4 #include "util/evlist.h"
5 #include "util/evsel.h"
7 #include "util/cache.h"
8 #include "util/symbol.h"
9 #include "util/thread.h"
10 #include "util/header.h"
12 #include "util/parse-options.h"
13 #include "util/trace-event.h"
15 #include "util/debug.h"
16 #include "util/session.h"
17 #include "util/tool.h"
19 #include <sys/types.h>
20 #include <sys/prctl.h>
21 #include <semaphore.h>
26 #include <linux/list.h>
27 #include <linux/hash.h>
29 static struct perf_session *session;
31 /* based on kernel/lockdep.c */
32 #define LOCKHASH_BITS 12
33 #define LOCKHASH_SIZE (1UL << LOCKHASH_BITS)
35 static struct list_head lockhash_table[LOCKHASH_SIZE];
37 #define __lockhashfn(key) hash_long((unsigned long)key, LOCKHASH_BITS)
38 #define lockhashentry(key) (lockhash_table + __lockhashfn((key)))
41 struct list_head hash_entry;
42 struct rb_node rb; /* used for sorting */
45 * FIXME: perf_evsel__intval() returns u64,
46 * so address of lockdep_map should be dealed as 64bit.
47 * Is there more better solution?
49 void *addr; /* address of lockdep_map, used as ID */
50 char *name; /* for strcpy(), we cannot use const */
52 unsigned int nr_acquire;
53 unsigned int nr_acquired;
54 unsigned int nr_contended;
55 unsigned int nr_release;
57 unsigned int nr_readlock;
58 unsigned int nr_trylock;
59 /* these times are in nano sec. */
64 int discard; /* flag of blacklist */
68 * States of lock_seq_stat
70 * UNINITIALIZED is required for detecting first event of acquire.
71 * As the nature of lock events, there is no guarantee
72 * that the first event for the locks are acquire,
73 * it can be acquired, contended or release.
75 #define SEQ_STATE_UNINITIALIZED 0 /* initial state */
76 #define SEQ_STATE_RELEASED 1
77 #define SEQ_STATE_ACQUIRING 2
78 #define SEQ_STATE_ACQUIRED 3
79 #define SEQ_STATE_READ_ACQUIRED 4
80 #define SEQ_STATE_CONTENDED 5
84 * Imported from include/linux/sched.h.
85 * Should this be synchronized?
87 #define MAX_LOCK_DEPTH 48
90 * struct lock_seq_stat:
91 * Place to put on state of one lock sequence
92 * 1) acquire -> acquired -> release
93 * 2) acquire -> contended -> acquired -> release
94 * 3) acquire (with read or try) -> release
95 * 4) Are there other patterns?
97 struct lock_seq_stat {
98 struct list_head list;
110 struct list_head seq_list;
113 static struct rb_root thread_stats;
115 static struct thread_stat *thread_stat_find(u32 tid)
117 struct rb_node *node;
118 struct thread_stat *st;
120 node = thread_stats.rb_node;
122 st = container_of(node, struct thread_stat, rb);
125 else if (tid < st->tid)
126 node = node->rb_left;
128 node = node->rb_right;
134 static void thread_stat_insert(struct thread_stat *new)
136 struct rb_node **rb = &thread_stats.rb_node;
137 struct rb_node *parent = NULL;
138 struct thread_stat *p;
141 p = container_of(*rb, struct thread_stat, rb);
144 if (new->tid < p->tid)
145 rb = &(*rb)->rb_left;
146 else if (new->tid > p->tid)
147 rb = &(*rb)->rb_right;
149 BUG_ON("inserting invalid thread_stat\n");
152 rb_link_node(&new->rb, parent, rb);
153 rb_insert_color(&new->rb, &thread_stats);
156 static struct thread_stat *thread_stat_findnew_after_first(u32 tid)
158 struct thread_stat *st;
160 st = thread_stat_find(tid);
164 st = zalloc(sizeof(struct thread_stat));
166 pr_err("memory allocation failed\n");
171 INIT_LIST_HEAD(&st->seq_list);
173 thread_stat_insert(st);
178 static struct thread_stat *thread_stat_findnew_first(u32 tid);
179 static struct thread_stat *(*thread_stat_findnew)(u32 tid) =
180 thread_stat_findnew_first;
182 static struct thread_stat *thread_stat_findnew_first(u32 tid)
184 struct thread_stat *st;
186 st = zalloc(sizeof(struct thread_stat));
188 pr_err("memory allocation failed\n");
192 INIT_LIST_HEAD(&st->seq_list);
194 rb_link_node(&st->rb, NULL, &thread_stats.rb_node);
195 rb_insert_color(&st->rb, &thread_stats);
197 thread_stat_findnew = thread_stat_findnew_after_first;
201 /* build simple key function one is bigger than two */
202 #define SINGLE_KEY(member) \
203 static int lock_stat_key_ ## member(struct lock_stat *one, \
204 struct lock_stat *two) \
206 return one->member > two->member; \
209 SINGLE_KEY(nr_acquired)
210 SINGLE_KEY(nr_contended)
211 SINGLE_KEY(wait_time_total)
212 SINGLE_KEY(wait_time_max)
214 static int lock_stat_key_wait_time_min(struct lock_stat *one,
215 struct lock_stat *two)
217 u64 s1 = one->wait_time_min;
218 u64 s2 = two->wait_time_min;
219 if (s1 == ULLONG_MAX)
221 if (s2 == ULLONG_MAX)
228 * name: the value for specify by user
229 * this should be simpler than raw name of member
230 * e.g. nr_acquired -> acquired, wait_time_total -> wait_total
233 int (*key)(struct lock_stat*, struct lock_stat*);
236 static const char *sort_key = "acquired";
238 static int (*compare)(struct lock_stat *, struct lock_stat *);
240 static struct rb_root result; /* place to store sorted data */
242 #define DEF_KEY_LOCK(name, fn_suffix) \
243 { #name, lock_stat_key_ ## fn_suffix }
244 struct lock_key keys[] = {
245 DEF_KEY_LOCK(acquired, nr_acquired),
246 DEF_KEY_LOCK(contended, nr_contended),
247 DEF_KEY_LOCK(wait_total, wait_time_total),
248 DEF_KEY_LOCK(wait_min, wait_time_min),
249 DEF_KEY_LOCK(wait_max, wait_time_max),
251 /* extra comparisons much complicated should be here */
256 static int select_key(void)
260 for (i = 0; keys[i].name; i++) {
261 if (!strcmp(keys[i].name, sort_key)) {
262 compare = keys[i].key;
267 pr_err("Unknown compare key: %s\n", sort_key);
272 static void insert_to_result(struct lock_stat *st,
273 int (*bigger)(struct lock_stat *, struct lock_stat *))
275 struct rb_node **rb = &result.rb_node;
276 struct rb_node *parent = NULL;
280 p = container_of(*rb, struct lock_stat, rb);
284 rb = &(*rb)->rb_left;
286 rb = &(*rb)->rb_right;
289 rb_link_node(&st->rb, parent, rb);
290 rb_insert_color(&st->rb, &result);
293 /* returns left most element of result, and erase it */
294 static struct lock_stat *pop_from_result(void)
296 struct rb_node *node = result.rb_node;
301 while (node->rb_left)
302 node = node->rb_left;
304 rb_erase(node, &result);
305 return container_of(node, struct lock_stat, rb);
308 static struct lock_stat *lock_stat_findnew(void *addr, const char *name)
310 struct list_head *entry = lockhashentry(addr);
311 struct lock_stat *ret, *new;
313 list_for_each_entry(ret, entry, hash_entry) {
314 if (ret->addr == addr)
318 new = zalloc(sizeof(struct lock_stat));
323 new->name = zalloc(sizeof(char) * strlen(name) + 1);
326 strcpy(new->name, name);
328 new->wait_time_min = ULLONG_MAX;
330 list_add(&new->hash_entry, entry);
334 pr_err("memory allocation failed\n");
338 struct trace_lock_handler {
339 int (*acquire_event)(struct perf_evsel *evsel,
340 struct perf_sample *sample);
342 int (*acquired_event)(struct perf_evsel *evsel,
343 struct perf_sample *sample);
345 int (*contended_event)(struct perf_evsel *evsel,
346 struct perf_sample *sample);
348 int (*release_event)(struct perf_evsel *evsel,
349 struct perf_sample *sample);
352 static struct lock_seq_stat *get_seq(struct thread_stat *ts, void *addr)
354 struct lock_seq_stat *seq;
356 list_for_each_entry(seq, &ts->seq_list, list) {
357 if (seq->addr == addr)
361 seq = zalloc(sizeof(struct lock_seq_stat));
363 pr_err("memory allocation failed\n");
366 seq->state = SEQ_STATE_UNINITIALIZED;
369 list_add(&seq->list, &ts->seq_list);
381 static int bad_hist[BROKEN_MAX];
388 static int report_lock_acquire_event(struct perf_evsel *evsel,
389 struct perf_sample *sample)
392 struct lock_stat *ls;
393 struct thread_stat *ts;
394 struct lock_seq_stat *seq;
395 const char *name = perf_evsel__strval(evsel, sample, "name");
396 u64 tmp = perf_evsel__intval(evsel, sample, "lockdep_addr");
397 int flag = perf_evsel__intval(evsel, sample, "flag");
399 memcpy(&addr, &tmp, sizeof(void *));
401 ls = lock_stat_findnew(addr, name);
407 ts = thread_stat_findnew(sample->tid);
411 seq = get_seq(ts, addr);
415 switch (seq->state) {
416 case SEQ_STATE_UNINITIALIZED:
417 case SEQ_STATE_RELEASED:
419 seq->state = SEQ_STATE_ACQUIRING;
423 if (flag & READ_LOCK)
425 seq->state = SEQ_STATE_READ_ACQUIRED;
430 case SEQ_STATE_READ_ACQUIRED:
431 if (flag & READ_LOCK) {
439 case SEQ_STATE_ACQUIRED:
440 case SEQ_STATE_ACQUIRING:
441 case SEQ_STATE_CONTENDED:
443 /* broken lock sequence, discard it */
445 bad_hist[BROKEN_ACQUIRE]++;
446 list_del(&seq->list);
450 BUG_ON("Unknown state of lock sequence found!\n");
455 seq->prev_event_time = sample->time;
460 static int report_lock_acquired_event(struct perf_evsel *evsel,
461 struct perf_sample *sample)
464 struct lock_stat *ls;
465 struct thread_stat *ts;
466 struct lock_seq_stat *seq;
468 const char *name = perf_evsel__strval(evsel, sample, "name");
469 u64 tmp = perf_evsel__intval(evsel, sample, "lockdep_addr");
471 memcpy(&addr, &tmp, sizeof(void *));
473 ls = lock_stat_findnew(addr, name);
479 ts = thread_stat_findnew(sample->tid);
483 seq = get_seq(ts, addr);
487 switch (seq->state) {
488 case SEQ_STATE_UNINITIALIZED:
489 /* orphan event, do nothing */
491 case SEQ_STATE_ACQUIRING:
493 case SEQ_STATE_CONTENDED:
494 contended_term = sample->time - seq->prev_event_time;
495 ls->wait_time_total += contended_term;
496 if (contended_term < ls->wait_time_min)
497 ls->wait_time_min = contended_term;
498 if (ls->wait_time_max < contended_term)
499 ls->wait_time_max = contended_term;
501 case SEQ_STATE_RELEASED:
502 case SEQ_STATE_ACQUIRED:
503 case SEQ_STATE_READ_ACQUIRED:
504 /* broken lock sequence, discard it */
506 bad_hist[BROKEN_ACQUIRED]++;
507 list_del(&seq->list);
511 BUG_ON("Unknown state of lock sequence found!\n");
515 seq->state = SEQ_STATE_ACQUIRED;
517 seq->prev_event_time = sample->time;
522 static int report_lock_contended_event(struct perf_evsel *evsel,
523 struct perf_sample *sample)
526 struct lock_stat *ls;
527 struct thread_stat *ts;
528 struct lock_seq_stat *seq;
529 const char *name = perf_evsel__strval(evsel, sample, "name");
530 u64 tmp = perf_evsel__intval(evsel, sample, "lockdep_addr");
532 memcpy(&addr, &tmp, sizeof(void *));
534 ls = lock_stat_findnew(addr, name);
540 ts = thread_stat_findnew(sample->tid);
544 seq = get_seq(ts, addr);
548 switch (seq->state) {
549 case SEQ_STATE_UNINITIALIZED:
550 /* orphan event, do nothing */
552 case SEQ_STATE_ACQUIRING:
554 case SEQ_STATE_RELEASED:
555 case SEQ_STATE_ACQUIRED:
556 case SEQ_STATE_READ_ACQUIRED:
557 case SEQ_STATE_CONTENDED:
558 /* broken lock sequence, discard it */
560 bad_hist[BROKEN_CONTENDED]++;
561 list_del(&seq->list);
565 BUG_ON("Unknown state of lock sequence found!\n");
569 seq->state = SEQ_STATE_CONTENDED;
571 seq->prev_event_time = sample->time;
576 static int report_lock_release_event(struct perf_evsel *evsel,
577 struct perf_sample *sample)
580 struct lock_stat *ls;
581 struct thread_stat *ts;
582 struct lock_seq_stat *seq;
583 const char *name = perf_evsel__strval(evsel, sample, "name");
584 u64 tmp = perf_evsel__intval(evsel, sample, "lockdep_addr");
586 memcpy(&addr, &tmp, sizeof(void *));
588 ls = lock_stat_findnew(addr, name);
594 ts = thread_stat_findnew(sample->tid);
598 seq = get_seq(ts, addr);
602 switch (seq->state) {
603 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]++;
623 BUG_ON("Unknown state of lock sequence found!\n");
629 list_del(&seq->list);
635 /* lock oriented handlers */
636 /* TODO: handlers for CPU oriented, thread oriented */
637 static struct trace_lock_handler report_lock_ops = {
638 .acquire_event = report_lock_acquire_event,
639 .acquired_event = report_lock_acquired_event,
640 .contended_event = report_lock_contended_event,
641 .release_event = report_lock_release_event,
644 static struct trace_lock_handler *trace_handler;
646 static int perf_evsel__process_lock_acquire(struct perf_evsel *evsel,
647 struct perf_sample *sample)
649 if (trace_handler->acquire_event)
650 return trace_handler->acquire_event(evsel, sample);
654 static int perf_evsel__process_lock_acquired(struct perf_evsel *evsel,
655 struct perf_sample *sample)
657 if (trace_handler->acquired_event)
658 return trace_handler->acquired_event(evsel, sample);
662 static int perf_evsel__process_lock_contended(struct perf_evsel *evsel,
663 struct perf_sample *sample)
665 if (trace_handler->contended_event)
666 return trace_handler->contended_event(evsel, sample);
670 static int perf_evsel__process_lock_release(struct perf_evsel *evsel,
671 struct perf_sample *sample)
673 if (trace_handler->release_event)
674 return trace_handler->release_event(evsel, sample);
678 static void print_bad_events(int bad, int total)
680 /* Output for debug, this have to be removed */
682 const char *name[4] =
683 { "acquire", "acquired", "contended", "release" };
685 pr_info("\n=== output for debug===\n\n");
686 pr_info("bad: %d, total: %d\n", bad, total);
687 pr_info("bad rate: %f %%\n", (double)bad / (double)total * 100);
688 pr_info("histogram of events caused bad sequence\n");
689 for (i = 0; i < BROKEN_MAX; i++)
690 pr_info(" %10s: %d\n", name[i], bad_hist[i]);
693 /* TODO: various way to print, coloring, nano or milli sec */
694 static void print_result(void)
696 struct lock_stat *st;
700 pr_info("%20s ", "Name");
701 pr_info("%10s ", "acquired");
702 pr_info("%10s ", "contended");
704 pr_info("%15s ", "total wait (ns)");
705 pr_info("%15s ", "max wait (ns)");
706 pr_info("%15s ", "min wait (ns)");
711 while ((st = pop_from_result())) {
719 if (strlen(st->name) < 16) {
720 /* output raw name */
721 pr_info("%20s ", st->name);
723 strncpy(cut_name, st->name, 16);
728 /* cut off name for saving output style */
729 pr_info("%20s ", cut_name);
732 pr_info("%10u ", st->nr_acquired);
733 pr_info("%10u ", st->nr_contended);
735 pr_info("%15" PRIu64 " ", st->wait_time_total);
736 pr_info("%15" PRIu64 " ", st->wait_time_max);
737 pr_info("%15" PRIu64 " ", st->wait_time_min == ULLONG_MAX ?
738 0 : st->wait_time_min);
742 print_bad_events(bad, total);
745 static bool info_threads, info_map;
747 static void dump_threads(void)
749 struct thread_stat *st;
750 struct rb_node *node;
753 pr_info("%10s: comm\n", "Thread ID");
755 node = rb_first(&thread_stats);
757 st = container_of(node, struct thread_stat, rb);
758 t = perf_session__findnew(session, st->tid);
759 pr_info("%10d: %s\n", st->tid, t->comm);
760 node = rb_next(node);
764 static void dump_map(void)
767 struct lock_stat *st;
769 pr_info("Address of instance: name of class\n");
770 for (i = 0; i < LOCKHASH_SIZE; i++) {
771 list_for_each_entry(st, &lockhash_table[i], hash_entry) {
772 pr_info(" %p: %s\n", st->addr, st->name);
777 static int dump_info(void)
787 pr_err("Unknown type of information\n");
793 typedef int (*tracepoint_handler)(struct perf_evsel *evsel,
794 struct perf_sample *sample);
796 static int process_sample_event(struct perf_tool *tool __maybe_unused,
797 union perf_event *event,
798 struct perf_sample *sample,
799 struct perf_evsel *evsel,
800 struct machine *machine)
802 struct thread *thread = machine__findnew_thread(machine, sample->pid,
805 if (thread == NULL) {
806 pr_debug("problem processing %d event, skipping it.\n",
811 if (evsel->handler.func != NULL) {
812 tracepoint_handler f = evsel->handler.func;
813 return f(evsel, sample);
819 static const struct perf_evsel_str_handler lock_tracepoints[] = {
820 { "lock:lock_acquire", perf_evsel__process_lock_acquire, }, /* CONFIG_LOCKDEP */
821 { "lock:lock_acquired", perf_evsel__process_lock_acquired, }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */
822 { "lock:lock_contended", perf_evsel__process_lock_contended, }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */
823 { "lock:lock_release", perf_evsel__process_lock_release, }, /* CONFIG_LOCKDEP */
826 static int read_events(void)
828 struct perf_tool eops = {
829 .sample = process_sample_event,
830 .comm = perf_event__process_comm,
831 .ordered_samples = true,
833 session = perf_session__new(input_name, O_RDONLY, 0, false, &eops);
835 pr_err("Initializing perf session failed\n");
839 if (perf_session__set_tracepoints_handlers(session, lock_tracepoints)) {
840 pr_err("Initializing perf session tracepoint handlers failed\n");
844 return perf_session__process_events(session, &eops);
847 static void sort_result(void)
850 struct lock_stat *st;
852 for (i = 0; i < LOCKHASH_SIZE; i++) {
853 list_for_each_entry(st, &lockhash_table[i], hash_entry) {
854 insert_to_result(st, compare);
859 static int __cmd_report(void)
863 if ((select_key() != 0) ||
864 (read_events() != 0))
873 static int __cmd_record(int argc, const char **argv)
875 const char *record_args[] = {
876 "record", "-R", "-m", "1024", "-c", "1",
878 unsigned int rec_argc, i, j;
879 const char **rec_argv;
881 for (i = 0; i < ARRAY_SIZE(lock_tracepoints); i++) {
882 if (!is_valid_tracepoint(lock_tracepoints[i].name)) {
883 pr_err("tracepoint %s is not enabled. "
884 "Are CONFIG_LOCKDEP and CONFIG_LOCK_STAT enabled?\n",
885 lock_tracepoints[i].name);
890 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
891 /* factor of 2 is for -e in front of each tracepoint */
892 rec_argc += 2 * ARRAY_SIZE(lock_tracepoints);
894 rec_argv = calloc(rec_argc + 1, sizeof(char *));
895 if (rec_argv == NULL)
898 for (i = 0; i < ARRAY_SIZE(record_args); i++)
899 rec_argv[i] = strdup(record_args[i]);
901 for (j = 0; j < ARRAY_SIZE(lock_tracepoints); j++) {
902 rec_argv[i++] = "-e";
903 rec_argv[i++] = strdup(lock_tracepoints[j].name);
906 for (j = 1; j < (unsigned int)argc; j++, i++)
907 rec_argv[i] = argv[j];
909 BUG_ON(i != rec_argc);
911 return cmd_record(i, rec_argv, NULL);
914 int cmd_lock(int argc, const char **argv, const char *prefix __maybe_unused)
916 const struct option info_options[] = {
917 OPT_BOOLEAN('t', "threads", &info_threads,
918 "dump thread list in perf.data"),
919 OPT_BOOLEAN('m', "map", &info_map,
920 "map of lock instances (address:name table)"),
923 const struct option lock_options[] = {
924 OPT_STRING('i', "input", &input_name, "file", "input file name"),
925 OPT_INCR('v', "verbose", &verbose, "be more verbose (show symbol address, etc)"),
926 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, "dump raw trace in ASCII"),
929 const struct option report_options[] = {
930 OPT_STRING('k', "key", &sort_key, "acquired",
931 "key for sorting (acquired / contended / wait_total / wait_max / wait_min)"),
935 const char * const info_usage[] = {
936 "perf lock info [<options>]",
939 const char * const lock_usage[] = {
940 "perf lock [<options>] {record|report|script|info}",
943 const char * const report_usage[] = {
944 "perf lock report [<options>]",
951 for (i = 0; i < LOCKHASH_SIZE; i++)
952 INIT_LIST_HEAD(lockhash_table + i);
954 argc = parse_options(argc, argv, lock_options, lock_usage,
955 PARSE_OPT_STOP_AT_NON_OPTION);
957 usage_with_options(lock_usage, lock_options);
959 if (!strncmp(argv[0], "rec", 3)) {
960 return __cmd_record(argc, argv);
961 } else if (!strncmp(argv[0], "report", 6)) {
962 trace_handler = &report_lock_ops;
964 argc = parse_options(argc, argv,
965 report_options, report_usage, 0);
967 usage_with_options(report_usage, report_options);
970 } else if (!strcmp(argv[0], "script")) {
971 /* Aliased to 'perf script' */
972 return cmd_script(argc, argv, prefix);
973 } else if (!strcmp(argv[0], "info")) {
975 argc = parse_options(argc, argv,
976 info_options, info_usage, 0);
978 usage_with_options(info_usage, info_options);
980 /* recycling report_lock_ops */
981 trace_handler = &report_lock_ops;
983 if (read_events() != 0)
988 usage_with_options(lock_usage, lock_options);