2 * builtin-timechart.c - make an svg timechart of system activity
4 * (C) Copyright 2009 Intel Corporation
7 * Arjan van de Ven <arjan@linux.intel.com>
9 * This program is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU General Public License
11 * as published by the Free Software Foundation; version 2
17 #include "util/util.h"
19 #include "util/color.h"
20 #include <linux/list.h>
21 #include "util/cache.h"
22 #include <linux/rbtree.h>
23 #include "util/symbol.h"
24 #include "util/string.h"
25 #include "util/callchain.h"
26 #include "util/strlist.h"
29 #include "util/header.h"
30 #include "util/parse-options.h"
31 #include "util/parse-events.h"
32 #include "util/event.h"
33 #include "util/session.h"
34 #include "util/svghelper.h"
36 static char const *input_name
= "perf.data";
37 static char const *output_name
= "output.svg";
40 static u64 sample_type
;
42 static unsigned int numcpus
;
43 static u64 min_freq
; /* Lowest CPU frequency seen */
44 static u64 max_freq
; /* Highest CPU frequency seen */
45 static u64 turbo_frequency
;
47 static u64 first_time
, last_time
;
49 static int power_only
;
59 struct sample_wrapper
;
62 * Datastructure layout:
63 * We keep an list of "pid"s, matching the kernels notion of a task struct.
64 * Each "pid" entry, has a list of "comm"s.
65 * this is because we want to track different programs different, while
66 * exec will reuse the original pid (by design).
67 * Each comm has a list of samples that will be used to draw
82 struct per_pidcomm
*all
;
83 struct per_pidcomm
*current
;
90 struct per_pidcomm
*next
;
104 struct cpu_sample
*samples
;
107 struct sample_wrapper
{
108 struct sample_wrapper
*next
;
111 unsigned char data
[0];
115 #define TYPE_RUNNING 1
116 #define TYPE_WAITING 2
117 #define TYPE_BLOCKED 3
120 struct cpu_sample
*next
;
128 static struct per_pid
*all_data
;
134 struct power_event
*next
;
143 struct wake_event
*next
;
149 static struct power_event
*power_events
;
150 static struct wake_event
*wake_events
;
152 struct sample_wrapper
*all_samples
;
155 struct process_filter
;
156 struct process_filter
{
159 struct process_filter
*next
;
162 static struct process_filter
*process_filter
;
165 static struct per_pid
*find_create_pid(int pid
)
167 struct per_pid
*cursor
= all_data
;
170 if (cursor
->pid
== pid
)
172 cursor
= cursor
->next
;
174 cursor
= malloc(sizeof(struct per_pid
));
175 assert(cursor
!= NULL
);
176 memset(cursor
, 0, sizeof(struct per_pid
));
178 cursor
->next
= all_data
;
183 static void pid_set_comm(int pid
, char *comm
)
186 struct per_pidcomm
*c
;
187 p
= find_create_pid(pid
);
190 if (c
->comm
&& strcmp(c
->comm
, comm
) == 0) {
195 c
->comm
= strdup(comm
);
201 c
= malloc(sizeof(struct per_pidcomm
));
203 memset(c
, 0, sizeof(struct per_pidcomm
));
204 c
->comm
= strdup(comm
);
210 static void pid_fork(int pid
, int ppid
, u64 timestamp
)
212 struct per_pid
*p
, *pp
;
213 p
= find_create_pid(pid
);
214 pp
= find_create_pid(ppid
);
216 if (pp
->current
&& pp
->current
->comm
&& !p
->current
)
217 pid_set_comm(pid
, pp
->current
->comm
);
219 p
->start_time
= timestamp
;
221 p
->current
->start_time
= timestamp
;
222 p
->current
->state_since
= timestamp
;
226 static void pid_exit(int pid
, u64 timestamp
)
229 p
= find_create_pid(pid
);
230 p
->end_time
= timestamp
;
232 p
->current
->end_time
= timestamp
;
236 pid_put_sample(int pid
, int type
, unsigned int cpu
, u64 start
, u64 end
)
239 struct per_pidcomm
*c
;
240 struct cpu_sample
*sample
;
242 p
= find_create_pid(pid
);
245 c
= malloc(sizeof(struct per_pidcomm
));
247 memset(c
, 0, sizeof(struct per_pidcomm
));
253 sample
= malloc(sizeof(struct cpu_sample
));
254 assert(sample
!= NULL
);
255 memset(sample
, 0, sizeof(struct cpu_sample
));
256 sample
->start_time
= start
;
257 sample
->end_time
= end
;
259 sample
->next
= c
->samples
;
263 if (sample
->type
== TYPE_RUNNING
&& end
> start
&& start
> 0) {
264 c
->total_time
+= (end
-start
);
265 p
->total_time
+= (end
-start
);
268 if (c
->start_time
== 0 || c
->start_time
> start
)
269 c
->start_time
= start
;
270 if (p
->start_time
== 0 || p
->start_time
> start
)
271 p
->start_time
= start
;
277 #define MAX_CPUS 4096
279 static u64 cpus_cstate_start_times
[MAX_CPUS
];
280 static int cpus_cstate_state
[MAX_CPUS
];
281 static u64 cpus_pstate_start_times
[MAX_CPUS
];
282 static u64 cpus_pstate_state
[MAX_CPUS
];
284 static int process_comm_event(event_t
*event
, struct perf_session
*session __used
)
286 pid_set_comm(event
->comm
.pid
, event
->comm
.comm
);
290 static int process_fork_event(event_t
*event
, struct perf_session
*session __used
)
292 pid_fork(event
->fork
.pid
, event
->fork
.ppid
, event
->fork
.time
);
296 static int process_exit_event(event_t
*event
, struct perf_session
*session __used
)
298 pid_exit(event
->fork
.pid
, event
->fork
.time
);
305 unsigned char preempt_count
;
311 struct trace_entry te
;
316 #define TASK_COMM_LEN 16
317 struct wakeup_entry
{
318 struct trace_entry te
;
319 char comm
[TASK_COMM_LEN
];
326 * trace_flag_type is an enumeration that holds different
327 * states when a trace occurs. These are:
328 * IRQS_OFF - interrupts were disabled
329 * IRQS_NOSUPPORT - arch does not support irqs_disabled_flags
330 * NEED_RESCED - reschedule is requested
331 * HARDIRQ - inside an interrupt handler
332 * SOFTIRQ - inside a softirq handler
334 enum trace_flag_type
{
335 TRACE_FLAG_IRQS_OFF
= 0x01,
336 TRACE_FLAG_IRQS_NOSUPPORT
= 0x02,
337 TRACE_FLAG_NEED_RESCHED
= 0x04,
338 TRACE_FLAG_HARDIRQ
= 0x08,
339 TRACE_FLAG_SOFTIRQ
= 0x10,
344 struct sched_switch
{
345 struct trace_entry te
;
346 char prev_comm
[TASK_COMM_LEN
];
349 long prev_state
; /* Arjan weeps. */
350 char next_comm
[TASK_COMM_LEN
];
355 static void c_state_start(int cpu
, u64 timestamp
, int state
)
357 cpus_cstate_start_times
[cpu
] = timestamp
;
358 cpus_cstate_state
[cpu
] = state
;
361 static void c_state_end(int cpu
, u64 timestamp
)
363 struct power_event
*pwr
;
364 pwr
= malloc(sizeof(struct power_event
));
367 memset(pwr
, 0, sizeof(struct power_event
));
369 pwr
->state
= cpus_cstate_state
[cpu
];
370 pwr
->start_time
= cpus_cstate_start_times
[cpu
];
371 pwr
->end_time
= timestamp
;
374 pwr
->next
= power_events
;
379 static void p_state_change(int cpu
, u64 timestamp
, u64 new_freq
)
381 struct power_event
*pwr
;
382 pwr
= malloc(sizeof(struct power_event
));
384 if (new_freq
> 8000000) /* detect invalid data */
389 memset(pwr
, 0, sizeof(struct power_event
));
391 pwr
->state
= cpus_pstate_state
[cpu
];
392 pwr
->start_time
= cpus_pstate_start_times
[cpu
];
393 pwr
->end_time
= timestamp
;
396 pwr
->next
= power_events
;
398 if (!pwr
->start_time
)
399 pwr
->start_time
= first_time
;
403 cpus_pstate_state
[cpu
] = new_freq
;
404 cpus_pstate_start_times
[cpu
] = timestamp
;
406 if ((u64
)new_freq
> max_freq
)
409 if (new_freq
< min_freq
|| min_freq
== 0)
412 if (new_freq
== max_freq
- 1000)
413 turbo_frequency
= max_freq
;
417 sched_wakeup(int cpu
, u64 timestamp
, int pid
, struct trace_entry
*te
)
419 struct wake_event
*we
;
421 struct wakeup_entry
*wake
= (void *)te
;
423 we
= malloc(sizeof(struct wake_event
));
427 memset(we
, 0, sizeof(struct wake_event
));
428 we
->time
= timestamp
;
431 if ((te
->flags
& TRACE_FLAG_HARDIRQ
) || (te
->flags
& TRACE_FLAG_SOFTIRQ
))
434 we
->wakee
= wake
->pid
;
435 we
->next
= wake_events
;
437 p
= find_create_pid(we
->wakee
);
439 if (p
&& p
->current
&& p
->current
->state
== TYPE_NONE
) {
440 p
->current
->state_since
= timestamp
;
441 p
->current
->state
= TYPE_WAITING
;
443 if (p
&& p
->current
&& p
->current
->state
== TYPE_BLOCKED
) {
444 pid_put_sample(p
->pid
, p
->current
->state
, cpu
, p
->current
->state_since
, timestamp
);
445 p
->current
->state_since
= timestamp
;
446 p
->current
->state
= TYPE_WAITING
;
450 static void sched_switch(int cpu
, u64 timestamp
, struct trace_entry
*te
)
452 struct per_pid
*p
= NULL
, *prev_p
;
453 struct sched_switch
*sw
= (void *)te
;
456 prev_p
= find_create_pid(sw
->prev_pid
);
458 p
= find_create_pid(sw
->next_pid
);
460 if (prev_p
->current
&& prev_p
->current
->state
!= TYPE_NONE
)
461 pid_put_sample(sw
->prev_pid
, TYPE_RUNNING
, cpu
, prev_p
->current
->state_since
, timestamp
);
462 if (p
&& p
->current
) {
463 if (p
->current
->state
!= TYPE_NONE
)
464 pid_put_sample(sw
->next_pid
, p
->current
->state
, cpu
, p
->current
->state_since
, timestamp
);
466 p
->current
->state_since
= timestamp
;
467 p
->current
->state
= TYPE_RUNNING
;
470 if (prev_p
->current
) {
471 prev_p
->current
->state
= TYPE_NONE
;
472 prev_p
->current
->state_since
= timestamp
;
473 if (sw
->prev_state
& 2)
474 prev_p
->current
->state
= TYPE_BLOCKED
;
475 if (sw
->prev_state
== 0)
476 prev_p
->current
->state
= TYPE_WAITING
;
482 process_sample_event(event_t
*event
)
484 struct sample_data data
;
485 struct trace_entry
*te
;
487 memset(&data
, 0, sizeof(data
));
489 event__parse_sample(event
, sample_type
, &data
);
491 if (sample_type
& PERF_SAMPLE_TIME
) {
492 if (!first_time
|| first_time
> data
.time
)
493 first_time
= data
.time
;
494 if (last_time
< data
.time
)
495 last_time
= data
.time
;
498 te
= (void *)data
.raw_data
;
499 if (sample_type
& PERF_SAMPLE_RAW
&& data
.raw_size
> 0) {
501 struct power_entry
*pe
;
505 event_str
= perf_header__find_event(te
->type
);
510 if (strcmp(event_str
, "power:power_start") == 0)
511 c_state_start(data
.cpu
, data
.time
, pe
->value
);
513 if (strcmp(event_str
, "power:power_end") == 0)
514 c_state_end(data
.cpu
, data
.time
);
516 if (strcmp(event_str
, "power:power_frequency") == 0)
517 p_state_change(data
.cpu
, data
.time
, pe
->value
);
519 if (strcmp(event_str
, "sched:sched_wakeup") == 0)
520 sched_wakeup(data
.cpu
, data
.time
, data
.pid
, te
);
522 if (strcmp(event_str
, "sched:sched_switch") == 0)
523 sched_switch(data
.cpu
, data
.time
, te
);
529 * After the last sample we need to wrap up the current C/P state
530 * and close out each CPU for these.
532 static void end_sample_processing(void)
535 struct power_event
*pwr
;
537 for (cpu
= 0; cpu
<= numcpus
; cpu
++) {
538 pwr
= malloc(sizeof(struct power_event
));
541 memset(pwr
, 0, sizeof(struct power_event
));
545 pwr
->state
= cpus_cstate_state
[cpu
];
546 pwr
->start_time
= cpus_cstate_start_times
[cpu
];
547 pwr
->end_time
= last_time
;
550 pwr
->next
= power_events
;
556 pwr
= malloc(sizeof(struct power_event
));
559 memset(pwr
, 0, sizeof(struct power_event
));
561 pwr
->state
= cpus_pstate_state
[cpu
];
562 pwr
->start_time
= cpus_pstate_start_times
[cpu
];
563 pwr
->end_time
= last_time
;
566 pwr
->next
= power_events
;
568 if (!pwr
->start_time
)
569 pwr
->start_time
= first_time
;
571 pwr
->state
= min_freq
;
576 static u64
sample_time(event_t
*event
)
581 if (sample_type
& PERF_SAMPLE_IP
)
583 if (sample_type
& PERF_SAMPLE_TID
)
585 if (sample_type
& PERF_SAMPLE_TIME
)
586 return event
->sample
.array
[cursor
];
592 * We first queue all events, sorted backwards by insertion.
593 * The order will get flipped later.
595 static int queue_sample_event(event_t
*event
, struct perf_session
*session __used
)
597 struct sample_wrapper
*copy
, *prev
;
600 size
= event
->sample
.header
.size
+ sizeof(struct sample_wrapper
) + 8;
606 memset(copy
, 0, size
);
609 copy
->timestamp
= sample_time(event
);
611 memcpy(©
->data
, event
, event
->sample
.header
.size
);
613 /* insert in the right place in the list */
616 /* first sample ever */
621 if (all_samples
->timestamp
< copy
->timestamp
) {
622 /* insert at the head of the list */
623 copy
->next
= all_samples
;
630 if (prev
->next
->timestamp
< copy
->timestamp
) {
631 copy
->next
= prev
->next
;
637 /* insert at the end of the list */
643 static void sort_queued_samples(void)
645 struct sample_wrapper
*cursor
, *next
;
647 cursor
= all_samples
;
652 cursor
->next
= all_samples
;
653 all_samples
= cursor
;
659 * Sort the pid datastructure
661 static void sort_pids(void)
663 struct per_pid
*new_list
, *p
, *cursor
, *prev
;
664 /* sort by ppid first, then by pid, lowest to highest */
673 if (new_list
== NULL
) {
681 if (cursor
->ppid
> p
->ppid
||
682 (cursor
->ppid
== p
->ppid
&& cursor
->pid
> p
->pid
)) {
683 /* must insert before */
685 p
->next
= prev
->next
;
698 cursor
= cursor
->next
;
707 static void draw_c_p_states(void)
709 struct power_event
*pwr
;
713 * two pass drawing so that the P state bars are on top of the C state blocks
716 if (pwr
->type
== CSTATE
)
717 svg_cstate(pwr
->cpu
, pwr
->start_time
, pwr
->end_time
, pwr
->state
);
723 if (pwr
->type
== PSTATE
) {
725 pwr
->state
= min_freq
;
726 svg_pstate(pwr
->cpu
, pwr
->start_time
, pwr
->end_time
, pwr
->state
);
732 static void draw_wakeups(void)
734 struct wake_event
*we
;
736 struct per_pidcomm
*c
;
740 int from
= 0, to
= 0;
741 char *task_from
= NULL
, *task_to
= NULL
;
743 /* locate the column of the waker and wakee */
746 if (p
->pid
== we
->waker
|| p
->pid
== we
->wakee
) {
749 if (c
->Y
&& c
->start_time
<= we
->time
&& c
->end_time
>= we
->time
) {
750 if (p
->pid
== we
->waker
&& !from
) {
752 task_from
= strdup(c
->comm
);
754 if (p
->pid
== we
->wakee
&& !to
) {
756 task_to
= strdup(c
->comm
);
763 if (p
->pid
== we
->waker
&& !from
) {
765 task_from
= strdup(c
->comm
);
767 if (p
->pid
== we
->wakee
&& !to
) {
769 task_to
= strdup(c
->comm
);
778 task_from
= malloc(40);
779 sprintf(task_from
, "[%i]", we
->waker
);
782 task_to
= malloc(40);
783 sprintf(task_to
, "[%i]", we
->wakee
);
787 svg_interrupt(we
->time
, to
);
788 else if (from
&& to
&& abs(from
- to
) == 1)
789 svg_wakeline(we
->time
, from
, to
);
791 svg_partial_wakeline(we
->time
, from
, task_from
, to
, task_to
);
799 static void draw_cpu_usage(void)
802 struct per_pidcomm
*c
;
803 struct cpu_sample
*sample
;
810 if (sample
->type
== TYPE_RUNNING
)
811 svg_process(sample
->cpu
, sample
->start_time
, sample
->end_time
, "sample", c
->comm
);
813 sample
= sample
->next
;
821 static void draw_process_bars(void)
824 struct per_pidcomm
*c
;
825 struct cpu_sample
*sample
;
840 svg_box(Y
, c
->start_time
, c
->end_time
, "process");
843 if (sample
->type
== TYPE_RUNNING
)
844 svg_sample(Y
, sample
->cpu
, sample
->start_time
, sample
->end_time
);
845 if (sample
->type
== TYPE_BLOCKED
)
846 svg_box(Y
, sample
->start_time
, sample
->end_time
, "blocked");
847 if (sample
->type
== TYPE_WAITING
)
848 svg_waiting(Y
, sample
->start_time
, sample
->end_time
);
849 sample
= sample
->next
;
854 if (c
->total_time
> 5000000000) /* 5 seconds */
855 sprintf(comm
, "%s:%i (%2.2fs)", c
->comm
, p
->pid
, c
->total_time
/ 1000000000.0);
857 sprintf(comm
, "%s:%i (%3.1fms)", c
->comm
, p
->pid
, c
->total_time
/ 1000000.0);
859 svg_text(Y
, c
->start_time
, comm
);
869 static void add_process_filter(const char *string
)
871 struct process_filter
*filt
;
874 pid
= strtoull(string
, NULL
, 10);
875 filt
= malloc(sizeof(struct process_filter
));
879 filt
->name
= strdup(string
);
881 filt
->next
= process_filter
;
883 process_filter
= filt
;
886 static int passes_filter(struct per_pid
*p
, struct per_pidcomm
*c
)
888 struct process_filter
*filt
;
892 filt
= process_filter
;
894 if (filt
->pid
&& p
->pid
== filt
->pid
)
896 if (strcmp(filt
->name
, c
->comm
) == 0)
903 static int determine_display_tasks_filtered(void)
906 struct per_pidcomm
*c
;
912 if (p
->start_time
== 1)
913 p
->start_time
= first_time
;
915 /* no exit marker, task kept running to the end */
916 if (p
->end_time
== 0)
917 p
->end_time
= last_time
;
924 if (c
->start_time
== 1)
925 c
->start_time
= first_time
;
927 if (passes_filter(p
, c
)) {
933 if (c
->end_time
== 0)
934 c
->end_time
= last_time
;
943 static int determine_display_tasks(u64 threshold
)
946 struct per_pidcomm
*c
;
950 return determine_display_tasks_filtered();
955 if (p
->start_time
== 1)
956 p
->start_time
= first_time
;
958 /* no exit marker, task kept running to the end */
959 if (p
->end_time
== 0)
960 p
->end_time
= last_time
;
961 if (p
->total_time
>= threshold
&& !power_only
)
969 if (c
->start_time
== 1)
970 c
->start_time
= first_time
;
972 if (c
->total_time
>= threshold
&& !power_only
) {
977 if (c
->end_time
== 0)
978 c
->end_time
= last_time
;
989 #define TIME_THRESH 10000000
991 static void write_svg_file(const char *filename
)
999 count
= determine_display_tasks(TIME_THRESH
);
1001 /* We'd like to show at least 15 tasks; be less picky if we have fewer */
1003 count
= determine_display_tasks(TIME_THRESH
/ 10);
1005 open_svg(filename
, numcpus
, count
, first_time
, last_time
);
1010 for (i
= 0; i
< numcpus
; i
++)
1011 svg_cpu_box(i
, max_freq
, turbo_frequency
);
1014 draw_process_bars();
1021 static void process_samples(void)
1023 struct sample_wrapper
*cursor
;
1026 sort_queued_samples();
1028 cursor
= all_samples
;
1030 event
= (void *)&cursor
->data
;
1031 cursor
= cursor
->next
;
1032 process_sample_event(event
);
1036 static int sample_type_check(u64 type
)
1040 if (!(sample_type
& PERF_SAMPLE_RAW
)) {
1041 fprintf(stderr
, "No trace samples found in the file.\n"
1042 "Have you used 'perf timechart record' to record it?\n");
1049 static struct perf_event_ops event_ops
= {
1050 .process_comm_event
= process_comm_event
,
1051 .process_fork_event
= process_fork_event
,
1052 .process_exit_event
= process_exit_event
,
1053 .process_sample_event
= queue_sample_event
,
1054 .sample_type_check
= sample_type_check
,
1057 static int __cmd_timechart(void)
1059 struct perf_session
*session
= perf_session__new(input_name
, O_RDONLY
, 0);
1062 if (session
== NULL
)
1065 ret
= perf_session__process_events(session
, &event_ops
, 0,
1066 &event__cwdlen
, &event__cwd
);
1072 end_sample_processing();
1076 write_svg_file(output_name
);
1078 pr_info("Written %2.1f seconds of trace to %s.\n",
1079 (last_time
- first_time
) / 1000000000.0, output_name
);
1081 perf_session__delete(session
);
1085 static const char * const timechart_usage
[] = {
1086 "perf timechart [<options>] {record}",
1090 static const char *record_args
[] = {
1097 "-e", "power:power_start",
1098 "-e", "power:power_end",
1099 "-e", "power:power_frequency",
1100 "-e", "sched:sched_wakeup",
1101 "-e", "sched:sched_switch",
1104 static int __cmd_record(int argc
, const char **argv
)
1106 unsigned int rec_argc
, i
, j
;
1107 const char **rec_argv
;
1109 rec_argc
= ARRAY_SIZE(record_args
) + argc
- 1;
1110 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1112 for (i
= 0; i
< ARRAY_SIZE(record_args
); i
++)
1113 rec_argv
[i
] = strdup(record_args
[i
]);
1115 for (j
= 1; j
< (unsigned int)argc
; j
++, i
++)
1116 rec_argv
[i
] = argv
[j
];
1118 return cmd_record(i
, rec_argv
, NULL
);
1122 parse_process(const struct option
*opt __used
, const char *arg
, int __used unset
)
1125 add_process_filter(arg
);
1129 static const struct option options
[] = {
1130 OPT_STRING('i', "input", &input_name
, "file",
1132 OPT_STRING('o', "output", &output_name
, "file",
1133 "output file name"),
1134 OPT_INTEGER('w', "width", &svg_page_width
,
1136 OPT_BOOLEAN('P', "power-only", &power_only
,
1137 "output power data only"),
1138 OPT_CALLBACK('p', "process", NULL
, "process",
1139 "process selector. Pass a pid or process name.",
1145 int cmd_timechart(int argc
, const char **argv
, const char *prefix __used
)
1149 argc
= parse_options(argc
, argv
, options
, timechart_usage
,
1150 PARSE_OPT_STOP_AT_NON_OPTION
);
1152 if (argc
&& !strncmp(argv
[0], "rec", 3))
1153 return __cmd_record(argc
, argv
);
1155 usage_with_options(timechart_usage
, options
);
1159 return __cmd_timechart();