5 #include "util/cache.h"
6 #include "util/symbol.h"
7 #include "util/thread.h"
8 #include "util/header.h"
10 #include "util/parse-options.h"
11 #include "util/trace-event.h"
13 #include "util/debug.h"
14 #include "util/session.h"
16 #include <sys/types.h>
17 #include <sys/prctl.h>
18 #include <semaphore.h>
23 #include <linux/list.h>
24 #include <linux/hash.h>
26 static struct perf_session *session;
28 /* based on kernel/lockdep.c */
29 #define LOCKHASH_BITS 12
30 #define LOCKHASH_SIZE (1UL << LOCKHASH_BITS)
32 static struct list_head lockhash_table[LOCKHASH_SIZE];
34 #define __lockhashfn(key) hash_long((unsigned long)key, LOCKHASH_BITS)
35 #define lockhashentry(key) (lockhash_table + __lockhashfn((key)))
38 struct list_head hash_entry;
39 struct rb_node rb; /* used for sorting */
42 * FIXME: raw_field_value() returns unsigned long long,
43 * so address of lockdep_map should be dealed as 64bit.
44 * Is there more better solution?
46 void *addr; /* address of lockdep_map, used as ID */
47 char *name; /* for strcpy(), we cannot use const */
49 unsigned int nr_acquire;
50 unsigned int nr_acquired;
51 unsigned int nr_contended;
52 unsigned int nr_release;
54 unsigned int nr_readlock;
55 unsigned int nr_trylock;
56 /* these times are in nano sec. */
61 int discard; /* flag of blacklist */
65 * States of lock_seq_stat
67 * UNINITIALIZED is required for detecting first event of acquire.
68 * As the nature of lock events, there is no guarantee
69 * that the first event for the locks are acquire,
70 * it can be acquired, contended or release.
72 #define SEQ_STATE_UNINITIALIZED 0 /* initial state */
73 #define SEQ_STATE_RELEASED 1
74 #define SEQ_STATE_ACQUIRING 2
75 #define SEQ_STATE_ACQUIRED 3
76 #define SEQ_STATE_READ_ACQUIRED 4
77 #define SEQ_STATE_CONTENDED 5
81 * Imported from include/linux/sched.h.
82 * Should this be synchronized?
84 #define MAX_LOCK_DEPTH 48
87 * struct lock_seq_stat:
88 * Place to put on state of one lock sequence
89 * 1) acquire -> acquired -> release
90 * 2) acquire -> contended -> acquired -> release
91 * 3) acquire (with read or try) -> release
92 * 4) Are there other patterns?
94 struct lock_seq_stat {
95 struct list_head list;
107 struct list_head seq_list;
110 static struct rb_root thread_stats;
112 static struct thread_stat *thread_stat_find(u32 tid)
114 struct rb_node *node;
115 struct thread_stat *st;
117 node = thread_stats.rb_node;
119 st = container_of(node, struct thread_stat, rb);
122 else if (tid < st->tid)
123 node = node->rb_left;
125 node = node->rb_right;
131 static void thread_stat_insert(struct thread_stat *new)
133 struct rb_node **rb = &thread_stats.rb_node;
134 struct rb_node *parent = NULL;
135 struct thread_stat *p;
138 p = container_of(*rb, struct thread_stat, rb);
141 if (new->tid < p->tid)
142 rb = &(*rb)->rb_left;
143 else if (new->tid > p->tid)
144 rb = &(*rb)->rb_right;
146 BUG_ON("inserting invalid thread_stat\n");
149 rb_link_node(&new->rb, parent, rb);
150 rb_insert_color(&new->rb, &thread_stats);
153 static struct thread_stat *thread_stat_findnew_after_first(u32 tid)
155 struct thread_stat *st;
157 st = thread_stat_find(tid);
161 st = zalloc(sizeof(struct thread_stat));
163 die("memory allocation failed\n");
166 INIT_LIST_HEAD(&st->seq_list);
168 thread_stat_insert(st);
173 static struct thread_stat *thread_stat_findnew_first(u32 tid);
174 static struct thread_stat *(*thread_stat_findnew)(u32 tid) =
175 thread_stat_findnew_first;
177 static struct thread_stat *thread_stat_findnew_first(u32 tid)
179 struct thread_stat *st;
181 st = zalloc(sizeof(struct thread_stat));
183 die("memory allocation failed\n");
185 INIT_LIST_HEAD(&st->seq_list);
187 rb_link_node(&st->rb, NULL, &thread_stats.rb_node);
188 rb_insert_color(&st->rb, &thread_stats);
190 thread_stat_findnew = thread_stat_findnew_after_first;
194 /* build simple key function one is bigger than two */
195 #define SINGLE_KEY(member) \
196 static int lock_stat_key_ ## member(struct lock_stat *one, \
197 struct lock_stat *two) \
199 return one->member > two->member; \
202 SINGLE_KEY(nr_acquired)
203 SINGLE_KEY(nr_contended)
204 SINGLE_KEY(wait_time_total)
205 SINGLE_KEY(wait_time_max)
207 static int lock_stat_key_wait_time_min(struct lock_stat *one,
208 struct lock_stat *two)
210 u64 s1 = one->wait_time_min;
211 u64 s2 = two->wait_time_min;
212 if (s1 == ULLONG_MAX)
214 if (s2 == ULLONG_MAX)
221 * name: the value for specify by user
222 * this should be simpler than raw name of member
223 * e.g. nr_acquired -> acquired, wait_time_total -> wait_total
226 int (*key)(struct lock_stat*, struct lock_stat*);
229 static const char *sort_key = "acquired";
231 static int (*compare)(struct lock_stat *, struct lock_stat *);
233 static struct rb_root result; /* place to store sorted data */
235 #define DEF_KEY_LOCK(name, fn_suffix) \
236 { #name, lock_stat_key_ ## fn_suffix }
237 struct lock_key keys[] = {
238 DEF_KEY_LOCK(acquired, nr_acquired),
239 DEF_KEY_LOCK(contended, nr_contended),
240 DEF_KEY_LOCK(wait_total, wait_time_total),
241 DEF_KEY_LOCK(wait_min, wait_time_min),
242 DEF_KEY_LOCK(wait_max, wait_time_max),
244 /* extra comparisons much complicated should be here */
249 static void select_key(void)
253 for (i = 0; keys[i].name; i++) {
254 if (!strcmp(keys[i].name, sort_key)) {
255 compare = keys[i].key;
260 die("Unknown compare key:%s\n", sort_key);
263 static void insert_to_result(struct lock_stat *st,
264 int (*bigger)(struct lock_stat *, struct lock_stat *))
266 struct rb_node **rb = &result.rb_node;
267 struct rb_node *parent = NULL;
271 p = container_of(*rb, struct lock_stat, rb);
275 rb = &(*rb)->rb_left;
277 rb = &(*rb)->rb_right;
280 rb_link_node(&st->rb, parent, rb);
281 rb_insert_color(&st->rb, &result);
284 /* returns left most element of result, and erase it */
285 static struct lock_stat *pop_from_result(void)
287 struct rb_node *node = result.rb_node;
292 while (node->rb_left)
293 node = node->rb_left;
295 rb_erase(node, &result);
296 return container_of(node, struct lock_stat, rb);
299 static struct lock_stat *lock_stat_findnew(void *addr, const char *name)
301 struct list_head *entry = lockhashentry(addr);
302 struct lock_stat *ret, *new;
304 list_for_each_entry(ret, entry, hash_entry) {
305 if (ret->addr == addr)
309 new = zalloc(sizeof(struct lock_stat));
314 new->name = zalloc(sizeof(char) * strlen(name) + 1);
317 strcpy(new->name, name);
319 new->wait_time_min = ULLONG_MAX;
321 list_add(&new->hash_entry, entry);
325 die("memory allocation failed\n");
328 static char const *input_name = "perf.data";
330 struct raw_event_sample {
335 struct trace_acquire_event {
341 struct trace_acquired_event {
346 struct trace_contended_event {
351 struct trace_release_event {
356 struct trace_lock_handler {
357 void (*acquire_event)(struct trace_acquire_event *,
361 struct thread *thread);
363 void (*acquired_event)(struct trace_acquired_event *,
367 struct thread *thread);
369 void (*contended_event)(struct trace_contended_event *,
373 struct thread *thread);
375 void (*release_event)(struct trace_release_event *,
379 struct thread *thread);
382 static struct lock_seq_stat *get_seq(struct thread_stat *ts, void *addr)
384 struct lock_seq_stat *seq;
386 list_for_each_entry(seq, &ts->seq_list, list) {
387 if (seq->addr == addr)
391 seq = zalloc(sizeof(struct lock_seq_stat));
393 die("Not enough memory\n");
394 seq->state = SEQ_STATE_UNINITIALIZED;
397 list_add(&seq->list, &ts->seq_list);
409 static int bad_hist[BROKEN_MAX];
417 report_lock_acquire_event(struct trace_acquire_event *acquire_event,
418 struct event *__event __used,
420 u64 timestamp __used,
421 struct thread *thread __used)
423 struct lock_stat *ls;
424 struct thread_stat *ts;
425 struct lock_seq_stat *seq;
427 ls = lock_stat_findnew(acquire_event->addr, acquire_event->name);
431 ts = thread_stat_findnew(thread->pid);
432 seq = get_seq(ts, acquire_event->addr);
434 switch (seq->state) {
435 case SEQ_STATE_UNINITIALIZED:
436 case SEQ_STATE_RELEASED:
437 if (!acquire_event->flag) {
438 seq->state = SEQ_STATE_ACQUIRING;
440 if (acquire_event->flag & TRY_LOCK)
442 if (acquire_event->flag & READ_LOCK)
444 seq->state = SEQ_STATE_READ_ACQUIRED;
449 case SEQ_STATE_READ_ACQUIRED:
450 if (acquire_event->flag & READ_LOCK) {
458 case SEQ_STATE_ACQUIRED:
459 case SEQ_STATE_ACQUIRING:
460 case SEQ_STATE_CONTENDED:
462 /* broken lock sequence, discard it */
464 bad_hist[BROKEN_ACQUIRE]++;
465 list_del(&seq->list);
470 BUG_ON("Unknown state of lock sequence found!\n");
475 seq->prev_event_time = timestamp;
481 report_lock_acquired_event(struct trace_acquired_event *acquired_event,
482 struct event *__event __used,
484 u64 timestamp __used,
485 struct thread *thread __used)
487 struct lock_stat *ls;
488 struct thread_stat *ts;
489 struct lock_seq_stat *seq;
492 ls = lock_stat_findnew(acquired_event->addr, acquired_event->name);
496 ts = thread_stat_findnew(thread->pid);
497 seq = get_seq(ts, acquired_event->addr);
499 switch (seq->state) {
500 case SEQ_STATE_UNINITIALIZED:
501 /* orphan event, do nothing */
503 case SEQ_STATE_ACQUIRING:
505 case SEQ_STATE_CONTENDED:
506 contended_term = timestamp - seq->prev_event_time;
507 ls->wait_time_total += contended_term;
508 if (contended_term < ls->wait_time_min)
509 ls->wait_time_min = contended_term;
510 if (ls->wait_time_max < contended_term)
511 ls->wait_time_max = contended_term;
513 case SEQ_STATE_RELEASED:
514 case SEQ_STATE_ACQUIRED:
515 case SEQ_STATE_READ_ACQUIRED:
516 /* broken lock sequence, discard it */
518 bad_hist[BROKEN_ACQUIRED]++;
519 list_del(&seq->list);
525 BUG_ON("Unknown state of lock sequence found!\n");
529 seq->state = SEQ_STATE_ACQUIRED;
531 seq->prev_event_time = timestamp;
537 report_lock_contended_event(struct trace_contended_event *contended_event,
538 struct event *__event __used,
540 u64 timestamp __used,
541 struct thread *thread __used)
543 struct lock_stat *ls;
544 struct thread_stat *ts;
545 struct lock_seq_stat *seq;
547 ls = lock_stat_findnew(contended_event->addr, contended_event->name);
551 ts = thread_stat_findnew(thread->pid);
552 seq = get_seq(ts, contended_event->addr);
554 switch (seq->state) {
555 case SEQ_STATE_UNINITIALIZED:
556 /* orphan event, do nothing */
558 case SEQ_STATE_ACQUIRING:
560 case SEQ_STATE_RELEASED:
561 case SEQ_STATE_ACQUIRED:
562 case SEQ_STATE_READ_ACQUIRED:
563 case SEQ_STATE_CONTENDED:
564 /* broken lock sequence, discard it */
566 bad_hist[BROKEN_CONTENDED]++;
567 list_del(&seq->list);
572 BUG_ON("Unknown state of lock sequence found!\n");
576 seq->state = SEQ_STATE_CONTENDED;
578 seq->prev_event_time = timestamp;
584 report_lock_release_event(struct trace_release_event *release_event,
585 struct event *__event __used,
587 u64 timestamp __used,
588 struct thread *thread __used)
590 struct lock_stat *ls;
591 struct thread_stat *ts;
592 struct lock_seq_stat *seq;
594 ls = lock_stat_findnew(release_event->addr, release_event->name);
598 ts = thread_stat_findnew(thread->pid);
599 seq = get_seq(ts, release_event->addr);
601 switch (seq->state) {
602 case SEQ_STATE_UNINITIALIZED:
605 case SEQ_STATE_ACQUIRED:
607 case SEQ_STATE_READ_ACQUIRED:
609 BUG_ON(seq->read_count < 0);
610 if (!seq->read_count) {
615 case SEQ_STATE_ACQUIRING:
616 case SEQ_STATE_CONTENDED:
617 case SEQ_STATE_RELEASED:
618 /* broken lock sequence, discard it */
620 bad_hist[BROKEN_RELEASE]++;
624 BUG_ON("Unknown state of lock sequence found!\n");
630 list_del(&seq->list);
636 /* lock oriented handlers */
637 /* TODO: handlers for CPU oriented, thread oriented */
638 static struct trace_lock_handler report_lock_ops = {
639 .acquire_event = report_lock_acquire_event,
640 .acquired_event = report_lock_acquired_event,
641 .contended_event = report_lock_contended_event,
642 .release_event = report_lock_release_event,
645 static struct trace_lock_handler *trace_handler;
648 process_lock_acquire_event(void *data,
649 struct event *event __used,
651 u64 timestamp __used,
652 struct thread *thread __used)
654 struct trace_acquire_event acquire_event;
655 u64 tmp; /* this is required for casting... */
657 tmp = raw_field_value(event, "lockdep_addr", data);
658 memcpy(&acquire_event.addr, &tmp, sizeof(void *));
659 acquire_event.name = (char *)raw_field_ptr(event, "name", data);
660 acquire_event.flag = (int)raw_field_value(event, "flag", data);
662 if (trace_handler->acquire_event)
663 trace_handler->acquire_event(&acquire_event, event, cpu, timestamp, thread);
667 process_lock_acquired_event(void *data,
668 struct event *event __used,
670 u64 timestamp __used,
671 struct thread *thread __used)
673 struct trace_acquired_event acquired_event;
674 u64 tmp; /* this is required for casting... */
676 tmp = raw_field_value(event, "lockdep_addr", data);
677 memcpy(&acquired_event.addr, &tmp, sizeof(void *));
678 acquired_event.name = (char *)raw_field_ptr(event, "name", data);
680 if (trace_handler->acquire_event)
681 trace_handler->acquired_event(&acquired_event, event, cpu, timestamp, thread);
685 process_lock_contended_event(void *data,
686 struct event *event __used,
688 u64 timestamp __used,
689 struct thread *thread __used)
691 struct trace_contended_event contended_event;
692 u64 tmp; /* this is required for casting... */
694 tmp = raw_field_value(event, "lockdep_addr", data);
695 memcpy(&contended_event.addr, &tmp, sizeof(void *));
696 contended_event.name = (char *)raw_field_ptr(event, "name", data);
698 if (trace_handler->acquire_event)
699 trace_handler->contended_event(&contended_event, event, cpu, timestamp, thread);
703 process_lock_release_event(void *data,
704 struct event *event __used,
706 u64 timestamp __used,
707 struct thread *thread __used)
709 struct trace_release_event release_event;
710 u64 tmp; /* this is required for casting... */
712 tmp = raw_field_value(event, "lockdep_addr", data);
713 memcpy(&release_event.addr, &tmp, sizeof(void *));
714 release_event.name = (char *)raw_field_ptr(event, "name", data);
716 if (trace_handler->acquire_event)
717 trace_handler->release_event(&release_event, event, cpu, timestamp, thread);
721 process_raw_event(void *data, int cpu, u64 timestamp, struct thread *thread)
726 type = trace_parse_common_type(data);
727 event = trace_find_event(type);
729 if (!strcmp(event->name, "lock_acquire"))
730 process_lock_acquire_event(data, event, cpu, timestamp, thread);
731 if (!strcmp(event->name, "lock_acquired"))
732 process_lock_acquired_event(data, event, cpu, timestamp, thread);
733 if (!strcmp(event->name, "lock_contended"))
734 process_lock_contended_event(data, event, cpu, timestamp, thread);
735 if (!strcmp(event->name, "lock_release"))
736 process_lock_release_event(data, event, cpu, timestamp, thread);
739 static void print_bad_events(int bad, int total)
741 /* Output for debug, this have to be removed */
743 const char *name[4] =
744 { "acquire", "acquired", "contended", "release" };
746 pr_info("\n=== output for debug===\n\n");
747 pr_info("bad: %d, total: %d\n", bad, total);
748 pr_info("bad rate: %f %%\n", (double)bad / (double)total * 100);
749 pr_info("histogram of events caused bad sequence\n");
750 for (i = 0; i < BROKEN_MAX; i++)
751 pr_info(" %10s: %d\n", name[i], bad_hist[i]);
754 /* TODO: various way to print, coloring, nano or milli sec */
755 static void print_result(void)
757 struct lock_stat *st;
761 pr_info("%20s ", "Name");
762 pr_info("%10s ", "acquired");
763 pr_info("%10s ", "contended");
765 pr_info("%15s ", "total wait (ns)");
766 pr_info("%15s ", "max wait (ns)");
767 pr_info("%15s ", "min wait (ns)");
772 while ((st = pop_from_result())) {
780 if (strlen(st->name) < 16) {
781 /* output raw name */
782 pr_info("%20s ", st->name);
784 strncpy(cut_name, st->name, 16);
789 /* cut off name for saving output style */
790 pr_info("%20s ", cut_name);
793 pr_info("%10u ", st->nr_acquired);
794 pr_info("%10u ", st->nr_contended);
796 pr_info("%15" PRIu64 " ", st->wait_time_total);
797 pr_info("%15" PRIu64 " ", st->wait_time_max);
798 pr_info("%15" PRIu64 " ", st->wait_time_min == ULLONG_MAX ?
799 0 : st->wait_time_min);
803 print_bad_events(bad, total);
806 static bool info_threads, info_map;
808 static void dump_threads(void)
810 struct thread_stat *st;
811 struct rb_node *node;
814 pr_info("%10s: comm\n", "Thread ID");
816 node = rb_first(&thread_stats);
818 st = container_of(node, struct thread_stat, rb);
819 t = perf_session__findnew(session, st->tid);
820 pr_info("%10d: %s\n", st->tid, t->comm);
821 node = rb_next(node);
825 static void dump_map(void)
828 struct lock_stat *st;
830 pr_info("Address of instance: name of class\n");
831 for (i = 0; i < LOCKHASH_SIZE; i++) {
832 list_for_each_entry(st, &lockhash_table[i], hash_entry) {
833 pr_info(" %p: %s\n", st->addr, st->name);
838 static void dump_info(void)
845 die("Unknown type of information\n");
848 static int process_sample_event(union perf_event *event,
849 struct perf_sample *sample,
850 struct perf_evsel *evsel __used,
851 struct perf_session *s)
853 struct thread *thread = perf_session__findnew(s, sample->tid);
855 if (thread == NULL) {
856 pr_debug("problem processing %d event, skipping it.\n",
861 process_raw_event(sample->raw_data, sample->cpu, sample->time, thread);
866 static struct perf_event_ops eops = {
867 .sample = process_sample_event,
868 .comm = perf_event__process_comm,
869 .ordered_samples = true,
872 static int read_events(void)
874 session = perf_session__new(input_name, O_RDONLY, 0, false, &eops);
876 die("Initializing perf session failed\n");
878 return perf_session__process_events(session, &eops);
881 static void sort_result(void)
884 struct lock_stat *st;
886 for (i = 0; i < LOCKHASH_SIZE; i++) {
887 list_for_each_entry(st, &lockhash_table[i], hash_entry) {
888 insert_to_result(st, compare);
893 static void __cmd_report(void)
902 static const char * const report_usage[] = {
903 "perf lock report [<options>]",
907 static const struct option report_options[] = {
908 OPT_STRING('k', "key", &sort_key, "acquired",
909 "key for sorting (acquired / contended / wait_total / wait_max / wait_min)"),
914 static const char * const info_usage[] = {
915 "perf lock info [<options>]",
919 static const struct option info_options[] = {
920 OPT_BOOLEAN('t', "threads", &info_threads,
921 "dump thread list in perf.data"),
922 OPT_BOOLEAN('m', "map", &info_map,
923 "map of lock instances (name:address table)"),
927 static const char * const lock_usage[] = {
928 "perf lock [<options>] {record|trace|report}",
932 static const struct option lock_options[] = {
933 OPT_STRING('i', "input", &input_name, "file", "input file name"),
934 OPT_INCR('v', "verbose", &verbose, "be more verbose (show symbol address, etc)"),
935 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, "dump raw trace in ASCII"),
939 static const char *record_args[] = {
945 "-e", "lock:lock_acquire:r",
946 "-e", "lock:lock_acquired:r",
947 "-e", "lock:lock_contended:r",
948 "-e", "lock:lock_release:r",
951 static int __cmd_record(int argc, const char **argv)
953 unsigned int rec_argc, i, j;
954 const char **rec_argv;
956 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
957 rec_argv = calloc(rec_argc + 1, sizeof(char *));
959 if (rec_argv == NULL)
962 for (i = 0; i < ARRAY_SIZE(record_args); i++)
963 rec_argv[i] = strdup(record_args[i]);
965 for (j = 1; j < (unsigned int)argc; j++, i++)
966 rec_argv[i] = argv[j];
968 BUG_ON(i != rec_argc);
970 return cmd_record(i, rec_argv, NULL);
973 int cmd_lock(int argc, const char **argv, const char *prefix __used)
978 for (i = 0; i < LOCKHASH_SIZE; i++)
979 INIT_LIST_HEAD(lockhash_table + i);
981 argc = parse_options(argc, argv, lock_options, lock_usage,
982 PARSE_OPT_STOP_AT_NON_OPTION);
984 usage_with_options(lock_usage, lock_options);
986 if (!strncmp(argv[0], "rec", 3)) {
987 return __cmd_record(argc, argv);
988 } else if (!strncmp(argv[0], "report", 6)) {
989 trace_handler = &report_lock_ops;
991 argc = parse_options(argc, argv,
992 report_options, report_usage, 0);
994 usage_with_options(report_usage, report_options);
997 } else if (!strcmp(argv[0], "script")) {
998 /* Aliased to 'perf script' */
999 return cmd_script(argc, argv, prefix);
1000 } else if (!strcmp(argv[0], "info")) {
1002 argc = parse_options(argc, argv,
1003 info_options, info_usage, 0);
1005 usage_with_options(info_usage, info_options);
1007 /* recycling report_lock_ops */
1008 trace_handler = &report_lock_ops;
1013 usage_with_options(lock_usage, lock_options);