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 static char const *input_name = "perf.data";
36 static char const *output_name = "output.svg";
38 static unsigned int numcpus;
39 static u64 min_freq; /* Lowest CPU frequency seen */
40 static u64 max_freq; /* Highest CPU frequency seen */
41 static u64 turbo_frequency;
43 static u64 first_time, last_time;
45 static bool power_only;
55 struct sample_wrapper;
58 * Datastructure layout:
59 * We keep an list of "pid"s, matching the kernels notion of a task struct.
60 * Each "pid" entry, has a list of "comm"s.
61 * this is because we want to track different programs different, while
62 * exec will reuse the original pid (by design).
63 * Each comm has a list of samples that will be used to draw
78 struct per_pidcomm *all;
79 struct per_pidcomm *current;
84 struct per_pidcomm *next;
98 struct cpu_sample *samples;
101 struct sample_wrapper {
102 struct sample_wrapper *next;
105 unsigned char data[0];
109 #define TYPE_RUNNING 1
110 #define TYPE_WAITING 2
111 #define TYPE_BLOCKED 3
114 struct cpu_sample *next;
122 static struct per_pid *all_data;
128 struct power_event *next;
137 struct wake_event *next;
143 static struct power_event *power_events;
144 static struct wake_event *wake_events;
146 struct process_filter;
147 struct process_filter {
150 struct process_filter *next;
153 static struct process_filter *process_filter;
156 static struct per_pid *find_create_pid(int pid)
158 struct per_pid *cursor = all_data;
161 if (cursor->pid == pid)
163 cursor = cursor->next;
165 cursor = malloc(sizeof(struct per_pid));
166 assert(cursor != NULL);
167 memset(cursor, 0, sizeof(struct per_pid));
169 cursor->next = all_data;
174 static void pid_set_comm(int pid, char *comm)
177 struct per_pidcomm *c;
178 p = find_create_pid(pid);
181 if (c->comm && strcmp(c->comm, comm) == 0) {
186 c->comm = strdup(comm);
192 c = malloc(sizeof(struct per_pidcomm));
194 memset(c, 0, sizeof(struct per_pidcomm));
195 c->comm = strdup(comm);
201 static void pid_fork(int pid, int ppid, u64 timestamp)
203 struct per_pid *p, *pp;
204 p = find_create_pid(pid);
205 pp = find_create_pid(ppid);
207 if (pp->current && pp->current->comm && !p->current)
208 pid_set_comm(pid, pp->current->comm);
210 p->start_time = timestamp;
212 p->current->start_time = timestamp;
213 p->current->state_since = timestamp;
217 static void pid_exit(int pid, u64 timestamp)
220 p = find_create_pid(pid);
221 p->end_time = timestamp;
223 p->current->end_time = timestamp;
227 pid_put_sample(int pid, int type, unsigned int cpu, u64 start, u64 end)
230 struct per_pidcomm *c;
231 struct cpu_sample *sample;
233 p = find_create_pid(pid);
236 c = malloc(sizeof(struct per_pidcomm));
238 memset(c, 0, sizeof(struct per_pidcomm));
244 sample = malloc(sizeof(struct cpu_sample));
245 assert(sample != NULL);
246 memset(sample, 0, sizeof(struct cpu_sample));
247 sample->start_time = start;
248 sample->end_time = end;
250 sample->next = c->samples;
254 if (sample->type == TYPE_RUNNING && end > start && start > 0) {
255 c->total_time += (end-start);
256 p->total_time += (end-start);
259 if (c->start_time == 0 || c->start_time > start)
260 c->start_time = start;
261 if (p->start_time == 0 || p->start_time > start)
262 p->start_time = start;
268 #define MAX_CPUS 4096
270 static u64 cpus_cstate_start_times[MAX_CPUS];
271 static int cpus_cstate_state[MAX_CPUS];
272 static u64 cpus_pstate_start_times[MAX_CPUS];
273 static u64 cpus_pstate_state[MAX_CPUS];
275 static int process_comm_event(event_t *event, struct perf_session *session __used)
277 pid_set_comm(event->comm.tid, event->comm.comm);
281 static int process_fork_event(event_t *event, struct perf_session *session __used)
283 pid_fork(event->fork.pid, event->fork.ppid, event->fork.time);
287 static int process_exit_event(event_t *event, struct perf_session *session __used)
289 pid_exit(event->fork.pid, event->fork.time);
296 unsigned char preempt_count;
302 struct trace_entry te;
308 #define TASK_COMM_LEN 16
309 struct wakeup_entry {
310 struct trace_entry te;
311 char comm[TASK_COMM_LEN];
318 * trace_flag_type is an enumeration that holds different
319 * states when a trace occurs. These are:
320 * IRQS_OFF - interrupts were disabled
321 * IRQS_NOSUPPORT - arch does not support irqs_disabled_flags
322 * NEED_RESCED - reschedule is requested
323 * HARDIRQ - inside an interrupt handler
324 * SOFTIRQ - inside a softirq handler
326 enum trace_flag_type {
327 TRACE_FLAG_IRQS_OFF = 0x01,
328 TRACE_FLAG_IRQS_NOSUPPORT = 0x02,
329 TRACE_FLAG_NEED_RESCHED = 0x04,
330 TRACE_FLAG_HARDIRQ = 0x08,
331 TRACE_FLAG_SOFTIRQ = 0x10,
336 struct sched_switch {
337 struct trace_entry te;
338 char prev_comm[TASK_COMM_LEN];
341 long prev_state; /* Arjan weeps. */
342 char next_comm[TASK_COMM_LEN];
347 static void c_state_start(int cpu, u64 timestamp, int state)
349 cpus_cstate_start_times[cpu] = timestamp;
350 cpus_cstate_state[cpu] = state;
353 static void c_state_end(int cpu, u64 timestamp)
355 struct power_event *pwr;
356 pwr = malloc(sizeof(struct power_event));
359 memset(pwr, 0, sizeof(struct power_event));
361 pwr->state = cpus_cstate_state[cpu];
362 pwr->start_time = cpus_cstate_start_times[cpu];
363 pwr->end_time = timestamp;
366 pwr->next = power_events;
371 static void p_state_change(int cpu, u64 timestamp, u64 new_freq)
373 struct power_event *pwr;
374 pwr = malloc(sizeof(struct power_event));
376 if (new_freq > 8000000) /* detect invalid data */
381 memset(pwr, 0, sizeof(struct power_event));
383 pwr->state = cpus_pstate_state[cpu];
384 pwr->start_time = cpus_pstate_start_times[cpu];
385 pwr->end_time = timestamp;
388 pwr->next = power_events;
390 if (!pwr->start_time)
391 pwr->start_time = first_time;
395 cpus_pstate_state[cpu] = new_freq;
396 cpus_pstate_start_times[cpu] = timestamp;
398 if ((u64)new_freq > max_freq)
401 if (new_freq < min_freq || min_freq == 0)
404 if (new_freq == max_freq - 1000)
405 turbo_frequency = max_freq;
409 sched_wakeup(int cpu, u64 timestamp, int pid, struct trace_entry *te)
411 struct wake_event *we;
413 struct wakeup_entry *wake = (void *)te;
415 we = malloc(sizeof(struct wake_event));
419 memset(we, 0, sizeof(struct wake_event));
420 we->time = timestamp;
423 if ((te->flags & TRACE_FLAG_HARDIRQ) || (te->flags & TRACE_FLAG_SOFTIRQ))
426 we->wakee = wake->pid;
427 we->next = wake_events;
429 p = find_create_pid(we->wakee);
431 if (p && p->current && p->current->state == TYPE_NONE) {
432 p->current->state_since = timestamp;
433 p->current->state = TYPE_WAITING;
435 if (p && p->current && p->current->state == TYPE_BLOCKED) {
436 pid_put_sample(p->pid, p->current->state, cpu, p->current->state_since, timestamp);
437 p->current->state_since = timestamp;
438 p->current->state = TYPE_WAITING;
442 static void sched_switch(int cpu, u64 timestamp, struct trace_entry *te)
444 struct per_pid *p = NULL, *prev_p;
445 struct sched_switch *sw = (void *)te;
448 prev_p = find_create_pid(sw->prev_pid);
450 p = find_create_pid(sw->next_pid);
452 if (prev_p->current && prev_p->current->state != TYPE_NONE)
453 pid_put_sample(sw->prev_pid, TYPE_RUNNING, cpu, prev_p->current->state_since, timestamp);
454 if (p && p->current) {
455 if (p->current->state != TYPE_NONE)
456 pid_put_sample(sw->next_pid, p->current->state, cpu, p->current->state_since, timestamp);
458 p->current->state_since = timestamp;
459 p->current->state = TYPE_RUNNING;
462 if (prev_p->current) {
463 prev_p->current->state = TYPE_NONE;
464 prev_p->current->state_since = timestamp;
465 if (sw->prev_state & 2)
466 prev_p->current->state = TYPE_BLOCKED;
467 if (sw->prev_state == 0)
468 prev_p->current->state = TYPE_WAITING;
473 static int process_sample_event(event_t *event, struct perf_session *session)
475 struct sample_data data;
476 struct trace_entry *te;
478 memset(&data, 0, sizeof(data));
480 event__parse_sample(event, session->sample_type, &data);
482 if (session->sample_type & PERF_SAMPLE_TIME) {
483 if (!first_time || first_time > data.time)
484 first_time = data.time;
485 if (last_time < data.time)
486 last_time = data.time;
489 te = (void *)data.raw_data;
490 if (session->sample_type & PERF_SAMPLE_RAW && data.raw_size > 0) {
492 struct power_entry *pe;
496 event_str = perf_header__find_event(te->type);
501 if (strcmp(event_str, "power:power_start") == 0)
502 c_state_start(pe->cpu_id, data.time, pe->value);
504 if (strcmp(event_str, "power:power_end") == 0)
505 c_state_end(pe->cpu_id, data.time);
507 if (strcmp(event_str, "power:power_frequency") == 0)
508 p_state_change(pe->cpu_id, data.time, pe->value);
510 if (strcmp(event_str, "sched:sched_wakeup") == 0)
511 sched_wakeup(data.cpu, data.time, data.pid, te);
513 if (strcmp(event_str, "sched:sched_switch") == 0)
514 sched_switch(data.cpu, data.time, te);
520 * After the last sample we need to wrap up the current C/P state
521 * and close out each CPU for these.
523 static void end_sample_processing(void)
526 struct power_event *pwr;
528 for (cpu = 0; cpu <= numcpus; cpu++) {
529 pwr = malloc(sizeof(struct power_event));
532 memset(pwr, 0, sizeof(struct power_event));
536 pwr->state = cpus_cstate_state[cpu];
537 pwr->start_time = cpus_cstate_start_times[cpu];
538 pwr->end_time = last_time;
541 pwr->next = power_events;
547 pwr = malloc(sizeof(struct power_event));
550 memset(pwr, 0, sizeof(struct power_event));
552 pwr->state = cpus_pstate_state[cpu];
553 pwr->start_time = cpus_pstate_start_times[cpu];
554 pwr->end_time = last_time;
557 pwr->next = power_events;
559 if (!pwr->start_time)
560 pwr->start_time = first_time;
562 pwr->state = min_freq;
568 * Sort the pid datastructure
570 static void sort_pids(void)
572 struct per_pid *new_list, *p, *cursor, *prev;
573 /* sort by ppid first, then by pid, lowest to highest */
582 if (new_list == NULL) {
590 if (cursor->ppid > p->ppid ||
591 (cursor->ppid == p->ppid && cursor->pid > p->pid)) {
592 /* must insert before */
594 p->next = prev->next;
607 cursor = cursor->next;
616 static void draw_c_p_states(void)
618 struct power_event *pwr;
622 * two pass drawing so that the P state bars are on top of the C state blocks
625 if (pwr->type == CSTATE)
626 svg_cstate(pwr->cpu, pwr->start_time, pwr->end_time, pwr->state);
632 if (pwr->type == PSTATE) {
634 pwr->state = min_freq;
635 svg_pstate(pwr->cpu, pwr->start_time, pwr->end_time, pwr->state);
641 static void draw_wakeups(void)
643 struct wake_event *we;
645 struct per_pidcomm *c;
649 int from = 0, to = 0;
650 char *task_from = NULL, *task_to = NULL;
652 /* locate the column of the waker and wakee */
655 if (p->pid == we->waker || p->pid == we->wakee) {
658 if (c->Y && c->start_time <= we->time && c->end_time >= we->time) {
659 if (p->pid == we->waker && !from) {
661 task_from = strdup(c->comm);
663 if (p->pid == we->wakee && !to) {
665 task_to = strdup(c->comm);
672 if (p->pid == we->waker && !from) {
674 task_from = strdup(c->comm);
676 if (p->pid == we->wakee && !to) {
678 task_to = strdup(c->comm);
687 task_from = malloc(40);
688 sprintf(task_from, "[%i]", we->waker);
691 task_to = malloc(40);
692 sprintf(task_to, "[%i]", we->wakee);
696 svg_interrupt(we->time, to);
697 else if (from && to && abs(from - to) == 1)
698 svg_wakeline(we->time, from, to);
700 svg_partial_wakeline(we->time, from, task_from, to, task_to);
708 static void draw_cpu_usage(void)
711 struct per_pidcomm *c;
712 struct cpu_sample *sample;
719 if (sample->type == TYPE_RUNNING)
720 svg_process(sample->cpu, sample->start_time, sample->end_time, "sample", c->comm);
722 sample = sample->next;
730 static void draw_process_bars(void)
733 struct per_pidcomm *c;
734 struct cpu_sample *sample;
749 svg_box(Y, c->start_time, c->end_time, "process");
752 if (sample->type == TYPE_RUNNING)
753 svg_sample(Y, sample->cpu, sample->start_time, sample->end_time);
754 if (sample->type == TYPE_BLOCKED)
755 svg_box(Y, sample->start_time, sample->end_time, "blocked");
756 if (sample->type == TYPE_WAITING)
757 svg_waiting(Y, sample->start_time, sample->end_time);
758 sample = sample->next;
763 if (c->total_time > 5000000000) /* 5 seconds */
764 sprintf(comm, "%s:%i (%2.2fs)", c->comm, p->pid, c->total_time / 1000000000.0);
766 sprintf(comm, "%s:%i (%3.1fms)", c->comm, p->pid, c->total_time / 1000000.0);
768 svg_text(Y, c->start_time, comm);
778 static void add_process_filter(const char *string)
780 struct process_filter *filt;
783 pid = strtoull(string, NULL, 10);
784 filt = malloc(sizeof(struct process_filter));
788 filt->name = strdup(string);
790 filt->next = process_filter;
792 process_filter = filt;
795 static int passes_filter(struct per_pid *p, struct per_pidcomm *c)
797 struct process_filter *filt;
801 filt = process_filter;
803 if (filt->pid && p->pid == filt->pid)
805 if (strcmp(filt->name, c->comm) == 0)
812 static int determine_display_tasks_filtered(void)
815 struct per_pidcomm *c;
821 if (p->start_time == 1)
822 p->start_time = first_time;
824 /* no exit marker, task kept running to the end */
825 if (p->end_time == 0)
826 p->end_time = last_time;
833 if (c->start_time == 1)
834 c->start_time = first_time;
836 if (passes_filter(p, c)) {
842 if (c->end_time == 0)
843 c->end_time = last_time;
852 static int determine_display_tasks(u64 threshold)
855 struct per_pidcomm *c;
859 return determine_display_tasks_filtered();
864 if (p->start_time == 1)
865 p->start_time = first_time;
867 /* no exit marker, task kept running to the end */
868 if (p->end_time == 0)
869 p->end_time = last_time;
870 if (p->total_time >= threshold && !power_only)
878 if (c->start_time == 1)
879 c->start_time = first_time;
881 if (c->total_time >= threshold && !power_only) {
886 if (c->end_time == 0)
887 c->end_time = last_time;
898 #define TIME_THRESH 10000000
900 static void write_svg_file(const char *filename)
908 count = determine_display_tasks(TIME_THRESH);
910 /* We'd like to show at least 15 tasks; be less picky if we have fewer */
912 count = determine_display_tasks(TIME_THRESH / 10);
914 open_svg(filename, numcpus, count, first_time, last_time);
919 for (i = 0; i < numcpus; i++)
920 svg_cpu_box(i, max_freq, turbo_frequency);
930 static struct perf_event_ops event_ops = {
931 .comm = process_comm_event,
932 .fork = process_fork_event,
933 .exit = process_exit_event,
934 .sample = process_sample_event,
935 .ordered_samples = true,
938 static int __cmd_timechart(void)
940 struct perf_session *session = perf_session__new(input_name, O_RDONLY, 0, false);
946 if (!perf_session__has_traces(session, "timechart record"))
949 ret = perf_session__process_events(session, &event_ops);
953 end_sample_processing();
957 write_svg_file(output_name);
959 pr_info("Written %2.1f seconds of trace to %s.\n",
960 (last_time - first_time) / 1000000000.0, output_name);
962 perf_session__delete(session);
966 static const char * const timechart_usage[] = {
967 "perf timechart [<options>] {record}",
971 static const char *record_args[] = {
977 "-e", "power:power_start",
978 "-e", "power:power_end",
979 "-e", "power:power_frequency",
980 "-e", "sched:sched_wakeup",
981 "-e", "sched:sched_switch",
984 static int __cmd_record(int argc, const char **argv)
986 unsigned int rec_argc, i, j;
987 const char **rec_argv;
989 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
990 rec_argv = calloc(rec_argc + 1, sizeof(char *));
992 for (i = 0; i < ARRAY_SIZE(record_args); i++)
993 rec_argv[i] = strdup(record_args[i]);
995 for (j = 1; j < (unsigned int)argc; j++, i++)
996 rec_argv[i] = argv[j];
998 return cmd_record(i, rec_argv, NULL);
1002 parse_process(const struct option *opt __used, const char *arg, int __used unset)
1005 add_process_filter(arg);
1009 static const struct option options[] = {
1010 OPT_STRING('i', "input", &input_name, "file",
1012 OPT_STRING('o', "output", &output_name, "file",
1013 "output file name"),
1014 OPT_INTEGER('w', "width", &svg_page_width,
1016 OPT_BOOLEAN('P', "power-only", &power_only,
1017 "output power data only"),
1018 OPT_CALLBACK('p', "process", NULL, "process",
1019 "process selector. Pass a pid or process name.",
1025 int cmd_timechart(int argc, const char **argv, const char *prefix __used)
1027 argc = parse_options(argc, argv, options, timechart_usage,
1028 PARSE_OPT_STOP_AT_NON_OPTION);
1032 if (argc && !strncmp(argv[0], "rec", 3))
1033 return __cmd_record(argc, argv);
1035 usage_with_options(timechart_usage, options);
1039 return __cmd_timechart();