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/callchain.h"
25 #include "util/strlist.h"
28 #include "util/header.h"
29 #include "util/parse-options.h"
30 #include "util/parse-events.h"
31 #include "util/event.h"
32 #include "util/session.h"
33 #include "util/svghelper.h"
35 #define SUPPORT_OLD_POWER_EVENTS 1
36 #define PWR_EVENT_EXIT -1
39 static char const *input_name
= "perf.data";
40 static char const *output_name
= "output.svg";
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 bool 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
;
88 struct per_pidcomm
*next
;
102 struct cpu_sample
*samples
;
105 struct sample_wrapper
{
106 struct sample_wrapper
*next
;
109 unsigned char data
[0];
113 #define TYPE_RUNNING 1
114 #define TYPE_WAITING 2
115 #define TYPE_BLOCKED 3
118 struct cpu_sample
*next
;
126 static struct per_pid
*all_data
;
132 struct power_event
*next
;
141 struct wake_event
*next
;
147 static struct power_event
*power_events
;
148 static struct wake_event
*wake_events
;
150 struct process_filter
;
151 struct process_filter
{
154 struct process_filter
*next
;
157 static struct process_filter
*process_filter
;
160 static struct per_pid
*find_create_pid(int pid
)
162 struct per_pid
*cursor
= all_data
;
165 if (cursor
->pid
== pid
)
167 cursor
= cursor
->next
;
169 cursor
= malloc(sizeof(struct per_pid
));
170 assert(cursor
!= NULL
);
171 memset(cursor
, 0, sizeof(struct per_pid
));
173 cursor
->next
= all_data
;
178 static void pid_set_comm(int pid
, char *comm
)
181 struct per_pidcomm
*c
;
182 p
= find_create_pid(pid
);
185 if (c
->comm
&& strcmp(c
->comm
, comm
) == 0) {
190 c
->comm
= strdup(comm
);
196 c
= malloc(sizeof(struct per_pidcomm
));
198 memset(c
, 0, sizeof(struct per_pidcomm
));
199 c
->comm
= strdup(comm
);
205 static void pid_fork(int pid
, int ppid
, u64 timestamp
)
207 struct per_pid
*p
, *pp
;
208 p
= find_create_pid(pid
);
209 pp
= find_create_pid(ppid
);
211 if (pp
->current
&& pp
->current
->comm
&& !p
->current
)
212 pid_set_comm(pid
, pp
->current
->comm
);
214 p
->start_time
= timestamp
;
216 p
->current
->start_time
= timestamp
;
217 p
->current
->state_since
= timestamp
;
221 static void pid_exit(int pid
, u64 timestamp
)
224 p
= find_create_pid(pid
);
225 p
->end_time
= timestamp
;
227 p
->current
->end_time
= timestamp
;
231 pid_put_sample(int pid
, int type
, unsigned int cpu
, u64 start
, u64 end
)
234 struct per_pidcomm
*c
;
235 struct cpu_sample
*sample
;
237 p
= find_create_pid(pid
);
240 c
= malloc(sizeof(struct per_pidcomm
));
242 memset(c
, 0, sizeof(struct per_pidcomm
));
248 sample
= malloc(sizeof(struct cpu_sample
));
249 assert(sample
!= NULL
);
250 memset(sample
, 0, sizeof(struct cpu_sample
));
251 sample
->start_time
= start
;
252 sample
->end_time
= end
;
254 sample
->next
= c
->samples
;
258 if (sample
->type
== TYPE_RUNNING
&& end
> start
&& start
> 0) {
259 c
->total_time
+= (end
-start
);
260 p
->total_time
+= (end
-start
);
263 if (c
->start_time
== 0 || c
->start_time
> start
)
264 c
->start_time
= start
;
265 if (p
->start_time
== 0 || p
->start_time
> start
)
266 p
->start_time
= start
;
269 #define MAX_CPUS 4096
271 static u64 cpus_cstate_start_times
[MAX_CPUS
];
272 static int cpus_cstate_state
[MAX_CPUS
];
273 static u64 cpus_pstate_start_times
[MAX_CPUS
];
274 static u64 cpus_pstate_state
[MAX_CPUS
];
276 static int process_comm_event(union perf_event
*event
,
277 struct perf_sample
*sample __used
,
278 struct perf_session
*session __used
)
280 pid_set_comm(event
->comm
.tid
, event
->comm
.comm
);
284 static int process_fork_event(union perf_event
*event
,
285 struct perf_sample
*sample __used
,
286 struct perf_session
*session __used
)
288 pid_fork(event
->fork
.pid
, event
->fork
.ppid
, event
->fork
.time
);
292 static int process_exit_event(union perf_event
*event
,
293 struct perf_sample
*sample __used
,
294 struct perf_session
*session __used
)
296 pid_exit(event
->fork
.pid
, event
->fork
.time
);
303 unsigned char preempt_count
;
308 #ifdef SUPPORT_OLD_POWER_EVENTS
309 static int use_old_power_events
;
310 struct power_entry_old
{
311 struct trace_entry te
;
318 struct power_processor_entry
{
319 struct trace_entry te
;
324 #define TASK_COMM_LEN 16
325 struct wakeup_entry
{
326 struct trace_entry te
;
327 char comm
[TASK_COMM_LEN
];
334 * trace_flag_type is an enumeration that holds different
335 * states when a trace occurs. These are:
336 * IRQS_OFF - interrupts were disabled
337 * IRQS_NOSUPPORT - arch does not support irqs_disabled_flags
338 * NEED_RESCED - reschedule is requested
339 * HARDIRQ - inside an interrupt handler
340 * SOFTIRQ - inside a softirq handler
342 enum trace_flag_type
{
343 TRACE_FLAG_IRQS_OFF
= 0x01,
344 TRACE_FLAG_IRQS_NOSUPPORT
= 0x02,
345 TRACE_FLAG_NEED_RESCHED
= 0x04,
346 TRACE_FLAG_HARDIRQ
= 0x08,
347 TRACE_FLAG_SOFTIRQ
= 0x10,
352 struct sched_switch
{
353 struct trace_entry te
;
354 char prev_comm
[TASK_COMM_LEN
];
357 long prev_state
; /* Arjan weeps. */
358 char next_comm
[TASK_COMM_LEN
];
363 static void c_state_start(int cpu
, u64 timestamp
, int state
)
365 cpus_cstate_start_times
[cpu
] = timestamp
;
366 cpus_cstate_state
[cpu
] = state
;
369 static void c_state_end(int cpu
, u64 timestamp
)
371 struct power_event
*pwr
;
372 pwr
= malloc(sizeof(struct power_event
));
375 memset(pwr
, 0, sizeof(struct power_event
));
377 pwr
->state
= cpus_cstate_state
[cpu
];
378 pwr
->start_time
= cpus_cstate_start_times
[cpu
];
379 pwr
->end_time
= timestamp
;
382 pwr
->next
= power_events
;
387 static void p_state_change(int cpu
, u64 timestamp
, u64 new_freq
)
389 struct power_event
*pwr
;
390 pwr
= malloc(sizeof(struct power_event
));
392 if (new_freq
> 8000000) /* detect invalid data */
397 memset(pwr
, 0, sizeof(struct power_event
));
399 pwr
->state
= cpus_pstate_state
[cpu
];
400 pwr
->start_time
= cpus_pstate_start_times
[cpu
];
401 pwr
->end_time
= timestamp
;
404 pwr
->next
= power_events
;
406 if (!pwr
->start_time
)
407 pwr
->start_time
= first_time
;
411 cpus_pstate_state
[cpu
] = new_freq
;
412 cpus_pstate_start_times
[cpu
] = timestamp
;
414 if ((u64
)new_freq
> max_freq
)
417 if (new_freq
< min_freq
|| min_freq
== 0)
420 if (new_freq
== max_freq
- 1000)
421 turbo_frequency
= max_freq
;
425 sched_wakeup(int cpu
, u64 timestamp
, int pid
, struct trace_entry
*te
)
427 struct wake_event
*we
;
429 struct wakeup_entry
*wake
= (void *)te
;
431 we
= malloc(sizeof(struct wake_event
));
435 memset(we
, 0, sizeof(struct wake_event
));
436 we
->time
= timestamp
;
439 if ((te
->flags
& TRACE_FLAG_HARDIRQ
) || (te
->flags
& TRACE_FLAG_SOFTIRQ
))
442 we
->wakee
= wake
->pid
;
443 we
->next
= wake_events
;
445 p
= find_create_pid(we
->wakee
);
447 if (p
&& p
->current
&& p
->current
->state
== TYPE_NONE
) {
448 p
->current
->state_since
= timestamp
;
449 p
->current
->state
= TYPE_WAITING
;
451 if (p
&& p
->current
&& p
->current
->state
== TYPE_BLOCKED
) {
452 pid_put_sample(p
->pid
, p
->current
->state
, cpu
, p
->current
->state_since
, timestamp
);
453 p
->current
->state_since
= timestamp
;
454 p
->current
->state
= TYPE_WAITING
;
458 static void sched_switch(int cpu
, u64 timestamp
, struct trace_entry
*te
)
460 struct per_pid
*p
= NULL
, *prev_p
;
461 struct sched_switch
*sw
= (void *)te
;
464 prev_p
= find_create_pid(sw
->prev_pid
);
466 p
= find_create_pid(sw
->next_pid
);
468 if (prev_p
->current
&& prev_p
->current
->state
!= TYPE_NONE
)
469 pid_put_sample(sw
->prev_pid
, TYPE_RUNNING
, cpu
, prev_p
->current
->state_since
, timestamp
);
470 if (p
&& p
->current
) {
471 if (p
->current
->state
!= TYPE_NONE
)
472 pid_put_sample(sw
->next_pid
, p
->current
->state
, cpu
, p
->current
->state_since
, timestamp
);
474 p
->current
->state_since
= timestamp
;
475 p
->current
->state
= TYPE_RUNNING
;
478 if (prev_p
->current
) {
479 prev_p
->current
->state
= TYPE_NONE
;
480 prev_p
->current
->state_since
= timestamp
;
481 if (sw
->prev_state
& 2)
482 prev_p
->current
->state
= TYPE_BLOCKED
;
483 if (sw
->prev_state
== 0)
484 prev_p
->current
->state
= TYPE_WAITING
;
489 static int process_sample_event(union perf_event
*event __used
,
490 struct perf_sample
*sample
,
491 struct perf_evsel
*evsel __used
,
492 struct perf_session
*session
)
494 struct trace_entry
*te
;
496 if (session
->sample_type
& PERF_SAMPLE_TIME
) {
497 if (!first_time
|| first_time
> sample
->time
)
498 first_time
= sample
->time
;
499 if (last_time
< sample
->time
)
500 last_time
= sample
->time
;
503 te
= (void *)sample
->raw_data
;
504 if (session
->sample_type
& PERF_SAMPLE_RAW
&& sample
->raw_size
> 0) {
506 #ifdef SUPPORT_OLD_POWER_EVENTS
507 struct power_entry_old
*peo
;
511 * FIXME: use evsel, its already mapped from id to perf_evsel,
512 * remove perf_header__find_event infrastructure bits.
513 * Mapping all these "power:cpu_idle" strings to the tracepoint
514 * ID and then just comparing against evsel->attr.config.
518 * if (evsel->attr.config == power_cpu_idle_id)
520 event_str
= perf_header__find_event(te
->type
);
525 if (sample
->cpu
> numcpus
)
526 numcpus
= sample
->cpu
;
528 if (strcmp(event_str
, "power:cpu_idle") == 0) {
529 struct power_processor_entry
*ppe
= (void *)te
;
530 if (ppe
->state
== (u32
)PWR_EVENT_EXIT
)
531 c_state_end(ppe
->cpu_id
, sample
->time
);
533 c_state_start(ppe
->cpu_id
, sample
->time
,
536 else if (strcmp(event_str
, "power:cpu_frequency") == 0) {
537 struct power_processor_entry
*ppe
= (void *)te
;
538 p_state_change(ppe
->cpu_id
, sample
->time
, ppe
->state
);
541 else if (strcmp(event_str
, "sched:sched_wakeup") == 0)
542 sched_wakeup(sample
->cpu
, sample
->time
, sample
->pid
, te
);
544 else if (strcmp(event_str
, "sched:sched_switch") == 0)
545 sched_switch(sample
->cpu
, sample
->time
, te
);
547 #ifdef SUPPORT_OLD_POWER_EVENTS
548 if (use_old_power_events
) {
549 if (strcmp(event_str
, "power:power_start") == 0)
550 c_state_start(peo
->cpu_id
, sample
->time
,
553 else if (strcmp(event_str
, "power:power_end") == 0)
554 c_state_end(sample
->cpu
, sample
->time
);
556 else if (strcmp(event_str
,
557 "power:power_frequency") == 0)
558 p_state_change(peo
->cpu_id
, sample
->time
,
567 * After the last sample we need to wrap up the current C/P state
568 * and close out each CPU for these.
570 static void end_sample_processing(void)
573 struct power_event
*pwr
;
575 for (cpu
= 0; cpu
<= numcpus
; cpu
++) {
576 pwr
= malloc(sizeof(struct power_event
));
579 memset(pwr
, 0, sizeof(struct power_event
));
583 pwr
->state
= cpus_cstate_state
[cpu
];
584 pwr
->start_time
= cpus_cstate_start_times
[cpu
];
585 pwr
->end_time
= last_time
;
588 pwr
->next
= power_events
;
594 pwr
= malloc(sizeof(struct power_event
));
597 memset(pwr
, 0, sizeof(struct power_event
));
599 pwr
->state
= cpus_pstate_state
[cpu
];
600 pwr
->start_time
= cpus_pstate_start_times
[cpu
];
601 pwr
->end_time
= last_time
;
604 pwr
->next
= power_events
;
606 if (!pwr
->start_time
)
607 pwr
->start_time
= first_time
;
609 pwr
->state
= min_freq
;
615 * Sort the pid datastructure
617 static void sort_pids(void)
619 struct per_pid
*new_list
, *p
, *cursor
, *prev
;
620 /* sort by ppid first, then by pid, lowest to highest */
629 if (new_list
== NULL
) {
637 if (cursor
->ppid
> p
->ppid
||
638 (cursor
->ppid
== p
->ppid
&& cursor
->pid
> p
->pid
)) {
639 /* must insert before */
641 p
->next
= prev
->next
;
654 cursor
= cursor
->next
;
663 static void draw_c_p_states(void)
665 struct power_event
*pwr
;
669 * two pass drawing so that the P state bars are on top of the C state blocks
672 if (pwr
->type
== CSTATE
)
673 svg_cstate(pwr
->cpu
, pwr
->start_time
, pwr
->end_time
, pwr
->state
);
679 if (pwr
->type
== PSTATE
) {
681 pwr
->state
= min_freq
;
682 svg_pstate(pwr
->cpu
, pwr
->start_time
, pwr
->end_time
, pwr
->state
);
688 static void draw_wakeups(void)
690 struct wake_event
*we
;
692 struct per_pidcomm
*c
;
696 int from
= 0, to
= 0;
697 char *task_from
= NULL
, *task_to
= NULL
;
699 /* locate the column of the waker and wakee */
702 if (p
->pid
== we
->waker
|| p
->pid
== we
->wakee
) {
705 if (c
->Y
&& c
->start_time
<= we
->time
&& c
->end_time
>= we
->time
) {
706 if (p
->pid
== we
->waker
&& !from
) {
708 task_from
= strdup(c
->comm
);
710 if (p
->pid
== we
->wakee
&& !to
) {
712 task_to
= strdup(c
->comm
);
719 if (p
->pid
== we
->waker
&& !from
) {
721 task_from
= strdup(c
->comm
);
723 if (p
->pid
== we
->wakee
&& !to
) {
725 task_to
= strdup(c
->comm
);
734 task_from
= malloc(40);
735 sprintf(task_from
, "[%i]", we
->waker
);
738 task_to
= malloc(40);
739 sprintf(task_to
, "[%i]", we
->wakee
);
743 svg_interrupt(we
->time
, to
);
744 else if (from
&& to
&& abs(from
- to
) == 1)
745 svg_wakeline(we
->time
, from
, to
);
747 svg_partial_wakeline(we
->time
, from
, task_from
, to
, task_to
);
755 static void draw_cpu_usage(void)
758 struct per_pidcomm
*c
;
759 struct cpu_sample
*sample
;
766 if (sample
->type
== TYPE_RUNNING
)
767 svg_process(sample
->cpu
, sample
->start_time
, sample
->end_time
, "sample", c
->comm
);
769 sample
= sample
->next
;
777 static void draw_process_bars(void)
780 struct per_pidcomm
*c
;
781 struct cpu_sample
*sample
;
796 svg_box(Y
, c
->start_time
, c
->end_time
, "process");
799 if (sample
->type
== TYPE_RUNNING
)
800 svg_sample(Y
, sample
->cpu
, sample
->start_time
, sample
->end_time
);
801 if (sample
->type
== TYPE_BLOCKED
)
802 svg_box(Y
, sample
->start_time
, sample
->end_time
, "blocked");
803 if (sample
->type
== TYPE_WAITING
)
804 svg_waiting(Y
, sample
->start_time
, sample
->end_time
);
805 sample
= sample
->next
;
810 if (c
->total_time
> 5000000000) /* 5 seconds */
811 sprintf(comm
, "%s:%i (%2.2fs)", c
->comm
, p
->pid
, c
->total_time
/ 1000000000.0);
813 sprintf(comm
, "%s:%i (%3.1fms)", c
->comm
, p
->pid
, c
->total_time
/ 1000000.0);
815 svg_text(Y
, c
->start_time
, comm
);
825 static void add_process_filter(const char *string
)
827 struct process_filter
*filt
;
830 pid
= strtoull(string
, NULL
, 10);
831 filt
= malloc(sizeof(struct process_filter
));
835 filt
->name
= strdup(string
);
837 filt
->next
= process_filter
;
839 process_filter
= filt
;
842 static int passes_filter(struct per_pid
*p
, struct per_pidcomm
*c
)
844 struct process_filter
*filt
;
848 filt
= process_filter
;
850 if (filt
->pid
&& p
->pid
== filt
->pid
)
852 if (strcmp(filt
->name
, c
->comm
) == 0)
859 static int determine_display_tasks_filtered(void)
862 struct per_pidcomm
*c
;
868 if (p
->start_time
== 1)
869 p
->start_time
= first_time
;
871 /* no exit marker, task kept running to the end */
872 if (p
->end_time
== 0)
873 p
->end_time
= last_time
;
880 if (c
->start_time
== 1)
881 c
->start_time
= first_time
;
883 if (passes_filter(p
, c
)) {
889 if (c
->end_time
== 0)
890 c
->end_time
= last_time
;
899 static int determine_display_tasks(u64 threshold
)
902 struct per_pidcomm
*c
;
906 return determine_display_tasks_filtered();
911 if (p
->start_time
== 1)
912 p
->start_time
= first_time
;
914 /* no exit marker, task kept running to the end */
915 if (p
->end_time
== 0)
916 p
->end_time
= last_time
;
917 if (p
->total_time
>= threshold
&& !power_only
)
925 if (c
->start_time
== 1)
926 c
->start_time
= first_time
;
928 if (c
->total_time
>= threshold
&& !power_only
) {
933 if (c
->end_time
== 0)
934 c
->end_time
= last_time
;
945 #define TIME_THRESH 10000000
947 static void write_svg_file(const char *filename
)
955 count
= determine_display_tasks(TIME_THRESH
);
957 /* We'd like to show at least 15 tasks; be less picky if we have fewer */
959 count
= determine_display_tasks(TIME_THRESH
/ 10);
961 open_svg(filename
, numcpus
, count
, first_time
, last_time
);
966 for (i
= 0; i
< numcpus
; i
++)
967 svg_cpu_box(i
, max_freq
, turbo_frequency
);
977 static struct perf_event_ops event_ops
= {
978 .comm
= process_comm_event
,
979 .fork
= process_fork_event
,
980 .exit
= process_exit_event
,
981 .sample
= process_sample_event
,
982 .ordered_samples
= true,
985 static int __cmd_timechart(void)
987 struct perf_session
*session
= perf_session__new(input_name
, O_RDONLY
,
988 0, false, &event_ops
);
994 if (!perf_session__has_traces(session
, "timechart record"))
997 ret
= perf_session__process_events(session
, &event_ops
);
1001 end_sample_processing();
1005 write_svg_file(output_name
);
1007 pr_info("Written %2.1f seconds of trace to %s.\n",
1008 (last_time
- first_time
) / 1000000000.0, output_name
);
1010 perf_session__delete(session
);
1014 static const char * const timechart_usage
[] = {
1015 "perf timechart [<options>] {record}",
1019 #ifdef SUPPORT_OLD_POWER_EVENTS
1020 static const char * const record_old_args
[] = {
1026 "-e", "power:power_start",
1027 "-e", "power:power_end",
1028 "-e", "power:power_frequency",
1029 "-e", "sched:sched_wakeup",
1030 "-e", "sched:sched_switch",
1034 static const char * const record_new_args
[] = {
1040 "-e", "power:cpu_frequency",
1041 "-e", "power:cpu_idle",
1042 "-e", "sched:sched_wakeup",
1043 "-e", "sched:sched_switch",
1046 static int __cmd_record(int argc
, const char **argv
)
1048 unsigned int rec_argc
, i
, j
;
1049 const char **rec_argv
;
1050 const char * const *record_args
= record_new_args
;
1051 unsigned int record_elems
= ARRAY_SIZE(record_new_args
);
1053 #ifdef SUPPORT_OLD_POWER_EVENTS
1054 if (!is_valid_tracepoint("power:cpu_idle") &&
1055 is_valid_tracepoint("power:power_start")) {
1056 use_old_power_events
= 1;
1057 record_args
= record_old_args
;
1058 record_elems
= ARRAY_SIZE(record_old_args
);
1062 rec_argc
= record_elems
+ argc
- 1;
1063 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1065 if (rec_argv
== NULL
)
1068 for (i
= 0; i
< record_elems
; i
++)
1069 rec_argv
[i
] = strdup(record_args
[i
]);
1071 for (j
= 1; j
< (unsigned int)argc
; j
++, i
++)
1072 rec_argv
[i
] = argv
[j
];
1074 return cmd_record(i
, rec_argv
, NULL
);
1078 parse_process(const struct option
*opt __used
, const char *arg
, int __used unset
)
1081 add_process_filter(arg
);
1085 static const struct option options
[] = {
1086 OPT_STRING('i', "input", &input_name
, "file",
1088 OPT_STRING('o', "output", &output_name
, "file",
1089 "output file name"),
1090 OPT_INTEGER('w', "width", &svg_page_width
,
1092 OPT_BOOLEAN('P', "power-only", &power_only
,
1093 "output power data only"),
1094 OPT_CALLBACK('p', "process", NULL
, "process",
1095 "process selector. Pass a pid or process name.",
1097 OPT_STRING(0, "symfs", &symbol_conf
.symfs
, "directory",
1098 "Look for files with symbols relative to this directory"),
1103 int cmd_timechart(int argc
, const char **argv
, const char *prefix __used
)
1105 argc
= parse_options(argc
, argv
, options
, timechart_usage
,
1106 PARSE_OPT_STOP_AT_NON_OPTION
);
1110 if (argc
&& !strncmp(argv
[0], "rec", 3))
1111 return __cmd_record(argc
, argv
);
1113 usage_with_options(timechart_usage
, options
);
1117 return __cmd_timechart();