Loading...
1// SPDX-License-Identifier: GPL-2.0
2#include <errno.h>
3#include <inttypes.h>
4#include "builtin.h"
5#include "perf.h"
6
7#include "util/evlist.h"
8#include "util/evsel.h"
9#include "util/util.h"
10#include "util/cache.h"
11#include "util/symbol.h"
12#include "util/thread.h"
13#include "util/header.h"
14
15#include <subcmd/parse-options.h>
16#include "util/trace-event.h"
17
18#include "util/debug.h"
19#include "util/session.h"
20#include "util/tool.h"
21#include "util/data.h"
22
23#include <sys/types.h>
24#include <sys/prctl.h>
25#include <semaphore.h>
26#include <pthread.h>
27#include <math.h>
28#include <limits.h>
29
30#include <linux/list.h>
31#include <linux/hash.h>
32#include <linux/kernel.h>
33
34static struct perf_session *session;
35
36/* based on kernel/lockdep.c */
37#define LOCKHASH_BITS 12
38#define LOCKHASH_SIZE (1UL << LOCKHASH_BITS)
39
40static struct list_head lockhash_table[LOCKHASH_SIZE];
41
42#define __lockhashfn(key) hash_long((unsigned long)key, LOCKHASH_BITS)
43#define lockhashentry(key) (lockhash_table + __lockhashfn((key)))
44
45struct lock_stat {
46 struct list_head hash_entry;
47 struct rb_node rb; /* used for sorting */
48
49 /*
50 * FIXME: perf_evsel__intval() returns u64,
51 * so address of lockdep_map should be dealed as 64bit.
52 * Is there more better solution?
53 */
54 void *addr; /* address of lockdep_map, used as ID */
55 char *name; /* for strcpy(), we cannot use const */
56
57 unsigned int nr_acquire;
58 unsigned int nr_acquired;
59 unsigned int nr_contended;
60 unsigned int nr_release;
61
62 unsigned int nr_readlock;
63 unsigned int nr_trylock;
64
65 /* these times are in nano sec. */
66 u64 avg_wait_time;
67 u64 wait_time_total;
68 u64 wait_time_min;
69 u64 wait_time_max;
70
71 int discard; /* flag of blacklist */
72};
73
74/*
75 * States of lock_seq_stat
76 *
77 * UNINITIALIZED is required for detecting first event of acquire.
78 * As the nature of lock events, there is no guarantee
79 * that the first event for the locks are acquire,
80 * it can be acquired, contended or release.
81 */
82#define SEQ_STATE_UNINITIALIZED 0 /* initial state */
83#define SEQ_STATE_RELEASED 1
84#define SEQ_STATE_ACQUIRING 2
85#define SEQ_STATE_ACQUIRED 3
86#define SEQ_STATE_READ_ACQUIRED 4
87#define SEQ_STATE_CONTENDED 5
88
89/*
90 * MAX_LOCK_DEPTH
91 * Imported from include/linux/sched.h.
92 * Should this be synchronized?
93 */
94#define MAX_LOCK_DEPTH 48
95
96/*
97 * struct lock_seq_stat:
98 * Place to put on state of one lock sequence
99 * 1) acquire -> acquired -> release
100 * 2) acquire -> contended -> acquired -> release
101 * 3) acquire (with read or try) -> release
102 * 4) Are there other patterns?
103 */
104struct lock_seq_stat {
105 struct list_head list;
106 int state;
107 u64 prev_event_time;
108 void *addr;
109
110 int read_count;
111};
112
113struct thread_stat {
114 struct rb_node rb;
115
116 u32 tid;
117 struct list_head seq_list;
118};
119
120static struct rb_root thread_stats;
121
122static struct thread_stat *thread_stat_find(u32 tid)
123{
124 struct rb_node *node;
125 struct thread_stat *st;
126
127 node = thread_stats.rb_node;
128 while (node) {
129 st = container_of(node, struct thread_stat, rb);
130 if (st->tid == tid)
131 return st;
132 else if (tid < st->tid)
133 node = node->rb_left;
134 else
135 node = node->rb_right;
136 }
137
138 return NULL;
139}
140
141static void thread_stat_insert(struct thread_stat *new)
142{
143 struct rb_node **rb = &thread_stats.rb_node;
144 struct rb_node *parent = NULL;
145 struct thread_stat *p;
146
147 while (*rb) {
148 p = container_of(*rb, struct thread_stat, rb);
149 parent = *rb;
150
151 if (new->tid < p->tid)
152 rb = &(*rb)->rb_left;
153 else if (new->tid > p->tid)
154 rb = &(*rb)->rb_right;
155 else
156 BUG_ON("inserting invalid thread_stat\n");
157 }
158
159 rb_link_node(&new->rb, parent, rb);
160 rb_insert_color(&new->rb, &thread_stats);
161}
162
163static struct thread_stat *thread_stat_findnew_after_first(u32 tid)
164{
165 struct thread_stat *st;
166
167 st = thread_stat_find(tid);
168 if (st)
169 return st;
170
171 st = zalloc(sizeof(struct thread_stat));
172 if (!st) {
173 pr_err("memory allocation failed\n");
174 return NULL;
175 }
176
177 st->tid = tid;
178 INIT_LIST_HEAD(&st->seq_list);
179
180 thread_stat_insert(st);
181
182 return st;
183}
184
185static struct thread_stat *thread_stat_findnew_first(u32 tid);
186static struct thread_stat *(*thread_stat_findnew)(u32 tid) =
187 thread_stat_findnew_first;
188
189static struct thread_stat *thread_stat_findnew_first(u32 tid)
190{
191 struct thread_stat *st;
192
193 st = zalloc(sizeof(struct thread_stat));
194 if (!st) {
195 pr_err("memory allocation failed\n");
196 return NULL;
197 }
198 st->tid = tid;
199 INIT_LIST_HEAD(&st->seq_list);
200
201 rb_link_node(&st->rb, NULL, &thread_stats.rb_node);
202 rb_insert_color(&st->rb, &thread_stats);
203
204 thread_stat_findnew = thread_stat_findnew_after_first;
205 return st;
206}
207
208/* build simple key function one is bigger than two */
209#define SINGLE_KEY(member) \
210 static int lock_stat_key_ ## member(struct lock_stat *one, \
211 struct lock_stat *two) \
212 { \
213 return one->member > two->member; \
214 }
215
216SINGLE_KEY(nr_acquired)
217SINGLE_KEY(nr_contended)
218SINGLE_KEY(avg_wait_time)
219SINGLE_KEY(wait_time_total)
220SINGLE_KEY(wait_time_max)
221
222static int lock_stat_key_wait_time_min(struct lock_stat *one,
223 struct lock_stat *two)
224{
225 u64 s1 = one->wait_time_min;
226 u64 s2 = two->wait_time_min;
227 if (s1 == ULLONG_MAX)
228 s1 = 0;
229 if (s2 == ULLONG_MAX)
230 s2 = 0;
231 return s1 > s2;
232}
233
234struct lock_key {
235 /*
236 * name: the value for specify by user
237 * this should be simpler than raw name of member
238 * e.g. nr_acquired -> acquired, wait_time_total -> wait_total
239 */
240 const char *name;
241 int (*key)(struct lock_stat*, struct lock_stat*);
242};
243
244static const char *sort_key = "acquired";
245
246static int (*compare)(struct lock_stat *, struct lock_stat *);
247
248static struct rb_root result; /* place to store sorted data */
249
250#define DEF_KEY_LOCK(name, fn_suffix) \
251 { #name, lock_stat_key_ ## fn_suffix }
252struct lock_key keys[] = {
253 DEF_KEY_LOCK(acquired, nr_acquired),
254 DEF_KEY_LOCK(contended, nr_contended),
255 DEF_KEY_LOCK(avg_wait, avg_wait_time),
256 DEF_KEY_LOCK(wait_total, wait_time_total),
257 DEF_KEY_LOCK(wait_min, wait_time_min),
258 DEF_KEY_LOCK(wait_max, wait_time_max),
259
260 /* extra comparisons much complicated should be here */
261
262 { NULL, NULL }
263};
264
265static int select_key(void)
266{
267 int i;
268
269 for (i = 0; keys[i].name; i++) {
270 if (!strcmp(keys[i].name, sort_key)) {
271 compare = keys[i].key;
272 return 0;
273 }
274 }
275
276 pr_err("Unknown compare key: %s\n", sort_key);
277
278 return -1;
279}
280
281static void insert_to_result(struct lock_stat *st,
282 int (*bigger)(struct lock_stat *, struct lock_stat *))
283{
284 struct rb_node **rb = &result.rb_node;
285 struct rb_node *parent = NULL;
286 struct lock_stat *p;
287
288 while (*rb) {
289 p = container_of(*rb, struct lock_stat, rb);
290 parent = *rb;
291
292 if (bigger(st, p))
293 rb = &(*rb)->rb_left;
294 else
295 rb = &(*rb)->rb_right;
296 }
297
298 rb_link_node(&st->rb, parent, rb);
299 rb_insert_color(&st->rb, &result);
300}
301
302/* returns left most element of result, and erase it */
303static struct lock_stat *pop_from_result(void)
304{
305 struct rb_node *node = result.rb_node;
306
307 if (!node)
308 return NULL;
309
310 while (node->rb_left)
311 node = node->rb_left;
312
313 rb_erase(node, &result);
314 return container_of(node, struct lock_stat, rb);
315}
316
317static struct lock_stat *lock_stat_findnew(void *addr, const char *name)
318{
319 struct list_head *entry = lockhashentry(addr);
320 struct lock_stat *ret, *new;
321
322 list_for_each_entry(ret, entry, hash_entry) {
323 if (ret->addr == addr)
324 return ret;
325 }
326
327 new = zalloc(sizeof(struct lock_stat));
328 if (!new)
329 goto alloc_failed;
330
331 new->addr = addr;
332 new->name = zalloc(sizeof(char) * strlen(name) + 1);
333 if (!new->name) {
334 free(new);
335 goto alloc_failed;
336 }
337
338 strcpy(new->name, name);
339 new->wait_time_min = ULLONG_MAX;
340
341 list_add(&new->hash_entry, entry);
342 return new;
343
344alloc_failed:
345 pr_err("memory allocation failed\n");
346 return NULL;
347}
348
349struct trace_lock_handler {
350 int (*acquire_event)(struct perf_evsel *evsel,
351 struct perf_sample *sample);
352
353 int (*acquired_event)(struct perf_evsel *evsel,
354 struct perf_sample *sample);
355
356 int (*contended_event)(struct perf_evsel *evsel,
357 struct perf_sample *sample);
358
359 int (*release_event)(struct perf_evsel *evsel,
360 struct perf_sample *sample);
361};
362
363static struct lock_seq_stat *get_seq(struct thread_stat *ts, void *addr)
364{
365 struct lock_seq_stat *seq;
366
367 list_for_each_entry(seq, &ts->seq_list, list) {
368 if (seq->addr == addr)
369 return seq;
370 }
371
372 seq = zalloc(sizeof(struct lock_seq_stat));
373 if (!seq) {
374 pr_err("memory allocation failed\n");
375 return NULL;
376 }
377 seq->state = SEQ_STATE_UNINITIALIZED;
378 seq->addr = addr;
379
380 list_add(&seq->list, &ts->seq_list);
381 return seq;
382}
383
384enum broken_state {
385 BROKEN_ACQUIRE,
386 BROKEN_ACQUIRED,
387 BROKEN_CONTENDED,
388 BROKEN_RELEASE,
389 BROKEN_MAX,
390};
391
392static int bad_hist[BROKEN_MAX];
393
394enum acquire_flags {
395 TRY_LOCK = 1,
396 READ_LOCK = 2,
397};
398
399static int report_lock_acquire_event(struct perf_evsel *evsel,
400 struct perf_sample *sample)
401{
402 void *addr;
403 struct lock_stat *ls;
404 struct thread_stat *ts;
405 struct lock_seq_stat *seq;
406 const char *name = perf_evsel__strval(evsel, sample, "name");
407 u64 tmp = perf_evsel__intval(evsel, sample, "lockdep_addr");
408 int flag = perf_evsel__intval(evsel, sample, "flag");
409
410 memcpy(&addr, &tmp, sizeof(void *));
411
412 ls = lock_stat_findnew(addr, name);
413 if (!ls)
414 return -ENOMEM;
415 if (ls->discard)
416 return 0;
417
418 ts = thread_stat_findnew(sample->tid);
419 if (!ts)
420 return -ENOMEM;
421
422 seq = get_seq(ts, addr);
423 if (!seq)
424 return -ENOMEM;
425
426 switch (seq->state) {
427 case SEQ_STATE_UNINITIALIZED:
428 case SEQ_STATE_RELEASED:
429 if (!flag) {
430 seq->state = SEQ_STATE_ACQUIRING;
431 } else {
432 if (flag & TRY_LOCK)
433 ls->nr_trylock++;
434 if (flag & READ_LOCK)
435 ls->nr_readlock++;
436 seq->state = SEQ_STATE_READ_ACQUIRED;
437 seq->read_count = 1;
438 ls->nr_acquired++;
439 }
440 break;
441 case SEQ_STATE_READ_ACQUIRED:
442 if (flag & READ_LOCK) {
443 seq->read_count++;
444 ls->nr_acquired++;
445 goto end;
446 } else {
447 goto broken;
448 }
449 break;
450 case SEQ_STATE_ACQUIRED:
451 case SEQ_STATE_ACQUIRING:
452 case SEQ_STATE_CONTENDED:
453broken:
454 /* broken lock sequence, discard it */
455 ls->discard = 1;
456 bad_hist[BROKEN_ACQUIRE]++;
457 list_del(&seq->list);
458 free(seq);
459 goto end;
460 default:
461 BUG_ON("Unknown state of lock sequence found!\n");
462 break;
463 }
464
465 ls->nr_acquire++;
466 seq->prev_event_time = sample->time;
467end:
468 return 0;
469}
470
471static int report_lock_acquired_event(struct perf_evsel *evsel,
472 struct perf_sample *sample)
473{
474 void *addr;
475 struct lock_stat *ls;
476 struct thread_stat *ts;
477 struct lock_seq_stat *seq;
478 u64 contended_term;
479 const char *name = perf_evsel__strval(evsel, sample, "name");
480 u64 tmp = perf_evsel__intval(evsel, sample, "lockdep_addr");
481
482 memcpy(&addr, &tmp, sizeof(void *));
483
484 ls = lock_stat_findnew(addr, name);
485 if (!ls)
486 return -ENOMEM;
487 if (ls->discard)
488 return 0;
489
490 ts = thread_stat_findnew(sample->tid);
491 if (!ts)
492 return -ENOMEM;
493
494 seq = get_seq(ts, addr);
495 if (!seq)
496 return -ENOMEM;
497
498 switch (seq->state) {
499 case SEQ_STATE_UNINITIALIZED:
500 /* orphan event, do nothing */
501 return 0;
502 case SEQ_STATE_ACQUIRING:
503 break;
504 case SEQ_STATE_CONTENDED:
505 contended_term = sample->time - seq->prev_event_time;
506 ls->wait_time_total += contended_term;
507 if (contended_term < ls->wait_time_min)
508 ls->wait_time_min = contended_term;
509 if (ls->wait_time_max < contended_term)
510 ls->wait_time_max = contended_term;
511 break;
512 case SEQ_STATE_RELEASED:
513 case SEQ_STATE_ACQUIRED:
514 case SEQ_STATE_READ_ACQUIRED:
515 /* broken lock sequence, discard it */
516 ls->discard = 1;
517 bad_hist[BROKEN_ACQUIRED]++;
518 list_del(&seq->list);
519 free(seq);
520 goto end;
521 default:
522 BUG_ON("Unknown state of lock sequence found!\n");
523 break;
524 }
525
526 seq->state = SEQ_STATE_ACQUIRED;
527 ls->nr_acquired++;
528 ls->avg_wait_time = ls->nr_contended ? ls->wait_time_total/ls->nr_contended : 0;
529 seq->prev_event_time = sample->time;
530end:
531 return 0;
532}
533
534static int report_lock_contended_event(struct perf_evsel *evsel,
535 struct perf_sample *sample)
536{
537 void *addr;
538 struct lock_stat *ls;
539 struct thread_stat *ts;
540 struct lock_seq_stat *seq;
541 const char *name = perf_evsel__strval(evsel, sample, "name");
542 u64 tmp = perf_evsel__intval(evsel, sample, "lockdep_addr");
543
544 memcpy(&addr, &tmp, sizeof(void *));
545
546 ls = lock_stat_findnew(addr, name);
547 if (!ls)
548 return -ENOMEM;
549 if (ls->discard)
550 return 0;
551
552 ts = thread_stat_findnew(sample->tid);
553 if (!ts)
554 return -ENOMEM;
555
556 seq = get_seq(ts, addr);
557 if (!seq)
558 return -ENOMEM;
559
560 switch (seq->state) {
561 case SEQ_STATE_UNINITIALIZED:
562 /* orphan event, do nothing */
563 return 0;
564 case SEQ_STATE_ACQUIRING:
565 break;
566 case SEQ_STATE_RELEASED:
567 case SEQ_STATE_ACQUIRED:
568 case SEQ_STATE_READ_ACQUIRED:
569 case SEQ_STATE_CONTENDED:
570 /* broken lock sequence, discard it */
571 ls->discard = 1;
572 bad_hist[BROKEN_CONTENDED]++;
573 list_del(&seq->list);
574 free(seq);
575 goto end;
576 default:
577 BUG_ON("Unknown state of lock sequence found!\n");
578 break;
579 }
580
581 seq->state = SEQ_STATE_CONTENDED;
582 ls->nr_contended++;
583 ls->avg_wait_time = ls->wait_time_total/ls->nr_contended;
584 seq->prev_event_time = sample->time;
585end:
586 return 0;
587}
588
589static int report_lock_release_event(struct perf_evsel *evsel,
590 struct perf_sample *sample)
591{
592 void *addr;
593 struct lock_stat *ls;
594 struct thread_stat *ts;
595 struct lock_seq_stat *seq;
596 const char *name = perf_evsel__strval(evsel, sample, "name");
597 u64 tmp = perf_evsel__intval(evsel, sample, "lockdep_addr");
598
599 memcpy(&addr, &tmp, sizeof(void *));
600
601 ls = lock_stat_findnew(addr, name);
602 if (!ls)
603 return -ENOMEM;
604 if (ls->discard)
605 return 0;
606
607 ts = thread_stat_findnew(sample->tid);
608 if (!ts)
609 return -ENOMEM;
610
611 seq = get_seq(ts, addr);
612 if (!seq)
613 return -ENOMEM;
614
615 switch (seq->state) {
616 case SEQ_STATE_UNINITIALIZED:
617 goto end;
618 case SEQ_STATE_ACQUIRED:
619 break;
620 case SEQ_STATE_READ_ACQUIRED:
621 seq->read_count--;
622 BUG_ON(seq->read_count < 0);
623 if (!seq->read_count) {
624 ls->nr_release++;
625 goto end;
626 }
627 break;
628 case SEQ_STATE_ACQUIRING:
629 case SEQ_STATE_CONTENDED:
630 case SEQ_STATE_RELEASED:
631 /* broken lock sequence, discard it */
632 ls->discard = 1;
633 bad_hist[BROKEN_RELEASE]++;
634 goto free_seq;
635 default:
636 BUG_ON("Unknown state of lock sequence found!\n");
637 break;
638 }
639
640 ls->nr_release++;
641free_seq:
642 list_del(&seq->list);
643 free(seq);
644end:
645 return 0;
646}
647
648/* lock oriented handlers */
649/* TODO: handlers for CPU oriented, thread oriented */
650static struct trace_lock_handler report_lock_ops = {
651 .acquire_event = report_lock_acquire_event,
652 .acquired_event = report_lock_acquired_event,
653 .contended_event = report_lock_contended_event,
654 .release_event = report_lock_release_event,
655};
656
657static struct trace_lock_handler *trace_handler;
658
659static int perf_evsel__process_lock_acquire(struct perf_evsel *evsel,
660 struct perf_sample *sample)
661{
662 if (trace_handler->acquire_event)
663 return trace_handler->acquire_event(evsel, sample);
664 return 0;
665}
666
667static int perf_evsel__process_lock_acquired(struct perf_evsel *evsel,
668 struct perf_sample *sample)
669{
670 if (trace_handler->acquired_event)
671 return trace_handler->acquired_event(evsel, sample);
672 return 0;
673}
674
675static int perf_evsel__process_lock_contended(struct perf_evsel *evsel,
676 struct perf_sample *sample)
677{
678 if (trace_handler->contended_event)
679 return trace_handler->contended_event(evsel, sample);
680 return 0;
681}
682
683static int perf_evsel__process_lock_release(struct perf_evsel *evsel,
684 struct perf_sample *sample)
685{
686 if (trace_handler->release_event)
687 return trace_handler->release_event(evsel, sample);
688 return 0;
689}
690
691static void print_bad_events(int bad, int total)
692{
693 /* Output for debug, this have to be removed */
694 int i;
695 const char *name[4] =
696 { "acquire", "acquired", "contended", "release" };
697
698 pr_info("\n=== output for debug===\n\n");
699 pr_info("bad: %d, total: %d\n", bad, total);
700 pr_info("bad rate: %.2f %%\n", (double)bad / (double)total * 100);
701 pr_info("histogram of events caused bad sequence\n");
702 for (i = 0; i < BROKEN_MAX; i++)
703 pr_info(" %10s: %d\n", name[i], bad_hist[i]);
704}
705
706/* TODO: various way to print, coloring, nano or milli sec */
707static void print_result(void)
708{
709 struct lock_stat *st;
710 char cut_name[20];
711 int bad, total;
712
713 pr_info("%20s ", "Name");
714 pr_info("%10s ", "acquired");
715 pr_info("%10s ", "contended");
716
717 pr_info("%15s ", "avg wait (ns)");
718 pr_info("%15s ", "total wait (ns)");
719 pr_info("%15s ", "max wait (ns)");
720 pr_info("%15s ", "min wait (ns)");
721
722 pr_info("\n\n");
723
724 bad = total = 0;
725 while ((st = pop_from_result())) {
726 total++;
727 if (st->discard) {
728 bad++;
729 continue;
730 }
731 bzero(cut_name, 20);
732
733 if (strlen(st->name) < 16) {
734 /* output raw name */
735 pr_info("%20s ", st->name);
736 } else {
737 strncpy(cut_name, st->name, 16);
738 cut_name[16] = '.';
739 cut_name[17] = '.';
740 cut_name[18] = '.';
741 cut_name[19] = '\0';
742 /* cut off name for saving output style */
743 pr_info("%20s ", cut_name);
744 }
745
746 pr_info("%10u ", st->nr_acquired);
747 pr_info("%10u ", st->nr_contended);
748
749 pr_info("%15" PRIu64 " ", st->avg_wait_time);
750 pr_info("%15" PRIu64 " ", st->wait_time_total);
751 pr_info("%15" PRIu64 " ", st->wait_time_max);
752 pr_info("%15" PRIu64 " ", st->wait_time_min == ULLONG_MAX ?
753 0 : st->wait_time_min);
754 pr_info("\n");
755 }
756
757 print_bad_events(bad, total);
758}
759
760static bool info_threads, info_map;
761
762static void dump_threads(void)
763{
764 struct thread_stat *st;
765 struct rb_node *node;
766 struct thread *t;
767
768 pr_info("%10s: comm\n", "Thread ID");
769
770 node = rb_first(&thread_stats);
771 while (node) {
772 st = container_of(node, struct thread_stat, rb);
773 t = perf_session__findnew(session, st->tid);
774 pr_info("%10d: %s\n", st->tid, thread__comm_str(t));
775 node = rb_next(node);
776 thread__put(t);
777 };
778}
779
780static void dump_map(void)
781{
782 unsigned int i;
783 struct lock_stat *st;
784
785 pr_info("Address of instance: name of class\n");
786 for (i = 0; i < LOCKHASH_SIZE; i++) {
787 list_for_each_entry(st, &lockhash_table[i], hash_entry) {
788 pr_info(" %p: %s\n", st->addr, st->name);
789 }
790 }
791}
792
793static int dump_info(void)
794{
795 int rc = 0;
796
797 if (info_threads)
798 dump_threads();
799 else if (info_map)
800 dump_map();
801 else {
802 rc = -1;
803 pr_err("Unknown type of information\n");
804 }
805
806 return rc;
807}
808
809typedef int (*tracepoint_handler)(struct perf_evsel *evsel,
810 struct perf_sample *sample);
811
812static int process_sample_event(struct perf_tool *tool __maybe_unused,
813 union perf_event *event,
814 struct perf_sample *sample,
815 struct perf_evsel *evsel,
816 struct machine *machine)
817{
818 int err = 0;
819 struct thread *thread = machine__findnew_thread(machine, sample->pid,
820 sample->tid);
821
822 if (thread == NULL) {
823 pr_debug("problem processing %d event, skipping it.\n",
824 event->header.type);
825 return -1;
826 }
827
828 if (evsel->handler != NULL) {
829 tracepoint_handler f = evsel->handler;
830 err = f(evsel, sample);
831 }
832
833 thread__put(thread);
834
835 return err;
836}
837
838static void sort_result(void)
839{
840 unsigned int i;
841 struct lock_stat *st;
842
843 for (i = 0; i < LOCKHASH_SIZE; i++) {
844 list_for_each_entry(st, &lockhash_table[i], hash_entry) {
845 insert_to_result(st, compare);
846 }
847 }
848}
849
850static const struct perf_evsel_str_handler lock_tracepoints[] = {
851 { "lock:lock_acquire", perf_evsel__process_lock_acquire, }, /* CONFIG_LOCKDEP */
852 { "lock:lock_acquired", perf_evsel__process_lock_acquired, }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */
853 { "lock:lock_contended", perf_evsel__process_lock_contended, }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */
854 { "lock:lock_release", perf_evsel__process_lock_release, }, /* CONFIG_LOCKDEP */
855};
856
857static bool force;
858
859static int __cmd_report(bool display_info)
860{
861 int err = -EINVAL;
862 struct perf_tool eops = {
863 .sample = process_sample_event,
864 .comm = perf_event__process_comm,
865 .namespaces = perf_event__process_namespaces,
866 .ordered_events = true,
867 };
868 struct perf_data data = {
869 .file = {
870 .path = input_name,
871 },
872 .mode = PERF_DATA_MODE_READ,
873 .force = force,
874 };
875
876 session = perf_session__new(&data, false, &eops);
877 if (!session) {
878 pr_err("Initializing perf session failed\n");
879 return -1;
880 }
881
882 symbol__init(&session->header.env);
883
884 if (!perf_session__has_traces(session, "lock record"))
885 goto out_delete;
886
887 if (perf_session__set_tracepoints_handlers(session, lock_tracepoints)) {
888 pr_err("Initializing perf session tracepoint handlers failed\n");
889 goto out_delete;
890 }
891
892 if (select_key())
893 goto out_delete;
894
895 err = perf_session__process_events(session);
896 if (err)
897 goto out_delete;
898
899 setup_pager();
900 if (display_info) /* used for info subcommand */
901 err = dump_info();
902 else {
903 sort_result();
904 print_result();
905 }
906
907out_delete:
908 perf_session__delete(session);
909 return err;
910}
911
912static int __cmd_record(int argc, const char **argv)
913{
914 const char *record_args[] = {
915 "record", "-R", "-m", "1024", "-c", "1",
916 };
917 unsigned int rec_argc, i, j, ret;
918 const char **rec_argv;
919
920 for (i = 0; i < ARRAY_SIZE(lock_tracepoints); i++) {
921 if (!is_valid_tracepoint(lock_tracepoints[i].name)) {
922 pr_err("tracepoint %s is not enabled. "
923 "Are CONFIG_LOCKDEP and CONFIG_LOCK_STAT enabled?\n",
924 lock_tracepoints[i].name);
925 return 1;
926 }
927 }
928
929 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
930 /* factor of 2 is for -e in front of each tracepoint */
931 rec_argc += 2 * ARRAY_SIZE(lock_tracepoints);
932
933 rec_argv = calloc(rec_argc + 1, sizeof(char *));
934 if (!rec_argv)
935 return -ENOMEM;
936
937 for (i = 0; i < ARRAY_SIZE(record_args); i++)
938 rec_argv[i] = strdup(record_args[i]);
939
940 for (j = 0; j < ARRAY_SIZE(lock_tracepoints); j++) {
941 rec_argv[i++] = "-e";
942 rec_argv[i++] = strdup(lock_tracepoints[j].name);
943 }
944
945 for (j = 1; j < (unsigned int)argc; j++, i++)
946 rec_argv[i] = argv[j];
947
948 BUG_ON(i != rec_argc);
949
950 ret = cmd_record(i, rec_argv);
951 free(rec_argv);
952 return ret;
953}
954
955int cmd_lock(int argc, const char **argv)
956{
957 const struct option lock_options[] = {
958 OPT_STRING('i', "input", &input_name, "file", "input file name"),
959 OPT_INCR('v', "verbose", &verbose, "be more verbose (show symbol address, etc)"),
960 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, "dump raw trace in ASCII"),
961 OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
962 OPT_END()
963 };
964
965 const struct option info_options[] = {
966 OPT_BOOLEAN('t', "threads", &info_threads,
967 "dump thread list in perf.data"),
968 OPT_BOOLEAN('m', "map", &info_map,
969 "map of lock instances (address:name table)"),
970 OPT_PARENT(lock_options)
971 };
972
973 const struct option report_options[] = {
974 OPT_STRING('k', "key", &sort_key, "acquired",
975 "key for sorting (acquired / contended / avg_wait / wait_total / wait_max / wait_min)"),
976 /* TODO: type */
977 OPT_PARENT(lock_options)
978 };
979
980 const char * const info_usage[] = {
981 "perf lock info [<options>]",
982 NULL
983 };
984 const char *const lock_subcommands[] = { "record", "report", "script",
985 "info", NULL };
986 const char *lock_usage[] = {
987 NULL,
988 NULL
989 };
990 const char * const report_usage[] = {
991 "perf lock report [<options>]",
992 NULL
993 };
994 unsigned int i;
995 int rc = 0;
996
997 for (i = 0; i < LOCKHASH_SIZE; i++)
998 INIT_LIST_HEAD(lockhash_table + i);
999
1000 argc = parse_options_subcommand(argc, argv, lock_options, lock_subcommands,
1001 lock_usage, PARSE_OPT_STOP_AT_NON_OPTION);
1002 if (!argc)
1003 usage_with_options(lock_usage, lock_options);
1004
1005 if (!strncmp(argv[0], "rec", 3)) {
1006 return __cmd_record(argc, argv);
1007 } else if (!strncmp(argv[0], "report", 6)) {
1008 trace_handler = &report_lock_ops;
1009 if (argc) {
1010 argc = parse_options(argc, argv,
1011 report_options, report_usage, 0);
1012 if (argc)
1013 usage_with_options(report_usage, report_options);
1014 }
1015 rc = __cmd_report(false);
1016 } else if (!strcmp(argv[0], "script")) {
1017 /* Aliased to 'perf script' */
1018 return cmd_script(argc, argv);
1019 } else if (!strcmp(argv[0], "info")) {
1020 if (argc) {
1021 argc = parse_options(argc, argv,
1022 info_options, info_usage, 0);
1023 if (argc)
1024 usage_with_options(info_usage, info_options);
1025 }
1026 /* recycling report_lock_ops */
1027 trace_handler = &report_lock_ops;
1028 rc = __cmd_report(true);
1029 } else {
1030 usage_with_options(lock_usage, lock_options);
1031 }
1032
1033 return rc;
1034}
1// SPDX-License-Identifier: GPL-2.0
2#include <errno.h>
3#include <inttypes.h>
4#include "builtin.h"
5#include "perf.h"
6
7#include "util/evlist.h" // for struct evsel_str_handler
8#include "util/evsel.h"
9#include "util/symbol.h"
10#include "util/thread.h"
11#include "util/header.h"
12
13#include <subcmd/pager.h>
14#include <subcmd/parse-options.h>
15#include "util/trace-event.h"
16
17#include "util/debug.h"
18#include "util/session.h"
19#include "util/tool.h"
20#include "util/data.h"
21
22#include <sys/types.h>
23#include <sys/prctl.h>
24#include <semaphore.h>
25#include <pthread.h>
26#include <math.h>
27#include <limits.h>
28
29#include <linux/list.h>
30#include <linux/hash.h>
31#include <linux/kernel.h>
32#include <linux/zalloc.h>
33#include <linux/err.h>
34
35static struct perf_session *session;
36
37/* based on kernel/lockdep.c */
38#define LOCKHASH_BITS 12
39#define LOCKHASH_SIZE (1UL << LOCKHASH_BITS)
40
41static struct list_head lockhash_table[LOCKHASH_SIZE];
42
43#define __lockhashfn(key) hash_long((unsigned long)key, LOCKHASH_BITS)
44#define lockhashentry(key) (lockhash_table + __lockhashfn((key)))
45
46struct lock_stat {
47 struct list_head hash_entry;
48 struct rb_node rb; /* used for sorting */
49
50 /*
51 * FIXME: evsel__intval() returns u64,
52 * so address of lockdep_map should be treated as 64bit.
53 * Is there more better solution?
54 */
55 void *addr; /* address of lockdep_map, used as ID */
56 char *name; /* for strcpy(), we cannot use const */
57
58 unsigned int nr_acquire;
59 unsigned int nr_acquired;
60 unsigned int nr_contended;
61 unsigned int nr_release;
62
63 unsigned int nr_readlock;
64 unsigned int nr_trylock;
65
66 /* these times are in nano sec. */
67 u64 avg_wait_time;
68 u64 wait_time_total;
69 u64 wait_time_min;
70 u64 wait_time_max;
71
72 int discard; /* flag of blacklist */
73};
74
75/*
76 * States of lock_seq_stat
77 *
78 * UNINITIALIZED is required for detecting first event of acquire.
79 * As the nature of lock events, there is no guarantee
80 * that the first event for the locks are acquire,
81 * it can be acquired, contended or release.
82 */
83#define SEQ_STATE_UNINITIALIZED 0 /* initial state */
84#define SEQ_STATE_RELEASED 1
85#define SEQ_STATE_ACQUIRING 2
86#define SEQ_STATE_ACQUIRED 3
87#define SEQ_STATE_READ_ACQUIRED 4
88#define SEQ_STATE_CONTENDED 5
89
90/*
91 * MAX_LOCK_DEPTH
92 * Imported from include/linux/sched.h.
93 * Should this be synchronized?
94 */
95#define MAX_LOCK_DEPTH 48
96
97/*
98 * struct lock_seq_stat:
99 * Place to put on state of one lock sequence
100 * 1) acquire -> acquired -> release
101 * 2) acquire -> contended -> acquired -> release
102 * 3) acquire (with read or try) -> release
103 * 4) Are there other patterns?
104 */
105struct lock_seq_stat {
106 struct list_head list;
107 int state;
108 u64 prev_event_time;
109 void *addr;
110
111 int read_count;
112};
113
114struct thread_stat {
115 struct rb_node rb;
116
117 u32 tid;
118 struct list_head seq_list;
119};
120
121static struct rb_root thread_stats;
122
123static struct thread_stat *thread_stat_find(u32 tid)
124{
125 struct rb_node *node;
126 struct thread_stat *st;
127
128 node = thread_stats.rb_node;
129 while (node) {
130 st = container_of(node, struct thread_stat, rb);
131 if (st->tid == tid)
132 return st;
133 else if (tid < st->tid)
134 node = node->rb_left;
135 else
136 node = node->rb_right;
137 }
138
139 return NULL;
140}
141
142static void thread_stat_insert(struct thread_stat *new)
143{
144 struct rb_node **rb = &thread_stats.rb_node;
145 struct rb_node *parent = NULL;
146 struct thread_stat *p;
147
148 while (*rb) {
149 p = container_of(*rb, struct thread_stat, rb);
150 parent = *rb;
151
152 if (new->tid < p->tid)
153 rb = &(*rb)->rb_left;
154 else if (new->tid > p->tid)
155 rb = &(*rb)->rb_right;
156 else
157 BUG_ON("inserting invalid thread_stat\n");
158 }
159
160 rb_link_node(&new->rb, parent, rb);
161 rb_insert_color(&new->rb, &thread_stats);
162}
163
164static struct thread_stat *thread_stat_findnew_after_first(u32 tid)
165{
166 struct thread_stat *st;
167
168 st = thread_stat_find(tid);
169 if (st)
170 return st;
171
172 st = zalloc(sizeof(struct thread_stat));
173 if (!st) {
174 pr_err("memory allocation failed\n");
175 return NULL;
176 }
177
178 st->tid = tid;
179 INIT_LIST_HEAD(&st->seq_list);
180
181 thread_stat_insert(st);
182
183 return st;
184}
185
186static struct thread_stat *thread_stat_findnew_first(u32 tid);
187static struct thread_stat *(*thread_stat_findnew)(u32 tid) =
188 thread_stat_findnew_first;
189
190static struct thread_stat *thread_stat_findnew_first(u32 tid)
191{
192 struct thread_stat *st;
193
194 st = zalloc(sizeof(struct thread_stat));
195 if (!st) {
196 pr_err("memory allocation failed\n");
197 return NULL;
198 }
199 st->tid = tid;
200 INIT_LIST_HEAD(&st->seq_list);
201
202 rb_link_node(&st->rb, NULL, &thread_stats.rb_node);
203 rb_insert_color(&st->rb, &thread_stats);
204
205 thread_stat_findnew = thread_stat_findnew_after_first;
206 return st;
207}
208
209/* build simple key function one is bigger than two */
210#define SINGLE_KEY(member) \
211 static int lock_stat_key_ ## member(struct lock_stat *one, \
212 struct lock_stat *two) \
213 { \
214 return one->member > two->member; \
215 }
216
217SINGLE_KEY(nr_acquired)
218SINGLE_KEY(nr_contended)
219SINGLE_KEY(avg_wait_time)
220SINGLE_KEY(wait_time_total)
221SINGLE_KEY(wait_time_max)
222
223static int lock_stat_key_wait_time_min(struct lock_stat *one,
224 struct lock_stat *two)
225{
226 u64 s1 = one->wait_time_min;
227 u64 s2 = two->wait_time_min;
228 if (s1 == ULLONG_MAX)
229 s1 = 0;
230 if (s2 == ULLONG_MAX)
231 s2 = 0;
232 return s1 > s2;
233}
234
235struct lock_key {
236 /*
237 * name: the value for specify by user
238 * this should be simpler than raw name of member
239 * e.g. nr_acquired -> acquired, wait_time_total -> wait_total
240 */
241 const char *name;
242 int (*key)(struct lock_stat*, struct lock_stat*);
243};
244
245static const char *sort_key = "acquired";
246
247static int (*compare)(struct lock_stat *, struct lock_stat *);
248
249static struct rb_root result; /* place to store sorted data */
250
251#define DEF_KEY_LOCK(name, fn_suffix) \
252 { #name, lock_stat_key_ ## fn_suffix }
253struct lock_key keys[] = {
254 DEF_KEY_LOCK(acquired, nr_acquired),
255 DEF_KEY_LOCK(contended, nr_contended),
256 DEF_KEY_LOCK(avg_wait, avg_wait_time),
257 DEF_KEY_LOCK(wait_total, wait_time_total),
258 DEF_KEY_LOCK(wait_min, wait_time_min),
259 DEF_KEY_LOCK(wait_max, wait_time_max),
260
261 /* extra comparisons much complicated should be here */
262
263 { NULL, NULL }
264};
265
266static int select_key(void)
267{
268 int i;
269
270 for (i = 0; keys[i].name; i++) {
271 if (!strcmp(keys[i].name, sort_key)) {
272 compare = keys[i].key;
273 return 0;
274 }
275 }
276
277 pr_err("Unknown compare key: %s\n", sort_key);
278
279 return -1;
280}
281
282static void insert_to_result(struct lock_stat *st,
283 int (*bigger)(struct lock_stat *, struct lock_stat *))
284{
285 struct rb_node **rb = &result.rb_node;
286 struct rb_node *parent = NULL;
287 struct lock_stat *p;
288
289 while (*rb) {
290 p = container_of(*rb, struct lock_stat, rb);
291 parent = *rb;
292
293 if (bigger(st, p))
294 rb = &(*rb)->rb_left;
295 else
296 rb = &(*rb)->rb_right;
297 }
298
299 rb_link_node(&st->rb, parent, rb);
300 rb_insert_color(&st->rb, &result);
301}
302
303/* returns left most element of result, and erase it */
304static struct lock_stat *pop_from_result(void)
305{
306 struct rb_node *node = result.rb_node;
307
308 if (!node)
309 return NULL;
310
311 while (node->rb_left)
312 node = node->rb_left;
313
314 rb_erase(node, &result);
315 return container_of(node, struct lock_stat, rb);
316}
317
318static struct lock_stat *lock_stat_findnew(void *addr, const char *name)
319{
320 struct list_head *entry = lockhashentry(addr);
321 struct lock_stat *ret, *new;
322
323 list_for_each_entry(ret, entry, hash_entry) {
324 if (ret->addr == addr)
325 return ret;
326 }
327
328 new = zalloc(sizeof(struct lock_stat));
329 if (!new)
330 goto alloc_failed;
331
332 new->addr = addr;
333 new->name = zalloc(sizeof(char) * strlen(name) + 1);
334 if (!new->name) {
335 free(new);
336 goto alloc_failed;
337 }
338
339 strcpy(new->name, name);
340 new->wait_time_min = ULLONG_MAX;
341
342 list_add(&new->hash_entry, entry);
343 return new;
344
345alloc_failed:
346 pr_err("memory allocation failed\n");
347 return NULL;
348}
349
350struct trace_lock_handler {
351 int (*acquire_event)(struct evsel *evsel,
352 struct perf_sample *sample);
353
354 int (*acquired_event)(struct evsel *evsel,
355 struct perf_sample *sample);
356
357 int (*contended_event)(struct evsel *evsel,
358 struct perf_sample *sample);
359
360 int (*release_event)(struct evsel *evsel,
361 struct perf_sample *sample);
362};
363
364static struct lock_seq_stat *get_seq(struct thread_stat *ts, void *addr)
365{
366 struct lock_seq_stat *seq;
367
368 list_for_each_entry(seq, &ts->seq_list, list) {
369 if (seq->addr == addr)
370 return seq;
371 }
372
373 seq = zalloc(sizeof(struct lock_seq_stat));
374 if (!seq) {
375 pr_err("memory allocation failed\n");
376 return NULL;
377 }
378 seq->state = SEQ_STATE_UNINITIALIZED;
379 seq->addr = addr;
380
381 list_add(&seq->list, &ts->seq_list);
382 return seq;
383}
384
385enum broken_state {
386 BROKEN_ACQUIRE,
387 BROKEN_ACQUIRED,
388 BROKEN_CONTENDED,
389 BROKEN_RELEASE,
390 BROKEN_MAX,
391};
392
393static int bad_hist[BROKEN_MAX];
394
395enum acquire_flags {
396 TRY_LOCK = 1,
397 READ_LOCK = 2,
398};
399
400static int report_lock_acquire_event(struct evsel *evsel,
401 struct perf_sample *sample)
402{
403 void *addr;
404 struct lock_stat *ls;
405 struct thread_stat *ts;
406 struct lock_seq_stat *seq;
407 const char *name = evsel__strval(evsel, sample, "name");
408 u64 tmp = evsel__intval(evsel, sample, "lockdep_addr");
409 int flag = evsel__intval(evsel, sample, "flags");
410
411 memcpy(&addr, &tmp, sizeof(void *));
412
413 ls = lock_stat_findnew(addr, name);
414 if (!ls)
415 return -ENOMEM;
416 if (ls->discard)
417 return 0;
418
419 ts = thread_stat_findnew(sample->tid);
420 if (!ts)
421 return -ENOMEM;
422
423 seq = get_seq(ts, addr);
424 if (!seq)
425 return -ENOMEM;
426
427 switch (seq->state) {
428 case SEQ_STATE_UNINITIALIZED:
429 case SEQ_STATE_RELEASED:
430 if (!flag) {
431 seq->state = SEQ_STATE_ACQUIRING;
432 } else {
433 if (flag & TRY_LOCK)
434 ls->nr_trylock++;
435 if (flag & READ_LOCK)
436 ls->nr_readlock++;
437 seq->state = SEQ_STATE_READ_ACQUIRED;
438 seq->read_count = 1;
439 ls->nr_acquired++;
440 }
441 break;
442 case SEQ_STATE_READ_ACQUIRED:
443 if (flag & READ_LOCK) {
444 seq->read_count++;
445 ls->nr_acquired++;
446 goto end;
447 } else {
448 goto broken;
449 }
450 break;
451 case SEQ_STATE_ACQUIRED:
452 case SEQ_STATE_ACQUIRING:
453 case SEQ_STATE_CONTENDED:
454broken:
455 /* broken lock sequence, discard it */
456 ls->discard = 1;
457 bad_hist[BROKEN_ACQUIRE]++;
458 list_del_init(&seq->list);
459 free(seq);
460 goto end;
461 default:
462 BUG_ON("Unknown state of lock sequence found!\n");
463 break;
464 }
465
466 ls->nr_acquire++;
467 seq->prev_event_time = sample->time;
468end:
469 return 0;
470}
471
472static int report_lock_acquired_event(struct evsel *evsel,
473 struct perf_sample *sample)
474{
475 void *addr;
476 struct lock_stat *ls;
477 struct thread_stat *ts;
478 struct lock_seq_stat *seq;
479 u64 contended_term;
480 const char *name = evsel__strval(evsel, sample, "name");
481 u64 tmp = evsel__intval(evsel, sample, "lockdep_addr");
482
483 memcpy(&addr, &tmp, sizeof(void *));
484
485 ls = lock_stat_findnew(addr, name);
486 if (!ls)
487 return -ENOMEM;
488 if (ls->discard)
489 return 0;
490
491 ts = thread_stat_findnew(sample->tid);
492 if (!ts)
493 return -ENOMEM;
494
495 seq = get_seq(ts, addr);
496 if (!seq)
497 return -ENOMEM;
498
499 switch (seq->state) {
500 case SEQ_STATE_UNINITIALIZED:
501 /* orphan event, do nothing */
502 return 0;
503 case SEQ_STATE_ACQUIRING:
504 break;
505 case SEQ_STATE_CONTENDED:
506 contended_term = sample->time - 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;
512 break;
513 case SEQ_STATE_RELEASED:
514 case SEQ_STATE_ACQUIRED:
515 case SEQ_STATE_READ_ACQUIRED:
516 /* broken lock sequence, discard it */
517 ls->discard = 1;
518 bad_hist[BROKEN_ACQUIRED]++;
519 list_del_init(&seq->list);
520 free(seq);
521 goto end;
522 default:
523 BUG_ON("Unknown state of lock sequence found!\n");
524 break;
525 }
526
527 seq->state = SEQ_STATE_ACQUIRED;
528 ls->nr_acquired++;
529 ls->avg_wait_time = ls->nr_contended ? ls->wait_time_total/ls->nr_contended : 0;
530 seq->prev_event_time = sample->time;
531end:
532 return 0;
533}
534
535static int report_lock_contended_event(struct evsel *evsel,
536 struct perf_sample *sample)
537{
538 void *addr;
539 struct lock_stat *ls;
540 struct thread_stat *ts;
541 struct lock_seq_stat *seq;
542 const char *name = evsel__strval(evsel, sample, "name");
543 u64 tmp = evsel__intval(evsel, sample, "lockdep_addr");
544
545 memcpy(&addr, &tmp, sizeof(void *));
546
547 ls = lock_stat_findnew(addr, name);
548 if (!ls)
549 return -ENOMEM;
550 if (ls->discard)
551 return 0;
552
553 ts = thread_stat_findnew(sample->tid);
554 if (!ts)
555 return -ENOMEM;
556
557 seq = get_seq(ts, addr);
558 if (!seq)
559 return -ENOMEM;
560
561 switch (seq->state) {
562 case SEQ_STATE_UNINITIALIZED:
563 /* orphan event, do nothing */
564 return 0;
565 case SEQ_STATE_ACQUIRING:
566 break;
567 case SEQ_STATE_RELEASED:
568 case SEQ_STATE_ACQUIRED:
569 case SEQ_STATE_READ_ACQUIRED:
570 case SEQ_STATE_CONTENDED:
571 /* broken lock sequence, discard it */
572 ls->discard = 1;
573 bad_hist[BROKEN_CONTENDED]++;
574 list_del_init(&seq->list);
575 free(seq);
576 goto end;
577 default:
578 BUG_ON("Unknown state of lock sequence found!\n");
579 break;
580 }
581
582 seq->state = SEQ_STATE_CONTENDED;
583 ls->nr_contended++;
584 ls->avg_wait_time = ls->wait_time_total/ls->nr_contended;
585 seq->prev_event_time = sample->time;
586end:
587 return 0;
588}
589
590static int report_lock_release_event(struct evsel *evsel,
591 struct perf_sample *sample)
592{
593 void *addr;
594 struct lock_stat *ls;
595 struct thread_stat *ts;
596 struct lock_seq_stat *seq;
597 const char *name = evsel__strval(evsel, sample, "name");
598 u64 tmp = evsel__intval(evsel, sample, "lockdep_addr");
599
600 memcpy(&addr, &tmp, sizeof(void *));
601
602 ls = lock_stat_findnew(addr, name);
603 if (!ls)
604 return -ENOMEM;
605 if (ls->discard)
606 return 0;
607
608 ts = thread_stat_findnew(sample->tid);
609 if (!ts)
610 return -ENOMEM;
611
612 seq = get_seq(ts, addr);
613 if (!seq)
614 return -ENOMEM;
615
616 switch (seq->state) {
617 case SEQ_STATE_UNINITIALIZED:
618 goto end;
619 case SEQ_STATE_ACQUIRED:
620 break;
621 case SEQ_STATE_READ_ACQUIRED:
622 seq->read_count--;
623 BUG_ON(seq->read_count < 0);
624 if (seq->read_count) {
625 ls->nr_release++;
626 goto end;
627 }
628 break;
629 case SEQ_STATE_ACQUIRING:
630 case SEQ_STATE_CONTENDED:
631 case SEQ_STATE_RELEASED:
632 /* broken lock sequence, discard it */
633 ls->discard = 1;
634 bad_hist[BROKEN_RELEASE]++;
635 goto free_seq;
636 default:
637 BUG_ON("Unknown state of lock sequence found!\n");
638 break;
639 }
640
641 ls->nr_release++;
642free_seq:
643 list_del_init(&seq->list);
644 free(seq);
645end:
646 return 0;
647}
648
649/* lock oriented handlers */
650/* TODO: handlers for CPU oriented, thread oriented */
651static struct trace_lock_handler report_lock_ops = {
652 .acquire_event = report_lock_acquire_event,
653 .acquired_event = report_lock_acquired_event,
654 .contended_event = report_lock_contended_event,
655 .release_event = report_lock_release_event,
656};
657
658static struct trace_lock_handler *trace_handler;
659
660static int evsel__process_lock_acquire(struct evsel *evsel, struct perf_sample *sample)
661{
662 if (trace_handler->acquire_event)
663 return trace_handler->acquire_event(evsel, sample);
664 return 0;
665}
666
667static int evsel__process_lock_acquired(struct evsel *evsel, struct perf_sample *sample)
668{
669 if (trace_handler->acquired_event)
670 return trace_handler->acquired_event(evsel, sample);
671 return 0;
672}
673
674static int evsel__process_lock_contended(struct evsel *evsel, struct perf_sample *sample)
675{
676 if (trace_handler->contended_event)
677 return trace_handler->contended_event(evsel, sample);
678 return 0;
679}
680
681static int evsel__process_lock_release(struct evsel *evsel, struct perf_sample *sample)
682{
683 if (trace_handler->release_event)
684 return trace_handler->release_event(evsel, sample);
685 return 0;
686}
687
688static void print_bad_events(int bad, int total)
689{
690 /* Output for debug, this have to be removed */
691 int i;
692 const char *name[4] =
693 { "acquire", "acquired", "contended", "release" };
694
695 pr_info("\n=== output for debug===\n\n");
696 pr_info("bad: %d, total: %d\n", bad, total);
697 pr_info("bad rate: %.2f %%\n", (double)bad / (double)total * 100);
698 pr_info("histogram of events caused bad sequence\n");
699 for (i = 0; i < BROKEN_MAX; i++)
700 pr_info(" %10s: %d\n", name[i], bad_hist[i]);
701}
702
703/* TODO: various way to print, coloring, nano or milli sec */
704static void print_result(void)
705{
706 struct lock_stat *st;
707 char cut_name[20];
708 int bad, total;
709
710 pr_info("%20s ", "Name");
711 pr_info("%10s ", "acquired");
712 pr_info("%10s ", "contended");
713
714 pr_info("%15s ", "avg wait (ns)");
715 pr_info("%15s ", "total wait (ns)");
716 pr_info("%15s ", "max wait (ns)");
717 pr_info("%15s ", "min wait (ns)");
718
719 pr_info("\n\n");
720
721 bad = total = 0;
722 while ((st = pop_from_result())) {
723 total++;
724 if (st->discard) {
725 bad++;
726 continue;
727 }
728 bzero(cut_name, 20);
729
730 if (strlen(st->name) < 16) {
731 /* output raw name */
732 pr_info("%20s ", st->name);
733 } else {
734 strncpy(cut_name, st->name, 16);
735 cut_name[16] = '.';
736 cut_name[17] = '.';
737 cut_name[18] = '.';
738 cut_name[19] = '\0';
739 /* cut off name for saving output style */
740 pr_info("%20s ", cut_name);
741 }
742
743 pr_info("%10u ", st->nr_acquired);
744 pr_info("%10u ", st->nr_contended);
745
746 pr_info("%15" PRIu64 " ", st->avg_wait_time);
747 pr_info("%15" PRIu64 " ", st->wait_time_total);
748 pr_info("%15" PRIu64 " ", st->wait_time_max);
749 pr_info("%15" PRIu64 " ", st->wait_time_min == ULLONG_MAX ?
750 0 : st->wait_time_min);
751 pr_info("\n");
752 }
753
754 print_bad_events(bad, total);
755}
756
757static bool info_threads, info_map;
758
759static void dump_threads(void)
760{
761 struct thread_stat *st;
762 struct rb_node *node;
763 struct thread *t;
764
765 pr_info("%10s: comm\n", "Thread ID");
766
767 node = rb_first(&thread_stats);
768 while (node) {
769 st = container_of(node, struct thread_stat, rb);
770 t = perf_session__findnew(session, st->tid);
771 pr_info("%10d: %s\n", st->tid, thread__comm_str(t));
772 node = rb_next(node);
773 thread__put(t);
774 }
775}
776
777static void dump_map(void)
778{
779 unsigned int i;
780 struct lock_stat *st;
781
782 pr_info("Address of instance: name of class\n");
783 for (i = 0; i < LOCKHASH_SIZE; i++) {
784 list_for_each_entry(st, &lockhash_table[i], hash_entry) {
785 pr_info(" %p: %s\n", st->addr, st->name);
786 }
787 }
788}
789
790static int dump_info(void)
791{
792 int rc = 0;
793
794 if (info_threads)
795 dump_threads();
796 else if (info_map)
797 dump_map();
798 else {
799 rc = -1;
800 pr_err("Unknown type of information\n");
801 }
802
803 return rc;
804}
805
806typedef int (*tracepoint_handler)(struct evsel *evsel,
807 struct perf_sample *sample);
808
809static int process_sample_event(struct perf_tool *tool __maybe_unused,
810 union perf_event *event,
811 struct perf_sample *sample,
812 struct evsel *evsel,
813 struct machine *machine)
814{
815 int err = 0;
816 struct thread *thread = machine__findnew_thread(machine, sample->pid,
817 sample->tid);
818
819 if (thread == NULL) {
820 pr_debug("problem processing %d event, skipping it.\n",
821 event->header.type);
822 return -1;
823 }
824
825 if (evsel->handler != NULL) {
826 tracepoint_handler f = evsel->handler;
827 err = f(evsel, sample);
828 }
829
830 thread__put(thread);
831
832 return err;
833}
834
835static void sort_result(void)
836{
837 unsigned int i;
838 struct lock_stat *st;
839
840 for (i = 0; i < LOCKHASH_SIZE; i++) {
841 list_for_each_entry(st, &lockhash_table[i], hash_entry) {
842 insert_to_result(st, compare);
843 }
844 }
845}
846
847static const struct evsel_str_handler lock_tracepoints[] = {
848 { "lock:lock_acquire", evsel__process_lock_acquire, }, /* CONFIG_LOCKDEP */
849 { "lock:lock_acquired", evsel__process_lock_acquired, }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */
850 { "lock:lock_contended", evsel__process_lock_contended, }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */
851 { "lock:lock_release", evsel__process_lock_release, }, /* CONFIG_LOCKDEP */
852};
853
854static bool force;
855
856static int __cmd_report(bool display_info)
857{
858 int err = -EINVAL;
859 struct perf_tool eops = {
860 .sample = process_sample_event,
861 .comm = perf_event__process_comm,
862 .namespaces = perf_event__process_namespaces,
863 .ordered_events = true,
864 };
865 struct perf_data data = {
866 .path = input_name,
867 .mode = PERF_DATA_MODE_READ,
868 .force = force,
869 };
870
871 session = perf_session__new(&data, false, &eops);
872 if (IS_ERR(session)) {
873 pr_err("Initializing perf session failed\n");
874 return PTR_ERR(session);
875 }
876
877 symbol__init(&session->header.env);
878
879 if (!perf_session__has_traces(session, "lock record"))
880 goto out_delete;
881
882 if (perf_session__set_tracepoints_handlers(session, lock_tracepoints)) {
883 pr_err("Initializing perf session tracepoint handlers failed\n");
884 goto out_delete;
885 }
886
887 if (select_key())
888 goto out_delete;
889
890 err = perf_session__process_events(session);
891 if (err)
892 goto out_delete;
893
894 setup_pager();
895 if (display_info) /* used for info subcommand */
896 err = dump_info();
897 else {
898 sort_result();
899 print_result();
900 }
901
902out_delete:
903 perf_session__delete(session);
904 return err;
905}
906
907static int __cmd_record(int argc, const char **argv)
908{
909 const char *record_args[] = {
910 "record", "-R", "-m", "1024", "-c", "1",
911 };
912 unsigned int rec_argc, i, j, ret;
913 const char **rec_argv;
914
915 for (i = 0; i < ARRAY_SIZE(lock_tracepoints); i++) {
916 if (!is_valid_tracepoint(lock_tracepoints[i].name)) {
917 pr_err("tracepoint %s is not enabled. "
918 "Are CONFIG_LOCKDEP and CONFIG_LOCK_STAT enabled?\n",
919 lock_tracepoints[i].name);
920 return 1;
921 }
922 }
923
924 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
925 /* factor of 2 is for -e in front of each tracepoint */
926 rec_argc += 2 * ARRAY_SIZE(lock_tracepoints);
927
928 rec_argv = calloc(rec_argc + 1, sizeof(char *));
929 if (!rec_argv)
930 return -ENOMEM;
931
932 for (i = 0; i < ARRAY_SIZE(record_args); i++)
933 rec_argv[i] = strdup(record_args[i]);
934
935 for (j = 0; j < ARRAY_SIZE(lock_tracepoints); j++) {
936 rec_argv[i++] = "-e";
937 rec_argv[i++] = strdup(lock_tracepoints[j].name);
938 }
939
940 for (j = 1; j < (unsigned int)argc; j++, i++)
941 rec_argv[i] = argv[j];
942
943 BUG_ON(i != rec_argc);
944
945 ret = cmd_record(i, rec_argv);
946 free(rec_argv);
947 return ret;
948}
949
950int cmd_lock(int argc, const char **argv)
951{
952 const struct option lock_options[] = {
953 OPT_STRING('i', "input", &input_name, "file", "input file name"),
954 OPT_INCR('v', "verbose", &verbose, "be more verbose (show symbol address, etc)"),
955 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, "dump raw trace in ASCII"),
956 OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
957 OPT_END()
958 };
959
960 const struct option info_options[] = {
961 OPT_BOOLEAN('t', "threads", &info_threads,
962 "dump thread list in perf.data"),
963 OPT_BOOLEAN('m', "map", &info_map,
964 "map of lock instances (address:name table)"),
965 OPT_PARENT(lock_options)
966 };
967
968 const struct option report_options[] = {
969 OPT_STRING('k', "key", &sort_key, "acquired",
970 "key for sorting (acquired / contended / avg_wait / wait_total / wait_max / wait_min)"),
971 /* TODO: type */
972 OPT_PARENT(lock_options)
973 };
974
975 const char * const info_usage[] = {
976 "perf lock info [<options>]",
977 NULL
978 };
979 const char *const lock_subcommands[] = { "record", "report", "script",
980 "info", NULL };
981 const char *lock_usage[] = {
982 NULL,
983 NULL
984 };
985 const char * const report_usage[] = {
986 "perf lock report [<options>]",
987 NULL
988 };
989 unsigned int i;
990 int rc = 0;
991
992 for (i = 0; i < LOCKHASH_SIZE; i++)
993 INIT_LIST_HEAD(lockhash_table + i);
994
995 argc = parse_options_subcommand(argc, argv, lock_options, lock_subcommands,
996 lock_usage, PARSE_OPT_STOP_AT_NON_OPTION);
997 if (!argc)
998 usage_with_options(lock_usage, lock_options);
999
1000 if (!strncmp(argv[0], "rec", 3)) {
1001 return __cmd_record(argc, argv);
1002 } else if (!strncmp(argv[0], "report", 6)) {
1003 trace_handler = &report_lock_ops;
1004 if (argc) {
1005 argc = parse_options(argc, argv,
1006 report_options, report_usage, 0);
1007 if (argc)
1008 usage_with_options(report_usage, report_options);
1009 }
1010 rc = __cmd_report(false);
1011 } else if (!strcmp(argv[0], "script")) {
1012 /* Aliased to 'perf script' */
1013 return cmd_script(argc, argv);
1014 } else if (!strcmp(argv[0], "info")) {
1015 if (argc) {
1016 argc = parse_options(argc, argv,
1017 info_options, info_usage, 0);
1018 if (argc)
1019 usage_with_options(info_usage, info_options);
1020 }
1021 /* recycling report_lock_ops */
1022 trace_handler = &report_lock_ops;
1023 rc = __cmd_report(true);
1024 } else {
1025 usage_with_options(lock_usage, lock_options);
1026 }
1027
1028 return rc;
1029}