Loading...
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#include "util/tool.h"
16
17#include <sys/types.h>
18#include <sys/prctl.h>
19#include <semaphore.h>
20#include <pthread.h>
21#include <math.h>
22#include <limits.h>
23
24#include <linux/list.h>
25#include <linux/hash.h>
26
27static struct perf_session *session;
28
29/* based on kernel/lockdep.c */
30#define LOCKHASH_BITS 12
31#define LOCKHASH_SIZE (1UL << LOCKHASH_BITS)
32
33static struct list_head lockhash_table[LOCKHASH_SIZE];
34
35#define __lockhashfn(key) hash_long((unsigned long)key, LOCKHASH_BITS)
36#define lockhashentry(key) (lockhash_table + __lockhashfn((key)))
37
38struct lock_stat {
39 struct list_head hash_entry;
40 struct rb_node rb; /* used for sorting */
41
42 /*
43 * FIXME: raw_field_value() returns unsigned long long,
44 * so address of lockdep_map should be dealed as 64bit.
45 * Is there more better solution?
46 */
47 void *addr; /* address of lockdep_map, used as ID */
48 char *name; /* for strcpy(), we cannot use const */
49
50 unsigned int nr_acquire;
51 unsigned int nr_acquired;
52 unsigned int nr_contended;
53 unsigned int nr_release;
54
55 unsigned int nr_readlock;
56 unsigned int nr_trylock;
57 /* these times are in nano sec. */
58 u64 wait_time_total;
59 u64 wait_time_min;
60 u64 wait_time_max;
61
62 int discard; /* flag of blacklist */
63};
64
65/*
66 * States of lock_seq_stat
67 *
68 * UNINITIALIZED is required for detecting first event of acquire.
69 * As the nature of lock events, there is no guarantee
70 * that the first event for the locks are acquire,
71 * it can be acquired, contended or release.
72 */
73#define SEQ_STATE_UNINITIALIZED 0 /* initial state */
74#define SEQ_STATE_RELEASED 1
75#define SEQ_STATE_ACQUIRING 2
76#define SEQ_STATE_ACQUIRED 3
77#define SEQ_STATE_READ_ACQUIRED 4
78#define SEQ_STATE_CONTENDED 5
79
80/*
81 * MAX_LOCK_DEPTH
82 * Imported from include/linux/sched.h.
83 * Should this be synchronized?
84 */
85#define MAX_LOCK_DEPTH 48
86
87/*
88 * struct lock_seq_stat:
89 * Place to put on state of one lock sequence
90 * 1) acquire -> acquired -> release
91 * 2) acquire -> contended -> acquired -> release
92 * 3) acquire (with read or try) -> release
93 * 4) Are there other patterns?
94 */
95struct lock_seq_stat {
96 struct list_head list;
97 int state;
98 u64 prev_event_time;
99 void *addr;
100
101 int read_count;
102};
103
104struct thread_stat {
105 struct rb_node rb;
106
107 u32 tid;
108 struct list_head seq_list;
109};
110
111static struct rb_root thread_stats;
112
113static struct thread_stat *thread_stat_find(u32 tid)
114{
115 struct rb_node *node;
116 struct thread_stat *st;
117
118 node = thread_stats.rb_node;
119 while (node) {
120 st = container_of(node, struct thread_stat, rb);
121 if (st->tid == tid)
122 return st;
123 else if (tid < st->tid)
124 node = node->rb_left;
125 else
126 node = node->rb_right;
127 }
128
129 return NULL;
130}
131
132static void thread_stat_insert(struct thread_stat *new)
133{
134 struct rb_node **rb = &thread_stats.rb_node;
135 struct rb_node *parent = NULL;
136 struct thread_stat *p;
137
138 while (*rb) {
139 p = container_of(*rb, struct thread_stat, rb);
140 parent = *rb;
141
142 if (new->tid < p->tid)
143 rb = &(*rb)->rb_left;
144 else if (new->tid > p->tid)
145 rb = &(*rb)->rb_right;
146 else
147 BUG_ON("inserting invalid thread_stat\n");
148 }
149
150 rb_link_node(&new->rb, parent, rb);
151 rb_insert_color(&new->rb, &thread_stats);
152}
153
154static struct thread_stat *thread_stat_findnew_after_first(u32 tid)
155{
156 struct thread_stat *st;
157
158 st = thread_stat_find(tid);
159 if (st)
160 return st;
161
162 st = zalloc(sizeof(struct thread_stat));
163 if (!st)
164 die("memory allocation failed\n");
165
166 st->tid = tid;
167 INIT_LIST_HEAD(&st->seq_list);
168
169 thread_stat_insert(st);
170
171 return st;
172}
173
174static struct thread_stat *thread_stat_findnew_first(u32 tid);
175static struct thread_stat *(*thread_stat_findnew)(u32 tid) =
176 thread_stat_findnew_first;
177
178static struct thread_stat *thread_stat_findnew_first(u32 tid)
179{
180 struct thread_stat *st;
181
182 st = zalloc(sizeof(struct thread_stat));
183 if (!st)
184 die("memory allocation failed\n");
185 st->tid = tid;
186 INIT_LIST_HEAD(&st->seq_list);
187
188 rb_link_node(&st->rb, NULL, &thread_stats.rb_node);
189 rb_insert_color(&st->rb, &thread_stats);
190
191 thread_stat_findnew = thread_stat_findnew_after_first;
192 return st;
193}
194
195/* build simple key function one is bigger than two */
196#define SINGLE_KEY(member) \
197 static int lock_stat_key_ ## member(struct lock_stat *one, \
198 struct lock_stat *two) \
199 { \
200 return one->member > two->member; \
201 }
202
203SINGLE_KEY(nr_acquired)
204SINGLE_KEY(nr_contended)
205SINGLE_KEY(wait_time_total)
206SINGLE_KEY(wait_time_max)
207
208static int lock_stat_key_wait_time_min(struct lock_stat *one,
209 struct lock_stat *two)
210{
211 u64 s1 = one->wait_time_min;
212 u64 s2 = two->wait_time_min;
213 if (s1 == ULLONG_MAX)
214 s1 = 0;
215 if (s2 == ULLONG_MAX)
216 s2 = 0;
217 return s1 > s2;
218}
219
220struct lock_key {
221 /*
222 * name: the value for specify by user
223 * this should be simpler than raw name of member
224 * e.g. nr_acquired -> acquired, wait_time_total -> wait_total
225 */
226 const char *name;
227 int (*key)(struct lock_stat*, struct lock_stat*);
228};
229
230static const char *sort_key = "acquired";
231
232static int (*compare)(struct lock_stat *, struct lock_stat *);
233
234static struct rb_root result; /* place to store sorted data */
235
236#define DEF_KEY_LOCK(name, fn_suffix) \
237 { #name, lock_stat_key_ ## fn_suffix }
238struct lock_key keys[] = {
239 DEF_KEY_LOCK(acquired, nr_acquired),
240 DEF_KEY_LOCK(contended, nr_contended),
241 DEF_KEY_LOCK(wait_total, wait_time_total),
242 DEF_KEY_LOCK(wait_min, wait_time_min),
243 DEF_KEY_LOCK(wait_max, wait_time_max),
244
245 /* extra comparisons much complicated should be here */
246
247 { NULL, NULL }
248};
249
250static void select_key(void)
251{
252 int i;
253
254 for (i = 0; keys[i].name; i++) {
255 if (!strcmp(keys[i].name, sort_key)) {
256 compare = keys[i].key;
257 return;
258 }
259 }
260
261 die("Unknown compare key:%s\n", sort_key);
262}
263
264static void insert_to_result(struct lock_stat *st,
265 int (*bigger)(struct lock_stat *, struct lock_stat *))
266{
267 struct rb_node **rb = &result.rb_node;
268 struct rb_node *parent = NULL;
269 struct lock_stat *p;
270
271 while (*rb) {
272 p = container_of(*rb, struct lock_stat, rb);
273 parent = *rb;
274
275 if (bigger(st, p))
276 rb = &(*rb)->rb_left;
277 else
278 rb = &(*rb)->rb_right;
279 }
280
281 rb_link_node(&st->rb, parent, rb);
282 rb_insert_color(&st->rb, &result);
283}
284
285/* returns left most element of result, and erase it */
286static struct lock_stat *pop_from_result(void)
287{
288 struct rb_node *node = result.rb_node;
289
290 if (!node)
291 return NULL;
292
293 while (node->rb_left)
294 node = node->rb_left;
295
296 rb_erase(node, &result);
297 return container_of(node, struct lock_stat, rb);
298}
299
300static struct lock_stat *lock_stat_findnew(void *addr, const char *name)
301{
302 struct list_head *entry = lockhashentry(addr);
303 struct lock_stat *ret, *new;
304
305 list_for_each_entry(ret, entry, hash_entry) {
306 if (ret->addr == addr)
307 return ret;
308 }
309
310 new = zalloc(sizeof(struct lock_stat));
311 if (!new)
312 goto alloc_failed;
313
314 new->addr = addr;
315 new->name = zalloc(sizeof(char) * strlen(name) + 1);
316 if (!new->name)
317 goto alloc_failed;
318 strcpy(new->name, name);
319
320 new->wait_time_min = ULLONG_MAX;
321
322 list_add(&new->hash_entry, entry);
323 return new;
324
325alloc_failed:
326 die("memory allocation failed\n");
327}
328
329static const char *input_name;
330
331struct raw_event_sample {
332 u32 size;
333 char data[0];
334};
335
336struct trace_acquire_event {
337 void *addr;
338 const char *name;
339 int flag;
340};
341
342struct trace_acquired_event {
343 void *addr;
344 const char *name;
345};
346
347struct trace_contended_event {
348 void *addr;
349 const char *name;
350};
351
352struct trace_release_event {
353 void *addr;
354 const char *name;
355};
356
357struct trace_lock_handler {
358 void (*acquire_event)(struct trace_acquire_event *,
359 struct event_format *,
360 int cpu,
361 u64 timestamp,
362 struct thread *thread);
363
364 void (*acquired_event)(struct trace_acquired_event *,
365 struct event_format *,
366 int cpu,
367 u64 timestamp,
368 struct thread *thread);
369
370 void (*contended_event)(struct trace_contended_event *,
371 struct event_format *,
372 int cpu,
373 u64 timestamp,
374 struct thread *thread);
375
376 void (*release_event)(struct trace_release_event *,
377 struct event_format *,
378 int cpu,
379 u64 timestamp,
380 struct thread *thread);
381};
382
383static struct lock_seq_stat *get_seq(struct thread_stat *ts, void *addr)
384{
385 struct lock_seq_stat *seq;
386
387 list_for_each_entry(seq, &ts->seq_list, list) {
388 if (seq->addr == addr)
389 return seq;
390 }
391
392 seq = zalloc(sizeof(struct lock_seq_stat));
393 if (!seq)
394 die("Not enough memory\n");
395 seq->state = SEQ_STATE_UNINITIALIZED;
396 seq->addr = addr;
397
398 list_add(&seq->list, &ts->seq_list);
399 return seq;
400}
401
402enum broken_state {
403 BROKEN_ACQUIRE,
404 BROKEN_ACQUIRED,
405 BROKEN_CONTENDED,
406 BROKEN_RELEASE,
407 BROKEN_MAX,
408};
409
410static int bad_hist[BROKEN_MAX];
411
412enum acquire_flags {
413 TRY_LOCK = 1,
414 READ_LOCK = 2,
415};
416
417static void
418report_lock_acquire_event(struct trace_acquire_event *acquire_event,
419 struct event_format *__event __used,
420 int cpu __used,
421 u64 timestamp __used,
422 struct thread *thread __used)
423{
424 struct lock_stat *ls;
425 struct thread_stat *ts;
426 struct lock_seq_stat *seq;
427
428 ls = lock_stat_findnew(acquire_event->addr, acquire_event->name);
429 if (ls->discard)
430 return;
431
432 ts = thread_stat_findnew(thread->pid);
433 seq = get_seq(ts, acquire_event->addr);
434
435 switch (seq->state) {
436 case SEQ_STATE_UNINITIALIZED:
437 case SEQ_STATE_RELEASED:
438 if (!acquire_event->flag) {
439 seq->state = SEQ_STATE_ACQUIRING;
440 } else {
441 if (acquire_event->flag & TRY_LOCK)
442 ls->nr_trylock++;
443 if (acquire_event->flag & READ_LOCK)
444 ls->nr_readlock++;
445 seq->state = SEQ_STATE_READ_ACQUIRED;
446 seq->read_count = 1;
447 ls->nr_acquired++;
448 }
449 break;
450 case SEQ_STATE_READ_ACQUIRED:
451 if (acquire_event->flag & READ_LOCK) {
452 seq->read_count++;
453 ls->nr_acquired++;
454 goto end;
455 } else {
456 goto broken;
457 }
458 break;
459 case SEQ_STATE_ACQUIRED:
460 case SEQ_STATE_ACQUIRING:
461 case SEQ_STATE_CONTENDED:
462broken:
463 /* broken lock sequence, discard it */
464 ls->discard = 1;
465 bad_hist[BROKEN_ACQUIRE]++;
466 list_del(&seq->list);
467 free(seq);
468 goto end;
469 break;
470 default:
471 BUG_ON("Unknown state of lock sequence found!\n");
472 break;
473 }
474
475 ls->nr_acquire++;
476 seq->prev_event_time = timestamp;
477end:
478 return;
479}
480
481static void
482report_lock_acquired_event(struct trace_acquired_event *acquired_event,
483 struct event_format *__event __used,
484 int cpu __used,
485 u64 timestamp __used,
486 struct thread *thread __used)
487{
488 struct lock_stat *ls;
489 struct thread_stat *ts;
490 struct lock_seq_stat *seq;
491 u64 contended_term;
492
493 ls = lock_stat_findnew(acquired_event->addr, acquired_event->name);
494 if (ls->discard)
495 return;
496
497 ts = thread_stat_findnew(thread->pid);
498 seq = get_seq(ts, acquired_event->addr);
499
500 switch (seq->state) {
501 case SEQ_STATE_UNINITIALIZED:
502 /* orphan event, do nothing */
503 return;
504 case SEQ_STATE_ACQUIRING:
505 break;
506 case SEQ_STATE_CONTENDED:
507 contended_term = timestamp - seq->prev_event_time;
508 ls->wait_time_total += contended_term;
509 if (contended_term < ls->wait_time_min)
510 ls->wait_time_min = contended_term;
511 if (ls->wait_time_max < contended_term)
512 ls->wait_time_max = contended_term;
513 break;
514 case SEQ_STATE_RELEASED:
515 case SEQ_STATE_ACQUIRED:
516 case SEQ_STATE_READ_ACQUIRED:
517 /* broken lock sequence, discard it */
518 ls->discard = 1;
519 bad_hist[BROKEN_ACQUIRED]++;
520 list_del(&seq->list);
521 free(seq);
522 goto end;
523 break;
524
525 default:
526 BUG_ON("Unknown state of lock sequence found!\n");
527 break;
528 }
529
530 seq->state = SEQ_STATE_ACQUIRED;
531 ls->nr_acquired++;
532 seq->prev_event_time = timestamp;
533end:
534 return;
535}
536
537static void
538report_lock_contended_event(struct trace_contended_event *contended_event,
539 struct event_format *__event __used,
540 int cpu __used,
541 u64 timestamp __used,
542 struct thread *thread __used)
543{
544 struct lock_stat *ls;
545 struct thread_stat *ts;
546 struct lock_seq_stat *seq;
547
548 ls = lock_stat_findnew(contended_event->addr, contended_event->name);
549 if (ls->discard)
550 return;
551
552 ts = thread_stat_findnew(thread->pid);
553 seq = get_seq(ts, contended_event->addr);
554
555 switch (seq->state) {
556 case SEQ_STATE_UNINITIALIZED:
557 /* orphan event, do nothing */
558 return;
559 case SEQ_STATE_ACQUIRING:
560 break;
561 case SEQ_STATE_RELEASED:
562 case SEQ_STATE_ACQUIRED:
563 case SEQ_STATE_READ_ACQUIRED:
564 case SEQ_STATE_CONTENDED:
565 /* broken lock sequence, discard it */
566 ls->discard = 1;
567 bad_hist[BROKEN_CONTENDED]++;
568 list_del(&seq->list);
569 free(seq);
570 goto end;
571 break;
572 default:
573 BUG_ON("Unknown state of lock sequence found!\n");
574 break;
575 }
576
577 seq->state = SEQ_STATE_CONTENDED;
578 ls->nr_contended++;
579 seq->prev_event_time = timestamp;
580end:
581 return;
582}
583
584static void
585report_lock_release_event(struct trace_release_event *release_event,
586 struct event_format *__event __used,
587 int cpu __used,
588 u64 timestamp __used,
589 struct thread *thread __used)
590{
591 struct lock_stat *ls;
592 struct thread_stat *ts;
593 struct lock_seq_stat *seq;
594
595 ls = lock_stat_findnew(release_event->addr, release_event->name);
596 if (ls->discard)
597 return;
598
599 ts = thread_stat_findnew(thread->pid);
600 seq = get_seq(ts, release_event->addr);
601
602 switch (seq->state) {
603 case SEQ_STATE_UNINITIALIZED:
604 goto end;
605 break;
606 case SEQ_STATE_ACQUIRED:
607 break;
608 case SEQ_STATE_READ_ACQUIRED:
609 seq->read_count--;
610 BUG_ON(seq->read_count < 0);
611 if (!seq->read_count) {
612 ls->nr_release++;
613 goto end;
614 }
615 break;
616 case SEQ_STATE_ACQUIRING:
617 case SEQ_STATE_CONTENDED:
618 case SEQ_STATE_RELEASED:
619 /* broken lock sequence, discard it */
620 ls->discard = 1;
621 bad_hist[BROKEN_RELEASE]++;
622 goto free_seq;
623 break;
624 default:
625 BUG_ON("Unknown state of lock sequence found!\n");
626 break;
627 }
628
629 ls->nr_release++;
630free_seq:
631 list_del(&seq->list);
632 free(seq);
633end:
634 return;
635}
636
637/* lock oriented handlers */
638/* TODO: handlers for CPU oriented, thread oriented */
639static struct trace_lock_handler report_lock_ops = {
640 .acquire_event = report_lock_acquire_event,
641 .acquired_event = report_lock_acquired_event,
642 .contended_event = report_lock_contended_event,
643 .release_event = report_lock_release_event,
644};
645
646static struct trace_lock_handler *trace_handler;
647
648static void
649process_lock_acquire_event(void *data,
650 struct event_format *event __used,
651 int cpu __used,
652 u64 timestamp __used,
653 struct thread *thread __used)
654{
655 struct trace_acquire_event acquire_event;
656 u64 tmp; /* this is required for casting... */
657
658 tmp = raw_field_value(event, "lockdep_addr", data);
659 memcpy(&acquire_event.addr, &tmp, sizeof(void *));
660 acquire_event.name = (char *)raw_field_ptr(event, "name", data);
661 acquire_event.flag = (int)raw_field_value(event, "flag", data);
662
663 if (trace_handler->acquire_event)
664 trace_handler->acquire_event(&acquire_event, event, cpu, timestamp, thread);
665}
666
667static void
668process_lock_acquired_event(void *data,
669 struct event_format *event __used,
670 int cpu __used,
671 u64 timestamp __used,
672 struct thread *thread __used)
673{
674 struct trace_acquired_event acquired_event;
675 u64 tmp; /* this is required for casting... */
676
677 tmp = raw_field_value(event, "lockdep_addr", data);
678 memcpy(&acquired_event.addr, &tmp, sizeof(void *));
679 acquired_event.name = (char *)raw_field_ptr(event, "name", data);
680
681 if (trace_handler->acquire_event)
682 trace_handler->acquired_event(&acquired_event, event, cpu, timestamp, thread);
683}
684
685static void
686process_lock_contended_event(void *data,
687 struct event_format *event __used,
688 int cpu __used,
689 u64 timestamp __used,
690 struct thread *thread __used)
691{
692 struct trace_contended_event contended_event;
693 u64 tmp; /* this is required for casting... */
694
695 tmp = raw_field_value(event, "lockdep_addr", data);
696 memcpy(&contended_event.addr, &tmp, sizeof(void *));
697 contended_event.name = (char *)raw_field_ptr(event, "name", data);
698
699 if (trace_handler->acquire_event)
700 trace_handler->contended_event(&contended_event, event, cpu, timestamp, thread);
701}
702
703static void
704process_lock_release_event(void *data,
705 struct event_format *event __used,
706 int cpu __used,
707 u64 timestamp __used,
708 struct thread *thread __used)
709{
710 struct trace_release_event release_event;
711 u64 tmp; /* this is required for casting... */
712
713 tmp = raw_field_value(event, "lockdep_addr", data);
714 memcpy(&release_event.addr, &tmp, sizeof(void *));
715 release_event.name = (char *)raw_field_ptr(event, "name", data);
716
717 if (trace_handler->acquire_event)
718 trace_handler->release_event(&release_event, event, cpu, timestamp, thread);
719}
720
721static void
722process_raw_event(void *data, int cpu, u64 timestamp, struct thread *thread)
723{
724 struct event_format *event;
725 int type;
726
727 type = trace_parse_common_type(data);
728 event = trace_find_event(type);
729
730 if (!strcmp(event->name, "lock_acquire"))
731 process_lock_acquire_event(data, event, cpu, timestamp, thread);
732 if (!strcmp(event->name, "lock_acquired"))
733 process_lock_acquired_event(data, event, cpu, timestamp, thread);
734 if (!strcmp(event->name, "lock_contended"))
735 process_lock_contended_event(data, event, cpu, timestamp, thread);
736 if (!strcmp(event->name, "lock_release"))
737 process_lock_release_event(data, event, cpu, timestamp, thread);
738}
739
740static void print_bad_events(int bad, int total)
741{
742 /* Output for debug, this have to be removed */
743 int i;
744 const char *name[4] =
745 { "acquire", "acquired", "contended", "release" };
746
747 pr_info("\n=== output for debug===\n\n");
748 pr_info("bad: %d, total: %d\n", bad, total);
749 pr_info("bad rate: %f %%\n", (double)bad / (double)total * 100);
750 pr_info("histogram of events caused bad sequence\n");
751 for (i = 0; i < BROKEN_MAX; i++)
752 pr_info(" %10s: %d\n", name[i], bad_hist[i]);
753}
754
755/* TODO: various way to print, coloring, nano or milli sec */
756static void print_result(void)
757{
758 struct lock_stat *st;
759 char cut_name[20];
760 int bad, total;
761
762 pr_info("%20s ", "Name");
763 pr_info("%10s ", "acquired");
764 pr_info("%10s ", "contended");
765
766 pr_info("%15s ", "total wait (ns)");
767 pr_info("%15s ", "max wait (ns)");
768 pr_info("%15s ", "min wait (ns)");
769
770 pr_info("\n\n");
771
772 bad = total = 0;
773 while ((st = pop_from_result())) {
774 total++;
775 if (st->discard) {
776 bad++;
777 continue;
778 }
779 bzero(cut_name, 20);
780
781 if (strlen(st->name) < 16) {
782 /* output raw name */
783 pr_info("%20s ", st->name);
784 } else {
785 strncpy(cut_name, st->name, 16);
786 cut_name[16] = '.';
787 cut_name[17] = '.';
788 cut_name[18] = '.';
789 cut_name[19] = '\0';
790 /* cut off name for saving output style */
791 pr_info("%20s ", cut_name);
792 }
793
794 pr_info("%10u ", st->nr_acquired);
795 pr_info("%10u ", st->nr_contended);
796
797 pr_info("%15" PRIu64 " ", st->wait_time_total);
798 pr_info("%15" PRIu64 " ", st->wait_time_max);
799 pr_info("%15" PRIu64 " ", st->wait_time_min == ULLONG_MAX ?
800 0 : st->wait_time_min);
801 pr_info("\n");
802 }
803
804 print_bad_events(bad, total);
805}
806
807static bool info_threads, info_map;
808
809static void dump_threads(void)
810{
811 struct thread_stat *st;
812 struct rb_node *node;
813 struct thread *t;
814
815 pr_info("%10s: comm\n", "Thread ID");
816
817 node = rb_first(&thread_stats);
818 while (node) {
819 st = container_of(node, struct thread_stat, rb);
820 t = perf_session__findnew(session, st->tid);
821 pr_info("%10d: %s\n", st->tid, t->comm);
822 node = rb_next(node);
823 };
824}
825
826static void dump_map(void)
827{
828 unsigned int i;
829 struct lock_stat *st;
830
831 pr_info("Address of instance: name of class\n");
832 for (i = 0; i < LOCKHASH_SIZE; i++) {
833 list_for_each_entry(st, &lockhash_table[i], hash_entry) {
834 pr_info(" %p: %s\n", st->addr, st->name);
835 }
836 }
837}
838
839static void dump_info(void)
840{
841 if (info_threads)
842 dump_threads();
843 else if (info_map)
844 dump_map();
845 else
846 die("Unknown type of information\n");
847}
848
849static int process_sample_event(struct perf_tool *tool __used,
850 union perf_event *event,
851 struct perf_sample *sample,
852 struct perf_evsel *evsel __used,
853 struct machine *machine)
854{
855 struct thread *thread = machine__findnew_thread(machine, sample->tid);
856
857 if (thread == NULL) {
858 pr_debug("problem processing %d event, skipping it.\n",
859 event->header.type);
860 return -1;
861 }
862
863 process_raw_event(sample->raw_data, sample->cpu, sample->time, thread);
864
865 return 0;
866}
867
868static struct perf_tool eops = {
869 .sample = process_sample_event,
870 .comm = perf_event__process_comm,
871 .ordered_samples = true,
872};
873
874static int read_events(void)
875{
876 session = perf_session__new(input_name, O_RDONLY, 0, false, &eops);
877 if (!session)
878 die("Initializing perf session failed\n");
879
880 return perf_session__process_events(session, &eops);
881}
882
883static void sort_result(void)
884{
885 unsigned int i;
886 struct lock_stat *st;
887
888 for (i = 0; i < LOCKHASH_SIZE; i++) {
889 list_for_each_entry(st, &lockhash_table[i], hash_entry) {
890 insert_to_result(st, compare);
891 }
892 }
893}
894
895static void __cmd_report(void)
896{
897 setup_pager();
898 select_key();
899 read_events();
900 sort_result();
901 print_result();
902}
903
904static const char * const report_usage[] = {
905 "perf lock report [<options>]",
906 NULL
907};
908
909static const struct option report_options[] = {
910 OPT_STRING('k', "key", &sort_key, "acquired",
911 "key for sorting (acquired / contended / wait_total / wait_max / wait_min)"),
912 /* TODO: type */
913 OPT_END()
914};
915
916static const char * const info_usage[] = {
917 "perf lock info [<options>]",
918 NULL
919};
920
921static const struct option info_options[] = {
922 OPT_BOOLEAN('t', "threads", &info_threads,
923 "dump thread list in perf.data"),
924 OPT_BOOLEAN('m', "map", &info_map,
925 "map of lock instances (address:name table)"),
926 OPT_END()
927};
928
929static const char * const lock_usage[] = {
930 "perf lock [<options>] {record|report|script|info}",
931 NULL
932};
933
934static const struct option lock_options[] = {
935 OPT_STRING('i', "input", &input_name, "file", "input file name"),
936 OPT_INCR('v', "verbose", &verbose, "be more verbose (show symbol address, etc)"),
937 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, "dump raw trace in ASCII"),
938 OPT_END()
939};
940
941static const char *record_args[] = {
942 "record",
943 "-R",
944 "-f",
945 "-m", "1024",
946 "-c", "1",
947 "-e", "lock:lock_acquire",
948 "-e", "lock:lock_acquired",
949 "-e", "lock:lock_contended",
950 "-e", "lock:lock_release",
951};
952
953static int __cmd_record(int argc, const char **argv)
954{
955 unsigned int rec_argc, i, j;
956 const char **rec_argv;
957
958 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
959 rec_argv = calloc(rec_argc + 1, sizeof(char *));
960
961 if (rec_argv == NULL)
962 return -ENOMEM;
963
964 for (i = 0; i < ARRAY_SIZE(record_args); i++)
965 rec_argv[i] = strdup(record_args[i]);
966
967 for (j = 1; j < (unsigned int)argc; j++, i++)
968 rec_argv[i] = argv[j];
969
970 BUG_ON(i != rec_argc);
971
972 return cmd_record(i, rec_argv, NULL);
973}
974
975int cmd_lock(int argc, const char **argv, const char *prefix __used)
976{
977 unsigned int i;
978
979 symbol__init();
980 for (i = 0; i < LOCKHASH_SIZE; i++)
981 INIT_LIST_HEAD(lockhash_table + i);
982
983 argc = parse_options(argc, argv, lock_options, lock_usage,
984 PARSE_OPT_STOP_AT_NON_OPTION);
985 if (!argc)
986 usage_with_options(lock_usage, lock_options);
987
988 if (!strncmp(argv[0], "rec", 3)) {
989 return __cmd_record(argc, argv);
990 } else if (!strncmp(argv[0], "report", 6)) {
991 trace_handler = &report_lock_ops;
992 if (argc) {
993 argc = parse_options(argc, argv,
994 report_options, report_usage, 0);
995 if (argc)
996 usage_with_options(report_usage, report_options);
997 }
998 __cmd_report();
999 } else if (!strcmp(argv[0], "script")) {
1000 /* Aliased to 'perf script' */
1001 return cmd_script(argc, argv, prefix);
1002 } else if (!strcmp(argv[0], "info")) {
1003 if (argc) {
1004 argc = parse_options(argc, argv,
1005 info_options, info_usage, 0);
1006 if (argc)
1007 usage_with_options(info_usage, info_options);
1008 }
1009 /* recycling report_lock_ops */
1010 trace_handler = &report_lock_ops;
1011 setup_pager();
1012 read_events();
1013 dump_info();
1014 } else {
1015 usage_with_options(lock_usage, lock_options);
1016 }
1017
1018 return 0;
1019}
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#include "util/target.h"
13#include "util/cgroup.h"
14#include "util/callchain.h"
15#include "util/lock-contention.h"
16#include "util/bpf_skel/lock_data.h"
17
18#include <subcmd/pager.h>
19#include <subcmd/parse-options.h>
20#include "util/trace-event.h"
21#include "util/tracepoint.h"
22
23#include "util/debug.h"
24#include "util/session.h"
25#include "util/tool.h"
26#include "util/data.h"
27#include "util/string2.h"
28#include "util/map.h"
29#include "util/util.h"
30
31#include <stdio.h>
32#include <sys/types.h>
33#include <sys/prctl.h>
34#include <semaphore.h>
35#include <math.h>
36#include <limits.h>
37#include <ctype.h>
38
39#include <linux/list.h>
40#include <linux/hash.h>
41#include <linux/kernel.h>
42#include <linux/zalloc.h>
43#include <linux/err.h>
44#include <linux/stringify.h>
45
46static struct perf_session *session;
47static struct target target;
48
49/* based on kernel/lockdep.c */
50#define LOCKHASH_BITS 12
51#define LOCKHASH_SIZE (1UL << LOCKHASH_BITS)
52
53static struct hlist_head *lockhash_table;
54
55#define __lockhashfn(key) hash_long((unsigned long)key, LOCKHASH_BITS)
56#define lockhashentry(key) (lockhash_table + __lockhashfn((key)))
57
58static struct rb_root thread_stats;
59
60static bool combine_locks;
61static bool show_thread_stats;
62static bool show_lock_addrs;
63static bool show_lock_owner;
64static bool show_lock_cgroups;
65static bool use_bpf;
66static unsigned long bpf_map_entries = MAX_ENTRIES;
67static int max_stack_depth = CONTENTION_STACK_DEPTH;
68static int stack_skip = CONTENTION_STACK_SKIP;
69static int print_nr_entries = INT_MAX / 2;
70static LIST_HEAD(callstack_filters);
71static const char *output_name = NULL;
72static FILE *lock_output;
73
74struct callstack_filter {
75 struct list_head list;
76 char name[];
77};
78
79static struct lock_filter filters;
80
81static enum lock_aggr_mode aggr_mode = LOCK_AGGR_ADDR;
82
83static bool needs_callstack(void)
84{
85 return !list_empty(&callstack_filters);
86}
87
88static struct thread_stat *thread_stat_find(u32 tid)
89{
90 struct rb_node *node;
91 struct thread_stat *st;
92
93 node = thread_stats.rb_node;
94 while (node) {
95 st = container_of(node, struct thread_stat, rb);
96 if (st->tid == tid)
97 return st;
98 else if (tid < st->tid)
99 node = node->rb_left;
100 else
101 node = node->rb_right;
102 }
103
104 return NULL;
105}
106
107static void thread_stat_insert(struct thread_stat *new)
108{
109 struct rb_node **rb = &thread_stats.rb_node;
110 struct rb_node *parent = NULL;
111 struct thread_stat *p;
112
113 while (*rb) {
114 p = container_of(*rb, struct thread_stat, rb);
115 parent = *rb;
116
117 if (new->tid < p->tid)
118 rb = &(*rb)->rb_left;
119 else if (new->tid > p->tid)
120 rb = &(*rb)->rb_right;
121 else
122 BUG_ON("inserting invalid thread_stat\n");
123 }
124
125 rb_link_node(&new->rb, parent, rb);
126 rb_insert_color(&new->rb, &thread_stats);
127}
128
129static struct thread_stat *thread_stat_findnew_after_first(u32 tid)
130{
131 struct thread_stat *st;
132
133 st = thread_stat_find(tid);
134 if (st)
135 return st;
136
137 st = zalloc(sizeof(struct thread_stat));
138 if (!st) {
139 pr_err("memory allocation failed\n");
140 return NULL;
141 }
142
143 st->tid = tid;
144 INIT_LIST_HEAD(&st->seq_list);
145
146 thread_stat_insert(st);
147
148 return st;
149}
150
151static struct thread_stat *thread_stat_findnew_first(u32 tid);
152static struct thread_stat *(*thread_stat_findnew)(u32 tid) =
153 thread_stat_findnew_first;
154
155static struct thread_stat *thread_stat_findnew_first(u32 tid)
156{
157 struct thread_stat *st;
158
159 st = zalloc(sizeof(struct thread_stat));
160 if (!st) {
161 pr_err("memory allocation failed\n");
162 return NULL;
163 }
164 st->tid = tid;
165 INIT_LIST_HEAD(&st->seq_list);
166
167 rb_link_node(&st->rb, NULL, &thread_stats.rb_node);
168 rb_insert_color(&st->rb, &thread_stats);
169
170 thread_stat_findnew = thread_stat_findnew_after_first;
171 return st;
172}
173
174/* build simple key function one is bigger than two */
175#define SINGLE_KEY(member) \
176 static int lock_stat_key_ ## member(struct lock_stat *one, \
177 struct lock_stat *two) \
178 { \
179 return one->member > two->member; \
180 }
181
182SINGLE_KEY(nr_acquired)
183SINGLE_KEY(nr_contended)
184SINGLE_KEY(avg_wait_time)
185SINGLE_KEY(wait_time_total)
186SINGLE_KEY(wait_time_max)
187
188static int lock_stat_key_wait_time_min(struct lock_stat *one,
189 struct lock_stat *two)
190{
191 u64 s1 = one->wait_time_min;
192 u64 s2 = two->wait_time_min;
193 if (s1 == ULLONG_MAX)
194 s1 = 0;
195 if (s2 == ULLONG_MAX)
196 s2 = 0;
197 return s1 > s2;
198}
199
200struct lock_key {
201 /*
202 * name: the value for specify by user
203 * this should be simpler than raw name of member
204 * e.g. nr_acquired -> acquired, wait_time_total -> wait_total
205 */
206 const char *name;
207 /* header: the string printed on the header line */
208 const char *header;
209 /* len: the printing width of the field */
210 int len;
211 /* key: a pointer to function to compare two lock stats for sorting */
212 int (*key)(struct lock_stat*, struct lock_stat*);
213 /* print: a pointer to function to print a given lock stats */
214 void (*print)(struct lock_key*, struct lock_stat*);
215 /* list: list entry to link this */
216 struct list_head list;
217};
218
219static void lock_stat_key_print_time(unsigned long long nsec, int len)
220{
221 static const struct {
222 float base;
223 const char *unit;
224 } table[] = {
225 { 1e9 * 3600, "h " },
226 { 1e9 * 60, "m " },
227 { 1e9, "s " },
228 { 1e6, "ms" },
229 { 1e3, "us" },
230 { 0, NULL },
231 };
232
233 /* for CSV output */
234 if (len == 0) {
235 fprintf(lock_output, "%llu", nsec);
236 return;
237 }
238
239 for (int i = 0; table[i].unit; i++) {
240 if (nsec < table[i].base)
241 continue;
242
243 fprintf(lock_output, "%*.2f %s", len - 3, nsec / table[i].base, table[i].unit);
244 return;
245 }
246
247 fprintf(lock_output, "%*llu %s", len - 3, nsec, "ns");
248}
249
250#define PRINT_KEY(member) \
251static void lock_stat_key_print_ ## member(struct lock_key *key, \
252 struct lock_stat *ls) \
253{ \
254 fprintf(lock_output, "%*llu", key->len, (unsigned long long)ls->member);\
255}
256
257#define PRINT_TIME(member) \
258static void lock_stat_key_print_ ## member(struct lock_key *key, \
259 struct lock_stat *ls) \
260{ \
261 lock_stat_key_print_time((unsigned long long)ls->member, key->len); \
262}
263
264PRINT_KEY(nr_acquired)
265PRINT_KEY(nr_contended)
266PRINT_TIME(avg_wait_time)
267PRINT_TIME(wait_time_total)
268PRINT_TIME(wait_time_max)
269
270static void lock_stat_key_print_wait_time_min(struct lock_key *key,
271 struct lock_stat *ls)
272{
273 u64 wait_time = ls->wait_time_min;
274
275 if (wait_time == ULLONG_MAX)
276 wait_time = 0;
277
278 lock_stat_key_print_time(wait_time, key->len);
279}
280
281
282static const char *sort_key = "acquired";
283
284static int (*compare)(struct lock_stat *, struct lock_stat *);
285
286static struct rb_root sorted; /* place to store intermediate data */
287static struct rb_root result; /* place to store sorted data */
288
289static LIST_HEAD(lock_keys);
290static const char *output_fields;
291
292#define DEF_KEY_LOCK(name, header, fn_suffix, len) \
293 { #name, header, len, lock_stat_key_ ## fn_suffix, lock_stat_key_print_ ## fn_suffix, {} }
294static struct lock_key report_keys[] = {
295 DEF_KEY_LOCK(acquired, "acquired", nr_acquired, 10),
296 DEF_KEY_LOCK(contended, "contended", nr_contended, 10),
297 DEF_KEY_LOCK(avg_wait, "avg wait", avg_wait_time, 12),
298 DEF_KEY_LOCK(wait_total, "total wait", wait_time_total, 12),
299 DEF_KEY_LOCK(wait_max, "max wait", wait_time_max, 12),
300 DEF_KEY_LOCK(wait_min, "min wait", wait_time_min, 12),
301
302 /* extra comparisons much complicated should be here */
303 { }
304};
305
306static struct lock_key contention_keys[] = {
307 DEF_KEY_LOCK(contended, "contended", nr_contended, 10),
308 DEF_KEY_LOCK(wait_total, "total wait", wait_time_total, 12),
309 DEF_KEY_LOCK(wait_max, "max wait", wait_time_max, 12),
310 DEF_KEY_LOCK(wait_min, "min wait", wait_time_min, 12),
311 DEF_KEY_LOCK(avg_wait, "avg wait", avg_wait_time, 12),
312
313 /* extra comparisons much complicated should be here */
314 { }
315};
316
317static int select_key(bool contention)
318{
319 int i;
320 struct lock_key *keys = report_keys;
321
322 if (contention)
323 keys = contention_keys;
324
325 for (i = 0; keys[i].name; i++) {
326 if (!strcmp(keys[i].name, sort_key)) {
327 compare = keys[i].key;
328
329 /* selected key should be in the output fields */
330 if (list_empty(&keys[i].list))
331 list_add_tail(&keys[i].list, &lock_keys);
332
333 return 0;
334 }
335 }
336
337 pr_err("Unknown compare key: %s\n", sort_key);
338 return -1;
339}
340
341static int add_output_field(bool contention, char *name)
342{
343 int i;
344 struct lock_key *keys = report_keys;
345
346 if (contention)
347 keys = contention_keys;
348
349 for (i = 0; keys[i].name; i++) {
350 if (strcmp(keys[i].name, name))
351 continue;
352
353 /* prevent double link */
354 if (list_empty(&keys[i].list))
355 list_add_tail(&keys[i].list, &lock_keys);
356
357 return 0;
358 }
359
360 pr_err("Unknown output field: %s\n", name);
361 return -1;
362}
363
364static int setup_output_field(bool contention, const char *str)
365{
366 char *tok, *tmp, *orig;
367 int i, ret = 0;
368 struct lock_key *keys = report_keys;
369
370 if (contention)
371 keys = contention_keys;
372
373 /* no output field given: use all of them */
374 if (str == NULL) {
375 for (i = 0; keys[i].name; i++)
376 list_add_tail(&keys[i].list, &lock_keys);
377 return 0;
378 }
379
380 for (i = 0; keys[i].name; i++)
381 INIT_LIST_HEAD(&keys[i].list);
382
383 orig = tmp = strdup(str);
384 if (orig == NULL)
385 return -ENOMEM;
386
387 while ((tok = strsep(&tmp, ",")) != NULL){
388 ret = add_output_field(contention, tok);
389 if (ret < 0)
390 break;
391 }
392 free(orig);
393
394 return ret;
395}
396
397static void combine_lock_stats(struct lock_stat *st)
398{
399 struct rb_node **rb = &sorted.rb_node;
400 struct rb_node *parent = NULL;
401 struct lock_stat *p;
402 int ret;
403
404 while (*rb) {
405 p = container_of(*rb, struct lock_stat, rb);
406 parent = *rb;
407
408 if (st->name && p->name)
409 ret = strcmp(st->name, p->name);
410 else
411 ret = !!st->name - !!p->name;
412
413 if (ret == 0) {
414 p->nr_acquired += st->nr_acquired;
415 p->nr_contended += st->nr_contended;
416 p->wait_time_total += st->wait_time_total;
417
418 if (p->nr_contended)
419 p->avg_wait_time = p->wait_time_total / p->nr_contended;
420
421 if (p->wait_time_min > st->wait_time_min)
422 p->wait_time_min = st->wait_time_min;
423 if (p->wait_time_max < st->wait_time_max)
424 p->wait_time_max = st->wait_time_max;
425
426 p->broken |= st->broken;
427 st->combined = 1;
428 return;
429 }
430
431 if (ret < 0)
432 rb = &(*rb)->rb_left;
433 else
434 rb = &(*rb)->rb_right;
435 }
436
437 rb_link_node(&st->rb, parent, rb);
438 rb_insert_color(&st->rb, &sorted);
439}
440
441static void insert_to_result(struct lock_stat *st,
442 int (*bigger)(struct lock_stat *, struct lock_stat *))
443{
444 struct rb_node **rb = &result.rb_node;
445 struct rb_node *parent = NULL;
446 struct lock_stat *p;
447
448 if (combine_locks && st->combined)
449 return;
450
451 while (*rb) {
452 p = container_of(*rb, struct lock_stat, rb);
453 parent = *rb;
454
455 if (bigger(st, p))
456 rb = &(*rb)->rb_left;
457 else
458 rb = &(*rb)->rb_right;
459 }
460
461 rb_link_node(&st->rb, parent, rb);
462 rb_insert_color(&st->rb, &result);
463}
464
465/* returns left most element of result, and erase it */
466static struct lock_stat *pop_from_result(void)
467{
468 struct rb_node *node = result.rb_node;
469
470 if (!node)
471 return NULL;
472
473 while (node->rb_left)
474 node = node->rb_left;
475
476 rb_erase(node, &result);
477 return container_of(node, struct lock_stat, rb);
478}
479
480struct lock_stat *lock_stat_find(u64 addr)
481{
482 struct hlist_head *entry = lockhashentry(addr);
483 struct lock_stat *ret;
484
485 hlist_for_each_entry(ret, entry, hash_entry) {
486 if (ret->addr == addr)
487 return ret;
488 }
489 return NULL;
490}
491
492struct lock_stat *lock_stat_findnew(u64 addr, const char *name, int flags)
493{
494 struct hlist_head *entry = lockhashentry(addr);
495 struct lock_stat *ret, *new;
496
497 hlist_for_each_entry(ret, entry, hash_entry) {
498 if (ret->addr == addr)
499 return ret;
500 }
501
502 new = zalloc(sizeof(struct lock_stat));
503 if (!new)
504 goto alloc_failed;
505
506 new->addr = addr;
507 new->name = strdup(name);
508 if (!new->name) {
509 free(new);
510 goto alloc_failed;
511 }
512
513 new->flags = flags;
514 new->wait_time_min = ULLONG_MAX;
515
516 hlist_add_head(&new->hash_entry, entry);
517 return new;
518
519alloc_failed:
520 pr_err("memory allocation failed\n");
521 return NULL;
522}
523
524bool match_callstack_filter(struct machine *machine, u64 *callstack)
525{
526 struct map *kmap;
527 struct symbol *sym;
528 u64 ip;
529 const char *arch = perf_env__arch(machine->env);
530
531 if (list_empty(&callstack_filters))
532 return true;
533
534 for (int i = 0; i < max_stack_depth; i++) {
535 struct callstack_filter *filter;
536
537 /*
538 * In powerpc, the callchain saved by kernel always includes
539 * first three entries as the NIP (next instruction pointer),
540 * LR (link register), and the contents of LR save area in the
541 * second stack frame. In certain scenarios its possible to have
542 * invalid kernel instruction addresses in either LR or the second
543 * stack frame's LR. In that case, kernel will store that address as
544 * zero.
545 *
546 * The below check will continue to look into callstack,
547 * incase first or second callstack index entry has 0
548 * address for powerpc.
549 */
550 if (!callstack || (!callstack[i] && (strcmp(arch, "powerpc") ||
551 (i != 1 && i != 2))))
552 break;
553
554 ip = callstack[i];
555 sym = machine__find_kernel_symbol(machine, ip, &kmap);
556 if (sym == NULL)
557 continue;
558
559 list_for_each_entry(filter, &callstack_filters, list) {
560 if (strstr(sym->name, filter->name))
561 return true;
562 }
563 }
564 return false;
565}
566
567struct trace_lock_handler {
568 /* it's used on CONFIG_LOCKDEP */
569 int (*acquire_event)(struct evsel *evsel,
570 struct perf_sample *sample);
571
572 /* it's used on CONFIG_LOCKDEP && CONFIG_LOCK_STAT */
573 int (*acquired_event)(struct evsel *evsel,
574 struct perf_sample *sample);
575
576 /* it's used on CONFIG_LOCKDEP && CONFIG_LOCK_STAT */
577 int (*contended_event)(struct evsel *evsel,
578 struct perf_sample *sample);
579
580 /* it's used on CONFIG_LOCKDEP */
581 int (*release_event)(struct evsel *evsel,
582 struct perf_sample *sample);
583
584 /* it's used when CONFIG_LOCKDEP is off */
585 int (*contention_begin_event)(struct evsel *evsel,
586 struct perf_sample *sample);
587
588 /* it's used when CONFIG_LOCKDEP is off */
589 int (*contention_end_event)(struct evsel *evsel,
590 struct perf_sample *sample);
591};
592
593static struct lock_seq_stat *get_seq(struct thread_stat *ts, u64 addr)
594{
595 struct lock_seq_stat *seq;
596
597 list_for_each_entry(seq, &ts->seq_list, list) {
598 if (seq->addr == addr)
599 return seq;
600 }
601
602 seq = zalloc(sizeof(struct lock_seq_stat));
603 if (!seq) {
604 pr_err("memory allocation failed\n");
605 return NULL;
606 }
607 seq->state = SEQ_STATE_UNINITIALIZED;
608 seq->addr = addr;
609
610 list_add(&seq->list, &ts->seq_list);
611 return seq;
612}
613
614enum broken_state {
615 BROKEN_ACQUIRE,
616 BROKEN_ACQUIRED,
617 BROKEN_CONTENDED,
618 BROKEN_RELEASE,
619 BROKEN_MAX,
620};
621
622static int bad_hist[BROKEN_MAX];
623
624enum acquire_flags {
625 TRY_LOCK = 1,
626 READ_LOCK = 2,
627};
628
629static int get_key_by_aggr_mode_simple(u64 *key, u64 addr, u32 tid)
630{
631 switch (aggr_mode) {
632 case LOCK_AGGR_ADDR:
633 *key = addr;
634 break;
635 case LOCK_AGGR_TASK:
636 *key = tid;
637 break;
638 case LOCK_AGGR_CALLER:
639 case LOCK_AGGR_CGROUP:
640 default:
641 pr_err("Invalid aggregation mode: %d\n", aggr_mode);
642 return -EINVAL;
643 }
644 return 0;
645}
646
647static u64 callchain_id(struct evsel *evsel, struct perf_sample *sample);
648
649static int get_key_by_aggr_mode(u64 *key, u64 addr, struct evsel *evsel,
650 struct perf_sample *sample)
651{
652 if (aggr_mode == LOCK_AGGR_CALLER) {
653 *key = callchain_id(evsel, sample);
654 return 0;
655 }
656 return get_key_by_aggr_mode_simple(key, addr, sample->tid);
657}
658
659static int report_lock_acquire_event(struct evsel *evsel,
660 struct perf_sample *sample)
661{
662 struct lock_stat *ls;
663 struct thread_stat *ts;
664 struct lock_seq_stat *seq;
665 const char *name = evsel__strval(evsel, sample, "name");
666 u64 addr = evsel__intval(evsel, sample, "lockdep_addr");
667 int flag = evsel__intval(evsel, sample, "flags");
668 u64 key;
669 int ret;
670
671 ret = get_key_by_aggr_mode_simple(&key, addr, sample->tid);
672 if (ret < 0)
673 return ret;
674
675 ls = lock_stat_findnew(key, name, 0);
676 if (!ls)
677 return -ENOMEM;
678
679 ts = thread_stat_findnew(sample->tid);
680 if (!ts)
681 return -ENOMEM;
682
683 seq = get_seq(ts, addr);
684 if (!seq)
685 return -ENOMEM;
686
687 switch (seq->state) {
688 case SEQ_STATE_UNINITIALIZED:
689 case SEQ_STATE_RELEASED:
690 if (!flag) {
691 seq->state = SEQ_STATE_ACQUIRING;
692 } else {
693 if (flag & TRY_LOCK)
694 ls->nr_trylock++;
695 if (flag & READ_LOCK)
696 ls->nr_readlock++;
697 seq->state = SEQ_STATE_READ_ACQUIRED;
698 seq->read_count = 1;
699 ls->nr_acquired++;
700 }
701 break;
702 case SEQ_STATE_READ_ACQUIRED:
703 if (flag & READ_LOCK) {
704 seq->read_count++;
705 ls->nr_acquired++;
706 goto end;
707 } else {
708 goto broken;
709 }
710 break;
711 case SEQ_STATE_ACQUIRED:
712 case SEQ_STATE_ACQUIRING:
713 case SEQ_STATE_CONTENDED:
714broken:
715 /* broken lock sequence */
716 if (!ls->broken) {
717 ls->broken = 1;
718 bad_hist[BROKEN_ACQUIRE]++;
719 }
720 list_del_init(&seq->list);
721 free(seq);
722 goto end;
723 default:
724 BUG_ON("Unknown state of lock sequence found!\n");
725 break;
726 }
727
728 ls->nr_acquire++;
729 seq->prev_event_time = sample->time;
730end:
731 return 0;
732}
733
734static int report_lock_acquired_event(struct evsel *evsel,
735 struct perf_sample *sample)
736{
737 struct lock_stat *ls;
738 struct thread_stat *ts;
739 struct lock_seq_stat *seq;
740 u64 contended_term;
741 const char *name = evsel__strval(evsel, sample, "name");
742 u64 addr = evsel__intval(evsel, sample, "lockdep_addr");
743 u64 key;
744 int ret;
745
746 ret = get_key_by_aggr_mode_simple(&key, addr, sample->tid);
747 if (ret < 0)
748 return ret;
749
750 ls = lock_stat_findnew(key, name, 0);
751 if (!ls)
752 return -ENOMEM;
753
754 ts = thread_stat_findnew(sample->tid);
755 if (!ts)
756 return -ENOMEM;
757
758 seq = get_seq(ts, addr);
759 if (!seq)
760 return -ENOMEM;
761
762 switch (seq->state) {
763 case SEQ_STATE_UNINITIALIZED:
764 /* orphan event, do nothing */
765 return 0;
766 case SEQ_STATE_ACQUIRING:
767 break;
768 case SEQ_STATE_CONTENDED:
769 contended_term = sample->time - seq->prev_event_time;
770 ls->wait_time_total += contended_term;
771 if (contended_term < ls->wait_time_min)
772 ls->wait_time_min = contended_term;
773 if (ls->wait_time_max < contended_term)
774 ls->wait_time_max = contended_term;
775 break;
776 case SEQ_STATE_RELEASED:
777 case SEQ_STATE_ACQUIRED:
778 case SEQ_STATE_READ_ACQUIRED:
779 /* broken lock sequence */
780 if (!ls->broken) {
781 ls->broken = 1;
782 bad_hist[BROKEN_ACQUIRED]++;
783 }
784 list_del_init(&seq->list);
785 free(seq);
786 goto end;
787 default:
788 BUG_ON("Unknown state of lock sequence found!\n");
789 break;
790 }
791
792 seq->state = SEQ_STATE_ACQUIRED;
793 ls->nr_acquired++;
794 ls->avg_wait_time = ls->nr_contended ? ls->wait_time_total/ls->nr_contended : 0;
795 seq->prev_event_time = sample->time;
796end:
797 return 0;
798}
799
800static int report_lock_contended_event(struct evsel *evsel,
801 struct perf_sample *sample)
802{
803 struct lock_stat *ls;
804 struct thread_stat *ts;
805 struct lock_seq_stat *seq;
806 const char *name = evsel__strval(evsel, sample, "name");
807 u64 addr = evsel__intval(evsel, sample, "lockdep_addr");
808 u64 key;
809 int ret;
810
811 ret = get_key_by_aggr_mode_simple(&key, addr, sample->tid);
812 if (ret < 0)
813 return ret;
814
815 ls = lock_stat_findnew(key, name, 0);
816 if (!ls)
817 return -ENOMEM;
818
819 ts = thread_stat_findnew(sample->tid);
820 if (!ts)
821 return -ENOMEM;
822
823 seq = get_seq(ts, addr);
824 if (!seq)
825 return -ENOMEM;
826
827 switch (seq->state) {
828 case SEQ_STATE_UNINITIALIZED:
829 /* orphan event, do nothing */
830 return 0;
831 case SEQ_STATE_ACQUIRING:
832 break;
833 case SEQ_STATE_RELEASED:
834 case SEQ_STATE_ACQUIRED:
835 case SEQ_STATE_READ_ACQUIRED:
836 case SEQ_STATE_CONTENDED:
837 /* broken lock sequence */
838 if (!ls->broken) {
839 ls->broken = 1;
840 bad_hist[BROKEN_CONTENDED]++;
841 }
842 list_del_init(&seq->list);
843 free(seq);
844 goto end;
845 default:
846 BUG_ON("Unknown state of lock sequence found!\n");
847 break;
848 }
849
850 seq->state = SEQ_STATE_CONTENDED;
851 ls->nr_contended++;
852 ls->avg_wait_time = ls->wait_time_total/ls->nr_contended;
853 seq->prev_event_time = sample->time;
854end:
855 return 0;
856}
857
858static int report_lock_release_event(struct evsel *evsel,
859 struct perf_sample *sample)
860{
861 struct lock_stat *ls;
862 struct thread_stat *ts;
863 struct lock_seq_stat *seq;
864 const char *name = evsel__strval(evsel, sample, "name");
865 u64 addr = evsel__intval(evsel, sample, "lockdep_addr");
866 u64 key;
867 int ret;
868
869 ret = get_key_by_aggr_mode_simple(&key, addr, sample->tid);
870 if (ret < 0)
871 return ret;
872
873 ls = lock_stat_findnew(key, name, 0);
874 if (!ls)
875 return -ENOMEM;
876
877 ts = thread_stat_findnew(sample->tid);
878 if (!ts)
879 return -ENOMEM;
880
881 seq = get_seq(ts, addr);
882 if (!seq)
883 return -ENOMEM;
884
885 switch (seq->state) {
886 case SEQ_STATE_UNINITIALIZED:
887 goto end;
888 case SEQ_STATE_ACQUIRED:
889 break;
890 case SEQ_STATE_READ_ACQUIRED:
891 seq->read_count--;
892 BUG_ON(seq->read_count < 0);
893 if (seq->read_count) {
894 ls->nr_release++;
895 goto end;
896 }
897 break;
898 case SEQ_STATE_ACQUIRING:
899 case SEQ_STATE_CONTENDED:
900 case SEQ_STATE_RELEASED:
901 /* broken lock sequence */
902 if (!ls->broken) {
903 ls->broken = 1;
904 bad_hist[BROKEN_RELEASE]++;
905 }
906 goto free_seq;
907 default:
908 BUG_ON("Unknown state of lock sequence found!\n");
909 break;
910 }
911
912 ls->nr_release++;
913free_seq:
914 list_del_init(&seq->list);
915 free(seq);
916end:
917 return 0;
918}
919
920static int get_symbol_name_offset(struct map *map, struct symbol *sym, u64 ip,
921 char *buf, int size)
922{
923 u64 offset;
924
925 if (map == NULL || sym == NULL) {
926 buf[0] = '\0';
927 return 0;
928 }
929
930 offset = map__map_ip(map, ip) - sym->start;
931
932 if (offset)
933 return scnprintf(buf, size, "%s+%#lx", sym->name, offset);
934 else
935 return strlcpy(buf, sym->name, size);
936}
937static int lock_contention_caller(struct evsel *evsel, struct perf_sample *sample,
938 char *buf, int size)
939{
940 struct thread *thread;
941 struct callchain_cursor *cursor;
942 struct machine *machine = &session->machines.host;
943 struct symbol *sym;
944 int skip = 0;
945 int ret;
946
947 /* lock names will be replaced to task name later */
948 if (show_thread_stats)
949 return -1;
950
951 thread = machine__findnew_thread(machine, -1, sample->pid);
952 if (thread == NULL)
953 return -1;
954
955 cursor = get_tls_callchain_cursor();
956
957 /* use caller function name from the callchain */
958 ret = thread__resolve_callchain(thread, cursor, evsel, sample,
959 NULL, NULL, max_stack_depth);
960 if (ret != 0) {
961 thread__put(thread);
962 return -1;
963 }
964
965 callchain_cursor_commit(cursor);
966 thread__put(thread);
967
968 while (true) {
969 struct callchain_cursor_node *node;
970
971 node = callchain_cursor_current(cursor);
972 if (node == NULL)
973 break;
974
975 /* skip first few entries - for lock functions */
976 if (++skip <= stack_skip)
977 goto next;
978
979 sym = node->ms.sym;
980 if (sym && !machine__is_lock_function(machine, node->ip)) {
981 get_symbol_name_offset(node->ms.map, sym, node->ip,
982 buf, size);
983 return 0;
984 }
985
986next:
987 callchain_cursor_advance(cursor);
988 }
989 return -1;
990}
991
992static u64 callchain_id(struct evsel *evsel, struct perf_sample *sample)
993{
994 struct callchain_cursor *cursor;
995 struct machine *machine = &session->machines.host;
996 struct thread *thread;
997 u64 hash = 0;
998 int skip = 0;
999 int ret;
1000
1001 thread = machine__findnew_thread(machine, -1, sample->pid);
1002 if (thread == NULL)
1003 return -1;
1004
1005 cursor = get_tls_callchain_cursor();
1006 /* use caller function name from the callchain */
1007 ret = thread__resolve_callchain(thread, cursor, evsel, sample,
1008 NULL, NULL, max_stack_depth);
1009 thread__put(thread);
1010
1011 if (ret != 0)
1012 return -1;
1013
1014 callchain_cursor_commit(cursor);
1015
1016 while (true) {
1017 struct callchain_cursor_node *node;
1018
1019 node = callchain_cursor_current(cursor);
1020 if (node == NULL)
1021 break;
1022
1023 /* skip first few entries - for lock functions */
1024 if (++skip <= stack_skip)
1025 goto next;
1026
1027 if (node->ms.sym && machine__is_lock_function(machine, node->ip))
1028 goto next;
1029
1030 hash ^= hash_long((unsigned long)node->ip, 64);
1031
1032next:
1033 callchain_cursor_advance(cursor);
1034 }
1035 return hash;
1036}
1037
1038static u64 *get_callstack(struct perf_sample *sample, int max_stack)
1039{
1040 u64 *callstack;
1041 u64 i;
1042 int c;
1043
1044 callstack = calloc(max_stack, sizeof(*callstack));
1045 if (callstack == NULL)
1046 return NULL;
1047
1048 for (i = 0, c = 0; i < sample->callchain->nr && c < max_stack; i++) {
1049 u64 ip = sample->callchain->ips[i];
1050
1051 if (ip >= PERF_CONTEXT_MAX)
1052 continue;
1053
1054 callstack[c++] = ip;
1055 }
1056 return callstack;
1057}
1058
1059static int report_lock_contention_begin_event(struct evsel *evsel,
1060 struct perf_sample *sample)
1061{
1062 struct lock_stat *ls;
1063 struct thread_stat *ts;
1064 struct lock_seq_stat *seq;
1065 u64 addr = evsel__intval(evsel, sample, "lock_addr");
1066 unsigned int flags = evsel__intval(evsel, sample, "flags");
1067 u64 key;
1068 int i, ret;
1069 static bool kmap_loaded;
1070 struct machine *machine = &session->machines.host;
1071 struct map *kmap;
1072 struct symbol *sym;
1073
1074 ret = get_key_by_aggr_mode(&key, addr, evsel, sample);
1075 if (ret < 0)
1076 return ret;
1077
1078 if (!kmap_loaded) {
1079 unsigned long *addrs;
1080
1081 /* make sure it loads the kernel map to find lock symbols */
1082 map__load(machine__kernel_map(machine));
1083 kmap_loaded = true;
1084
1085 /* convert (kernel) symbols to addresses */
1086 for (i = 0; i < filters.nr_syms; i++) {
1087 sym = machine__find_kernel_symbol_by_name(machine,
1088 filters.syms[i],
1089 &kmap);
1090 if (sym == NULL) {
1091 pr_warning("ignore unknown symbol: %s\n",
1092 filters.syms[i]);
1093 continue;
1094 }
1095
1096 addrs = realloc(filters.addrs,
1097 (filters.nr_addrs + 1) * sizeof(*addrs));
1098 if (addrs == NULL) {
1099 pr_warning("memory allocation failure\n");
1100 return -ENOMEM;
1101 }
1102
1103 addrs[filters.nr_addrs++] = map__unmap_ip(kmap, sym->start);
1104 filters.addrs = addrs;
1105 }
1106 }
1107
1108 ls = lock_stat_find(key);
1109 if (!ls) {
1110 char buf[128];
1111 const char *name = "";
1112
1113 switch (aggr_mode) {
1114 case LOCK_AGGR_ADDR:
1115 sym = machine__find_kernel_symbol(machine, key, &kmap);
1116 if (sym)
1117 name = sym->name;
1118 break;
1119 case LOCK_AGGR_CALLER:
1120 name = buf;
1121 if (lock_contention_caller(evsel, sample, buf, sizeof(buf)) < 0)
1122 name = "Unknown";
1123 break;
1124 case LOCK_AGGR_CGROUP:
1125 case LOCK_AGGR_TASK:
1126 default:
1127 break;
1128 }
1129
1130 ls = lock_stat_findnew(key, name, flags);
1131 if (!ls)
1132 return -ENOMEM;
1133 }
1134
1135 if (filters.nr_types) {
1136 bool found = false;
1137
1138 for (i = 0; i < filters.nr_types; i++) {
1139 if (flags == filters.types[i]) {
1140 found = true;
1141 break;
1142 }
1143 }
1144
1145 if (!found)
1146 return 0;
1147 }
1148
1149 if (filters.nr_addrs) {
1150 bool found = false;
1151
1152 for (i = 0; i < filters.nr_addrs; i++) {
1153 if (addr == filters.addrs[i]) {
1154 found = true;
1155 break;
1156 }
1157 }
1158
1159 if (!found)
1160 return 0;
1161 }
1162
1163 if (needs_callstack()) {
1164 u64 *callstack = get_callstack(sample, max_stack_depth);
1165 if (callstack == NULL)
1166 return -ENOMEM;
1167
1168 if (!match_callstack_filter(machine, callstack)) {
1169 free(callstack);
1170 return 0;
1171 }
1172
1173 if (ls->callstack == NULL)
1174 ls->callstack = callstack;
1175 else
1176 free(callstack);
1177 }
1178
1179 ts = thread_stat_findnew(sample->tid);
1180 if (!ts)
1181 return -ENOMEM;
1182
1183 seq = get_seq(ts, addr);
1184 if (!seq)
1185 return -ENOMEM;
1186
1187 switch (seq->state) {
1188 case SEQ_STATE_UNINITIALIZED:
1189 case SEQ_STATE_ACQUIRED:
1190 break;
1191 case SEQ_STATE_CONTENDED:
1192 /*
1193 * It can have nested contention begin with mutex spinning,
1194 * then we would use the original contention begin event and
1195 * ignore the second one.
1196 */
1197 goto end;
1198 case SEQ_STATE_ACQUIRING:
1199 case SEQ_STATE_READ_ACQUIRED:
1200 case SEQ_STATE_RELEASED:
1201 /* broken lock sequence */
1202 if (!ls->broken) {
1203 ls->broken = 1;
1204 bad_hist[BROKEN_CONTENDED]++;
1205 }
1206 list_del_init(&seq->list);
1207 free(seq);
1208 goto end;
1209 default:
1210 BUG_ON("Unknown state of lock sequence found!\n");
1211 break;
1212 }
1213
1214 if (seq->state != SEQ_STATE_CONTENDED) {
1215 seq->state = SEQ_STATE_CONTENDED;
1216 seq->prev_event_time = sample->time;
1217 ls->nr_contended++;
1218 }
1219end:
1220 return 0;
1221}
1222
1223static int report_lock_contention_end_event(struct evsel *evsel,
1224 struct perf_sample *sample)
1225{
1226 struct lock_stat *ls;
1227 struct thread_stat *ts;
1228 struct lock_seq_stat *seq;
1229 u64 contended_term;
1230 u64 addr = evsel__intval(evsel, sample, "lock_addr");
1231 u64 key;
1232 int ret;
1233
1234 ret = get_key_by_aggr_mode(&key, addr, evsel, sample);
1235 if (ret < 0)
1236 return ret;
1237
1238 ls = lock_stat_find(key);
1239 if (!ls)
1240 return 0;
1241
1242 ts = thread_stat_find(sample->tid);
1243 if (!ts)
1244 return 0;
1245
1246 seq = get_seq(ts, addr);
1247 if (!seq)
1248 return -ENOMEM;
1249
1250 switch (seq->state) {
1251 case SEQ_STATE_UNINITIALIZED:
1252 goto end;
1253 case SEQ_STATE_CONTENDED:
1254 contended_term = sample->time - seq->prev_event_time;
1255 ls->wait_time_total += contended_term;
1256 if (contended_term < ls->wait_time_min)
1257 ls->wait_time_min = contended_term;
1258 if (ls->wait_time_max < contended_term)
1259 ls->wait_time_max = contended_term;
1260 break;
1261 case SEQ_STATE_ACQUIRING:
1262 case SEQ_STATE_ACQUIRED:
1263 case SEQ_STATE_READ_ACQUIRED:
1264 case SEQ_STATE_RELEASED:
1265 /* broken lock sequence */
1266 if (!ls->broken) {
1267 ls->broken = 1;
1268 bad_hist[BROKEN_ACQUIRED]++;
1269 }
1270 list_del_init(&seq->list);
1271 free(seq);
1272 goto end;
1273 default:
1274 BUG_ON("Unknown state of lock sequence found!\n");
1275 break;
1276 }
1277
1278 seq->state = SEQ_STATE_ACQUIRED;
1279 ls->nr_acquired++;
1280 ls->avg_wait_time = ls->wait_time_total/ls->nr_acquired;
1281end:
1282 return 0;
1283}
1284
1285/* lock oriented handlers */
1286/* TODO: handlers for CPU oriented, thread oriented */
1287static struct trace_lock_handler report_lock_ops = {
1288 .acquire_event = report_lock_acquire_event,
1289 .acquired_event = report_lock_acquired_event,
1290 .contended_event = report_lock_contended_event,
1291 .release_event = report_lock_release_event,
1292 .contention_begin_event = report_lock_contention_begin_event,
1293 .contention_end_event = report_lock_contention_end_event,
1294};
1295
1296static struct trace_lock_handler contention_lock_ops = {
1297 .contention_begin_event = report_lock_contention_begin_event,
1298 .contention_end_event = report_lock_contention_end_event,
1299};
1300
1301
1302static struct trace_lock_handler *trace_handler;
1303
1304static int evsel__process_lock_acquire(struct evsel *evsel, struct perf_sample *sample)
1305{
1306 if (trace_handler->acquire_event)
1307 return trace_handler->acquire_event(evsel, sample);
1308 return 0;
1309}
1310
1311static int evsel__process_lock_acquired(struct evsel *evsel, struct perf_sample *sample)
1312{
1313 if (trace_handler->acquired_event)
1314 return trace_handler->acquired_event(evsel, sample);
1315 return 0;
1316}
1317
1318static int evsel__process_lock_contended(struct evsel *evsel, struct perf_sample *sample)
1319{
1320 if (trace_handler->contended_event)
1321 return trace_handler->contended_event(evsel, sample);
1322 return 0;
1323}
1324
1325static int evsel__process_lock_release(struct evsel *evsel, struct perf_sample *sample)
1326{
1327 if (trace_handler->release_event)
1328 return trace_handler->release_event(evsel, sample);
1329 return 0;
1330}
1331
1332static int evsel__process_contention_begin(struct evsel *evsel, struct perf_sample *sample)
1333{
1334 if (trace_handler->contention_begin_event)
1335 return trace_handler->contention_begin_event(evsel, sample);
1336 return 0;
1337}
1338
1339static int evsel__process_contention_end(struct evsel *evsel, struct perf_sample *sample)
1340{
1341 if (trace_handler->contention_end_event)
1342 return trace_handler->contention_end_event(evsel, sample);
1343 return 0;
1344}
1345
1346static void print_bad_events(int bad, int total)
1347{
1348 /* Output for debug, this have to be removed */
1349 int i;
1350 int broken = 0;
1351 const char *name[4] =
1352 { "acquire", "acquired", "contended", "release" };
1353
1354 for (i = 0; i < BROKEN_MAX; i++)
1355 broken += bad_hist[i];
1356
1357 if (quiet || total == 0 || (broken == 0 && verbose <= 0))
1358 return;
1359
1360 fprintf(lock_output, "\n=== output for debug ===\n\n");
1361 fprintf(lock_output, "bad: %d, total: %d\n", bad, total);
1362 fprintf(lock_output, "bad rate: %.2f %%\n", (double)bad / (double)total * 100);
1363 fprintf(lock_output, "histogram of events caused bad sequence\n");
1364 for (i = 0; i < BROKEN_MAX; i++)
1365 fprintf(lock_output, " %10s: %d\n", name[i], bad_hist[i]);
1366}
1367
1368/* TODO: various way to print, coloring, nano or milli sec */
1369static void print_result(void)
1370{
1371 struct lock_stat *st;
1372 struct lock_key *key;
1373 char cut_name[20];
1374 int bad, total, printed;
1375
1376 if (!quiet) {
1377 fprintf(lock_output, "%20s ", "Name");
1378 list_for_each_entry(key, &lock_keys, list)
1379 fprintf(lock_output, "%*s ", key->len, key->header);
1380 fprintf(lock_output, "\n\n");
1381 }
1382
1383 bad = total = printed = 0;
1384 while ((st = pop_from_result())) {
1385 total++;
1386 if (st->broken)
1387 bad++;
1388 if (!st->nr_acquired)
1389 continue;
1390
1391 bzero(cut_name, 20);
1392
1393 if (strlen(st->name) < 20) {
1394 /* output raw name */
1395 const char *name = st->name;
1396
1397 if (show_thread_stats) {
1398 struct thread *t;
1399
1400 /* st->addr contains tid of thread */
1401 t = perf_session__findnew(session, st->addr);
1402 name = thread__comm_str(t);
1403 }
1404
1405 fprintf(lock_output, "%20s ", name);
1406 } else {
1407 strncpy(cut_name, st->name, 16);
1408 cut_name[16] = '.';
1409 cut_name[17] = '.';
1410 cut_name[18] = '.';
1411 cut_name[19] = '\0';
1412 /* cut off name for saving output style */
1413 fprintf(lock_output, "%20s ", cut_name);
1414 }
1415
1416 list_for_each_entry(key, &lock_keys, list) {
1417 key->print(key, st);
1418 fprintf(lock_output, " ");
1419 }
1420 fprintf(lock_output, "\n");
1421
1422 if (++printed >= print_nr_entries)
1423 break;
1424 }
1425
1426 print_bad_events(bad, total);
1427}
1428
1429static bool info_threads, info_map;
1430
1431static void dump_threads(void)
1432{
1433 struct thread_stat *st;
1434 struct rb_node *node;
1435 struct thread *t;
1436
1437 fprintf(lock_output, "%10s: comm\n", "Thread ID");
1438
1439 node = rb_first(&thread_stats);
1440 while (node) {
1441 st = container_of(node, struct thread_stat, rb);
1442 t = perf_session__findnew(session, st->tid);
1443 fprintf(lock_output, "%10d: %s\n", st->tid, thread__comm_str(t));
1444 node = rb_next(node);
1445 thread__put(t);
1446 }
1447}
1448
1449static int compare_maps(struct lock_stat *a, struct lock_stat *b)
1450{
1451 int ret;
1452
1453 if (a->name && b->name)
1454 ret = strcmp(a->name, b->name);
1455 else
1456 ret = !!a->name - !!b->name;
1457
1458 if (!ret)
1459 return a->addr < b->addr;
1460 else
1461 return ret < 0;
1462}
1463
1464static void dump_map(void)
1465{
1466 unsigned int i;
1467 struct lock_stat *st;
1468
1469 fprintf(lock_output, "Address of instance: name of class\n");
1470 for (i = 0; i < LOCKHASH_SIZE; i++) {
1471 hlist_for_each_entry(st, &lockhash_table[i], hash_entry) {
1472 insert_to_result(st, compare_maps);
1473 }
1474 }
1475
1476 while ((st = pop_from_result()))
1477 fprintf(lock_output, " %#llx: %s\n", (unsigned long long)st->addr, st->name);
1478}
1479
1480static void dump_info(void)
1481{
1482 if (info_threads)
1483 dump_threads();
1484
1485 if (info_map) {
1486 if (info_threads)
1487 fputc('\n', lock_output);
1488 dump_map();
1489 }
1490}
1491
1492static const struct evsel_str_handler lock_tracepoints[] = {
1493 { "lock:lock_acquire", evsel__process_lock_acquire, }, /* CONFIG_LOCKDEP */
1494 { "lock:lock_acquired", evsel__process_lock_acquired, }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */
1495 { "lock:lock_contended", evsel__process_lock_contended, }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */
1496 { "lock:lock_release", evsel__process_lock_release, }, /* CONFIG_LOCKDEP */
1497};
1498
1499static const struct evsel_str_handler contention_tracepoints[] = {
1500 { "lock:contention_begin", evsel__process_contention_begin, },
1501 { "lock:contention_end", evsel__process_contention_end, },
1502};
1503
1504static int process_event_update(const struct perf_tool *tool,
1505 union perf_event *event,
1506 struct evlist **pevlist)
1507{
1508 int ret;
1509
1510 ret = perf_event__process_event_update(tool, event, pevlist);
1511 if (ret < 0)
1512 return ret;
1513
1514 /* this can return -EEXIST since we call it for each evsel */
1515 perf_session__set_tracepoints_handlers(session, lock_tracepoints);
1516 perf_session__set_tracepoints_handlers(session, contention_tracepoints);
1517 return 0;
1518}
1519
1520typedef int (*tracepoint_handler)(struct evsel *evsel,
1521 struct perf_sample *sample);
1522
1523static int process_sample_event(const struct perf_tool *tool __maybe_unused,
1524 union perf_event *event,
1525 struct perf_sample *sample,
1526 struct evsel *evsel,
1527 struct machine *machine)
1528{
1529 int err = 0;
1530 struct thread *thread = machine__findnew_thread(machine, sample->pid,
1531 sample->tid);
1532
1533 if (thread == NULL) {
1534 pr_debug("problem processing %d event, skipping it.\n",
1535 event->header.type);
1536 return -1;
1537 }
1538
1539 if (evsel->handler != NULL) {
1540 tracepoint_handler f = evsel->handler;
1541 err = f(evsel, sample);
1542 }
1543
1544 thread__put(thread);
1545
1546 return err;
1547}
1548
1549static void combine_result(void)
1550{
1551 unsigned int i;
1552 struct lock_stat *st;
1553
1554 if (!combine_locks)
1555 return;
1556
1557 for (i = 0; i < LOCKHASH_SIZE; i++) {
1558 hlist_for_each_entry(st, &lockhash_table[i], hash_entry) {
1559 combine_lock_stats(st);
1560 }
1561 }
1562}
1563
1564static void sort_result(void)
1565{
1566 unsigned int i;
1567 struct lock_stat *st;
1568
1569 for (i = 0; i < LOCKHASH_SIZE; i++) {
1570 hlist_for_each_entry(st, &lockhash_table[i], hash_entry) {
1571 insert_to_result(st, compare);
1572 }
1573 }
1574}
1575
1576static const struct {
1577 unsigned int flags;
1578 const char *str;
1579 const char *name;
1580} lock_type_table[] = {
1581 { 0, "semaphore", "semaphore" },
1582 { LCB_F_SPIN, "spinlock", "spinlock" },
1583 { LCB_F_SPIN | LCB_F_READ, "rwlock:R", "rwlock" },
1584 { LCB_F_SPIN | LCB_F_WRITE, "rwlock:W", "rwlock" },
1585 { LCB_F_READ, "rwsem:R", "rwsem" },
1586 { LCB_F_WRITE, "rwsem:W", "rwsem" },
1587 { LCB_F_RT, "rt-mutex", "rt-mutex" },
1588 { LCB_F_RT | LCB_F_READ, "rwlock-rt:R", "rwlock-rt" },
1589 { LCB_F_RT | LCB_F_WRITE, "rwlock-rt:W", "rwlock-rt" },
1590 { LCB_F_PERCPU | LCB_F_READ, "pcpu-sem:R", "percpu-rwsem" },
1591 { LCB_F_PERCPU | LCB_F_WRITE, "pcpu-sem:W", "percpu-rwsem" },
1592 { LCB_F_MUTEX, "mutex", "mutex" },
1593 { LCB_F_MUTEX | LCB_F_SPIN, "mutex", "mutex" },
1594 /* alias for optimistic spinning only */
1595 { LCB_F_MUTEX | LCB_F_SPIN, "mutex:spin", "mutex-spin" },
1596};
1597
1598static const char *get_type_str(unsigned int flags)
1599{
1600 flags &= LCB_F_MAX_FLAGS - 1;
1601
1602 for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) {
1603 if (lock_type_table[i].flags == flags)
1604 return lock_type_table[i].str;
1605 }
1606 return "unknown";
1607}
1608
1609static const char *get_type_name(unsigned int flags)
1610{
1611 flags &= LCB_F_MAX_FLAGS - 1;
1612
1613 for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) {
1614 if (lock_type_table[i].flags == flags)
1615 return lock_type_table[i].name;
1616 }
1617 return "unknown";
1618}
1619
1620static void lock_filter_finish(void)
1621{
1622 zfree(&filters.types);
1623 filters.nr_types = 0;
1624
1625 zfree(&filters.addrs);
1626 filters.nr_addrs = 0;
1627
1628 for (int i = 0; i < filters.nr_syms; i++)
1629 free(filters.syms[i]);
1630
1631 zfree(&filters.syms);
1632 filters.nr_syms = 0;
1633
1634 zfree(&filters.cgrps);
1635 filters.nr_cgrps = 0;
1636}
1637
1638static void sort_contention_result(void)
1639{
1640 sort_result();
1641}
1642
1643static void print_header_stdio(void)
1644{
1645 struct lock_key *key;
1646
1647 list_for_each_entry(key, &lock_keys, list)
1648 fprintf(lock_output, "%*s ", key->len, key->header);
1649
1650 switch (aggr_mode) {
1651 case LOCK_AGGR_TASK:
1652 fprintf(lock_output, " %10s %s\n\n", "pid",
1653 show_lock_owner ? "owner" : "comm");
1654 break;
1655 case LOCK_AGGR_CALLER:
1656 fprintf(lock_output, " %10s %s\n\n", "type", "caller");
1657 break;
1658 case LOCK_AGGR_ADDR:
1659 fprintf(lock_output, " %16s %s\n\n", "address", "symbol");
1660 break;
1661 case LOCK_AGGR_CGROUP:
1662 fprintf(lock_output, " %s\n\n", "cgroup");
1663 break;
1664 default:
1665 break;
1666 }
1667}
1668
1669static void print_header_csv(const char *sep)
1670{
1671 struct lock_key *key;
1672
1673 fprintf(lock_output, "# output: ");
1674 list_for_each_entry(key, &lock_keys, list)
1675 fprintf(lock_output, "%s%s ", key->header, sep);
1676
1677 switch (aggr_mode) {
1678 case LOCK_AGGR_TASK:
1679 fprintf(lock_output, "%s%s %s\n", "pid", sep,
1680 show_lock_owner ? "owner" : "comm");
1681 break;
1682 case LOCK_AGGR_CALLER:
1683 fprintf(lock_output, "%s%s %s", "type", sep, "caller");
1684 if (verbose > 0)
1685 fprintf(lock_output, "%s %s", sep, "stacktrace");
1686 fprintf(lock_output, "\n");
1687 break;
1688 case LOCK_AGGR_ADDR:
1689 fprintf(lock_output, "%s%s %s%s %s\n", "address", sep, "symbol", sep, "type");
1690 break;
1691 case LOCK_AGGR_CGROUP:
1692 fprintf(lock_output, "%s\n", "cgroup");
1693 break;
1694 default:
1695 break;
1696 }
1697}
1698
1699static void print_header(void)
1700{
1701 if (!quiet) {
1702 if (symbol_conf.field_sep)
1703 print_header_csv(symbol_conf.field_sep);
1704 else
1705 print_header_stdio();
1706 }
1707}
1708
1709static void print_lock_stat_stdio(struct lock_contention *con, struct lock_stat *st)
1710{
1711 struct lock_key *key;
1712 struct thread *t;
1713 int pid;
1714
1715 list_for_each_entry(key, &lock_keys, list) {
1716 key->print(key, st);
1717 fprintf(lock_output, " ");
1718 }
1719
1720 switch (aggr_mode) {
1721 case LOCK_AGGR_CALLER:
1722 fprintf(lock_output, " %10s %s\n", get_type_str(st->flags), st->name);
1723 break;
1724 case LOCK_AGGR_TASK:
1725 pid = st->addr;
1726 t = perf_session__findnew(session, pid);
1727 fprintf(lock_output, " %10d %s\n",
1728 pid, pid == -1 ? "Unknown" : thread__comm_str(t));
1729 break;
1730 case LOCK_AGGR_ADDR:
1731 fprintf(lock_output, " %016llx %s (%s)\n", (unsigned long long)st->addr,
1732 st->name, get_type_name(st->flags));
1733 break;
1734 case LOCK_AGGR_CGROUP:
1735 fprintf(lock_output, " %s\n", st->name);
1736 break;
1737 default:
1738 break;
1739 }
1740
1741 if (aggr_mode == LOCK_AGGR_CALLER && verbose > 0) {
1742 struct map *kmap;
1743 struct symbol *sym;
1744 char buf[128];
1745 u64 ip;
1746
1747 for (int i = 0; i < max_stack_depth; i++) {
1748 if (!st->callstack || !st->callstack[i])
1749 break;
1750
1751 ip = st->callstack[i];
1752 sym = machine__find_kernel_symbol(con->machine, ip, &kmap);
1753 get_symbol_name_offset(kmap, sym, ip, buf, sizeof(buf));
1754 fprintf(lock_output, "\t\t\t%#lx %s\n", (unsigned long)ip, buf);
1755 }
1756 }
1757}
1758
1759static void print_lock_stat_csv(struct lock_contention *con, struct lock_stat *st,
1760 const char *sep)
1761{
1762 struct lock_key *key;
1763 struct thread *t;
1764 int pid;
1765
1766 list_for_each_entry(key, &lock_keys, list) {
1767 key->print(key, st);
1768 fprintf(lock_output, "%s ", sep);
1769 }
1770
1771 switch (aggr_mode) {
1772 case LOCK_AGGR_CALLER:
1773 fprintf(lock_output, "%s%s %s", get_type_str(st->flags), sep, st->name);
1774 if (verbose <= 0)
1775 fprintf(lock_output, "\n");
1776 break;
1777 case LOCK_AGGR_TASK:
1778 pid = st->addr;
1779 t = perf_session__findnew(session, pid);
1780 fprintf(lock_output, "%d%s %s\n", pid, sep,
1781 pid == -1 ? "Unknown" : thread__comm_str(t));
1782 break;
1783 case LOCK_AGGR_ADDR:
1784 fprintf(lock_output, "%llx%s %s%s %s\n", (unsigned long long)st->addr, sep,
1785 st->name, sep, get_type_name(st->flags));
1786 break;
1787 case LOCK_AGGR_CGROUP:
1788 fprintf(lock_output, "%s\n",st->name);
1789 break;
1790 default:
1791 break;
1792 }
1793
1794 if (aggr_mode == LOCK_AGGR_CALLER && verbose > 0) {
1795 struct map *kmap;
1796 struct symbol *sym;
1797 char buf[128];
1798 u64 ip;
1799
1800 for (int i = 0; i < max_stack_depth; i++) {
1801 if (!st->callstack || !st->callstack[i])
1802 break;
1803
1804 ip = st->callstack[i];
1805 sym = machine__find_kernel_symbol(con->machine, ip, &kmap);
1806 get_symbol_name_offset(kmap, sym, ip, buf, sizeof(buf));
1807 fprintf(lock_output, "%s %#lx %s", i ? ":" : sep, (unsigned long) ip, buf);
1808 }
1809 fprintf(lock_output, "\n");
1810 }
1811}
1812
1813static void print_lock_stat(struct lock_contention *con, struct lock_stat *st)
1814{
1815 if (symbol_conf.field_sep)
1816 print_lock_stat_csv(con, st, symbol_conf.field_sep);
1817 else
1818 print_lock_stat_stdio(con, st);
1819}
1820
1821static void print_footer_stdio(int total, int bad, struct lock_contention_fails *fails)
1822{
1823 /* Output for debug, this have to be removed */
1824 int broken = fails->task + fails->stack + fails->time + fails->data;
1825
1826 if (!use_bpf)
1827 print_bad_events(bad, total);
1828
1829 if (quiet || total == 0 || (broken == 0 && verbose <= 0))
1830 return;
1831
1832 total += broken;
1833 fprintf(lock_output, "\n=== output for debug ===\n\n");
1834 fprintf(lock_output, "bad: %d, total: %d\n", broken, total);
1835 fprintf(lock_output, "bad rate: %.2f %%\n", 100.0 * broken / total);
1836
1837 fprintf(lock_output, "histogram of failure reasons\n");
1838 fprintf(lock_output, " %10s: %d\n", "task", fails->task);
1839 fprintf(lock_output, " %10s: %d\n", "stack", fails->stack);
1840 fprintf(lock_output, " %10s: %d\n", "time", fails->time);
1841 fprintf(lock_output, " %10s: %d\n", "data", fails->data);
1842}
1843
1844static void print_footer_csv(int total, int bad, struct lock_contention_fails *fails,
1845 const char *sep)
1846{
1847 /* Output for debug, this have to be removed */
1848 if (use_bpf)
1849 bad = fails->task + fails->stack + fails->time + fails->data;
1850
1851 if (quiet || total == 0 || (bad == 0 && verbose <= 0))
1852 return;
1853
1854 total += bad;
1855 fprintf(lock_output, "# debug: total=%d%s bad=%d", total, sep, bad);
1856
1857 if (use_bpf) {
1858 fprintf(lock_output, "%s bad_%s=%d", sep, "task", fails->task);
1859 fprintf(lock_output, "%s bad_%s=%d", sep, "stack", fails->stack);
1860 fprintf(lock_output, "%s bad_%s=%d", sep, "time", fails->time);
1861 fprintf(lock_output, "%s bad_%s=%d", sep, "data", fails->data);
1862 } else {
1863 int i;
1864 const char *name[4] = { "acquire", "acquired", "contended", "release" };
1865
1866 for (i = 0; i < BROKEN_MAX; i++)
1867 fprintf(lock_output, "%s bad_%s=%d", sep, name[i], bad_hist[i]);
1868 }
1869 fprintf(lock_output, "\n");
1870}
1871
1872static void print_footer(int total, int bad, struct lock_contention_fails *fails)
1873{
1874 if (symbol_conf.field_sep)
1875 print_footer_csv(total, bad, fails, symbol_conf.field_sep);
1876 else
1877 print_footer_stdio(total, bad, fails);
1878}
1879
1880static void print_contention_result(struct lock_contention *con)
1881{
1882 struct lock_stat *st;
1883 int bad, total, printed;
1884
1885 if (!quiet)
1886 print_header();
1887
1888 bad = total = printed = 0;
1889
1890 while ((st = pop_from_result())) {
1891 total += use_bpf ? st->nr_contended : 1;
1892 if (st->broken)
1893 bad++;
1894
1895 if (!st->wait_time_total)
1896 continue;
1897
1898 print_lock_stat(con, st);
1899
1900 if (++printed >= print_nr_entries)
1901 break;
1902 }
1903
1904 if (print_nr_entries) {
1905 /* update the total/bad stats */
1906 while ((st = pop_from_result())) {
1907 total += use_bpf ? st->nr_contended : 1;
1908 if (st->broken)
1909 bad++;
1910 }
1911 }
1912 /* some entries are collected but hidden by the callstack filter */
1913 total += con->nr_filtered;
1914
1915 print_footer(total, bad, &con->fails);
1916}
1917
1918static bool force;
1919
1920static int __cmd_report(bool display_info)
1921{
1922 int err = -EINVAL;
1923 struct perf_tool eops;
1924 struct perf_data data = {
1925 .path = input_name,
1926 .mode = PERF_DATA_MODE_READ,
1927 .force = force,
1928 };
1929
1930 perf_tool__init(&eops, /*ordered_events=*/true);
1931 eops.attr = perf_event__process_attr;
1932 eops.event_update = process_event_update;
1933 eops.sample = process_sample_event;
1934 eops.comm = perf_event__process_comm;
1935 eops.mmap = perf_event__process_mmap;
1936 eops.namespaces = perf_event__process_namespaces;
1937 eops.tracing_data = perf_event__process_tracing_data;
1938 session = perf_session__new(&data, &eops);
1939 if (IS_ERR(session)) {
1940 pr_err("Initializing perf session failed\n");
1941 return PTR_ERR(session);
1942 }
1943
1944 symbol_conf.allow_aliases = true;
1945 symbol__init(&session->header.env);
1946
1947 if (!data.is_pipe) {
1948 if (!perf_session__has_traces(session, "lock record"))
1949 goto out_delete;
1950
1951 if (perf_session__set_tracepoints_handlers(session, lock_tracepoints)) {
1952 pr_err("Initializing perf session tracepoint handlers failed\n");
1953 goto out_delete;
1954 }
1955
1956 if (perf_session__set_tracepoints_handlers(session, contention_tracepoints)) {
1957 pr_err("Initializing perf session tracepoint handlers failed\n");
1958 goto out_delete;
1959 }
1960 }
1961
1962 if (setup_output_field(false, output_fields))
1963 goto out_delete;
1964
1965 if (select_key(false))
1966 goto out_delete;
1967
1968 if (show_thread_stats)
1969 aggr_mode = LOCK_AGGR_TASK;
1970
1971 err = perf_session__process_events(session);
1972 if (err)
1973 goto out_delete;
1974
1975 setup_pager();
1976 if (display_info) /* used for info subcommand */
1977 dump_info();
1978 else {
1979 combine_result();
1980 sort_result();
1981 print_result();
1982 }
1983
1984out_delete:
1985 perf_session__delete(session);
1986 return err;
1987}
1988
1989static void sighandler(int sig __maybe_unused)
1990{
1991}
1992
1993static int check_lock_contention_options(const struct option *options,
1994 const char * const *usage)
1995
1996{
1997 if (show_thread_stats && show_lock_addrs) {
1998 pr_err("Cannot use thread and addr mode together\n");
1999 parse_options_usage(usage, options, "threads", 0);
2000 parse_options_usage(NULL, options, "lock-addr", 0);
2001 return -1;
2002 }
2003
2004 if (show_lock_owner && !use_bpf) {
2005 pr_err("Lock owners are available only with BPF\n");
2006 parse_options_usage(usage, options, "lock-owner", 0);
2007 parse_options_usage(NULL, options, "use-bpf", 0);
2008 return -1;
2009 }
2010
2011 if (show_lock_owner && show_lock_addrs) {
2012 pr_err("Cannot use owner and addr mode together\n");
2013 parse_options_usage(usage, options, "lock-owner", 0);
2014 parse_options_usage(NULL, options, "lock-addr", 0);
2015 return -1;
2016 }
2017
2018 if (show_lock_cgroups && !use_bpf) {
2019 pr_err("Cgroups are available only with BPF\n");
2020 parse_options_usage(usage, options, "lock-cgroup", 0);
2021 parse_options_usage(NULL, options, "use-bpf", 0);
2022 return -1;
2023 }
2024
2025 if (show_lock_cgroups && show_lock_addrs) {
2026 pr_err("Cannot use cgroup and addr mode together\n");
2027 parse_options_usage(usage, options, "lock-cgroup", 0);
2028 parse_options_usage(NULL, options, "lock-addr", 0);
2029 return -1;
2030 }
2031
2032 if (show_lock_cgroups && show_thread_stats) {
2033 pr_err("Cannot use cgroup and thread mode together\n");
2034 parse_options_usage(usage, options, "lock-cgroup", 0);
2035 parse_options_usage(NULL, options, "threads", 0);
2036 return -1;
2037 }
2038
2039 if (symbol_conf.field_sep) {
2040 if (strstr(symbol_conf.field_sep, ":") || /* part of type flags */
2041 strstr(symbol_conf.field_sep, "+") || /* part of caller offset */
2042 strstr(symbol_conf.field_sep, ".")) { /* can be in a symbol name */
2043 pr_err("Cannot use the separator that is already used\n");
2044 parse_options_usage(usage, options, "x", 1);
2045 return -1;
2046 }
2047 }
2048
2049 if (show_lock_owner)
2050 show_thread_stats = true;
2051
2052 return 0;
2053}
2054
2055static int __cmd_contention(int argc, const char **argv)
2056{
2057 int err = -EINVAL;
2058 struct perf_tool eops;
2059 struct perf_data data = {
2060 .path = input_name,
2061 .mode = PERF_DATA_MODE_READ,
2062 .force = force,
2063 };
2064 struct lock_contention con = {
2065 .target = &target,
2066 .map_nr_entries = bpf_map_entries,
2067 .max_stack = max_stack_depth,
2068 .stack_skip = stack_skip,
2069 .filters = &filters,
2070 .save_callstack = needs_callstack(),
2071 .owner = show_lock_owner,
2072 .cgroups = RB_ROOT,
2073 };
2074
2075 lockhash_table = calloc(LOCKHASH_SIZE, sizeof(*lockhash_table));
2076 if (!lockhash_table)
2077 return -ENOMEM;
2078
2079 con.result = &lockhash_table[0];
2080
2081 perf_tool__init(&eops, /*ordered_events=*/true);
2082 eops.attr = perf_event__process_attr;
2083 eops.event_update = process_event_update;
2084 eops.sample = process_sample_event;
2085 eops.comm = perf_event__process_comm;
2086 eops.mmap = perf_event__process_mmap;
2087 eops.tracing_data = perf_event__process_tracing_data;
2088
2089 session = perf_session__new(use_bpf ? NULL : &data, &eops);
2090 if (IS_ERR(session)) {
2091 pr_err("Initializing perf session failed\n");
2092 err = PTR_ERR(session);
2093 session = NULL;
2094 goto out_delete;
2095 }
2096
2097 con.machine = &session->machines.host;
2098
2099 con.aggr_mode = aggr_mode = show_thread_stats ? LOCK_AGGR_TASK :
2100 show_lock_addrs ? LOCK_AGGR_ADDR :
2101 show_lock_cgroups ? LOCK_AGGR_CGROUP : LOCK_AGGR_CALLER;
2102
2103 if (con.aggr_mode == LOCK_AGGR_CALLER)
2104 con.save_callstack = true;
2105
2106 symbol_conf.allow_aliases = true;
2107 symbol__init(&session->header.env);
2108
2109 if (use_bpf) {
2110 err = target__validate(&target);
2111 if (err) {
2112 char errbuf[512];
2113
2114 target__strerror(&target, err, errbuf, 512);
2115 pr_err("%s\n", errbuf);
2116 goto out_delete;
2117 }
2118
2119 signal(SIGINT, sighandler);
2120 signal(SIGCHLD, sighandler);
2121 signal(SIGTERM, sighandler);
2122
2123 con.evlist = evlist__new();
2124 if (con.evlist == NULL) {
2125 err = -ENOMEM;
2126 goto out_delete;
2127 }
2128
2129 err = evlist__create_maps(con.evlist, &target);
2130 if (err < 0)
2131 goto out_delete;
2132
2133 if (argc) {
2134 err = evlist__prepare_workload(con.evlist, &target,
2135 argv, false, NULL);
2136 if (err < 0)
2137 goto out_delete;
2138 }
2139
2140 if (lock_contention_prepare(&con) < 0) {
2141 pr_err("lock contention BPF setup failed\n");
2142 goto out_delete;
2143 }
2144 } else if (!data.is_pipe) {
2145 if (!perf_session__has_traces(session, "lock record"))
2146 goto out_delete;
2147
2148 if (!evlist__find_evsel_by_str(session->evlist,
2149 "lock:contention_begin")) {
2150 pr_err("lock contention evsel not found\n");
2151 goto out_delete;
2152 }
2153
2154 if (perf_session__set_tracepoints_handlers(session,
2155 contention_tracepoints)) {
2156 pr_err("Initializing perf session tracepoint handlers failed\n");
2157 goto out_delete;
2158 }
2159 }
2160
2161 if (setup_output_field(true, output_fields))
2162 goto out_delete;
2163
2164 if (select_key(true))
2165 goto out_delete;
2166
2167 if (symbol_conf.field_sep) {
2168 int i;
2169 struct lock_key *keys = contention_keys;
2170
2171 /* do not align output in CSV format */
2172 for (i = 0; keys[i].name; i++)
2173 keys[i].len = 0;
2174 }
2175
2176 if (use_bpf) {
2177 lock_contention_start();
2178 if (argc)
2179 evlist__start_workload(con.evlist);
2180
2181 /* wait for signal */
2182 pause();
2183
2184 lock_contention_stop();
2185 lock_contention_read(&con);
2186 } else {
2187 err = perf_session__process_events(session);
2188 if (err)
2189 goto out_delete;
2190 }
2191
2192 setup_pager();
2193
2194 sort_contention_result();
2195 print_contention_result(&con);
2196
2197out_delete:
2198 lock_filter_finish();
2199 evlist__delete(con.evlist);
2200 lock_contention_finish(&con);
2201 perf_session__delete(session);
2202 zfree(&lockhash_table);
2203 return err;
2204}
2205
2206
2207static int __cmd_record(int argc, const char **argv)
2208{
2209 const char *record_args[] = {
2210 "record", "-R", "-m", "1024", "-c", "1", "--synth", "task",
2211 };
2212 const char *callgraph_args[] = {
2213 "--call-graph", "fp," __stringify(CONTENTION_STACK_DEPTH),
2214 };
2215 unsigned int rec_argc, i, j, ret;
2216 unsigned int nr_tracepoints;
2217 unsigned int nr_callgraph_args = 0;
2218 const char **rec_argv;
2219 bool has_lock_stat = true;
2220
2221 for (i = 0; i < ARRAY_SIZE(lock_tracepoints); i++) {
2222 if (!is_valid_tracepoint(lock_tracepoints[i].name)) {
2223 pr_debug("tracepoint %s is not enabled. "
2224 "Are CONFIG_LOCKDEP and CONFIG_LOCK_STAT enabled?\n",
2225 lock_tracepoints[i].name);
2226 has_lock_stat = false;
2227 break;
2228 }
2229 }
2230
2231 if (has_lock_stat)
2232 goto setup_args;
2233
2234 for (i = 0; i < ARRAY_SIZE(contention_tracepoints); i++) {
2235 if (!is_valid_tracepoint(contention_tracepoints[i].name)) {
2236 pr_err("tracepoint %s is not enabled.\n",
2237 contention_tracepoints[i].name);
2238 return 1;
2239 }
2240 }
2241
2242 nr_callgraph_args = ARRAY_SIZE(callgraph_args);
2243
2244setup_args:
2245 rec_argc = ARRAY_SIZE(record_args) + nr_callgraph_args + argc - 1;
2246
2247 if (has_lock_stat)
2248 nr_tracepoints = ARRAY_SIZE(lock_tracepoints);
2249 else
2250 nr_tracepoints = ARRAY_SIZE(contention_tracepoints);
2251
2252 /* factor of 2 is for -e in front of each tracepoint */
2253 rec_argc += 2 * nr_tracepoints;
2254
2255 rec_argv = calloc(rec_argc + 1, sizeof(char *));
2256 if (!rec_argv)
2257 return -ENOMEM;
2258
2259 for (i = 0; i < ARRAY_SIZE(record_args); i++)
2260 rec_argv[i] = record_args[i];
2261
2262 for (j = 0; j < nr_tracepoints; j++) {
2263 rec_argv[i++] = "-e";
2264 rec_argv[i++] = has_lock_stat
2265 ? lock_tracepoints[j].name
2266 : contention_tracepoints[j].name;
2267 }
2268
2269 for (j = 0; j < nr_callgraph_args; j++, i++)
2270 rec_argv[i] = callgraph_args[j];
2271
2272 for (j = 1; j < (unsigned int)argc; j++, i++)
2273 rec_argv[i] = argv[j];
2274
2275 BUG_ON(i != rec_argc);
2276
2277 ret = cmd_record(i, rec_argv);
2278 free(rec_argv);
2279 return ret;
2280}
2281
2282static int parse_map_entry(const struct option *opt, const char *str,
2283 int unset __maybe_unused)
2284{
2285 unsigned long *len = (unsigned long *)opt->value;
2286 unsigned long val;
2287 char *endptr;
2288
2289 errno = 0;
2290 val = strtoul(str, &endptr, 0);
2291 if (*endptr != '\0' || errno != 0) {
2292 pr_err("invalid BPF map length: %s\n", str);
2293 return -1;
2294 }
2295
2296 *len = val;
2297 return 0;
2298}
2299
2300static int parse_max_stack(const struct option *opt, const char *str,
2301 int unset __maybe_unused)
2302{
2303 unsigned long *len = (unsigned long *)opt->value;
2304 long val;
2305 char *endptr;
2306
2307 errno = 0;
2308 val = strtol(str, &endptr, 0);
2309 if (*endptr != '\0' || errno != 0) {
2310 pr_err("invalid max stack depth: %s\n", str);
2311 return -1;
2312 }
2313
2314 if (val < 0 || val > sysctl__max_stack()) {
2315 pr_err("invalid max stack depth: %ld\n", val);
2316 return -1;
2317 }
2318
2319 *len = val;
2320 return 0;
2321}
2322
2323static bool add_lock_type(unsigned int flags)
2324{
2325 unsigned int *tmp;
2326
2327 tmp = realloc(filters.types, (filters.nr_types + 1) * sizeof(*filters.types));
2328 if (tmp == NULL)
2329 return false;
2330
2331 tmp[filters.nr_types++] = flags;
2332 filters.types = tmp;
2333 return true;
2334}
2335
2336static int parse_lock_type(const struct option *opt __maybe_unused, const char *str,
2337 int unset __maybe_unused)
2338{
2339 char *s, *tmp, *tok;
2340
2341 s = strdup(str);
2342 if (s == NULL)
2343 return -1;
2344
2345 for (tok = strtok_r(s, ", ", &tmp); tok; tok = strtok_r(NULL, ", ", &tmp)) {
2346 bool found = false;
2347
2348 /* `tok` is `str` in `lock_type_table` if it contains ':'. */
2349 if (strchr(tok, ':')) {
2350 for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) {
2351 if (!strcmp(lock_type_table[i].str, tok) &&
2352 add_lock_type(lock_type_table[i].flags)) {
2353 found = true;
2354 break;
2355 }
2356 }
2357
2358 if (!found) {
2359 pr_err("Unknown lock flags name: %s\n", tok);
2360 free(s);
2361 return -1;
2362 }
2363
2364 continue;
2365 }
2366
2367 /*
2368 * Otherwise `tok` is `name` in `lock_type_table`.
2369 * Single lock name could contain multiple flags.
2370 */
2371 for (unsigned int i = 0; i < ARRAY_SIZE(lock_type_table); i++) {
2372 if (!strcmp(lock_type_table[i].name, tok)) {
2373 if (add_lock_type(lock_type_table[i].flags)) {
2374 found = true;
2375 } else {
2376 free(s);
2377 return -1;
2378 }
2379 }
2380 }
2381
2382 if (!found) {
2383 pr_err("Unknown lock name: %s\n", tok);
2384 free(s);
2385 return -1;
2386 }
2387
2388 }
2389
2390 free(s);
2391 return 0;
2392}
2393
2394static bool add_lock_addr(unsigned long addr)
2395{
2396 unsigned long *tmp;
2397
2398 tmp = realloc(filters.addrs, (filters.nr_addrs + 1) * sizeof(*filters.addrs));
2399 if (tmp == NULL) {
2400 pr_err("Memory allocation failure\n");
2401 return false;
2402 }
2403
2404 tmp[filters.nr_addrs++] = addr;
2405 filters.addrs = tmp;
2406 return true;
2407}
2408
2409static bool add_lock_sym(char *name)
2410{
2411 char **tmp;
2412 char *sym = strdup(name);
2413
2414 if (sym == NULL) {
2415 pr_err("Memory allocation failure\n");
2416 return false;
2417 }
2418
2419 tmp = realloc(filters.syms, (filters.nr_syms + 1) * sizeof(*filters.syms));
2420 if (tmp == NULL) {
2421 pr_err("Memory allocation failure\n");
2422 free(sym);
2423 return false;
2424 }
2425
2426 tmp[filters.nr_syms++] = sym;
2427 filters.syms = tmp;
2428 return true;
2429}
2430
2431static int parse_lock_addr(const struct option *opt __maybe_unused, const char *str,
2432 int unset __maybe_unused)
2433{
2434 char *s, *tmp, *tok;
2435 int ret = 0;
2436 u64 addr;
2437
2438 s = strdup(str);
2439 if (s == NULL)
2440 return -1;
2441
2442 for (tok = strtok_r(s, ", ", &tmp); tok; tok = strtok_r(NULL, ", ", &tmp)) {
2443 char *end;
2444
2445 addr = strtoul(tok, &end, 16);
2446 if (*end == '\0') {
2447 if (!add_lock_addr(addr)) {
2448 ret = -1;
2449 break;
2450 }
2451 continue;
2452 }
2453
2454 /*
2455 * At this moment, we don't have kernel symbols. Save the symbols
2456 * in a separate list and resolve them to addresses later.
2457 */
2458 if (!add_lock_sym(tok)) {
2459 ret = -1;
2460 break;
2461 }
2462 }
2463
2464 free(s);
2465 return ret;
2466}
2467
2468static int parse_call_stack(const struct option *opt __maybe_unused, const char *str,
2469 int unset __maybe_unused)
2470{
2471 char *s, *tmp, *tok;
2472 int ret = 0;
2473
2474 s = strdup(str);
2475 if (s == NULL)
2476 return -1;
2477
2478 for (tok = strtok_r(s, ", ", &tmp); tok; tok = strtok_r(NULL, ", ", &tmp)) {
2479 struct callstack_filter *entry;
2480
2481 entry = malloc(sizeof(*entry) + strlen(tok) + 1);
2482 if (entry == NULL) {
2483 pr_err("Memory allocation failure\n");
2484 free(s);
2485 return -1;
2486 }
2487
2488 strcpy(entry->name, tok);
2489 list_add_tail(&entry->list, &callstack_filters);
2490 }
2491
2492 free(s);
2493 return ret;
2494}
2495
2496static int parse_output(const struct option *opt __maybe_unused, const char *str,
2497 int unset __maybe_unused)
2498{
2499 const char **name = (const char **)opt->value;
2500
2501 if (str == NULL)
2502 return -1;
2503
2504 lock_output = fopen(str, "w");
2505 if (lock_output == NULL) {
2506 pr_err("Cannot open %s\n", str);
2507 return -1;
2508 }
2509
2510 *name = str;
2511 return 0;
2512}
2513
2514static bool add_lock_cgroup(char *name)
2515{
2516 u64 *tmp;
2517 struct cgroup *cgrp;
2518
2519 cgrp = cgroup__new(name, /*do_open=*/false);
2520 if (cgrp == NULL) {
2521 pr_err("Failed to create cgroup: %s\n", name);
2522 return false;
2523 }
2524
2525 if (read_cgroup_id(cgrp) < 0) {
2526 pr_err("Failed to read cgroup id for %s\n", name);
2527 cgroup__put(cgrp);
2528 return false;
2529 }
2530
2531 tmp = realloc(filters.cgrps, (filters.nr_cgrps + 1) * sizeof(*filters.cgrps));
2532 if (tmp == NULL) {
2533 pr_err("Memory allocation failure\n");
2534 return false;
2535 }
2536
2537 tmp[filters.nr_cgrps++] = cgrp->id;
2538 filters.cgrps = tmp;
2539 cgroup__put(cgrp);
2540 return true;
2541}
2542
2543static int parse_cgroup_filter(const struct option *opt __maybe_unused, const char *str,
2544 int unset __maybe_unused)
2545{
2546 char *s, *tmp, *tok;
2547 int ret = 0;
2548
2549 s = strdup(str);
2550 if (s == NULL)
2551 return -1;
2552
2553 for (tok = strtok_r(s, ", ", &tmp); tok; tok = strtok_r(NULL, ", ", &tmp)) {
2554 if (!add_lock_cgroup(tok)) {
2555 ret = -1;
2556 break;
2557 }
2558 }
2559
2560 free(s);
2561 return ret;
2562}
2563
2564int cmd_lock(int argc, const char **argv)
2565{
2566 const struct option lock_options[] = {
2567 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2568 OPT_CALLBACK(0, "output", &output_name, "file", "output file name", parse_output),
2569 OPT_INCR('v', "verbose", &verbose, "be more verbose (show symbol address, etc)"),
2570 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, "dump raw trace in ASCII"),
2571 OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
2572 OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
2573 "file", "vmlinux pathname"),
2574 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
2575 "file", "kallsyms pathname"),
2576 OPT_BOOLEAN('q', "quiet", &quiet, "Do not show any warnings or messages"),
2577 OPT_END()
2578 };
2579
2580 const struct option info_options[] = {
2581 OPT_BOOLEAN('t', "threads", &info_threads,
2582 "dump the thread list in perf.data"),
2583 OPT_BOOLEAN('m', "map", &info_map,
2584 "dump the map of lock instances (address:name table)"),
2585 OPT_PARENT(lock_options)
2586 };
2587
2588 const struct option report_options[] = {
2589 OPT_STRING('k', "key", &sort_key, "acquired",
2590 "key for sorting (acquired / contended / avg_wait / wait_total / wait_max / wait_min)"),
2591 OPT_STRING('F', "field", &output_fields, NULL,
2592 "output fields (acquired / contended / avg_wait / wait_total / wait_max / wait_min)"),
2593 /* TODO: type */
2594 OPT_BOOLEAN('c', "combine-locks", &combine_locks,
2595 "combine locks in the same class"),
2596 OPT_BOOLEAN('t', "threads", &show_thread_stats,
2597 "show per-thread lock stats"),
2598 OPT_INTEGER('E', "entries", &print_nr_entries, "display this many functions"),
2599 OPT_PARENT(lock_options)
2600 };
2601
2602 struct option contention_options[] = {
2603 OPT_STRING('k', "key", &sort_key, "wait_total",
2604 "key for sorting (contended / wait_total / wait_max / wait_min / avg_wait)"),
2605 OPT_STRING('F', "field", &output_fields, "contended,wait_total,wait_max,avg_wait",
2606 "output fields (contended / wait_total / wait_max / wait_min / avg_wait)"),
2607 OPT_BOOLEAN('t', "threads", &show_thread_stats,
2608 "show per-thread lock stats"),
2609 OPT_BOOLEAN('b', "use-bpf", &use_bpf, "use BPF program to collect lock contention stats"),
2610 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
2611 "System-wide collection from all CPUs"),
2612 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
2613 "List of cpus to monitor"),
2614 OPT_STRING('p', "pid", &target.pid, "pid",
2615 "Trace on existing process id"),
2616 OPT_STRING(0, "tid", &target.tid, "tid",
2617 "Trace on existing thread id (exclusive to --pid)"),
2618 OPT_CALLBACK('M', "map-nr-entries", &bpf_map_entries, "num",
2619 "Max number of BPF map entries", parse_map_entry),
2620 OPT_CALLBACK(0, "max-stack", &max_stack_depth, "num",
2621 "Set the maximum stack depth when collecting lock contention, "
2622 "Default: " __stringify(CONTENTION_STACK_DEPTH), parse_max_stack),
2623 OPT_INTEGER(0, "stack-skip", &stack_skip,
2624 "Set the number of stack depth to skip when finding a lock caller, "
2625 "Default: " __stringify(CONTENTION_STACK_SKIP)),
2626 OPT_INTEGER('E', "entries", &print_nr_entries, "display this many functions"),
2627 OPT_BOOLEAN('l', "lock-addr", &show_lock_addrs, "show lock stats by address"),
2628 OPT_CALLBACK('Y', "type-filter", NULL, "FLAGS",
2629 "Filter specific type of locks", parse_lock_type),
2630 OPT_CALLBACK('L', "lock-filter", NULL, "ADDRS/NAMES",
2631 "Filter specific address/symbol of locks", parse_lock_addr),
2632 OPT_CALLBACK('S', "callstack-filter", NULL, "NAMES",
2633 "Filter specific function in the callstack", parse_call_stack),
2634 OPT_BOOLEAN('o', "lock-owner", &show_lock_owner, "show lock owners instead of waiters"),
2635 OPT_STRING_NOEMPTY('x', "field-separator", &symbol_conf.field_sep, "separator",
2636 "print result in CSV format with custom separator"),
2637 OPT_BOOLEAN(0, "lock-cgroup", &show_lock_cgroups, "show lock stats by cgroup"),
2638 OPT_CALLBACK('G', "cgroup-filter", NULL, "CGROUPS",
2639 "Filter specific cgroups", parse_cgroup_filter),
2640 OPT_PARENT(lock_options)
2641 };
2642
2643 const char * const info_usage[] = {
2644 "perf lock info [<options>]",
2645 NULL
2646 };
2647 const char *const lock_subcommands[] = { "record", "report", "script",
2648 "info", "contention", NULL };
2649 const char *lock_usage[] = {
2650 NULL,
2651 NULL
2652 };
2653 const char * const report_usage[] = {
2654 "perf lock report [<options>]",
2655 NULL
2656 };
2657 const char * const contention_usage[] = {
2658 "perf lock contention [<options>]",
2659 NULL
2660 };
2661 unsigned int i;
2662 int rc = 0;
2663
2664 lockhash_table = calloc(LOCKHASH_SIZE, sizeof(*lockhash_table));
2665 if (!lockhash_table)
2666 return -ENOMEM;
2667
2668 for (i = 0; i < LOCKHASH_SIZE; i++)
2669 INIT_HLIST_HEAD(lockhash_table + i);
2670
2671 lock_output = stderr;
2672 argc = parse_options_subcommand(argc, argv, lock_options, lock_subcommands,
2673 lock_usage, PARSE_OPT_STOP_AT_NON_OPTION);
2674 if (!argc)
2675 usage_with_options(lock_usage, lock_options);
2676
2677 if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2678 return __cmd_record(argc, argv);
2679 } else if (strlen(argv[0]) > 2 && strstarts("report", argv[0])) {
2680 trace_handler = &report_lock_ops;
2681 if (argc) {
2682 argc = parse_options(argc, argv,
2683 report_options, report_usage, 0);
2684 if (argc)
2685 usage_with_options(report_usage, report_options);
2686 }
2687 rc = __cmd_report(false);
2688 } else if (!strcmp(argv[0], "script")) {
2689 /* Aliased to 'perf script' */
2690 rc = cmd_script(argc, argv);
2691 } else if (!strcmp(argv[0], "info")) {
2692 if (argc) {
2693 argc = parse_options(argc, argv,
2694 info_options, info_usage, 0);
2695 if (argc)
2696 usage_with_options(info_usage, info_options);
2697 }
2698
2699 /* If neither threads nor map requested, display both */
2700 if (!info_threads && !info_map) {
2701 info_threads = true;
2702 info_map = true;
2703 }
2704
2705 /* recycling report_lock_ops */
2706 trace_handler = &report_lock_ops;
2707 rc = __cmd_report(true);
2708 } else if (strlen(argv[0]) > 2 && strstarts("contention", argv[0])) {
2709 trace_handler = &contention_lock_ops;
2710 sort_key = "wait_total";
2711 output_fields = "contended,wait_total,wait_max,avg_wait";
2712
2713#ifndef HAVE_BPF_SKEL
2714 set_option_nobuild(contention_options, 'b', "use-bpf",
2715 "no BUILD_BPF_SKEL=1", false);
2716#endif
2717 if (argc) {
2718 argc = parse_options(argc, argv, contention_options,
2719 contention_usage, 0);
2720 }
2721
2722 if (check_lock_contention_options(contention_options,
2723 contention_usage) < 0)
2724 return -1;
2725
2726 rc = __cmd_contention(argc, argv);
2727 } else {
2728 usage_with_options(lock_usage, lock_options);
2729 }
2730
2731 /* free usage string allocated by parse_options_subcommand */
2732 free((void *)lock_usage[0]);
2733
2734 zfree(&lockhash_table);
2735 return rc;
2736}