2 * builtin-timechart.c - make an svg timechart of system activity
4 * (C) Copyright 2009 Intel Corporation
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 const char *input_name;
42 static const char *output_name = "output.svg";
44 static unsigned int numcpus;
45 static u64 min_freq; /* Lowest CPU frequency seen */
46 static u64 max_freq; /* Highest CPU frequency seen */
47 static u64 turbo_frequency;
49 static u64 first_time, last_time;
51 static bool power_only;
61 struct sample_wrapper;
64 * Datastructure layout:
65 * We keep an list of "pid"s, matching the kernels notion of a task struct.
66 * Each "pid" entry, has a list of "comm"s.
67 * this is because we want to track different programs different, while
68 * exec will reuse the original pid (by design).
69 * Each comm has a list of samples that will be used to draw
84 struct per_pidcomm *all;
85 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 process_filter;
153 struct process_filter {
156 struct process_filter *next;
159 static struct process_filter *process_filter;
162 static struct per_pid *find_create_pid(int pid)
164 struct per_pid *cursor = all_data;
167 if (cursor->pid == pid)
169 cursor = cursor->next;
171 cursor = malloc(sizeof(struct per_pid));
172 assert(cursor != NULL);
173 memset(cursor, 0, sizeof(struct per_pid));
175 cursor->next = all_data;
180 static void pid_set_comm(int pid, char *comm)
183 struct per_pidcomm *c;
184 p = find_create_pid(pid);
187 if (c->comm && strcmp(c->comm, comm) == 0) {
192 c->comm = strdup(comm);
198 c = malloc(sizeof(struct per_pidcomm));
200 memset(c, 0, sizeof(struct per_pidcomm));
201 c->comm = strdup(comm);
207 static void pid_fork(int pid, int ppid, u64 timestamp)
209 struct per_pid *p, *pp;
210 p = find_create_pid(pid);
211 pp = find_create_pid(ppid);
213 if (pp->current && pp->current->comm && !p->current)
214 pid_set_comm(pid, pp->current->comm);
216 p->start_time = timestamp;
218 p->current->start_time = timestamp;
219 p->current->state_since = timestamp;
223 static void pid_exit(int pid, u64 timestamp)
226 p = find_create_pid(pid);
227 p->end_time = timestamp;
229 p->current->end_time = timestamp;
233 pid_put_sample(int pid, int type, unsigned int cpu, u64 start, u64 end)
236 struct per_pidcomm *c;
237 struct cpu_sample *sample;
239 p = find_create_pid(pid);
242 c = malloc(sizeof(struct per_pidcomm));
244 memset(c, 0, sizeof(struct per_pidcomm));
250 sample = malloc(sizeof(struct cpu_sample));
251 assert(sample != NULL);
252 memset(sample, 0, sizeof(struct cpu_sample));
253 sample->start_time = start;
254 sample->end_time = end;
256 sample->next = c->samples;
260 if (sample->type == TYPE_RUNNING && end > start && start > 0) {
261 c->total_time += (end-start);
262 p->total_time += (end-start);
265 if (c->start_time == 0 || c->start_time > start)
266 c->start_time = start;
267 if (p->start_time == 0 || p->start_time > start)
268 p->start_time = start;
271 #define MAX_CPUS 4096
273 static u64 cpus_cstate_start_times[MAX_CPUS];
274 static int cpus_cstate_state[MAX_CPUS];
275 static u64 cpus_pstate_start_times[MAX_CPUS];
276 static u64 cpus_pstate_state[MAX_CPUS];
278 static int process_comm_event(struct perf_tool *tool __used,
279 union perf_event *event,
280 struct perf_sample *sample __used,
281 struct machine *machine __used)
283 pid_set_comm(event->comm.tid, event->comm.comm);
287 static int process_fork_event(struct perf_tool *tool __used,
288 union perf_event *event,
289 struct perf_sample *sample __used,
290 struct machine *machine __used)
292 pid_fork(event->fork.pid, event->fork.ppid, event->fork.time);
296 static int process_exit_event(struct perf_tool *tool __used,
297 union perf_event *event,
298 struct perf_sample *sample __used,
299 struct machine *machine __used)
301 pid_exit(event->fork.pid, event->fork.time);
308 unsigned char preempt_count;
313 #ifdef SUPPORT_OLD_POWER_EVENTS
314 static int use_old_power_events;
315 struct power_entry_old {
316 struct trace_entry te;
323 struct power_processor_entry {
324 struct trace_entry te;
329 #define TASK_COMM_LEN 16
330 struct wakeup_entry {
331 struct trace_entry te;
332 char comm[TASK_COMM_LEN];
339 * trace_flag_type is an enumeration that holds different
340 * states when a trace occurs. These are:
341 * IRQS_OFF - interrupts were disabled
342 * IRQS_NOSUPPORT - arch does not support irqs_disabled_flags
343 * NEED_RESCED - reschedule is requested
344 * HARDIRQ - inside an interrupt handler
345 * SOFTIRQ - inside a softirq handler
347 enum trace_flag_type {
348 TRACE_FLAG_IRQS_OFF = 0x01,
349 TRACE_FLAG_IRQS_NOSUPPORT = 0x02,
350 TRACE_FLAG_NEED_RESCHED = 0x04,
351 TRACE_FLAG_HARDIRQ = 0x08,
352 TRACE_FLAG_SOFTIRQ = 0x10,
357 struct sched_switch {
358 struct trace_entry te;
359 char prev_comm[TASK_COMM_LEN];
362 long prev_state; /* Arjan weeps. */
363 char next_comm[TASK_COMM_LEN];
368 static void c_state_start(int cpu, u64 timestamp, int state)
370 cpus_cstate_start_times[cpu] = timestamp;
371 cpus_cstate_state[cpu] = state;
374 static void c_state_end(int cpu, u64 timestamp)
376 struct power_event *pwr;
377 pwr = malloc(sizeof(struct power_event));
380 memset(pwr, 0, sizeof(struct power_event));
382 pwr->state = cpus_cstate_state[cpu];
383 pwr->start_time = cpus_cstate_start_times[cpu];
384 pwr->end_time = timestamp;
387 pwr->next = power_events;
392 static void p_state_change(int cpu, u64 timestamp, u64 new_freq)
394 struct power_event *pwr;
395 pwr = malloc(sizeof(struct power_event));
397 if (new_freq > 8000000) /* detect invalid data */
402 memset(pwr, 0, sizeof(struct power_event));
404 pwr->state = cpus_pstate_state[cpu];
405 pwr->start_time = cpus_pstate_start_times[cpu];
406 pwr->end_time = timestamp;
409 pwr->next = power_events;
411 if (!pwr->start_time)
412 pwr->start_time = first_time;
416 cpus_pstate_state[cpu] = new_freq;
417 cpus_pstate_start_times[cpu] = timestamp;
419 if ((u64)new_freq > max_freq)
422 if (new_freq < min_freq || min_freq == 0)
425 if (new_freq == max_freq - 1000)
426 turbo_frequency = max_freq;
430 sched_wakeup(int cpu, u64 timestamp, int pid, struct trace_entry *te)
432 struct wake_event *we;
434 struct wakeup_entry *wake = (void *)te;
436 we = malloc(sizeof(struct wake_event));
440 memset(we, 0, sizeof(struct wake_event));
441 we->time = timestamp;
444 if ((te->flags & TRACE_FLAG_HARDIRQ) || (te->flags & TRACE_FLAG_SOFTIRQ))
447 we->wakee = wake->pid;
448 we->next = wake_events;
450 p = find_create_pid(we->wakee);
452 if (p && p->current && p->current->state == TYPE_NONE) {
453 p->current->state_since = timestamp;
454 p->current->state = TYPE_WAITING;
456 if (p && p->current && p->current->state == TYPE_BLOCKED) {
457 pid_put_sample(p->pid, p->current->state, cpu, p->current->state_since, timestamp);
458 p->current->state_since = timestamp;
459 p->current->state = TYPE_WAITING;
463 static void sched_switch(int cpu, u64 timestamp, struct trace_entry *te)
465 struct per_pid *p = NULL, *prev_p;
466 struct sched_switch *sw = (void *)te;
469 prev_p = find_create_pid(sw->prev_pid);
471 p = find_create_pid(sw->next_pid);
473 if (prev_p->current && prev_p->current->state != TYPE_NONE)
474 pid_put_sample(sw->prev_pid, TYPE_RUNNING, cpu, prev_p->current->state_since, timestamp);
475 if (p && p->current) {
476 if (p->current->state != TYPE_NONE)
477 pid_put_sample(sw->next_pid, p->current->state, cpu, p->current->state_since, timestamp);
479 p->current->state_since = timestamp;
480 p->current->state = TYPE_RUNNING;
483 if (prev_p->current) {
484 prev_p->current->state = TYPE_NONE;
485 prev_p->current->state_since = timestamp;
486 if (sw->prev_state & 2)
487 prev_p->current->state = TYPE_BLOCKED;
488 if (sw->prev_state == 0)
489 prev_p->current->state = TYPE_WAITING;
494 static int process_sample_event(struct perf_tool *tool __used,
495 union perf_event *event __used,
496 struct perf_sample *sample,
497 struct perf_evsel *evsel,
498 struct machine *machine __used)
500 struct trace_entry *te;
502 if (evsel->attr.sample_type & PERF_SAMPLE_TIME) {
503 if (!first_time || first_time > sample->time)
504 first_time = sample->time;
505 if (last_time < sample->time)
506 last_time = sample->time;
509 te = (void *)sample->raw_data;
510 if ((evsel->attr.sample_type & PERF_SAMPLE_RAW) && sample->raw_size > 0) {
512 #ifdef SUPPORT_OLD_POWER_EVENTS
513 struct power_entry_old *peo;
517 * FIXME: use evsel, its already mapped from id to perf_evsel,
518 * remove perf_header__find_event infrastructure bits.
519 * Mapping all these "power:cpu_idle" strings to the tracepoint
520 * ID and then just comparing against evsel->attr.config.
524 * if (evsel->attr.config == power_cpu_idle_id)
526 event_str = perf_header__find_event(te->type);
531 if (sample->cpu > numcpus)
532 numcpus = sample->cpu;
534 if (strcmp(event_str, "power:cpu_idle") == 0) {
535 struct power_processor_entry *ppe = (void *)te;
536 if (ppe->state == (u32)PWR_EVENT_EXIT)
537 c_state_end(ppe->cpu_id, sample->time);
539 c_state_start(ppe->cpu_id, sample->time,
542 else if (strcmp(event_str, "power:cpu_frequency") == 0) {
543 struct power_processor_entry *ppe = (void *)te;
544 p_state_change(ppe->cpu_id, sample->time, ppe->state);
547 else if (strcmp(event_str, "sched:sched_wakeup") == 0)
548 sched_wakeup(sample->cpu, sample->time, sample->pid, te);
550 else if (strcmp(event_str, "sched:sched_switch") == 0)
551 sched_switch(sample->cpu, sample->time, te);
553 #ifdef SUPPORT_OLD_POWER_EVENTS
554 if (use_old_power_events) {
555 if (strcmp(event_str, "power:power_start") == 0)
556 c_state_start(peo->cpu_id, sample->time,
559 else if (strcmp(event_str, "power:power_end") == 0)
560 c_state_end(sample->cpu, sample->time);
562 else if (strcmp(event_str,
563 "power:power_frequency") == 0)
564 p_state_change(peo->cpu_id, sample->time,
573 * After the last sample we need to wrap up the current C/P state
574 * and close out each CPU for these.
576 static void end_sample_processing(void)
579 struct power_event *pwr;
581 for (cpu = 0; cpu <= numcpus; cpu++) {
582 pwr = malloc(sizeof(struct power_event));
585 memset(pwr, 0, sizeof(struct power_event));
589 pwr->state = cpus_cstate_state[cpu];
590 pwr->start_time = cpus_cstate_start_times[cpu];
591 pwr->end_time = last_time;
594 pwr->next = power_events;
600 pwr = malloc(sizeof(struct power_event));
603 memset(pwr, 0, sizeof(struct power_event));
605 pwr->state = cpus_pstate_state[cpu];
606 pwr->start_time = cpus_pstate_start_times[cpu];
607 pwr->end_time = last_time;
610 pwr->next = power_events;
612 if (!pwr->start_time)
613 pwr->start_time = first_time;
615 pwr->state = min_freq;
621 * Sort the pid datastructure
623 static void sort_pids(void)
625 struct per_pid *new_list, *p, *cursor, *prev;
626 /* sort by ppid first, then by pid, lowest to highest */
635 if (new_list == NULL) {
643 if (cursor->ppid > p->ppid ||
644 (cursor->ppid == p->ppid && cursor->pid > p->pid)) {
645 /* must insert before */
647 p->next = prev->next;
660 cursor = cursor->next;
669 static void draw_c_p_states(void)
671 struct power_event *pwr;
675 * two pass drawing so that the P state bars are on top of the C state blocks
678 if (pwr->type == CSTATE)
679 svg_cstate(pwr->cpu, pwr->start_time, pwr->end_time, pwr->state);
685 if (pwr->type == PSTATE) {
687 pwr->state = min_freq;
688 svg_pstate(pwr->cpu, pwr->start_time, pwr->end_time, pwr->state);
694 static void draw_wakeups(void)
696 struct wake_event *we;
698 struct per_pidcomm *c;
702 int from = 0, to = 0;
703 char *task_from = NULL, *task_to = NULL;
705 /* locate the column of the waker and wakee */
708 if (p->pid == we->waker || p->pid == we->wakee) {
711 if (c->Y && c->start_time <= we->time && c->end_time >= we->time) {
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);
725 if (p->pid == we->waker && !from) {
727 task_from = strdup(c->comm);
729 if (p->pid == we->wakee && !to) {
731 task_to = strdup(c->comm);
740 task_from = malloc(40);
741 sprintf(task_from, "[%i]", we->waker);
744 task_to = malloc(40);
745 sprintf(task_to, "[%i]", we->wakee);
749 svg_interrupt(we->time, to);
750 else if (from && to && abs(from - to) == 1)
751 svg_wakeline(we->time, from, to);
753 svg_partial_wakeline(we->time, from, task_from, to, task_to);
761 static void draw_cpu_usage(void)
764 struct per_pidcomm *c;
765 struct cpu_sample *sample;
772 if (sample->type == TYPE_RUNNING)
773 svg_process(sample->cpu, sample->start_time, sample->end_time, "sample", c->comm);
775 sample = sample->next;
783 static void draw_process_bars(void)
786 struct per_pidcomm *c;
787 struct cpu_sample *sample;
802 svg_box(Y, c->start_time, c->end_time, "process");
805 if (sample->type == TYPE_RUNNING)
806 svg_sample(Y, sample->cpu, sample->start_time, sample->end_time);
807 if (sample->type == TYPE_BLOCKED)
808 svg_box(Y, sample->start_time, sample->end_time, "blocked");
809 if (sample->type == TYPE_WAITING)
810 svg_waiting(Y, sample->start_time, sample->end_time);
811 sample = sample->next;
816 if (c->total_time > 5000000000) /* 5 seconds */
817 sprintf(comm, "%s:%i (%2.2fs)", c->comm, p->pid, c->total_time / 1000000000.0);
819 sprintf(comm, "%s:%i (%3.1fms)", c->comm, p->pid, c->total_time / 1000000.0);
821 svg_text(Y, c->start_time, comm);
831 static void add_process_filter(const char *string)
833 struct process_filter *filt;
836 pid = strtoull(string, NULL, 10);
837 filt = malloc(sizeof(struct process_filter));
841 filt->name = strdup(string);
843 filt->next = process_filter;
845 process_filter = filt;
848 static int passes_filter(struct per_pid *p, struct per_pidcomm *c)
850 struct process_filter *filt;
854 filt = process_filter;
856 if (filt->pid && p->pid == filt->pid)
858 if (strcmp(filt->name, c->comm) == 0)
865 static int determine_display_tasks_filtered(void)
868 struct per_pidcomm *c;
874 if (p->start_time == 1)
875 p->start_time = first_time;
877 /* no exit marker, task kept running to the end */
878 if (p->end_time == 0)
879 p->end_time = last_time;
886 if (c->start_time == 1)
887 c->start_time = first_time;
889 if (passes_filter(p, c)) {
895 if (c->end_time == 0)
896 c->end_time = last_time;
905 static int determine_display_tasks(u64 threshold)
908 struct per_pidcomm *c;
912 return determine_display_tasks_filtered();
917 if (p->start_time == 1)
918 p->start_time = first_time;
920 /* no exit marker, task kept running to the end */
921 if (p->end_time == 0)
922 p->end_time = last_time;
923 if (p->total_time >= threshold && !power_only)
931 if (c->start_time == 1)
932 c->start_time = first_time;
934 if (c->total_time >= threshold && !power_only) {
939 if (c->end_time == 0)
940 c->end_time = last_time;
951 #define TIME_THRESH 10000000
953 static void write_svg_file(const char *filename)
961 count = determine_display_tasks(TIME_THRESH);
963 /* We'd like to show at least 15 tasks; be less picky if we have fewer */
965 count = determine_display_tasks(TIME_THRESH / 10);
967 open_svg(filename, numcpus, count, first_time, last_time);
972 for (i = 0; i < numcpus; i++)
973 svg_cpu_box(i, max_freq, turbo_frequency);
983 static struct perf_tool perf_timechart = {
984 .comm = process_comm_event,
985 .fork = process_fork_event,
986 .exit = process_exit_event,
987 .sample = process_sample_event,
988 .ordered_samples = true,
991 static int __cmd_timechart(void)
993 struct perf_session *session = perf_session__new(input_name, O_RDONLY,
994 0, false, &perf_timechart);
1000 if (!perf_session__has_traces(session, "timechart record"))
1003 ret = perf_session__process_events(session, &perf_timechart);
1007 end_sample_processing();
1011 write_svg_file(output_name);
1013 pr_info("Written %2.1f seconds of trace to %s.\n",
1014 (last_time - first_time) / 1000000000.0, output_name);
1016 perf_session__delete(session);
1020 static const char * const timechart_usage[] = {
1021 "perf timechart [<options>] {record}",
1025 #ifdef SUPPORT_OLD_POWER_EVENTS
1026 static const char * const record_old_args[] = {
1032 "-e", "power:power_start",
1033 "-e", "power:power_end",
1034 "-e", "power:power_frequency",
1035 "-e", "sched:sched_wakeup",
1036 "-e", "sched:sched_switch",
1040 static const char * const record_new_args[] = {
1046 "-e", "power:cpu_frequency",
1047 "-e", "power:cpu_idle",
1048 "-e", "sched:sched_wakeup",
1049 "-e", "sched:sched_switch",
1052 static int __cmd_record(int argc, const char **argv)
1054 unsigned int rec_argc, i, j;
1055 const char **rec_argv;
1056 const char * const *record_args = record_new_args;
1057 unsigned int record_elems = ARRAY_SIZE(record_new_args);
1059 #ifdef SUPPORT_OLD_POWER_EVENTS
1060 if (!is_valid_tracepoint("power:cpu_idle") &&
1061 is_valid_tracepoint("power:power_start")) {
1062 use_old_power_events = 1;
1063 record_args = record_old_args;
1064 record_elems = ARRAY_SIZE(record_old_args);
1068 rec_argc = record_elems + argc - 1;
1069 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1071 if (rec_argv == NULL)
1074 for (i = 0; i < record_elems; i++)
1075 rec_argv[i] = strdup(record_args[i]);
1077 for (j = 1; j < (unsigned int)argc; j++, i++)
1078 rec_argv[i] = argv[j];
1080 return cmd_record(i, rec_argv, NULL);
1084 parse_process(const struct option *opt __used, const char *arg, int __used unset)
1087 add_process_filter(arg);
1091 static const struct option options[] = {
1092 OPT_STRING('i', "input", &input_name, "file",
1094 OPT_STRING('o', "output", &output_name, "file",
1095 "output file name"),
1096 OPT_INTEGER('w', "width", &svg_page_width,
1098 OPT_BOOLEAN('P', "power-only", &power_only,
1099 "output power data only"),
1100 OPT_CALLBACK('p', "process", NULL, "process",
1101 "process selector. Pass a pid or process name.",
1103 OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
1104 "Look for files with symbols relative to this directory"),
1109 int cmd_timechart(int argc, const char **argv, const char *prefix __used)
1111 argc = parse_options(argc, argv, options, timechart_usage,
1112 PARSE_OPT_STOP_AT_NON_OPTION);
1116 if (argc && !strncmp(argv[0], "rec", 3))
1117 return __cmd_record(argc, argv);
1119 usage_with_options(timechart_usage, options);
1123 return __cmd_timechart();