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/data_map.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
];
285 process_comm_event(event_t
*event
)
287 pid_set_comm(event
->comm
.pid
, event
->comm
.comm
);
291 process_fork_event(event_t
*event
)
293 pid_fork(event
->fork
.pid
, event
->fork
.ppid
, event
->fork
.time
);
298 process_exit_event(event_t
*event
)
300 pid_exit(event
->fork
.pid
, event
->fork
.time
);
307 unsigned char preempt_count
;
313 struct trace_entry te
;
318 #define TASK_COMM_LEN 16
319 struct wakeup_entry
{
320 struct trace_entry te
;
321 char comm
[TASK_COMM_LEN
];
328 * trace_flag_type is an enumeration that holds different
329 * states when a trace occurs. These are:
330 * IRQS_OFF - interrupts were disabled
331 * IRQS_NOSUPPORT - arch does not support irqs_disabled_flags
332 * NEED_RESCED - reschedule is requested
333 * HARDIRQ - inside an interrupt handler
334 * SOFTIRQ - inside a softirq handler
336 enum trace_flag_type
{
337 TRACE_FLAG_IRQS_OFF
= 0x01,
338 TRACE_FLAG_IRQS_NOSUPPORT
= 0x02,
339 TRACE_FLAG_NEED_RESCHED
= 0x04,
340 TRACE_FLAG_HARDIRQ
= 0x08,
341 TRACE_FLAG_SOFTIRQ
= 0x10,
346 struct sched_switch
{
347 struct trace_entry te
;
348 char prev_comm
[TASK_COMM_LEN
];
351 long prev_state
; /* Arjan weeps. */
352 char next_comm
[TASK_COMM_LEN
];
357 static void c_state_start(int cpu
, u64 timestamp
, int state
)
359 cpus_cstate_start_times
[cpu
] = timestamp
;
360 cpus_cstate_state
[cpu
] = state
;
363 static void c_state_end(int cpu
, u64 timestamp
)
365 struct power_event
*pwr
;
366 pwr
= malloc(sizeof(struct power_event
));
369 memset(pwr
, 0, sizeof(struct power_event
));
371 pwr
->state
= cpus_cstate_state
[cpu
];
372 pwr
->start_time
= cpus_cstate_start_times
[cpu
];
373 pwr
->end_time
= timestamp
;
376 pwr
->next
= power_events
;
381 static void p_state_change(int cpu
, u64 timestamp
, u64 new_freq
)
383 struct power_event
*pwr
;
384 pwr
= malloc(sizeof(struct power_event
));
386 if (new_freq
> 8000000) /* detect invalid data */
391 memset(pwr
, 0, sizeof(struct power_event
));
393 pwr
->state
= cpus_pstate_state
[cpu
];
394 pwr
->start_time
= cpus_pstate_start_times
[cpu
];
395 pwr
->end_time
= timestamp
;
398 pwr
->next
= power_events
;
400 if (!pwr
->start_time
)
401 pwr
->start_time
= first_time
;
405 cpus_pstate_state
[cpu
] = new_freq
;
406 cpus_pstate_start_times
[cpu
] = timestamp
;
408 if ((u64
)new_freq
> max_freq
)
411 if (new_freq
< min_freq
|| min_freq
== 0)
414 if (new_freq
== max_freq
- 1000)
415 turbo_frequency
= max_freq
;
419 sched_wakeup(int cpu
, u64 timestamp
, int pid
, struct trace_entry
*te
)
421 struct wake_event
*we
;
423 struct wakeup_entry
*wake
= (void *)te
;
425 we
= malloc(sizeof(struct wake_event
));
429 memset(we
, 0, sizeof(struct wake_event
));
430 we
->time
= timestamp
;
433 if ((te
->flags
& TRACE_FLAG_HARDIRQ
) || (te
->flags
& TRACE_FLAG_SOFTIRQ
))
436 we
->wakee
= wake
->pid
;
437 we
->next
= wake_events
;
439 p
= find_create_pid(we
->wakee
);
441 if (p
&& p
->current
&& p
->current
->state
== TYPE_NONE
) {
442 p
->current
->state_since
= timestamp
;
443 p
->current
->state
= TYPE_WAITING
;
445 if (p
&& p
->current
&& p
->current
->state
== TYPE_BLOCKED
) {
446 pid_put_sample(p
->pid
, p
->current
->state
, cpu
, p
->current
->state_since
, timestamp
);
447 p
->current
->state_since
= timestamp
;
448 p
->current
->state
= TYPE_WAITING
;
452 static void sched_switch(int cpu
, u64 timestamp
, struct trace_entry
*te
)
454 struct per_pid
*p
= NULL
, *prev_p
;
455 struct sched_switch
*sw
= (void *)te
;
458 prev_p
= find_create_pid(sw
->prev_pid
);
460 p
= find_create_pid(sw
->next_pid
);
462 if (prev_p
->current
&& prev_p
->current
->state
!= TYPE_NONE
)
463 pid_put_sample(sw
->prev_pid
, TYPE_RUNNING
, cpu
, prev_p
->current
->state_since
, timestamp
);
464 if (p
&& p
->current
) {
465 if (p
->current
->state
!= TYPE_NONE
)
466 pid_put_sample(sw
->next_pid
, p
->current
->state
, cpu
, p
->current
->state_since
, timestamp
);
468 p
->current
->state_since
= timestamp
;
469 p
->current
->state
= TYPE_RUNNING
;
472 if (prev_p
->current
) {
473 prev_p
->current
->state
= TYPE_NONE
;
474 prev_p
->current
->state_since
= timestamp
;
475 if (sw
->prev_state
& 2)
476 prev_p
->current
->state
= TYPE_BLOCKED
;
477 if (sw
->prev_state
== 0)
478 prev_p
->current
->state
= TYPE_WAITING
;
484 process_sample_event(event_t
*event
)
492 struct trace_entry
*te
;
494 if (sample_type
& PERF_SAMPLE_IP
)
497 if (sample_type
& PERF_SAMPLE_TID
) {
498 pid
= event
->sample
.array
[cursor
]>>32;
501 if (sample_type
& PERF_SAMPLE_TIME
) {
502 stamp
= event
->sample
.array
[cursor
++];
504 if (!first_time
|| first_time
> stamp
)
506 if (last_time
< stamp
)
510 if (sample_type
& PERF_SAMPLE_ADDR
)
511 addr
= event
->sample
.array
[cursor
++];
512 if (sample_type
& PERF_SAMPLE_ID
)
514 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
516 if (sample_type
& PERF_SAMPLE_CPU
)
517 cpu
= event
->sample
.array
[cursor
++] & 0xFFFFFFFF;
518 if (sample_type
& PERF_SAMPLE_PERIOD
)
521 size_ptr
= (void *)&event
->sample
.array
[cursor
];
526 te
= (void *)size_ptr
;
527 if (sample_type
& PERF_SAMPLE_RAW
&& size
> 0) {
529 struct power_entry
*pe
;
533 event_str
= perf_header__find_event(te
->type
);
538 if (strcmp(event_str
, "power:power_start") == 0)
539 c_state_start(cpu
, stamp
, pe
->value
);
541 if (strcmp(event_str
, "power:power_end") == 0)
542 c_state_end(cpu
, stamp
);
544 if (strcmp(event_str
, "power:power_frequency") == 0)
545 p_state_change(cpu
, stamp
, pe
->value
);
547 if (strcmp(event_str
, "sched:sched_wakeup") == 0)
548 sched_wakeup(cpu
, stamp
, pid
, te
);
550 if (strcmp(event_str
, "sched:sched_switch") == 0)
551 sched_switch(cpu
, stamp
, te
);
557 * After the last sample we need to wrap up the current C/P state
558 * and close out each CPU for these.
560 static void end_sample_processing(void)
563 struct power_event
*pwr
;
565 for (cpu
= 0; cpu
<= numcpus
; cpu
++) {
566 pwr
= malloc(sizeof(struct power_event
));
569 memset(pwr
, 0, sizeof(struct power_event
));
573 pwr
->state
= cpus_cstate_state
[cpu
];
574 pwr
->start_time
= cpus_cstate_start_times
[cpu
];
575 pwr
->end_time
= last_time
;
578 pwr
->next
= power_events
;
584 pwr
= malloc(sizeof(struct power_event
));
587 memset(pwr
, 0, sizeof(struct power_event
));
589 pwr
->state
= cpus_pstate_state
[cpu
];
590 pwr
->start_time
= cpus_pstate_start_times
[cpu
];
591 pwr
->end_time
= last_time
;
594 pwr
->next
= power_events
;
596 if (!pwr
->start_time
)
597 pwr
->start_time
= first_time
;
599 pwr
->state
= min_freq
;
604 static u64
sample_time(event_t
*event
)
609 if (sample_type
& PERF_SAMPLE_IP
)
611 if (sample_type
& PERF_SAMPLE_TID
)
613 if (sample_type
& PERF_SAMPLE_TIME
)
614 return event
->sample
.array
[cursor
];
620 * We first queue all events, sorted backwards by insertion.
621 * The order will get flipped later.
624 queue_sample_event(event_t
*event
)
626 struct sample_wrapper
*copy
, *prev
;
629 size
= event
->sample
.header
.size
+ sizeof(struct sample_wrapper
) + 8;
635 memset(copy
, 0, size
);
638 copy
->timestamp
= sample_time(event
);
640 memcpy(©
->data
, event
, event
->sample
.header
.size
);
642 /* insert in the right place in the list */
645 /* first sample ever */
650 if (all_samples
->timestamp
< copy
->timestamp
) {
651 /* insert at the head of the list */
652 copy
->next
= all_samples
;
659 if (prev
->next
->timestamp
< copy
->timestamp
) {
660 copy
->next
= prev
->next
;
666 /* insert at the end of the list */
672 static void sort_queued_samples(void)
674 struct sample_wrapper
*cursor
, *next
;
676 cursor
= all_samples
;
681 cursor
->next
= all_samples
;
682 all_samples
= cursor
;
688 * Sort the pid datastructure
690 static void sort_pids(void)
692 struct per_pid
*new_list
, *p
, *cursor
, *prev
;
693 /* sort by ppid first, then by pid, lowest to highest */
702 if (new_list
== NULL
) {
710 if (cursor
->ppid
> p
->ppid
||
711 (cursor
->ppid
== p
->ppid
&& cursor
->pid
> p
->pid
)) {
712 /* must insert before */
714 p
->next
= prev
->next
;
727 cursor
= cursor
->next
;
736 static void draw_c_p_states(void)
738 struct power_event
*pwr
;
742 * two pass drawing so that the P state bars are on top of the C state blocks
745 if (pwr
->type
== CSTATE
)
746 svg_cstate(pwr
->cpu
, pwr
->start_time
, pwr
->end_time
, pwr
->state
);
752 if (pwr
->type
== PSTATE
) {
754 pwr
->state
= min_freq
;
755 svg_pstate(pwr
->cpu
, pwr
->start_time
, pwr
->end_time
, pwr
->state
);
761 static void draw_wakeups(void)
763 struct wake_event
*we
;
765 struct per_pidcomm
*c
;
769 int from
= 0, to
= 0;
770 char *task_from
= NULL
, *task_to
= NULL
;
772 /* locate the column of the waker and wakee */
775 if (p
->pid
== we
->waker
|| p
->pid
== we
->wakee
) {
778 if (c
->Y
&& c
->start_time
<= we
->time
&& c
->end_time
>= we
->time
) {
779 if (p
->pid
== we
->waker
&& !from
) {
781 task_from
= strdup(c
->comm
);
783 if (p
->pid
== we
->wakee
&& !to
) {
785 task_to
= strdup(c
->comm
);
792 if (p
->pid
== we
->waker
&& !from
) {
794 task_from
= strdup(c
->comm
);
796 if (p
->pid
== we
->wakee
&& !to
) {
798 task_to
= strdup(c
->comm
);
807 task_from
= malloc(40);
808 sprintf(task_from
, "[%i]", we
->waker
);
811 task_to
= malloc(40);
812 sprintf(task_to
, "[%i]", we
->wakee
);
816 svg_interrupt(we
->time
, to
);
817 else if (from
&& to
&& abs(from
- to
) == 1)
818 svg_wakeline(we
->time
, from
, to
);
820 svg_partial_wakeline(we
->time
, from
, task_from
, to
, task_to
);
828 static void draw_cpu_usage(void)
831 struct per_pidcomm
*c
;
832 struct cpu_sample
*sample
;
839 if (sample
->type
== TYPE_RUNNING
)
840 svg_process(sample
->cpu
, sample
->start_time
, sample
->end_time
, "sample", c
->comm
);
842 sample
= sample
->next
;
850 static void draw_process_bars(void)
853 struct per_pidcomm
*c
;
854 struct cpu_sample
*sample
;
869 svg_box(Y
, c
->start_time
, c
->end_time
, "process");
872 if (sample
->type
== TYPE_RUNNING
)
873 svg_sample(Y
, sample
->cpu
, sample
->start_time
, sample
->end_time
);
874 if (sample
->type
== TYPE_BLOCKED
)
875 svg_box(Y
, sample
->start_time
, sample
->end_time
, "blocked");
876 if (sample
->type
== TYPE_WAITING
)
877 svg_waiting(Y
, sample
->start_time
, sample
->end_time
);
878 sample
= sample
->next
;
883 if (c
->total_time
> 5000000000) /* 5 seconds */
884 sprintf(comm
, "%s:%i (%2.2fs)", c
->comm
, p
->pid
, c
->total_time
/ 1000000000.0);
886 sprintf(comm
, "%s:%i (%3.1fms)", c
->comm
, p
->pid
, c
->total_time
/ 1000000.0);
888 svg_text(Y
, c
->start_time
, comm
);
898 static void add_process_filter(const char *string
)
900 struct process_filter
*filt
;
903 pid
= strtoull(string
, NULL
, 10);
904 filt
= malloc(sizeof(struct process_filter
));
908 filt
->name
= strdup(string
);
910 filt
->next
= process_filter
;
912 process_filter
= filt
;
915 static int passes_filter(struct per_pid
*p
, struct per_pidcomm
*c
)
917 struct process_filter
*filt
;
921 filt
= process_filter
;
923 if (filt
->pid
&& p
->pid
== filt
->pid
)
925 if (strcmp(filt
->name
, c
->comm
) == 0)
932 static int determine_display_tasks_filtered(void)
935 struct per_pidcomm
*c
;
941 if (p
->start_time
== 1)
942 p
->start_time
= first_time
;
944 /* no exit marker, task kept running to the end */
945 if (p
->end_time
== 0)
946 p
->end_time
= last_time
;
953 if (c
->start_time
== 1)
954 c
->start_time
= first_time
;
956 if (passes_filter(p
, c
)) {
962 if (c
->end_time
== 0)
963 c
->end_time
= last_time
;
972 static int determine_display_tasks(u64 threshold
)
975 struct per_pidcomm
*c
;
979 return determine_display_tasks_filtered();
984 if (p
->start_time
== 1)
985 p
->start_time
= first_time
;
987 /* no exit marker, task kept running to the end */
988 if (p
->end_time
== 0)
989 p
->end_time
= last_time
;
990 if (p
->total_time
>= threshold
&& !power_only
)
998 if (c
->start_time
== 1)
999 c
->start_time
= first_time
;
1001 if (c
->total_time
>= threshold
&& !power_only
) {
1006 if (c
->end_time
== 0)
1007 c
->end_time
= last_time
;
1018 #define TIME_THRESH 10000000
1020 static void write_svg_file(const char *filename
)
1028 count
= determine_display_tasks(TIME_THRESH
);
1030 /* We'd like to show at least 15 tasks; be less picky if we have fewer */
1032 count
= determine_display_tasks(TIME_THRESH
/ 10);
1034 open_svg(filename
, numcpus
, count
, first_time
, last_time
);
1039 for (i
= 0; i
< numcpus
; i
++)
1040 svg_cpu_box(i
, max_freq
, turbo_frequency
);
1043 draw_process_bars();
1050 static void process_samples(void)
1052 struct sample_wrapper
*cursor
;
1055 sort_queued_samples();
1057 cursor
= all_samples
;
1059 event
= (void *)&cursor
->data
;
1060 cursor
= cursor
->next
;
1061 process_sample_event(event
);
1065 static int sample_type_check(u64 type
)
1069 if (!(sample_type
& PERF_SAMPLE_RAW
)) {
1070 fprintf(stderr
, "No trace samples found in the file.\n"
1071 "Have you used 'perf timechart record' to record it?\n");
1078 static struct perf_file_handler file_handler
= {
1079 .process_comm_event
= process_comm_event
,
1080 .process_fork_event
= process_fork_event
,
1081 .process_exit_event
= process_exit_event
,
1082 .process_sample_event
= queue_sample_event
,
1083 .sample_type_check
= sample_type_check
,
1086 static int __cmd_timechart(void)
1088 struct perf_header
*header
;
1091 register_perf_file_handler(&file_handler
);
1093 ret
= mmap_dispatch_perf_file(&header
, input_name
, 0, 0,
1094 &event__cwdlen
, &event__cwd
);
1096 return EXIT_FAILURE
;
1100 end_sample_processing();
1104 write_svg_file(output_name
);
1106 pr_info("Written %2.1f seconds of trace to %s.\n",
1107 (last_time
- first_time
) / 1000000000.0, output_name
);
1109 return EXIT_SUCCESS
;
1112 static const char * const timechart_usage
[] = {
1113 "perf timechart [<options>] {record}",
1117 static const char *record_args
[] = {
1124 "-e", "power:power_start",
1125 "-e", "power:power_end",
1126 "-e", "power:power_frequency",
1127 "-e", "sched:sched_wakeup",
1128 "-e", "sched:sched_switch",
1131 static int __cmd_record(int argc
, const char **argv
)
1133 unsigned int rec_argc
, i
, j
;
1134 const char **rec_argv
;
1136 rec_argc
= ARRAY_SIZE(record_args
) + argc
- 1;
1137 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1139 for (i
= 0; i
< ARRAY_SIZE(record_args
); i
++)
1140 rec_argv
[i
] = strdup(record_args
[i
]);
1142 for (j
= 1; j
< (unsigned int)argc
; j
++, i
++)
1143 rec_argv
[i
] = argv
[j
];
1145 return cmd_record(i
, rec_argv
, NULL
);
1149 parse_process(const struct option
*opt __used
, const char *arg
, int __used unset
)
1152 add_process_filter(arg
);
1156 static const struct option options
[] = {
1157 OPT_STRING('i', "input", &input_name
, "file",
1159 OPT_STRING('o', "output", &output_name
, "file",
1160 "output file name"),
1161 OPT_INTEGER('w', "width", &svg_page_width
,
1163 OPT_BOOLEAN('P', "power-only", &power_only
,
1164 "output power data only"),
1165 OPT_CALLBACK('p', "process", NULL
, "process",
1166 "process selector. Pass a pid or process name.",
1172 int cmd_timechart(int argc
, const char **argv
, const char *prefix __used
)
1176 argc
= parse_options(argc
, argv
, options
, timechart_usage
,
1177 PARSE_OPT_STOP_AT_NON_OPTION
);
1179 if (argc
&& !strncmp(argv
[0], "rec", 3))
1180 return __cmd_record(argc
, argv
);
1182 usage_with_options(timechart_usage
, options
);
1186 return __cmd_timechart();