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 "util/evsel.h"
23 #include <linux/rbtree.h>
24 #include "util/symbol.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"
35 #include "util/tool.h"
37 #define SUPPORT_OLD_POWER_EVENTS 1
38 #define PWR_EVENT_EXIT -1
41 static unsigned int numcpus;
42 static u64 min_freq; /* Lowest CPU frequency seen */
43 static u64 max_freq; /* Highest CPU frequency seen */
44 static u64 turbo_frequency;
46 static u64 first_time, last_time;
48 static bool power_only;
58 struct sample_wrapper;
61 * Datastructure layout:
62 * We keep an list of "pid"s, matching the kernels notion of a task struct.
63 * Each "pid" entry, has a list of "comm"s.
64 * this is because we want to track different programs different, while
65 * exec will reuse the original pid (by design).
66 * Each comm has a list of samples that will be used to draw
81 struct per_pidcomm *all;
82 struct per_pidcomm *current;
87 struct per_pidcomm *next;
101 struct cpu_sample *samples;
104 struct sample_wrapper {
105 struct sample_wrapper *next;
108 unsigned char data[0];
112 #define TYPE_RUNNING 1
113 #define TYPE_WAITING 2
114 #define TYPE_BLOCKED 3
117 struct cpu_sample *next;
125 static struct per_pid *all_data;
131 struct power_event *next;
140 struct wake_event *next;
146 static struct power_event *power_events;
147 static struct wake_event *wake_events;
149 struct process_filter;
150 struct process_filter {
153 struct process_filter *next;
156 static struct process_filter *process_filter;
159 static struct per_pid *find_create_pid(int pid)
161 struct per_pid *cursor = all_data;
164 if (cursor->pid == pid)
166 cursor = cursor->next;
168 cursor = zalloc(sizeof(*cursor));
169 assert(cursor != NULL);
171 cursor->next = all_data;
176 static void pid_set_comm(int pid, char *comm)
179 struct per_pidcomm *c;
180 p = find_create_pid(pid);
183 if (c->comm && strcmp(c->comm, comm) == 0) {
188 c->comm = strdup(comm);
194 c = zalloc(sizeof(*c));
196 c->comm = strdup(comm);
202 static void pid_fork(int pid, int ppid, u64 timestamp)
204 struct per_pid *p, *pp;
205 p = find_create_pid(pid);
206 pp = find_create_pid(ppid);
208 if (pp->current && pp->current->comm && !p->current)
209 pid_set_comm(pid, pp->current->comm);
211 p->start_time = timestamp;
213 p->current->start_time = timestamp;
214 p->current->state_since = timestamp;
218 static void pid_exit(int pid, u64 timestamp)
221 p = find_create_pid(pid);
222 p->end_time = timestamp;
224 p->current->end_time = timestamp;
228 pid_put_sample(int pid, int type, unsigned int cpu, u64 start, u64 end)
231 struct per_pidcomm *c;
232 struct cpu_sample *sample;
234 p = find_create_pid(pid);
237 c = zalloc(sizeof(*c));
244 sample = zalloc(sizeof(*sample));
245 assert(sample != NULL);
246 sample->start_time = start;
247 sample->end_time = end;
249 sample->next = c->samples;
253 if (sample->type == TYPE_RUNNING && end > start && start > 0) {
254 c->total_time += (end-start);
255 p->total_time += (end-start);
258 if (c->start_time == 0 || c->start_time > start)
259 c->start_time = start;
260 if (p->start_time == 0 || p->start_time > start)
261 p->start_time = start;
264 #define MAX_CPUS 4096
266 static u64 cpus_cstate_start_times[MAX_CPUS];
267 static int cpus_cstate_state[MAX_CPUS];
268 static u64 cpus_pstate_start_times[MAX_CPUS];
269 static u64 cpus_pstate_state[MAX_CPUS];
271 static int process_comm_event(struct perf_tool *tool __maybe_unused,
272 union perf_event *event,
273 struct perf_sample *sample __maybe_unused,
274 struct machine *machine __maybe_unused)
276 pid_set_comm(event->comm.tid, event->comm.comm);
280 static int process_fork_event(struct perf_tool *tool __maybe_unused,
281 union perf_event *event,
282 struct perf_sample *sample __maybe_unused,
283 struct machine *machine __maybe_unused)
285 pid_fork(event->fork.pid, event->fork.ppid, event->fork.time);
289 static int process_exit_event(struct perf_tool *tool __maybe_unused,
290 union perf_event *event,
291 struct perf_sample *sample __maybe_unused,
292 struct machine *machine __maybe_unused)
294 pid_exit(event->fork.pid, event->fork.time);
301 unsigned char preempt_count;
306 #ifdef SUPPORT_OLD_POWER_EVENTS
307 static int use_old_power_events;
308 struct power_entry_old {
309 struct trace_entry te;
316 struct power_processor_entry {
317 struct trace_entry te;
322 #define TASK_COMM_LEN 16
323 struct wakeup_entry {
324 struct trace_entry te;
325 char comm[TASK_COMM_LEN];
332 * trace_flag_type is an enumeration that holds different
333 * states when a trace occurs. These are:
334 * IRQS_OFF - interrupts were disabled
335 * IRQS_NOSUPPORT - arch does not support irqs_disabled_flags
336 * NEED_RESCED - reschedule is requested
337 * HARDIRQ - inside an interrupt handler
338 * SOFTIRQ - inside a softirq handler
340 enum trace_flag_type {
341 TRACE_FLAG_IRQS_OFF = 0x01,
342 TRACE_FLAG_IRQS_NOSUPPORT = 0x02,
343 TRACE_FLAG_NEED_RESCHED = 0x04,
344 TRACE_FLAG_HARDIRQ = 0x08,
345 TRACE_FLAG_SOFTIRQ = 0x10,
350 struct sched_switch {
351 struct trace_entry te;
352 char prev_comm[TASK_COMM_LEN];
355 long prev_state; /* Arjan weeps. */
356 char next_comm[TASK_COMM_LEN];
361 static void c_state_start(int cpu, u64 timestamp, int state)
363 cpus_cstate_start_times[cpu] = timestamp;
364 cpus_cstate_state[cpu] = state;
367 static void c_state_end(int cpu, u64 timestamp)
369 struct power_event *pwr = zalloc(sizeof(*pwr));
374 pwr->state = cpus_cstate_state[cpu];
375 pwr->start_time = cpus_cstate_start_times[cpu];
376 pwr->end_time = timestamp;
379 pwr->next = power_events;
384 static void p_state_change(int cpu, u64 timestamp, u64 new_freq)
386 struct power_event *pwr;
388 if (new_freq > 8000000) /* detect invalid data */
391 pwr = zalloc(sizeof(*pwr));
395 pwr->state = cpus_pstate_state[cpu];
396 pwr->start_time = cpus_pstate_start_times[cpu];
397 pwr->end_time = timestamp;
400 pwr->next = power_events;
402 if (!pwr->start_time)
403 pwr->start_time = first_time;
407 cpus_pstate_state[cpu] = new_freq;
408 cpus_pstate_start_times[cpu] = timestamp;
410 if ((u64)new_freq > max_freq)
413 if (new_freq < min_freq || min_freq == 0)
416 if (new_freq == max_freq - 1000)
417 turbo_frequency = max_freq;
421 sched_wakeup(int cpu, u64 timestamp, int pid, struct trace_entry *te)
424 struct wakeup_entry *wake = (void *)te;
425 struct wake_event *we = zalloc(sizeof(*we));
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;
483 static int process_sample_event(struct perf_tool *tool __maybe_unused,
484 union perf_event *event __maybe_unused,
485 struct perf_sample *sample,
486 struct perf_evsel *evsel,
487 struct machine *machine __maybe_unused)
489 struct trace_entry *te;
491 if (evsel->attr.sample_type & PERF_SAMPLE_TIME) {
492 if (!first_time || first_time > sample->time)
493 first_time = sample->time;
494 if (last_time < sample->time)
495 last_time = sample->time;
498 te = (void *)sample->raw_data;
499 if ((evsel->attr.sample_type & PERF_SAMPLE_RAW) && sample->raw_size > 0) {
501 #ifdef SUPPORT_OLD_POWER_EVENTS
502 struct power_entry_old *peo;
506 * FIXME: use evsel, its already mapped from id to perf_evsel,
507 * remove perf_header__find_event infrastructure bits.
508 * Mapping all these "power:cpu_idle" strings to the tracepoint
509 * ID and then just comparing against evsel->attr.config.
513 * if (evsel->attr.config == power_cpu_idle_id)
515 event_str = perf_header__find_event(te->type);
520 if (sample->cpu > numcpus)
521 numcpus = sample->cpu;
523 if (strcmp(event_str, "power:cpu_idle") == 0) {
524 struct power_processor_entry *ppe = (void *)te;
525 if (ppe->state == (u32)PWR_EVENT_EXIT)
526 c_state_end(ppe->cpu_id, sample->time);
528 c_state_start(ppe->cpu_id, sample->time,
531 else if (strcmp(event_str, "power:cpu_frequency") == 0) {
532 struct power_processor_entry *ppe = (void *)te;
533 p_state_change(ppe->cpu_id, sample->time, ppe->state);
536 else if (strcmp(event_str, "sched:sched_wakeup") == 0)
537 sched_wakeup(sample->cpu, sample->time, sample->pid, te);
539 else if (strcmp(event_str, "sched:sched_switch") == 0)
540 sched_switch(sample->cpu, sample->time, te);
542 #ifdef SUPPORT_OLD_POWER_EVENTS
543 if (use_old_power_events) {
544 if (strcmp(event_str, "power:power_start") == 0)
545 c_state_start(peo->cpu_id, sample->time,
548 else if (strcmp(event_str, "power:power_end") == 0)
549 c_state_end(sample->cpu, sample->time);
551 else if (strcmp(event_str,
552 "power:power_frequency") == 0)
553 p_state_change(peo->cpu_id, sample->time,
562 * After the last sample we need to wrap up the current C/P state
563 * and close out each CPU for these.
565 static void end_sample_processing(void)
568 struct power_event *pwr;
570 for (cpu = 0; cpu <= numcpus; cpu++) {
573 pwr = zalloc(sizeof(*pwr));
577 pwr->state = cpus_cstate_state[cpu];
578 pwr->start_time = cpus_cstate_start_times[cpu];
579 pwr->end_time = last_time;
582 pwr->next = power_events;
588 pwr = zalloc(sizeof(*pwr));
592 pwr->state = cpus_pstate_state[cpu];
593 pwr->start_time = cpus_pstate_start_times[cpu];
594 pwr->end_time = last_time;
597 pwr->next = power_events;
599 if (!pwr->start_time)
600 pwr->start_time = first_time;
602 pwr->state = min_freq;
608 * Sort the pid datastructure
610 static void sort_pids(void)
612 struct per_pid *new_list, *p, *cursor, *prev;
613 /* sort by ppid first, then by pid, lowest to highest */
622 if (new_list == NULL) {
630 if (cursor->ppid > p->ppid ||
631 (cursor->ppid == p->ppid && cursor->pid > p->pid)) {
632 /* must insert before */
634 p->next = prev->next;
647 cursor = cursor->next;
656 static void draw_c_p_states(void)
658 struct power_event *pwr;
662 * two pass drawing so that the P state bars are on top of the C state blocks
665 if (pwr->type == CSTATE)
666 svg_cstate(pwr->cpu, pwr->start_time, pwr->end_time, pwr->state);
672 if (pwr->type == PSTATE) {
674 pwr->state = min_freq;
675 svg_pstate(pwr->cpu, pwr->start_time, pwr->end_time, pwr->state);
681 static void draw_wakeups(void)
683 struct wake_event *we;
685 struct per_pidcomm *c;
689 int from = 0, to = 0;
690 char *task_from = NULL, *task_to = NULL;
692 /* locate the column of the waker and wakee */
695 if (p->pid == we->waker || p->pid == we->wakee) {
698 if (c->Y && c->start_time <= we->time && c->end_time >= we->time) {
699 if (p->pid == we->waker && !from) {
701 task_from = strdup(c->comm);
703 if (p->pid == we->wakee && !to) {
705 task_to = strdup(c->comm);
712 if (p->pid == we->waker && !from) {
714 task_from = strdup(c->comm);
716 if (p->pid == we->wakee && !to) {
718 task_to = strdup(c->comm);
727 task_from = malloc(40);
728 sprintf(task_from, "[%i]", we->waker);
731 task_to = malloc(40);
732 sprintf(task_to, "[%i]", we->wakee);
736 svg_interrupt(we->time, to);
737 else if (from && to && abs(from - to) == 1)
738 svg_wakeline(we->time, from, to);
740 svg_partial_wakeline(we->time, from, task_from, to, task_to);
748 static void draw_cpu_usage(void)
751 struct per_pidcomm *c;
752 struct cpu_sample *sample;
759 if (sample->type == TYPE_RUNNING)
760 svg_process(sample->cpu, sample->start_time, sample->end_time, "sample", c->comm);
762 sample = sample->next;
770 static void draw_process_bars(void)
773 struct per_pidcomm *c;
774 struct cpu_sample *sample;
789 svg_box(Y, c->start_time, c->end_time, "process");
792 if (sample->type == TYPE_RUNNING)
793 svg_sample(Y, sample->cpu, sample->start_time, sample->end_time);
794 if (sample->type == TYPE_BLOCKED)
795 svg_box(Y, sample->start_time, sample->end_time, "blocked");
796 if (sample->type == TYPE_WAITING)
797 svg_waiting(Y, sample->start_time, sample->end_time);
798 sample = sample->next;
803 if (c->total_time > 5000000000) /* 5 seconds */
804 sprintf(comm, "%s:%i (%2.2fs)", c->comm, p->pid, c->total_time / 1000000000.0);
806 sprintf(comm, "%s:%i (%3.1fms)", c->comm, p->pid, c->total_time / 1000000.0);
808 svg_text(Y, c->start_time, comm);
818 static void add_process_filter(const char *string)
820 int pid = strtoull(string, NULL, 10);
821 struct process_filter *filt = malloc(sizeof(*filt));
826 filt->name = strdup(string);
828 filt->next = process_filter;
830 process_filter = filt;
833 static int passes_filter(struct per_pid *p, struct per_pidcomm *c)
835 struct process_filter *filt;
839 filt = process_filter;
841 if (filt->pid && p->pid == filt->pid)
843 if (strcmp(filt->name, c->comm) == 0)
850 static int determine_display_tasks_filtered(void)
853 struct per_pidcomm *c;
859 if (p->start_time == 1)
860 p->start_time = first_time;
862 /* no exit marker, task kept running to the end */
863 if (p->end_time == 0)
864 p->end_time = last_time;
871 if (c->start_time == 1)
872 c->start_time = first_time;
874 if (passes_filter(p, c)) {
880 if (c->end_time == 0)
881 c->end_time = last_time;
890 static int determine_display_tasks(u64 threshold)
893 struct per_pidcomm *c;
897 return determine_display_tasks_filtered();
902 if (p->start_time == 1)
903 p->start_time = first_time;
905 /* no exit marker, task kept running to the end */
906 if (p->end_time == 0)
907 p->end_time = last_time;
908 if (p->total_time >= threshold && !power_only)
916 if (c->start_time == 1)
917 c->start_time = first_time;
919 if (c->total_time >= threshold && !power_only) {
924 if (c->end_time == 0)
925 c->end_time = last_time;
936 #define TIME_THRESH 10000000
938 static void write_svg_file(const char *filename)
946 count = determine_display_tasks(TIME_THRESH);
948 /* We'd like to show at least 15 tasks; be less picky if we have fewer */
950 count = determine_display_tasks(TIME_THRESH / 10);
952 open_svg(filename, numcpus, count, first_time, last_time);
957 for (i = 0; i < numcpus; i++)
958 svg_cpu_box(i, max_freq, turbo_frequency);
968 static int __cmd_timechart(const char *output_name)
970 struct perf_tool perf_timechart = {
971 .comm = process_comm_event,
972 .fork = process_fork_event,
973 .exit = process_exit_event,
974 .sample = process_sample_event,
975 .ordered_samples = true,
977 struct perf_session *session = perf_session__new(input_name, O_RDONLY,
978 0, false, &perf_timechart);
984 if (!perf_session__has_traces(session, "timechart record"))
987 ret = perf_session__process_events(session, &perf_timechart);
991 end_sample_processing();
995 write_svg_file(output_name);
997 pr_info("Written %2.1f seconds of trace to %s.\n",
998 (last_time - first_time) / 1000000000.0, output_name);
1000 perf_session__delete(session);
1004 static int __cmd_record(int argc, const char **argv)
1006 #ifdef SUPPORT_OLD_POWER_EVENTS
1007 const char * const record_old_args[] = {
1008 "record", "-a", "-R", "-f", "-c", "1",
1009 "-e", "power:power_start",
1010 "-e", "power:power_end",
1011 "-e", "power:power_frequency",
1012 "-e", "sched:sched_wakeup",
1013 "-e", "sched:sched_switch",
1016 const char * const record_new_args[] = {
1017 "record", "-a", "-R", "-f", "-c", "1",
1018 "-e", "power:cpu_frequency",
1019 "-e", "power:cpu_idle",
1020 "-e", "sched:sched_wakeup",
1021 "-e", "sched:sched_switch",
1023 unsigned int rec_argc, i, j;
1024 const char **rec_argv;
1025 const char * const *record_args = record_new_args;
1026 unsigned int record_elems = ARRAY_SIZE(record_new_args);
1028 #ifdef SUPPORT_OLD_POWER_EVENTS
1029 if (!is_valid_tracepoint("power:cpu_idle") &&
1030 is_valid_tracepoint("power:power_start")) {
1031 use_old_power_events = 1;
1032 record_args = record_old_args;
1033 record_elems = ARRAY_SIZE(record_old_args);
1037 rec_argc = record_elems + argc - 1;
1038 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1040 if (rec_argv == NULL)
1043 for (i = 0; i < record_elems; i++)
1044 rec_argv[i] = strdup(record_args[i]);
1046 for (j = 1; j < (unsigned int)argc; j++, i++)
1047 rec_argv[i] = argv[j];
1049 return cmd_record(i, rec_argv, NULL);
1053 parse_process(const struct option *opt __maybe_unused, const char *arg,
1054 int __maybe_unused unset)
1057 add_process_filter(arg);
1061 int cmd_timechart(int argc, const char **argv,
1062 const char *prefix __maybe_unused)
1064 const char *output_name = "output.svg";
1065 const struct option options[] = {
1066 OPT_STRING('i', "input", &input_name, "file", "input file name"),
1067 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1068 OPT_INTEGER('w', "width", &svg_page_width, "page width"),
1069 OPT_BOOLEAN('P', "power-only", &power_only, "output power data only"),
1070 OPT_CALLBACK('p', "process", NULL, "process",
1071 "process selector. Pass a pid or process name.",
1073 OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
1074 "Look for files with symbols relative to this directory"),
1077 const char * const timechart_usage[] = {
1078 "perf timechart [<options>] {record}",
1082 argc = parse_options(argc, argv, options, timechart_usage,
1083 PARSE_OPT_STOP_AT_NON_OPTION);
1087 if (argc && !strncmp(argv[0], "rec", 3))
1088 return __cmd_record(argc, argv);
1090 usage_with_options(timechart_usage, options);
1094 return __cmd_timechart(output_name);