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);
451 BUG_ON("Unknown state of lock sequence found!\n");
456 seq->prev_event_time = sample->time;
461 static int report_lock_acquired_event(struct perf_evsel *evsel,
462 struct perf_sample *sample)
465 struct lock_stat *ls;
466 struct thread_stat *ts;
467 struct lock_seq_stat *seq;
469 const char *name = perf_evsel__strval(evsel, sample, "name");
470 u64 tmp = perf_evsel__intval(evsel, sample, "lockdep_addr");
472 memcpy(&addr, &tmp, sizeof(void *));
474 ls = lock_stat_findnew(addr, name);
480 ts = thread_stat_findnew(sample->tid);
484 seq = get_seq(ts, addr);
488 switch (seq->state) {
489 case SEQ_STATE_UNINITIALIZED:
490 /* orphan event, do nothing */
492 case SEQ_STATE_ACQUIRING:
494 case SEQ_STATE_CONTENDED:
495 contended_term = sample->time - seq->prev_event_time;
496 ls->wait_time_total += contended_term;
497 if (contended_term < ls->wait_time_min)
498 ls->wait_time_min = contended_term;
499 if (ls->wait_time_max < contended_term)
500 ls->wait_time_max = contended_term;
502 case SEQ_STATE_RELEASED:
503 case SEQ_STATE_ACQUIRED:
504 case SEQ_STATE_READ_ACQUIRED:
505 /* broken lock sequence, discard it */
507 bad_hist[BROKEN_ACQUIRED]++;
508 list_del(&seq->list);
514 BUG_ON("Unknown state of lock sequence found!\n");
518 seq->state = SEQ_STATE_ACQUIRED;
520 seq->prev_event_time = sample->time;
525 static int report_lock_contended_event(struct perf_evsel *evsel,
526 struct perf_sample *sample)
529 struct lock_stat *ls;
530 struct thread_stat *ts;
531 struct lock_seq_stat *seq;
532 const char *name = perf_evsel__strval(evsel, sample, "name");
533 u64 tmp = perf_evsel__intval(evsel, sample, "lockdep_addr");
535 memcpy(&addr, &tmp, sizeof(void *));
537 ls = lock_stat_findnew(addr, name);
543 ts = thread_stat_findnew(sample->tid);
547 seq = get_seq(ts, addr);
551 switch (seq->state) {
552 case SEQ_STATE_UNINITIALIZED:
553 /* orphan event, do nothing */
555 case SEQ_STATE_ACQUIRING:
557 case SEQ_STATE_RELEASED:
558 case SEQ_STATE_ACQUIRED:
559 case SEQ_STATE_READ_ACQUIRED:
560 case SEQ_STATE_CONTENDED:
561 /* broken lock sequence, discard it */
563 bad_hist[BROKEN_CONTENDED]++;
564 list_del(&seq->list);
569 BUG_ON("Unknown state of lock sequence found!\n");
573 seq->state = SEQ_STATE_CONTENDED;
575 seq->prev_event_time = sample->time;
580 static int report_lock_release_event(struct perf_evsel *evsel,
581 struct perf_sample *sample)
584 struct lock_stat *ls;
585 struct thread_stat *ts;
586 struct lock_seq_stat *seq;
587 const char *name = perf_evsel__strval(evsel, sample, "name");
588 u64 tmp = perf_evsel__intval(evsel, sample, "lockdep_addr");
590 memcpy(&addr, &tmp, sizeof(void *));
592 ls = lock_stat_findnew(addr, name);
598 ts = thread_stat_findnew(sample->tid);
602 seq = get_seq(ts, addr);
606 switch (seq->state) {
607 case SEQ_STATE_UNINITIALIZED:
610 case SEQ_STATE_ACQUIRED:
612 case SEQ_STATE_READ_ACQUIRED:
614 BUG_ON(seq->read_count < 0);
615 if (!seq->read_count) {
620 case SEQ_STATE_ACQUIRING:
621 case SEQ_STATE_CONTENDED:
622 case SEQ_STATE_RELEASED:
623 /* broken lock sequence, discard it */
625 bad_hist[BROKEN_RELEASE]++;
629 BUG_ON("Unknown state of lock sequence found!\n");
635 list_del(&seq->list);
641 /* lock oriented handlers */
642 /* TODO: handlers for CPU oriented, thread oriented */
643 static struct trace_lock_handler report_lock_ops = {
644 .acquire_event = report_lock_acquire_event,
645 .acquired_event = report_lock_acquired_event,
646 .contended_event = report_lock_contended_event,
647 .release_event = report_lock_release_event,
650 static struct trace_lock_handler *trace_handler;
652 static int perf_evsel__process_lock_acquire(struct perf_evsel *evsel,
653 struct perf_sample *sample)
655 if (trace_handler->acquire_event)
656 return trace_handler->acquire_event(evsel, sample);
660 static int perf_evsel__process_lock_acquired(struct perf_evsel *evsel,
661 struct perf_sample *sample)
663 if (trace_handler->acquired_event)
664 return trace_handler->acquired_event(evsel, sample);
668 static int perf_evsel__process_lock_contended(struct perf_evsel *evsel,
669 struct perf_sample *sample)
671 if (trace_handler->contended_event)
672 return trace_handler->contended_event(evsel, sample);
676 static int perf_evsel__process_lock_release(struct perf_evsel *evsel,
677 struct perf_sample *sample)
679 if (trace_handler->release_event)
680 return trace_handler->release_event(evsel, sample);
684 static void print_bad_events(int bad, int total)
686 /* Output for debug, this have to be removed */
688 const char *name[4] =
689 { "acquire", "acquired", "contended", "release" };
691 pr_info("\n=== output for debug===\n\n");
692 pr_info("bad: %d, total: %d\n", bad, total);
693 pr_info("bad rate: %f %%\n", (double)bad / (double)total * 100);
694 pr_info("histogram of events caused bad sequence\n");
695 for (i = 0; i < BROKEN_MAX; i++)
696 pr_info(" %10s: %d\n", name[i], bad_hist[i]);
699 /* TODO: various way to print, coloring, nano or milli sec */
700 static void print_result(void)
702 struct lock_stat *st;
706 pr_info("%20s ", "Name");
707 pr_info("%10s ", "acquired");
708 pr_info("%10s ", "contended");
710 pr_info("%15s ", "total wait (ns)");
711 pr_info("%15s ", "max wait (ns)");
712 pr_info("%15s ", "min wait (ns)");
717 while ((st = pop_from_result())) {
725 if (strlen(st->name) < 16) {
726 /* output raw name */
727 pr_info("%20s ", st->name);
729 strncpy(cut_name, st->name, 16);
734 /* cut off name for saving output style */
735 pr_info("%20s ", cut_name);
738 pr_info("%10u ", st->nr_acquired);
739 pr_info("%10u ", st->nr_contended);
741 pr_info("%15" PRIu64 " ", st->wait_time_total);
742 pr_info("%15" PRIu64 " ", st->wait_time_max);
743 pr_info("%15" PRIu64 " ", st->wait_time_min == ULLONG_MAX ?
744 0 : st->wait_time_min);
748 print_bad_events(bad, total);
751 static bool info_threads, info_map;
753 static void dump_threads(void)
755 struct thread_stat *st;
756 struct rb_node *node;
759 pr_info("%10s: comm\n", "Thread ID");
761 node = rb_first(&thread_stats);
763 st = container_of(node, struct thread_stat, rb);
764 t = perf_session__findnew(session, st->tid);
765 pr_info("%10d: %s\n", st->tid, t->comm);
766 node = rb_next(node);
770 static void dump_map(void)
773 struct lock_stat *st;
775 pr_info("Address of instance: name of class\n");
776 for (i = 0; i < LOCKHASH_SIZE; i++) {
777 list_for_each_entry(st, &lockhash_table[i], hash_entry) {
778 pr_info(" %p: %s\n", st->addr, st->name);
783 static int dump_info(void)
793 pr_err("Unknown type of information\n");
799 typedef int (*tracepoint_handler)(struct perf_evsel *evsel,
800 struct perf_sample *sample);
802 static int process_sample_event(struct perf_tool *tool __maybe_unused,
803 union perf_event *event,
804 struct perf_sample *sample,
805 struct perf_evsel *evsel,
806 struct machine *machine)
808 struct thread *thread = machine__findnew_thread(machine, sample->tid);
810 if (thread == NULL) {
811 pr_debug("problem processing %d event, skipping it.\n",
816 if (evsel->handler.func != NULL) {
817 tracepoint_handler f = evsel->handler.func;
818 return f(evsel, sample);
824 static const struct perf_evsel_str_handler lock_tracepoints[] = {
825 { "lock:lock_acquire", perf_evsel__process_lock_acquire, }, /* CONFIG_LOCKDEP */
826 { "lock:lock_acquired", perf_evsel__process_lock_acquired, }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */
827 { "lock:lock_contended", perf_evsel__process_lock_contended, }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */
828 { "lock:lock_release", perf_evsel__process_lock_release, }, /* CONFIG_LOCKDEP */
831 static int read_events(void)
833 struct perf_tool eops = {
834 .sample = process_sample_event,
835 .comm = perf_event__process_comm,
836 .ordered_samples = true,
838 session = perf_session__new(input_name, O_RDONLY, 0, false, &eops);
840 pr_err("Initializing perf session failed\n");
844 if (perf_session__set_tracepoints_handlers(session, lock_tracepoints)) {
845 pr_err("Initializing perf session tracepoint handlers failed\n");
849 return perf_session__process_events(session, &eops);
852 static void sort_result(void)
855 struct lock_stat *st;
857 for (i = 0; i < LOCKHASH_SIZE; i++) {
858 list_for_each_entry(st, &lockhash_table[i], hash_entry) {
859 insert_to_result(st, compare);
864 static int __cmd_report(void)
868 if ((select_key() != 0) ||
869 (read_events() != 0))
878 static int __cmd_record(int argc, const char **argv)
880 const char *record_args[] = {
881 "record", "-R", "-f", "-m", "1024", "-c", "1",
883 unsigned int rec_argc, i, j;
884 const char **rec_argv;
886 for (i = 0; i < ARRAY_SIZE(lock_tracepoints); i++) {
887 if (!is_valid_tracepoint(lock_tracepoints[i].name)) {
888 pr_err("tracepoint %s is not enabled. "
889 "Are CONFIG_LOCKDEP and CONFIG_LOCK_STAT enabled?\n",
890 lock_tracepoints[i].name);
895 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
896 /* factor of 2 is for -e in front of each tracepoint */
897 rec_argc += 2 * ARRAY_SIZE(lock_tracepoints);
899 rec_argv = calloc(rec_argc + 1, sizeof(char *));
900 if (rec_argv == NULL)
903 for (i = 0; i < ARRAY_SIZE(record_args); i++)
904 rec_argv[i] = strdup(record_args[i]);
906 for (j = 0; j < ARRAY_SIZE(lock_tracepoints); j++) {
907 rec_argv[i++] = "-e";
908 rec_argv[i++] = strdup(lock_tracepoints[j].name);
911 for (j = 1; j < (unsigned int)argc; j++, i++)
912 rec_argv[i] = argv[j];
914 BUG_ON(i != rec_argc);
916 return cmd_record(i, rec_argv, NULL);
919 int cmd_lock(int argc, const char **argv, const char *prefix __maybe_unused)
921 const struct option info_options[] = {
922 OPT_BOOLEAN('t', "threads", &info_threads,
923 "dump thread list in perf.data"),
924 OPT_BOOLEAN('m', "map", &info_map,
925 "map of lock instances (address:name table)"),
928 const struct option lock_options[] = {
929 OPT_STRING('i', "input", &input_name, "file", "input file name"),
930 OPT_INCR('v', "verbose", &verbose, "be more verbose (show symbol address, etc)"),
931 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, "dump raw trace in ASCII"),
934 const struct option report_options[] = {
935 OPT_STRING('k', "key", &sort_key, "acquired",
936 "key for sorting (acquired / contended / wait_total / wait_max / wait_min)"),
940 const char * const info_usage[] = {
941 "perf lock info [<options>]",
944 const char * const lock_usage[] = {
945 "perf lock [<options>] {record|report|script|info}",
948 const char * const report_usage[] = {
949 "perf lock report [<options>]",
956 for (i = 0; i < LOCKHASH_SIZE; i++)
957 INIT_LIST_HEAD(lockhash_table + i);
959 argc = parse_options(argc, argv, lock_options, lock_usage,
960 PARSE_OPT_STOP_AT_NON_OPTION);
962 usage_with_options(lock_usage, lock_options);
964 if (!strncmp(argv[0], "rec", 3)) {
965 return __cmd_record(argc, argv);
966 } else if (!strncmp(argv[0], "report", 6)) {
967 trace_handler = &report_lock_ops;
969 argc = parse_options(argc, argv,
970 report_options, report_usage, 0);
972 usage_with_options(report_usage, report_options);
975 } else if (!strcmp(argv[0], "script")) {
976 /* Aliased to 'perf script' */
977 return cmd_script(argc, argv, prefix);
978 } else if (!strcmp(argv[0], "info")) {
980 argc = parse_options(argc, argv,
981 info_options, info_usage, 0);
983 usage_with_options(info_usage, info_options);
985 /* recycling report_lock_ops */
986 trace_handler = &report_lock_ops;
988 if (read_events() != 0)
993 usage_with_options(lock_usage, lock_options);