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