(this);
2376         ret = find_usage_backwards(&root, bit, &target_entry);
2377         if (ret < 0)
2378                 return print_bfs_bug(ret);
2379         if (ret == 1)
2380                 return ret;
2381
2382         return print_irq_inversion_bug(curr, &root, target_entry,
2383                                         this, 0, irqclass);
2384 }
2385
2386 void print_irqtrace_events(struct task_struct *curr)
2387 {
2388         printk("irq event stamp: %u\n", curr->irq_events);
2389         printk("hardirqs last  enabled at (%u): ", curr->hardirq_enable_event);
2390         print_ip_sym(curr->hardirq_enable_ip);
2391         printk("hardirqs last disabled at (%u): ", curr->hardirq_disable_event);
2392         print_ip_sym(curr->hardirq_disable_ip);
2393         printk("softirqs last  enabled at (%u): ", curr->softirq_enable_event);
2394         print_ip_sym(curr->softirq_enable_ip);
2395         printk("softirqs last disabled at (%u): ", curr->softirq_disable_event);
2396         print_ip_sym(curr->softirq_disable_ip);
2397 }
2398
2399 static int HARDIRQ_verbose(struct lock_class *class)
2400 {
2401 #if HARDIRQ_VERBOSE
2402         return class_filter(class);
2403 #endif
2404         return 0;
2405 }
2406
2407 static int SOFTIRQ_verbose(struct lock_class *class)
2408 {
2409 #if SOFTIRQ_VERBOSE
2410         return class_filter(class);
2411 #endif
2412         return 0;
2413 }
2414
2415 static int RECLAIM_FS_verbose(struct lock_class *class)
2416 {
2417 #if RECLAIM_VERBOSE
2418         return class_filter(class);
2419 #endif
2420         return 0;
2421 }
2422
2423 #define STRICT_READ_CHECKS      1
2424
2425 static int (*state_verbose_f[])(struct lock_class *class) = {
2426 #define LOCKDEP_STATE(__STATE) \
2427         __STATE##_verbose,
2428 #include "lockdep_states.h"
2429 #undef LOCKDEP_STATE
2430 };
2431
2432 static inline int state_verbose(enum lock_usage_bit bit,
2433                                 struct lock_class *class)
2434 {
2435         return state_verbose_f[bit >> 2](class);
2436 }
2437
2438 typedef int (*check_usage_f)(struct task_struct *, struct held_lock *,
2439                              enum lock_usage_bit bit, const char *name);
2440
2441 static int
2442 mark_lock_irq(struct task_struct *curr, struct held_lock *this,
2443                 enum lock_usage_bit new_bit)
2444 {
2445         int excl_bit = exclusive_bit(new_bit);
2446         int read = new_bit & 1;
2447         int dir = new_bit & 2;
2448
2449         /*
2450          * mark USED_IN has to look forwards -- to ensure no dependency
2451          * has ENABLED state, which would allow recursion deadlocks.
2452          *
2453          * mark ENABLED has to look backwards -- to ensure no dependee
2454          * has USED_IN state, which, again, would allow  recursion deadlocks.
2455          */
2456         check_usage_f usage = dir ?
2457                 check_usage_backwards : check_usage_forwards;
2458
2459         /*
2460          * Validate that this particular lock does not have conflicting
2461          * usage states.
2462          */
2463         if (!valid_state(curr, this, new_bit, excl_bit))
2464                 return 0;
2465
2466         /*
2467          * Validate that the lock dependencies don't have conflicting usage
2468          * states.
2469          */
2470         if ((!read || !dir || STRICT_READ_CHECKS) &&
2471                         !usage(curr, this, excl_bit, state_name(new_bit & ~1)))
2472                 return 0;
2473
2474         /*
2475          * Check for read in write conflicts
2476          */
2477         if (!read) {
2478                 if (!valid_state(curr, this, new_bit, excl_bit + 1))
2479                         return 0;
2480
2481                 if (STRICT_READ_CHECKS &&
2482                         !usage(curr, this, excl_bit + 1,
2483                                 state_name(new_bit + 1)))
2484                         return 0;
2485         }
2486
2487         if (state_verbose(new_bit, hlock_class(this)))
2488                 return 2;
2489
2490         return 1;
2491 }
2492
2493 enum mark_type {
2494 #define LOCKDEP_STATE(__STATE)  __STATE,
2495 #include "lockdep_states.h"
2496 #undef LOCKDEP_STATE
2497 };
2498
2499 /*
2500  * Mark all held locks with a usage bit:
2501  */
2502 static int
2503 mark_held_locks(struct task_struct *curr, enum mark_type mark)
2504 {
2505         enum lock_usage_bit usage_bit;
2506         struct held_lock *hlock;
2507         int i;
2508
2509         for (i = 0; i < curr->lockdep_depth; i++) {
2510                 hlock = curr->held_locks + i;
2511
2512                 usage_bit = 2 + (mark << 2); /* ENABLED */
2513                 if (hlock->read)
2514                         usage_bit += 1; /* READ */
2515
2516                 BUG_ON(usage_bit >= LOCK_USAGE_STATES);
2517
2518                 if (hlock_class(hlock)->key == __lockdep_no_validate__.subkeys)
2519                         continue;
2520
2521                 if (!mark_lock(curr, hlock, usage_bit))
2522                         return 0;
2523         }
2524
2525         return 1;
2526 }
2527
2528 /*
2529  * Hardirqs will be enabled:
2530  */
2531 static void __trace_hardirqs_on_caller(unsigned long ip)
2532 {
2533         struct task_struct *curr = current;
2534
2535         /* we'll do an OFF -> ON transition: */
2536         curr->hardirqs_enabled = 1;
2537
2538         /*
2539          * We are going to turn hardirqs on, so set the
2540          * usage bit for all held locks:
2541          */
2542         if (!mark_held_locks(curr, HARDIRQ))
2543                 return;
2544         /*
2545          * If we have softirqs enabled, then set the usage
2546          * bit for all held locks. (disabled hardirqs prevented
2547          * this bit from being set before)
2548          */
2549         if (curr->softirqs_enabled)
2550                 if (!mark_held_locks(curr, SOFTIRQ))
2551                         return;
2552
2553         curr->hardirq_enable_ip = ip;
2554         curr->hardirq_enable_event = ++curr->irq_events;
2555         debug_atomic_inc(hardirqs_on_events);
2556 }
2557
2558 void trace_hardirqs_on_caller(unsigned long ip)
2559 {
2560         time_hardirqs_on(CALLER_ADDR0, ip);
2561
2562         if (unlikely(!debug_locks || current->lockdep_recursion))
2563                 return;
2564
2565         if (unlikely(current->hardirqs_enabled)) {
2566                 /*
2567                  * Neither irq nor preemption are disabled here
2568                  * so this is racy by nature but losing one hit
2569                  * in a stat is not a big deal.
2570                  */
2571                 __debug_atomic_inc(redundant_hardirqs_on);
2572                 return;
2573         }
2574
2575         /*
2576          * We're enabling irqs and according to our state above irqs weren't
2577          * already enabled, yet we find the hardware thinks they are in fact
2578          * enabled.. someone messed up their IRQ state tracing.
2579          */
2580         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2581                 return;
2582
2583         /*
2584          * See the fine text that goes along with this variable definition.
2585          */
2586         if (DEBUG_LOCKS_WARN_ON(unlikely(early_boot_irqs_disabled)))
2587                 return;
2588
2589         /*
2590          * Can't allow enabling interrupts while in an interrupt handler,
2591          * that's general bad form and such. Recursion, limited stack etc..
2592          */
2593         if (DEBUG_LOCKS_WARN_ON(current->hardirq_context))
2594                 return;
2595
2596         current->lockdep_recursion = 1;
2597         __trace_hardirqs_on_caller(ip);
2598         current->lockdep_recursion = 0;
2599 }
2600 EXPORT_SYMBOL(trace_hardirqs_on_caller);
2601
2602 void trace_hardirqs_on(void)
2603 {
2604         trace_hardirqs_on_caller(CALLER_ADDR0);
2605 }
2606 EXPORT_SYMBOL(trace_hardirqs_on);
2607
2608 /*
2609  * Hardirqs were disabled:
2610  */
2611 void trace_hardirqs_off_caller(unsigned long ip)
2612 {
2613         struct task_struct *curr = current;
2614
2615         time_hardirqs_off(CALLER_ADDR0, ip);
2616
2617         if (unlikely(!debug_locks || current->lockdep_recursion))
2618                 return;
2619
2620         /*
2621          * So we're supposed to get called after you mask local IRQs, but for
2622          * some reason the hardware doesn't quite think you did a proper job.
2623          */
2624         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2625                 return;
2626
2627         if (curr->hardirqs_enabled) {
2628                 /*
2629                  * We have done an ON -> OFF transition:
2630                  */
2631                 curr->hardirqs_enabled = 0;
2632                 curr->hardirq_disable_ip = ip;
2633                 curr->hardirq_disable_event = ++curr->irq_events;
2634                 debug_atomic_inc(hardirqs_off_events);
2635         } else
2636                 debug_atomic_inc(redundant_hardirqs_off);
2637 }
2638 EXPORT_SYMBOL(trace_hardirqs_off_caller);
2639
2640 void trace_hardirqs_off(void)
2641 {
2642         trace_hardirqs_off_caller(CALLER_ADDR0);
2643 }
2644 EXPORT_SYMBOL(trace_hardirqs_off);
2645
2646 /*
2647  * Softirqs will be enabled:
2648  */
2649 void trace_softirqs_on(unsigned long ip)
2650 {
2651         struct task_struct *curr = current;
2652
2653         if (unlikely(!debug_locks || current->lockdep_recursion))
2654                 return;
2655
2656         /*
2657          * We fancy IRQs being disabled here, see softirq.c, avoids
2658          * funny state and nesting things.
2659          */
2660         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2661                 return;
2662
2663         if (curr->softirqs_enabled) {
2664                 debug_atomic_inc(redundant_softirqs_on);
2665                 return;
2666         }
2667
2668         current->lockdep_recursion = 1;
2669         /*
2670          * We'll do an OFF -> ON transition:
2671          */
2672         curr->softirqs_enabled = 1;
2673         curr->softirq_enable_ip = ip;
2674         curr->softirq_enable_event = ++curr->irq_events;
2675         debug_atomic_inc(softirqs_on_events);
2676         /*
2677          * We are going to turn softirqs on, so set the
2678          * usage bit for all held locks, if hardirqs are
2679          * enabled too:
2680          */
2681         if (curr->hardirqs_enabled)
2682                 mark_held_locks(curr, SOFTIRQ);
2683         current->lockdep_recursion = 0;
2684 }
2685
2686 /*
2687  * Softirqs were disabled:
2688  */
2689 void trace_softirqs_off(unsigned long ip)
2690 {
2691         struct task_struct *curr = current;
2692
2693         if (unlikely(!debug_locks || current->lockdep_recursion))
2694                 return;
2695
2696         /*
2697          * We fancy IRQs being disabled here, see softirq.c
2698          */
2699         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
2700                 return;
2701
2702         if (curr->softirqs_enabled) {
2703                 /*
2704                  * We have done an ON -> OFF transition:
2705                  */
2706                 curr->softirqs_enabled = 0;
2707                 curr->softirq_disable_ip = ip;
2708                 curr->softirq_disable_event = ++curr->irq_events;
2709                 debug_atomic_inc(softirqs_off_events);
2710                 /*
2711                  * Whoops, we wanted softirqs off, so why aren't they?
2712                  */
2713                 DEBUG_LOCKS_WARN_ON(!softirq_count());
2714         } else
2715                 debug_atomic_inc(redundant_softirqs_off);
2716 }
2717
2718 static void __lockdep_trace_alloc(gfp_t gfp_mask, unsigned long flags)
2719 {
2720         struct task_struct *curr = current;
2721
2722         if (unlikely(!debug_locks))
2723                 return;
2724
2725         /* no reclaim without waiting on it */
2726         if (!(gfp_mask & __GFP_WAIT))
2727                 return;
2728
2729         /* this guy won't enter reclaim */
2730         if ((curr->flags & PF_MEMALLOC) && !(gfp_mask & __GFP_NOMEMALLOC))
2731                 return;
2732
2733         /* We're only interested __GFP_FS allocations for now */
2734         if (!(gfp_mask & __GFP_FS))
2735                 return;
2736
2737         /*
2738          * Oi! Can't be having __GFP_FS allocations with IRQs disabled.
2739          */
2740         if (DEBUG_LOCKS_WARN_ON(irqs_disabled_flags(flags)))
2741                 return;
2742
2743         mark_held_locks(curr, RECLAIM_FS);
2744 }
2745
2746 static void check_flags(unsigned long flags);
2747
2748 void lockdep_trace_alloc(gfp_t gfp_mask)
2749 {
2750         unsigned long flags;
2751
2752         if (unlikely(current->lockdep_recursion))
2753                 return;
2754
2755         raw_local_irq_save(flags);
2756         check_flags(flags);
2757         current->lockdep_recursion = 1;
2758         __lockdep_trace_alloc(gfp_mask, flags);
2759         current->lockdep_recursion = 0;
2760         raw_local_irq_restore(flags);
2761 }
2762
2763 static int mark_irqflags(struct task_struct *curr, struct held_lock *hlock)
2764 {
2765         /*
2766          * If non-trylock use in a hardirq or softirq context, then
2767          * mark the lock as used in these contexts:
2768          */
2769         if (!hlock->trylock) {
2770                 if (hlock->read) {
2771                         if (curr->hardirq_context)
2772                                 if (!mark_lock(curr, hlock,
2773                                                 LOCK_USED_IN_HARDIRQ_READ))
2774                                         return 0;
2775                         if (curr->softirq_context)
2776                                 if (!mark_lock(curr, hlock,
2777                                                 LOCK_USED_IN_SOFTIRQ_READ))
2778                                         return 0;
2779                 } else {
2780                         if (curr->hardirq_context)
2781                                 if (!mark_lock(curr, hlock, LOCK_USED_IN_HARDIRQ))
2782                                         return 0;
2783                         if (curr->softirq_context)
2784                                 if (!mark_lock(curr, hlock, LOCK_USED_IN_SOFTIRQ))
2785                                         return 0;
2786                 }
2787         }
2788         if (!hlock->hardirqs_off) {
2789                 if (hlock->read) {
2790                         if (!mark_lock(curr, hlock,
2791                                         LOCK_ENABLED_HARDIRQ_READ))
2792                                 return 0;
2793                         if (curr->softirqs_enabled)
2794                                 if (!mark_lock(curr, hlock,
2795                                                 LOCK_ENABLED_SOFTIRQ_READ))
2796                                         return 0;
2797                 } else {
2798                         if (!mark_lock(curr, hlock,
2799                                         LOCK_ENABLED_HARDIRQ))
2800                                 return 0;
2801                         if (curr->softirqs_enabled)
2802                                 if (!mark_lock(curr, hlock,
2803                                                 LOCK_ENABLED_SOFTIRQ))
2804                                         return 0;
2805                 }
2806         }
2807
2808         /*
2809          * We reuse the irq context infrastructure more broadly as a general
2810          * context checking code. This tests GFP_FS recursion (a lock taken
2811          * during reclaim for a GFP_FS allocation is held over a GFP_FS
2812          * allocation).
2813          */
2814         if (!hlock->trylock && (curr->lockdep_reclaim_gfp & __GFP_FS)) {
2815                 if (hlock->read) {
2816                         if (!mark_lock(curr, hlock, LOCK_USED_IN_RECLAIM_FS_READ))
2817                                         return 0;
2818                 } else {
2819                         if (!mark_lock(curr, hlock, LOCK_USED_IN_RECLAIM_FS))
2820                                         return 0;
2821                 }
2822         }
2823
2824         return 1;
2825 }
2826
2827 static int separate_irq_context(struct task_struct *curr,
2828                 struct held_lock *hlock)
2829 {
2830         unsigned int depth = curr->lockdep_depth;
2831
2832         /*
2833          * Keep track of points where we cross into an interrupt context:
2834          */
2835         hlock->irq_context = 2*(curr->hardirq_context ? 1 : 0) +
2836                                 curr->softirq_context;
2837         if (depth) {
2838                 struct held_lock *prev_hlock;
2839
2840                 prev_hlock = curr->held_locks + depth-1;
2841                 /*
2842                  * If we cross into another context, reset the
2843                  * hash key (this also prevents the checking and the
2844                  * adding of the dependency to 'prev'):
2845                  */
2846                 if (prev_hlock->irq_context != hlock->irq_context)
2847                         return 1;
2848         }
2849         return 0;
2850 }
2851
2852 #else /* defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING) */
2853
2854 static inline
2855 int mark_lock_irq(struct task_struct *curr, struct held_lock *this,
2856                 enum lock_usage_bit new_bit)
2857 {
2858         WARN_ON(1); /* Impossible innit? when we don't have TRACE_IRQFLAG */
2859         return 1;
2860 }
2861
2862 static inline int mark_irqflags(struct task_struct *curr,
2863                 struct held_lock *hlock)
2864 {
2865         return 1;
2866 }
2867
2868 static inline int separate_irq_context(struct task_struct *curr,
2869                 struct held_lock *hlock)
2870 {
2871         return 0;
2872 }
2873
2874 void lockdep_trace_alloc(gfp_t gfp_mask)
2875 {
2876 }
2877
2878 #endif /* defined(CONFIG_TRACE_IRQFLAGS) && defined(CONFIG_PROVE_LOCKING) */
2879
2880 /*
2881  * Mark a lock with a usage bit, and validate the state transition:
2882  */
2883 static int mark_lock(struct task_struct *curr, struct held_lock *this,
2884                              enum lock_usage_bit new_bit)
2885 {
2886         unsigned int new_mask = 1 << new_bit, ret = 1;
2887
2888         /*
2889          * If already set then do not dirty the cacheline,
2890          * nor do any checks:
2891          */
2892         if (likely(hlock_class(this)->usage_mask & new_mask))
2893                 return 1;
2894
2895         if (!graph_lock())
2896                 return 0;
2897         /*
2898          * Make sure we didn't race:
2899          */
2900         if (unlikely(hlock_class(this)->usage_mask & new_mask)) {
2901                 graph_unlock();
2902                 return 1;
2903         }
2904
2905         hlock_class(this)->usage_mask |= new_mask;
2906
2907         if (!save_trace(hlock_class(this)->usage_traces + new_bit))
2908                 return 0;
2909
2910         switch (new_bit) {
2911 #define LOCKDEP_STATE(__STATE)                  \
2912         case LOCK_USED_IN_##__STATE:            \
2913         case LOCK_USED_IN_##__STATE##_READ:     \
2914         case LOCK_ENABLED_##__STATE:            \
2915         case LOCK_ENABLED_##__STATE##_READ:
2916 #include "lockdep_states.h"
2917 #undef LOCKDEP_STATE
2918                 ret = mark_lock_irq(curr, this, new_bit);
2919                 if (!ret)
2920                         return 0;
2921                 break;
2922         case LOCK_USED:
2923                 debug_atomic_dec(nr_unused_locks);
2924                 break;
2925         default:
2926                 if (!debug_locks_off_graph_unlock())
2927                         return 0;
2928                 WARN_ON(1);
2929                 return 0;
2930         }
2931
2932         graph_unlock();
2933
2934         /*
2935          * We must printk outside of the graph_lock:
2936          */
2937         if (ret == 2) {
2938                 printk("\nmarked lock as {%s}:\n", usage_str[new_bit]);
2939                 print_lock(this);
2940                 print_irqtrace_events(curr);
2941                 dump_stack();
2942         }
2943
2944         return ret;
2945 }
2946
2947 /*
2948  * Initialize a lock instance's lock-class mapping info:
2949  */
2950 void lockdep_init_map(struct lockdep_map *lock, const char *name,
2951                       struct lock_class_key *key, int subclass)
2952 {
2953         int i;
2954
2955         kmemcheck_mark_initialized(lock, sizeof(*lock));
2956
2957         for (i = 0; i < NR_LOCKDEP_CACHING_CLASSES; i++)
2958                 lock->class_cache[i] = NULL;
2959
2960 #ifdef CONFIG_LOCK_STAT
2961         lock->cpu = raw_smp_processor_id();
2962 #endif
2963
2964         /*
2965          * Can't be having no nameless bastards around this place!
2966          */
2967         if (DEBUG_LOCKS_WARN_ON(!name)) {
2968                 lock->name = "NULL";
2969                 return;
2970         }
2971
2972         lock->name = name;
2973
2974         /*
2975          * No key, no joy, we need to hash something.
2976          */
2977         if (DEBUG_LOCKS_WARN_ON(!key))
2978                 return;
2979         /*
2980          * Sanity check, the lock-class key must be persistent:
2981          */
2982         if (!static_obj(key)) {
2983                 printk("BUG: key %p not in .data!\n", key);
2984                 /*
2985                  * What it says above ^^^^^, I suggest you read it.
2986                  */
2987                 DEBUG_LOCKS_WARN_ON(1);
2988                 return;
2989         }
2990         lock->key = key;
2991
2992         if (unlikely(!debug_locks))
2993                 return;
2994
2995         if (subclass)
2996                 register_lock_class(lock, subclass, 1);
2997 }
2998 EXPORT_SYMBOL_GPL(lockdep_init_map);
2999
3000 struct lock_class_key __lockdep_no_validate__;
3001 EXPORT_SYMBOL_GPL(__lockdep_no_validate__);
3002
3003 static int
3004 print_lock_nested_lock_not_held(struct task_struct *curr,
3005                                 struct held_lock *hlock,
3006                                 unsigned long ip)
3007 {
3008         if (!debug_locks_off())
3009                 return 0;
3010         if (debug_locks_silent)
3011                 return 0;
3012
3013         printk("\n");
3014         printk("==================================\n");
3015         printk("[ BUG: Nested lock was not taken ]\n");
3016         print_kernel_ident();
3017         printk("----------------------------------\n");
3018
3019         printk("%s/%d is trying to lock:\n", curr->comm, task_pid_nr(curr));
3020         print_lock(hlock);
3021
3022         printk("\nbut this task is not holding:\n");
3023         printk("%s\n", hlock->nest_lock->name);
3024
3025         printk("\nstack backtrace:\n");
3026         dump_stack();
3027
3028         printk("\nother info that might help us debug this:\n");
3029         lockdep_print_held_locks(curr);
3030
3031         printk("\nstack backtrace:\n");
3032         dump_stack();
3033
3034         return 0;
3035 }
3036
3037 static int __lock_is_held(struct lockdep_map *lock);
3038
3039 /*
3040  * This gets called for every mutex_lock*()/spin_lock*() operation.
3041  * We maintain the dependency maps and validate the locking attempt:
3042  */
3043 static int __lock_acquire(struct lockdep_map *lock, unsigned int subclass,
3044                           int trylock, int read, int check, int hardirqs_off,
3045                           struct lockdep_map *nest_lock, unsigned long ip,
3046                           int references)
3047 {
3048         struct task_struct *curr = current;
3049         struct lock_class *class = NULL;
3050         struct held_lock *hlock;
3051         unsigned int depth, id;
3052         int chain_head = 0;
3053         int class_idx;
3054         u64 chain_key;
3055
3056         if (!prove_locking)
3057                 check = 1;
3058
3059         if (unlikely(!debug_locks))
3060                 return 0;
3061
3062         /*
3063          * Lockdep should run with IRQs disabled, otherwise we could
3064          * get an interrupt which would want to take locks, which would
3065          * end up in lockdep and have you got a head-ache already?
3066          */
3067         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
3068                 return 0;
3069
3070         if (lock->key == &__lockdep_no_validate__)
3071                 check = 1;
3072
3073         if (subclass < NR_LOCKDEP_CACHING_CLASSES)
3074                 class = lock->class_cache[subclass];
3075         /*
3076          * Not cached?
3077          */
3078         if (unlikely(!class)) {
3079                 class = register_lock_class(lock, subclass, 0);
3080                 if (!class)
3081                         return 0;
3082         }
3083         atomic_inc((atomic_t *)&class->ops);
3084         if (very_verbose(class)) {
3085                 printk("\nacquire class [%p] %s", class->key, class->name);
3086                 if (class->name_version > 1)
3087                         printk("#%d", class->name_version);
3088                 printk("\n");
3089                 dump_stack();
3090         }
3091
3092         /*
3093          * Add the lock to the list of currently held locks.
3094          * (we dont increase the depth just yet, up until the
3095          * dependency checks are done)
3096          */
3097         depth = curr->lockdep_depth;
3098         /*
3099          * Ran out of static storage for our per-task lock stack again have we?
3100          */
3101         if (DEBUG_LOCKS_WARN_ON(depth >= MAX_LOCK_DEPTH))
3102                 return 0;
3103
3104         class_idx = class - lock_classes + 1;
3105
3106         if (depth) {
3107                 hlock = curr->held_locks + depth - 1;
3108                 if (hlock->class_idx == class_idx && nest_lock) {
3109                         if (hlock->references)
3110                                 hlock->references++;
3111                         else
3112                                 hlock->references = 2;
3113
3114                         return 1;
3115                 }
3116         }
3117
3118         hlock = curr->held_locks + depth;
3119         /*
3120          * Plain impossible, we just registered it and checked it weren't no
3121          * NULL like.. I bet this mushroom I ate was good!
3122          */
3123         if (DEBUG_LOCKS_WARN_ON(!class))
3124                 return 0;
3125         hlock->class_idx = class_idx;
3126         hlock->acquire_ip = ip;
3127         hlock->instance = lock;
3128         hlock->nest_lock = nest_lock;
3129         hlock->trylock = trylock;
3130         hlock->read = read;
3131         hlock->check = check;
3132         hlock->hardirqs_off = !!hardirqs_off;
3133         hlock->references = references;
3134 #ifdef CONFIG_LOCK_STAT
3135         hlock->waittime_stamp = 0;
3136         hlock->holdtime_stamp = lockstat_clock();
3137 #endif
3138
3139         if (check == 2 && !mark_irqflags(curr, hlock))
3140                 return 0;
3141
3142         /* mark it as used: */
3143         if (!mark_lock(curr, hlock, LOCK_USED))
3144                 return 0;
3145
3146         /*
3147          * Calculate the chain hash: it's the combined hash of all the
3148          * lock keys along the dependency chain. We save the hash value
3149          * at every step so that we can get the current hash easily
3150          * after unlock. The chain hash is then used to cache dependency
3151          * results.
3152          *
3153          * The 'key ID' is what is the most compact key value to drive
3154          * the hash, not class->key.
3155          */
3156         id = class - lock_classes;
3157         /*
3158          * Whoops, we did it again.. ran straight out of our static allocation.
3159          */
3160         if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
3161                 return 0;
3162
3163         chain_key = curr->curr_chain_key;
3164         if (!depth) {
3165                 /*
3166                  * How can we have a chain hash when we ain't got no keys?!
3167                  */
3168                 if (DEBUG_LOCKS_WARN_ON(chain_key != 0))
3169                         return 0;
3170                 chain_head = 1;
3171         }
3172
3173         hlock->prev_chain_key = chain_key;
3174         if (separate_irq_context(curr, hlock)) {
3175                 chain_key = 0;
3176                 chain_head = 1;
3177         }
3178         chain_key = iterate_chain_key(chain_key, id);
3179
3180         if (nest_lock && !__lock_is_held(nest_lock))
3181                 return print_lock_nested_lock_not_held(curr, hlock, ip);
3182
3183         if (!validate_chain(curr, lock, hlock, chain_head, chain_key))
3184                 return 0;
3185
3186         curr->curr_chain_key = chain_key;
3187         curr->lockdep_depth++;
3188         check_chain_key(curr);
3189 #ifdef CONFIG_DEBUG_LOCKDEP
3190         if (unlikely(!debug_locks))
3191                 return 0;
3192 #endif
3193         if (unlikely(curr->lockdep_depth >= MAX_LOCK_DEPTH)) {
3194                 debug_locks_off();
3195                 print_lockdep_off("BUG: MAX_LOCK_DEPTH too low!");
3196                 printk(KERN_DEBUG "depth: %i  max: %lu!\n",
3197                        curr->lockdep_depth, MAX_LOCK_DEPTH);
3198
3199                 lockdep_print_held_locks(current);
3200                 debug_show_all_locks();
3201                 dump_stack();
3202
3203                 return 0;
3204         }
3205
3206         if (unlikely(curr->lockdep_depth > max_lockdep_depth))
3207                 max_lockdep_depth = curr->lockdep_depth;
3208
3209         return 1;
3210 }
3211
3212 static int
3213 print_unlock_imbalance_bug(struct task_struct *curr, struct lockdep_map *lock,
3214                            unsigned long ip)
3215 {
3216         if (!debug_locks_off())
3217                 return 0;
3218         if (debug_locks_silent)
3219                 return 0;
3220
3221         printk("\n");
3222         printk("=====================================\n");
3223         printk("[ BUG: bad unlock balance detected! ]\n");
3224         print_kernel_ident();
3225         printk("-------------------------------------\n");
3226         printk("%s/%d is trying to release lock (",
3227                 curr->comm, task_pid_nr(curr));
3228         print_lockdep_cache(lock);
3229         printk(") at:\n");
3230         print_ip_sym(ip);
3231         printk("but there are no more locks to release!\n");
3232         printk("\nother info that might help us debug this:\n");
3233         lockdep_print_held_locks(curr);
3234
3235         printk("\nstack backtrace:\n");
3236         dump_stack();
3237
3238         return 0;
3239 }
3240
3241 /*
3242  * Common debugging checks for both nested and non-nested unlock:
3243  */
3244 static int check_unlock(struct task_struct *curr, struct lockdep_map *lock,
3245                         unsigned long ip)
3246 {
3247         if (unlikely(!debug_locks))
3248                 return 0;
3249         /*
3250          * Lockdep should run with IRQs disabled, recursion, head-ache, etc..
3251          */
3252         if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
3253                 return 0;
3254
3255         if (curr->lockdep_depth <= 0)
3256                 return print_unlock_imbalance_bug(curr, lock, ip);
3257
3258         return 1;
3259 }
3260
3261 static int match_held_lock(struct held_lock *hlock, struct lockdep_map *lock)
3262 {
3263         if (hlock->instance == lock)
3264                 return 1;
3265
3266         if (hlock->references) {
3267                 struct lock_class *class = lock->class_cache[0];
3268
3269                 if (!class)
3270                         class = look_up_lock_class(lock, 0);
3271
3272                 /*
3273                  * If look_up_lock_class() failed to find a class, we're trying
3274                  * to test if we hold a lock that has never yet been acquired.
3275                  * Clearly if the lock hasn't been acquired _ever_, we're not
3276                  * holding it either, so report failure.
3277                  */
3278                 if (!class)
3279                         return 0;
3280
3281                 /*
3282                  * References, but not a lock we're actually ref-counting?
3283                  * State got messed up, follow the sites that change ->references
3284                  * and try to make sense of it.
3285                  */
3286                 if (DEBUG_LOCKS_WARN_ON(!hlock->nest_lock))
3287                         return 0;
3288
3289                 if (hlock->class_idx == class - lock_classes + 1)
3290                         return 1;
3291         }
3292
3293         return 0;
3294 }
3295
3296 static int
3297 __lock_set_class(struct lockdep_map *lock, const char *name,
3298                  struct lock_class_key *key, unsigned int subclass,
3299                  unsigned long ip)
3300 {
3301         struct task_struct *curr = current;
3302         struct held_lock *hlock, *prev_hlock;
3303         struct lock_class *class;
3304         unsigned int depth;
3305         int i;
3306
3307         depth = curr->lockdep_depth;
3308         /*
3309          * This function is about (re)setting the class of a held lock,
3310          * yet we're not actually holding any locks. Naughty user!
3311          */
3312         if (DEBUG_LOCKS_WARN_ON(!depth))
3313                 return 0;
3314
3315         prev_hlock = NULL;
3316         for (i = depth-1; i >= 0; i--) {
3317                 hlock = curr->held_locks + i;
3318                 /*
3319                  * We must not cross into another context:
3320                  */
3321                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3322                         break;
3323                 if (match_held_lock(hlock, lock))
3324                         goto found_it;
3325                 prev_hlock = hlock;
3326         }
3327         return print_unlock_imbalance_bug(curr, lock, ip);
3328
3329 found_it:
3330         lockdep_init_map(lock, name, key, 0);
3331         class = register_lock_class(lock, subclass, 0);
3332         hlock->class_idx = class - lock_classes + 1;
3333
3334         curr->lockdep_depth = i;
3335         curr->curr_chain_key = hlock->prev_chain_key;
3336
3337         for (; i < depth; i++) {
3338                 hlock = curr->held_locks + i;
3339                 if (!__lock_acquire(hlock->instance,
3340                         hlock_class(hlock)->subclass, hlock->trylock,
3341                                 hlock->read, hlock->check, hlock->hardirqs_off,
3342                                 hlock->nest_lock, hlock->acquire_ip,
3343                                 hlock->references))
3344                         return 0;
3345         }
3346
3347         /*
3348          * I took it apart and put it back together again, except now I have
3349          * these 'spare' parts.. where shall I put them.
3350          */
3351         if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth))
3352                 return 0;
3353         return 1;
3354 }
3355
3356 /*
3357  * Remove the lock to the list of currently held locks in a
3358  * potentially non-nested (out of order) manner. This is a
3359  * relatively rare operation, as all the unlock APIs default
3360  * to nested mode (which uses lock_release()):
3361  */
3362 static int
3363 lock_release_non_nested(struct task_struct *curr,
3364                         struct lockdep_map *lock, unsigned long ip)
3365 {
3366         struct held_lock *hlock, *prev_hlock;
3367         unsigned int depth;
3368         int i;
3369
3370         /*
3371          * Check whether the lock exists in the current stack
3372          * of held locks:
3373          */
3374         depth = curr->lockdep_depth;
3375         /*
3376          * So we're all set to release this lock.. wait what lock? We don't
3377          * own any locks, you've been drinking again?
3378          */
3379         if (DEBUG_LOCKS_WARN_ON(!depth))
3380                 return 0;
3381
3382         prev_hlock = NULL;
3383         for (i = depth-1; i >= 0; i--) {
3384                 hlock = curr->held_locks + i;
3385                 /*
3386                  * We must not cross into another context:
3387                  */
3388                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3389                         break;
3390                 if (match_held_lock(hlock, lock))
3391                         goto found_it;
3392                 prev_hlock = hlock;
3393         }
3394         return print_unlock_imbalance_bug(curr, lock, ip);
3395
3396 found_it:
3397         if (hlock->instance == lock)
3398                 lock_release_holdtime(hlock);
3399
3400         if (hlock->references) {
3401                 hlock->references--;
3402                 if (hlock->references) {
3403                         /*
3404                          * We had, and after removing one, still have
3405                          * references, the current lock stack is still
3406                          * valid. We're done!
3407                          */
3408                         return 1;
3409                 }
3410         }
3411
3412         /*
3413          * We have the right lock to unlock, 'hlock' points to it.
3414          * Now we remove it from the stack, and add back the other
3415          * entries (if any), recalculating the hash along the way:
3416          */
3417
3418         curr->lockdep_depth = i;
3419         curr->curr_chain_key = hlock->prev_chain_key;
3420
3421         for (i++; i < depth; i++) {
3422                 hlock = curr->held_locks + i;
3423                 if (!__lock_acquire(hlock->instance,
3424                         hlock_class(hlock)->subclass, hlock->trylock,
3425                                 hlock->read, hlock->check, hlock->hardirqs_off,
3426                                 hlock->nest_lock, hlock->acquire_ip,
3427                                 hlock->references))
3428                         return 0;
3429         }
3430
3431         /*
3432          * We had N bottles of beer on the wall, we drank one, but now
3433          * there's not N-1 bottles of beer left on the wall...
3434          */
3435         if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth - 1))
3436                 return 0;
3437         return 1;
3438 }
3439
3440 /*
3441  * Remove the lock to the list of currently held locks - this gets
3442  * called on mutex_unlock()/spin_unlock*() (or on a failed
3443  * mutex_lock_interruptible()). This is done for unlocks that nest
3444  * perfectly. (i.e. the current top of the lock-stack is unlocked)
3445  */
3446 static int lock_release_nested(struct task_struct *curr,
3447                                struct lockdep_map *lock, unsigned long ip)
3448 {
3449         struct held_lock *hlock;
3450         unsigned int depth;
3451
3452         /*
3453          * Pop off the top of the lock stack:
3454          */
3455         depth = curr->lockdep_depth - 1;
3456         hlock = curr->held_locks + depth;
3457
3458         /*
3459          * Is the unlock non-nested:
3460          */
3461         if (hlock->instance != lock || hlock->references)
3462                 return lock_release_non_nested(curr, lock, ip);
3463         curr->lockdep_depth--;
3464
3465         /*
3466          * No more locks, but somehow we've got hash left over, who left it?
3467          */
3468         if (DEBUG_LOCKS_WARN_ON(!depth && (hlock->prev_chain_key != 0)))
3469                 return 0;
3470
3471         curr->curr_chain_key = hlock->prev_chain_key;
3472
3473         lock_release_holdtime(hlock);
3474
3475 #ifdef CONFIG_DEBUG_LOCKDEP
3476         hlock->prev_chain_key = 0;
3477         hlock->class_idx = 0;
3478         hlock->acquire_ip = 0;
3479         hlock->irq_context = 0;
3480 #endif
3481         return 1;
3482 }
3483
3484 /*
3485  * Remove the lock to the list of currently held locks - this gets
3486  * called on mutex_unlock()/spin_unlock*() (or on a failed
3487  * mutex_lock_interruptible()). This is done for unlocks that nest
3488  * perfectly. (i.e. the current top of the lock-stack is unlocked)
3489  */
3490 static void
3491 __lock_release(struct lockdep_map *lock, int nested, unsigned long ip)
3492 {
3493         struct task_struct *curr = current;
3494
3495         if (!check_unlock(curr, lock, ip))
3496                 return;
3497
3498         if (nested) {
3499                 if (!lock_release_nested(curr, lock, ip))
3500                         return;
3501         } else {
3502                 if (!lock_release_non_nested(curr, lock, ip))
3503                         return;
3504         }
3505
3506         check_chain_key(curr);
3507 }
3508
3509 static int __lock_is_held(struct lockdep_map *lock)
3510 {
3511         struct task_struct *curr = current;
3512         int i;
3513
3514         for (i = 0; i < curr->lockdep_depth; i++) {
3515                 struct held_lock *hlock = curr->held_locks + i;
3516
3517                 if (match_held_lock(hlock, lock))
3518                         return 1;
3519         }
3520
3521         return 0;
3522 }
3523
3524 /*
3525  * Check whether we follow the irq-flags state precisely:
3526  */
3527 static void check_flags(unsigned long flags)
3528 {
3529 #if defined(CONFIG_PROVE_LOCKING) && defined(CONFIG_DEBUG_LOCKDEP) && \
3530     defined(CONFIG_TRACE_IRQFLAGS)
3531         if (!debug_locks)
3532                 return;
3533
3534         if (irqs_disabled_flags(flags)) {
3535                 if (DEBUG_LOCKS_WARN_ON(current->hardirqs_enabled)) {
3536                         printk("possible reason: unannotated irqs-off.\n");
3537                 }
3538         } else {
3539                 if (DEBUG_LOCKS_WARN_ON(!current->hardirqs_enabled)) {
3540                         printk("possible reason: unannotated irqs-on.\n");
3541                 }
3542         }
3543
3544         /*
3545          * We dont accurately track softirq state in e.g.
3546          * hardirq contexts (such as on 4KSTACKS), so only
3547          * check if not in hardirq contexts:
3548          */
3549         if (!hardirq_count()) {
3550                 if (softirq_count()) {
3551                         /* like the above, but with softirqs */
3552                         DEBUG_LOCKS_WARN_ON(current->softirqs_enabled);
3553                 } else {
3554                         /* lick the above, does it taste good? */
3555                         DEBUG_LOCKS_WARN_ON(!current->softirqs_enabled);
3556                 }
3557         }
3558
3559         if (!debug_locks)
3560                 print_irqtrace_events(current);
3561 #endif
3562 }
3563
3564 void lock_set_class(struct lockdep_map *lock, const char *name,
3565                     struct lock_class_key *key, unsigned int subclass,
3566                     unsigned long ip)
3567 {
3568         unsigned long flags;
3569
3570         if (unlikely(current->lockdep_recursion))
3571                 return;
3572
3573         raw_local_irq_save(flags);
3574         current->lockdep_recursion = 1;
3575         check_flags(flags);
3576         if (__lock_set_class(lock, name, key, subclass, ip))
3577                 check_chain_key(current);
3578         current->lockdep_recursion = 0;
3579         raw_local_irq_restore(flags);
3580 }
3581 EXPORT_SYMBOL_GPL(lock_set_class);
3582
3583 /*
3584  * We are not always called with irqs disabled - do that here,
3585  * and also avoid lockdep recursion:
3586  */
3587 void lock_acquire(struct lockdep_map *lock, unsigned int subclass,
3588                           int trylock, int read, int check,
3589                           struct lockdep_map *nest_lock, unsigned long ip)
3590 {
3591         unsigned long flags;
3592
3593         if (unlikely(current->lockdep_recursion))
3594                 return;
3595
3596         raw_local_irq_save(flags);
3597         check_flags(flags);
3598
3599         current->lockdep_recursion = 1;
3600         trace_lock_acquire(lock, subclass, trylock, read, check, nest_lock, ip);
3601         __lock_acquire(lock, subclass, trylock, read, check,
3602                        irqs_disabled_flags(flags), nest_lock, ip, 0);
3603         current->lockdep_recursion = 0;
3604         raw_local_irq_restore(flags);
3605 }
3606 EXPORT_SYMBOL_GPL(lock_acquire);
3607
3608 void lock_release(struct lockdep_map *lock, int nested,
3609                           unsigned long ip)
3610 {
3611         unsigned long flags;
3612
3613         if (unlikely(current->lockdep_recursion))
3614                 return;
3615
3616         raw_local_irq_save(flags);
3617         check_flags(flags);
3618         current->lockdep_recursion = 1;
3619         trace_lock_release(lock, ip);
3620         __lock_release(lock, nested, ip);
3621         current->lockdep_recursion = 0;
3622         raw_local_irq_restore(flags);
3623 }
3624 EXPORT_SYMBOL_GPL(lock_release);
3625
3626 int lock_is_held(struct lockdep_map *lock)
3627 {
3628         unsigned long flags;
3629         int ret = 0;
3630
3631         if (unlikely(current->lockdep_recursion))
3632                 return 1; /* avoid false negative lockdep_assert_held() */
3633
3634         raw_local_irq_save(flags);
3635         check_flags(flags);
3636
3637         current->lockdep_recursion = 1;
3638         ret = __lock_is_held(lock);
3639         current->lockdep_recursion = 0;
3640         raw_local_irq_restore(flags);
3641
3642         return ret;
3643 }
3644 EXPORT_SYMBOL_GPL(lock_is_held);
3645
3646 void lockdep_set_current_reclaim_state(gfp_t gfp_mask)
3647 {
3648         current->lockdep_reclaim_gfp = gfp_mask;
3649 }
3650
3651 void lockdep_clear_current_reclaim_state(void)
3652 {
3653         current->lockdep_reclaim_gfp = 0;
3654 }
3655
3656 #ifdef CONFIG_LOCK_STAT
3657 static int
3658 print_lock_contention_bug(struct task_struct *curr, struct lockdep_map *lock,
3659                            unsigned long ip)
3660 {
3661         if (!debug_locks_off())
3662                 return 0;
3663         if (debug_locks_silent)
3664                 return 0;
3665
3666         printk("\n");
3667         printk("=================================\n");
3668         printk("[ BUG: bad contention detected! ]\n");
3669         print_kernel_ident();
3670         printk("---------------------------------\n");
3671         printk("%s/%d is trying to contend lock (",
3672                 curr->comm, task_pid_nr(curr));
3673         print_lockdep_cache(lock);
3674         printk(") at:\n");
3675         print_ip_sym(ip);
3676         printk("but there are no locks held!\n");
3677         printk("\nother info that might help us debug this:\n");
3678         lockdep_print_held_locks(curr);
3679
3680         printk("\nstack backtrace:\n");
3681         dump_stack();
3682
3683         return 0;
3684 }
3685
3686 static void
3687 __lock_contended(struct lockdep_map *lock, unsigned long ip)
3688 {
3689         struct task_struct *curr = current;
3690         struct held_lock *hlock, *prev_hlock;
3691         struct lock_class_stats *stats;
3692         unsigned int depth;
3693         int i, contention_point, contending_point;
3694
3695         depth = curr->lockdep_depth;
3696         /*
3697          * Whee, we contended on this lock, except it seems we're not
3698          * actually trying to acquire anything much at all..
3699          */
3700         if (DEBUG_LOCKS_WARN_ON(!depth))
3701                 return;
3702
3703         prev_hlock = NULL;
3704         for (i = depth-1; i >= 0; i--) {
3705                 hlock = curr->held_locks + i;
3706                 /*
3707                  * We must not cross into another context:
3708                  */
3709                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3710                         break;
3711                 if (match_held_lock(hlock, lock))
3712                         goto found_it;
3713                 prev_hlock = hlock;
3714         }
3715         print_lock_contention_bug(curr, lock, ip);
3716         return;
3717
3718 found_it:
3719         if (hlock->instance != lock)
3720                 return;
3721
3722         hlock->waittime_stamp = lockstat_clock();
3723
3724         contention_point = lock_point(hlock_class(hlock)->contention_point, ip);
3725         contending_point = lock_point(hlock_class(hlock)->contending_point,
3726                                       lock->ip);
3727
3728         stats = get_lock_stats(hlock_class(hlock));
3729         if (contention_point < LOCKSTAT_POINTS)
3730                 stats->contention_point[contention_point]++;
3731         if (contending_point < LOCKSTAT_POINTS)
3732                 stats->contending_point[contending_point]++;
3733         if (lock->cpu != smp_processor_id())
3734                 stats->bounces[bounce_contended + !!hlock->read]++;
3735         put_lock_stats(stats);
3736 }
3737
3738 static void
3739 __lock_acquired(struct lockdep_map *lock, unsigned long ip)
3740 {
3741         struct task_struct *curr = current;
3742         struct held_lock *hlock, *prev_hlock;
3743         struct lock_class_stats *stats;
3744         unsigned int depth;
3745         u64 now, waittime = 0;
3746         int i, cpu;
3747
3748         depth = curr->lockdep_depth;
3749         /*
3750          * Yay, we acquired ownership of this lock we didn't try to
3751          * acquire, how the heck did that happen?
3752          */
3753         if (DEBUG_LOCKS_WARN_ON(!depth))
3754                 return;
3755
3756         prev_hlock = NULL;
3757         for (i = depth-1; i >= 0; i--) {
3758                 hlock = curr->held_locks + i;
3759                 /*
3760                  * We must not cross into another context:
3761                  */
3762                 if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
3763                         break;
3764                 if (match_held_lock(hlock, lock))
3765                         goto found_it;
3766                 prev_hlock = hlock;
3767         }
3768         print_lock_contention_bug(curr, lock, _RET_IP_);
3769         return;
3770
3771 found_it:
3772         if (hlock->instance != lock)
3773                 return;
3774
3775         cpu = smp_processor_id();
3776         if (hlock->waittime_stamp) {
3777                 now = lockstat_clock();
3778                 waittime = now - hlock->waittime_stamp;
3779                 hlock->holdtime_stamp = now;
3780         }
3781
3782         trace_lock_acquired(lock, ip);
3783
3784         stats = get_lock_stats(hlock_class(hlock));
3785         if (waittime) {
3786                 if (hlock->read)
3787                         lock_time_inc(&stats->read_waittime, waittime);
3788                 else
3789                         lock_time_inc(&stats->write_waittime, waittime);
3790         }
3791         if (lock->cpu != cpu)
3792                 stats->bounces[bounce_acquired + !!hlock->read]++;
3793         put_lock_stats(stats);
3794
3795         lock->cpu = cpu;
3796         lock->ip = ip;
3797 }
3798
3799 void lock_contended(struct lockdep_map *lock, unsigned long ip)
3800 {
3801         unsigned long flags;
3802
3803         if (unlikely(!lock_stat))
3804                 return;
3805
3806         if (unlikely(current->lockdep_recursion))
3807                 return;
3808
3809         raw_local_irq_save(flags);
3810         check_flags(flags);
3811         current->lockdep_recursion = 1;
3812         trace_lock_contended(lock, ip);
3813         __lock_contended(lock, ip);
3814         current->lockdep_recursion = 0;
3815         raw_local_irq_restore(flags);
3816 }
3817 EXPORT_SYMBOL_GPL(lock_contended);
3818
3819 void lock_acquired(struct lockdep_map *lock, unsigned long ip)
3820 {
3821         unsigned long flags;
3822
3823         if (unlikely(!lock_stat))
3824                 return;
3825
3826         if (unlikely(current->lockdep_recursion))
3827                 return;
3828
3829         raw_local_irq_save(flags);
3830         check_flags(flags);
3831         current->lockdep_recursion = 1;
3832         __lock_acquired(lock, ip);
3833         current->lockdep_recursion = 0;
3834         raw_local_irq_restore(flags);
3835 }
3836 EXPORT_SYMBOL_GPL(lock_acquired);
3837 #endif
3838
3839 /*
3840  * Used by the testsuite, sanitize the validator state
3841  * after a simulated failure:
3842  */
3843
3844 void lockdep_reset(void)
3845 {
3846         unsigned long flags;
3847         int i;
3848
3849         raw_local_irq_save(flags);
3850         current->curr_chain_key = 0;
3851         current->lockdep_depth = 0;
3852         current->lockdep_recursion = 0;
3853         memset(current->held_locks, 0, MAX_LOCK_DEPTH*sizeof(struct held_lock));
3854         nr_hardirq_chains = 0;
3855         nr_softirq_chains = 0;
3856         nr_process_chains = 0;
3857         debug_locks = 1;
3858         for (i = 0; i < CHAINHASH_SIZE; i++)
3859                 INIT_LIST_HEAD(chainhash_table + i);
3860         raw_local_irq_restore(flags);
3861 }
3862
3863 static void zap_class(struct lock_class *class)
3864 {
3865         int i;
3866
3867         /*
3868          * Remove all dependencies this lock is
3869          * involved in:
3870          */
3871         for (i = 0; i < nr_list_entries; i++) {
3872                 if (list_entries[i].class == class)
3873                         list_del_rcu(&list_entries[i].entry);
3874         }
3875         /*
3876          * Unhash the class and remove it from the all_lock_classes list:
3877          */
3878         list_del_rcu(&class->hash_entry);
3879         list_del_rcu(&class->lock_entry);
3880
3881         class->key = NULL;
3882 }
3883
3884 static inline int within(const void *addr, void *start, unsigned long size)
3885 {
3886         return addr >= start && addr < start + size;
3887 }
3888
3889 void lockdep_free_key_range(void *start, unsigned long size)
3890 {
3891         struct lock_class *class, *next;
3892         struct list_head *head;
3893         unsigned long flags;
3894         int i;
3895         int locked;
3896
3897         raw_local_irq_save(flags);
3898         locked = graph_lock();
3899
3900         /*
3901          * Unhash all classes that were created by this module:
3902          */
3903         for (i = 0; i < CLASSHASH_SIZE; i++) {
3904                 head = classhash_table + i;
3905                 if (list_empty(head))
3906                         continue;
3907                 list_for_each_entry_safe(class, next, head, hash_entry) {
3908                         if (within(class->key, start, size))
3909                                 zap_class(class);
3910                         else if (within(class->name, start, size))
3911                                 zap_class(class);
3912                 }
3913         }
3914
3915         if (locked)
3916                 graph_unlock();
3917         raw_local_irq_restore(flags);
3918 }
3919
3920 void lockdep_reset_lock(struct lockdep_map *lock)
3921 {
3922         struct lock_class *class, *next;
3923         struct list_head *head;
3924         unsigned long flags;
3925         int i, j;
3926         int locked;
3927
3928         raw_local_irq_save(flags);
3929
3930         /*
3931          * Remove all classes this lock might have:
3932          */
3933         for (j = 0; j < MAX_LOCKDEP_SUBCLASSES; j++) {
3934                 /*
3935                  * If the class exists we look it up and zap it:
3936                  */
3937                 class = look_up_lock_class(lock, j);
3938                 if (class)
3939                         zap_class(class);
3940         }
3941         /*
3942          * Debug check: in the end all mapped classes should
3943          * be gone.
3944          */
3945         locked = graph_lock();
3946         for (i = 0; i < CLASSHASH_SIZE; i++) {
3947                 head = classhash_table + i;
3948                 if (list_empty(head))
3949                         continue;
3950                 list_for_each_entry_safe(class, next, head, hash_entry) {
3951                         int match = 0;
3952
3953                         for (j = 0; j < NR_LOCKDEP_CACHING_CLASSES; j++)
3954                                 match |= class == lock->class_cache[j];
3955
3956                         if (unlikely(match)) {
3957                                 if (debug_locks_off_graph_unlock()) {
3958                                         /*
3959                                          * We all just reset everything, how did it match?
3960                                          */
3961                                         WARN_ON(1);
3962                                 }
3963                                 goto out_restore;
3964                         }
3965                 }
3966         }
3967         if (locked)
3968                 graph_unlock();
3969
3970 out_restore:
3971         raw_local_irq_restore(flags);
3972 }
3973
3974 void lockdep_init(void)
3975 {
3976         int i;
3977
3978         /*
3979          * Some architectures have their own start_kernel()
3980          * code which calls lockdep_init(), while we also
3981          * call lockdep_init() from the start_kernel() itself,
3982          * and we want to initialize the hashes only once:
3983          */
3984         if (lockdep_initialized)
3985                 return;
3986
3987         for (i = 0; i < CLASSHASH_SIZE; i++)
3988                 INIT_LIST_HEAD(classhash_table + i);
3989
3990         for (i = 0; i < CHAINHASH_SIZE; i++)
3991                 INIT_LIST_HEAD(chainhash_table + i);
3992
3993         lockdep_initialized = 1;
3994 }
3995
3996 void __init lockdep_info(void)
3997 {
3998         printk("Lock dependency validator: Copyright (c) 2006 Red Hat, Inc., Ingo Molnar\n");
3999
4000         printk("... MAX_LOCKDEP_SUBCLASSES:  %lu\n", MAX_LOCKDEP_SUBCLASSES);
4001         printk("... MAX_LOCK_DEPTH:          %lu\n", MAX_LOCK_DEPTH);
4002         printk("... MAX_LOCKDEP_KEYS:        %lu\n", MAX_LOCKDEP_KEYS);
4003         printk("... CLASSHASH_SIZE:          %lu\n", CLASSHASH_SIZE);
4004         printk("... MAX_LOCKDEP_ENTRIES:     %lu\n", MAX_LOCKDEP_ENTRIES);
4005         printk("... MAX_LOCKDEP_CHAINS:      %lu\n", MAX_LOCKDEP_CHAINS);
4006         printk("... CHAINHASH_SIZE:          %lu\n", CHAINHASH_SIZE);
4007
4008         printk(" memory used by lock dependency info: %lu kB\n",
4009                 (sizeof(struct lock_class) * MAX_LOCKDEP_KEYS +
4010                 sizeof(struct list_head) * CLASSHASH_SIZE +
4011                 sizeof(struct lock_list) * MAX_LOCKDEP_ENTRIES +
4012                 sizeof(struct lock_chain) * MAX_LOCKDEP_CHAINS +
4013                 sizeof(struct list_head) * CHAINHASH_SIZE
4014 #ifdef CONFIG_PROVE_LOCKING
4015                 + sizeof(struct circular_queue)
4016 #endif
4017                 ) / 1024
4018                 );
4019
4020         printk(" per task-struct memory footprint: %lu bytes\n",
4021                 sizeof(struct held_lock) * MAX_LOCK_DEPTH);
4022
4023 #ifdef CONFIG_DEBUG_LOCKDEP
4024         if (lockdep_init_error) {
4025                 printk("WARNING: lockdep init error! lock-%s was acquired"
4026                         "before lockdep_init\n", lock_init_error);
4027                 printk("Call stack leading to lockdep invocation was:\n");
4028                 print_stack_trace(&lockdep_init_trace, 0);
4029         }
4030 #endif
4031 }
4032
4033 static void
4034 print_freed_lock_bug(struct task_struct *curr, const void *mem_from,
4035                      const void *mem_to, struct held_lock *hlock)
4036 {
4037         if (!debug_locks_off())
4038                 return;
4039         if (debug_locks_silent)
4040                 return;
4041
4042         printk("\n");
4043         printk("=========================\n");
4044         printk("[ BUG: held lock freed! ]\n");
4045         print_kernel_ident();
4046         printk("-------------------------\n");
4047         printk("%s/%d is freeing memory %p-%p, with a lock still held there!\n",
4048                 curr->comm, task_pid_nr(curr), mem_from, mem_to-1);
4049         print_lock(hlock);
4050         lockdep_print_held_locks(curr);
4051
4052         printk("\nstack backtrace:\n");
4053         dump_stack();
4054 }
4055
4056 static inline int not_in_range(const void* mem_from, unsigned long mem_len,
4057                                 const void* lock_from, unsigned long lock_len)
4058 {
4059         return lock_from + lock_len <= mem_from ||
4060                 mem_from + mem_len <= lock_from;
4061 }
4062
4063 /*
4064  * Called when kernel memory is freed (or unmapped), or if a lock
4065  * is destroyed or reinitialized - this code checks whether there is
4066  * any held lock in the memory range of <from> to <to>:
4067  */
4068 void debug_check_no_locks_freed(const void *mem_from, unsigned long mem_len)
4069 {
4070         struct task_struct *curr = current;
4071         struct held_lock *hlock;
4072         unsigned long flags;
4073         int i;
4074
4075         if (unlikely(!debug_locks))
4076                 return;
4077
4078         local_irq_save(flags);
4079         for (i = 0; i < curr->lockdep_depth; i++) {
4080                 hlock = curr->held_locks + i;
4081
4082                 if (not_in_range(mem_from, mem_len, hlock->instance,
4083                                         sizeof(*hlock->instance)))
4084                         continue;
4085
4086                 print_freed_lock_bug(curr, mem_from, mem_from + mem_len, hlock);
4087                 break;
4088         }
4089         local_irq_restore(flags);
4090 }
4091 EXPORT_SYMBOL_GPL(debug_check_no_locks_freed);
4092
4093 static void print_held_locks_bug(void)
4094 {
4095         if (!debug_locks_off())
4096                 return;
4097         if (debug_locks_silent)
4098                 return;
4099
4100         printk("\n");
4101         printk("=====================================\n");
4102         printk("[ BUG: %s/%d still has locks held! ]\n",
4103                current->comm, task_pid_nr(current));
4104         print_kernel_ident();
4105         printk("-------------------------------------\n");
4106         lockdep_print_held_locks(current);
4107         printk("\nstack backtrace:\n");
4108         dump_stack();
4109 }
4110
4111 void debug_check_no_locks_held(void)
4112 {
4113         if (unlikely(current->lockdep_depth > 0))
4114                 print_held_locks_bug();
4115 }
4116 EXPORT_SYMBOL_GPL(debug_check_no_locks_held);
4117
4118 void debug_show_all_locks(void)
4119 {
4120         struct task_struct *g, *p;
4121         int count = 10;
4122         int unlock = 1;
4123
4124         if (unlikely(!debug_locks)) {
4125                 printk("INFO: lockdep is turned off.\n");
4126                 return;
4127         }
4128         printk("\nShowing all locks held in the system:\n");
4129
4130         /*
4131          * Here we try to get the tasklist_lock as hard as possible,
4132          * if not successful after 2 seconds we ignore it (but keep
4133          * trying). This is to enable a debug printout even if a
4134          * tasklist_lock-holding task deadlocks or crashes.
4135          */
4136 retry:
4137         if (!read_trylock(&tasklist_lock)) {
4138                 if (count == 10)
4139                         printk("hm, tasklist_lock locked, retrying... ");
4140                 if (count) {
4141                         count--;
4142                         printk(" #%d", 10-count);
4143                         mdelay(200);
4144                         goto retry;
4145                 }
4146                 printk(" ignoring it.\n");
4147                 unlock = 0;
4148         } else {
4149                 if (count != 10)
4150                         printk(KERN_CONT " locked it.\n");
4151         }
4152
4153         do_each_thread(g, p) {
4154                 /*
4155                  * It's not reliable to print a task's held locks
4156                  * if it's not sleeping (or if it's not the current
4157                  * task):
4158                  */
4159                 if (p->state == TASK_RUNNING && p != current)
4160                         continue;
4161                 if (p->lockdep_depth)
4162                         lockdep_print_held_locks(p);
4163                 if (!unlock)
4164                         if (read_trylock(&tasklist_lock))
4165                                 unlock = 1;
4166         } while_each_thread(g, p);
4167
4168         printk("\n");
4169         printk("=============================================\n\n");
4170
4171         if (unlock)
4172                 read_unlock(&tasklist_lock);
4173 }
4174 EXPORT_SYMBOL_GPL(debug_show_all_locks);
4175
4176 /*
4177  * Careful: only use this function if you are sure that
4178  * the task cannot run in parallel!
4179  */
4180 void debug_show_held_locks(struct task_struct *task)
4181 {
4182         if (unlikely(!debug_locks)) {
4183                 printk("INFO: lockdep is turned off.\n");
4184                 return;
4185         }
4186         lockdep_print_held_locks(task);
4187 }
4188 EXPORT_SYMBOL_GPL(debug_show_held_locks);
4189
4190 void lockdep_sys_exit(void)
4191 {
4192         struct task_struct *curr = current;
4193
4194         if (unlikely(curr->lockdep_depth)) {
4195                 if (!debug_locks_off())
4196                         return;
4197                 printk("\n");
4198                 printk("================================================\n");
4199                 printk("[ BUG: lock held when returning to user space! ]\n");
4200                 print_kernel_ident();
4201                 printk("------------------------------------------------\n");
4202                 printk("%s/%d is leaving the kernel with locks still held!\n",
4203                                 curr->comm, curr->pid);
4204                 lockdep_print_held_locks(curr);
4205         }
4206 }
4207
4208 void lockdep_rcu_suspicious(const char *file, const int line, const char *s)
4209 {
4210         struct task_struct *curr = current;
4211
4212 #ifndef CONFIG_PROVE_RCU_REPEATEDLY
4213         if (!debug_locks_off())
4214                 return;
4215 #endif /* #ifdef CONFIG_PROVE_RCU_REPEATEDLY */
4216         /* Note: the following can be executed concurrently, so be careful. */
4217         printk("\n");
4218         printk("===============================\n");
4219         printk("[ INFO: suspicious RCU usage. ]\n");
4220         print_kernel_ident();
4221         printk("-------------------------------\n");
4222         printk("%s:%d %s!\n", file, line, s);
4223         printk("\nother info that might help us debug this:\n\n");
4224         printk("\n%srcu_scheduler_active = %d, debug_locks = %d\n",
4225                !rcu_lockdep_current_cpu_online()
4226                         ? "RCU used illegally from offline CPU!\n"
4227                         : rcu_is_cpu_idle()
4228                                 ? "RCU used illegally from idle CPU!\n"
4229                                 : "",
4230                rcu_scheduler_active, debug_locks);
4231
4232         /*
4233          * If a CPU is in the RCU-free window in idle (ie: in the section
4234          * between rcu_idle_enter() and rcu_idle_exit(), then RCU
4235          * considers that CPU to be in an "extended quiescent state",
4236          * which means that RCU will be completely ignoring that CPU.
4237          * Therefore, rcu_read_lock() and friends have absolutely no
4238          * effect on a CPU running in that state. In other words, even if
4239          * such an RCU-idle CPU has called rcu_read_lock(), RCU might well
4240          * delete data structures out from under it.  RCU really has no
4241          * choice here: we need to keep an RCU-free window in idle where
4242          * the CPU may possibly enter into low power mode. This way we can
4243          * notice an extended quiescent state to other CPUs that started a grace
4244          * period. Otherwise we would delay any grace period as long as we run
4245          * in the idle task.
4246          *
4247          * So complain bitterly if someone does call rcu_read_lock(),
4248          * rcu_read_lock_bh() and so on from extended quiescent states.
4249          */
4250         if (rcu_is_cpu_idle())
4251                 printk("RCU used illegally from extended quiescent state!\n");
4252
4253         lockdep_print_held_locks(curr);
4254         printk("\nstack backtrace:\n");
4255         dump_stack();
4256 }
4257 EXPORT_SYMBOL_GPL(lockdep_rcu_suspicious);