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(event_t *event, struct sample_data *sample __used,
277 struct perf_session *session __used)
279 pid_set_comm(event->comm.tid, event->comm.comm);
283 static int process_fork_event(event_t *event, struct sample_data *sample __used,
284 struct perf_session *session __used)
286 pid_fork(event->fork.pid, event->fork.ppid, event->fork.time);
290 static int process_exit_event(event_t *event, struct sample_data *sample __used,
291 struct perf_session *session __used)
293 pid_exit(event->fork.pid, event->fork.time);
300 unsigned char preempt_count;
305 #ifdef SUPPORT_OLD_POWER_EVENTS
306 static int use_old_power_events;
307 struct power_entry_old {
308 struct trace_entry te;
315 struct power_processor_entry {
316 struct trace_entry te;
321 #define TASK_COMM_LEN 16
322 struct wakeup_entry {
323 struct trace_entry te;
324 char comm[TASK_COMM_LEN];
331 * trace_flag_type is an enumeration that holds different
332 * states when a trace occurs. These are:
333 * IRQS_OFF - interrupts were disabled
334 * IRQS_NOSUPPORT - arch does not support irqs_disabled_flags
335 * NEED_RESCED - reschedule is requested
336 * HARDIRQ - inside an interrupt handler
337 * SOFTIRQ - inside a softirq handler
339 enum trace_flag_type {
340 TRACE_FLAG_IRQS_OFF = 0x01,
341 TRACE_FLAG_IRQS_NOSUPPORT = 0x02,
342 TRACE_FLAG_NEED_RESCHED = 0x04,
343 TRACE_FLAG_HARDIRQ = 0x08,
344 TRACE_FLAG_SOFTIRQ = 0x10,
349 struct sched_switch {
350 struct trace_entry te;
351 char prev_comm[TASK_COMM_LEN];
354 long prev_state; /* Arjan weeps. */
355 char next_comm[TASK_COMM_LEN];
360 static void c_state_start(int cpu, u64 timestamp, int state)
362 cpus_cstate_start_times[cpu] = timestamp;
363 cpus_cstate_state[cpu] = state;
366 static void c_state_end(int cpu, u64 timestamp)
368 struct power_event *pwr;
369 pwr = malloc(sizeof(struct power_event));
372 memset(pwr, 0, sizeof(struct power_event));
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;
387 pwr = malloc(sizeof(struct power_event));
389 if (new_freq > 8000000) /* detect invalid data */
394 memset(pwr, 0, sizeof(struct power_event));
396 pwr->state = cpus_pstate_state[cpu];
397 pwr->start_time = cpus_pstate_start_times[cpu];
398 pwr->end_time = timestamp;
401 pwr->next = power_events;
403 if (!pwr->start_time)
404 pwr->start_time = first_time;
408 cpus_pstate_state[cpu] = new_freq;
409 cpus_pstate_start_times[cpu] = timestamp;
411 if ((u64)new_freq > max_freq)
414 if (new_freq < min_freq || min_freq == 0)
417 if (new_freq == max_freq - 1000)
418 turbo_frequency = max_freq;
422 sched_wakeup(int cpu, u64 timestamp, int pid, struct trace_entry *te)
424 struct wake_event *we;
426 struct wakeup_entry *wake = (void *)te;
428 we = malloc(sizeof(struct wake_event));
432 memset(we, 0, sizeof(struct wake_event));
433 we->time = timestamp;
436 if ((te->flags & TRACE_FLAG_HARDIRQ) || (te->flags & TRACE_FLAG_SOFTIRQ))
439 we->wakee = wake->pid;
440 we->next = wake_events;
442 p = find_create_pid(we->wakee);
444 if (p && p->current && p->current->state == TYPE_NONE) {
445 p->current->state_since = timestamp;
446 p->current->state = TYPE_WAITING;
448 if (p && p->current && p->current->state == TYPE_BLOCKED) {
449 pid_put_sample(p->pid, p->current->state, cpu, p->current->state_since, timestamp);
450 p->current->state_since = timestamp;
451 p->current->state = TYPE_WAITING;
455 static void sched_switch(int cpu, u64 timestamp, struct trace_entry *te)
457 struct per_pid *p = NULL, *prev_p;
458 struct sched_switch *sw = (void *)te;
461 prev_p = find_create_pid(sw->prev_pid);
463 p = find_create_pid(sw->next_pid);
465 if (prev_p->current && prev_p->current->state != TYPE_NONE)
466 pid_put_sample(sw->prev_pid, TYPE_RUNNING, cpu, prev_p->current->state_since, timestamp);
467 if (p && p->current) {
468 if (p->current->state != TYPE_NONE)
469 pid_put_sample(sw->next_pid, p->current->state, cpu, p->current->state_since, timestamp);
471 p->current->state_since = timestamp;
472 p->current->state = TYPE_RUNNING;
475 if (prev_p->current) {
476 prev_p->current->state = TYPE_NONE;
477 prev_p->current->state_since = timestamp;
478 if (sw->prev_state & 2)
479 prev_p->current->state = TYPE_BLOCKED;
480 if (sw->prev_state == 0)
481 prev_p->current->state = TYPE_WAITING;
486 static int process_sample_event(event_t *event __used,
487 struct sample_data *sample,
488 struct perf_session *session)
490 struct trace_entry *te;
492 if (session->sample_type & PERF_SAMPLE_TIME) {
493 if (!first_time || first_time > sample->time)
494 first_time = sample->time;
495 if (last_time < sample->time)
496 last_time = sample->time;
499 te = (void *)sample->raw_data;
500 if (session->sample_type & PERF_SAMPLE_RAW && sample->raw_size > 0) {
502 #ifdef SUPPORT_OLD_POWER_EVENTS
503 struct power_entry_old *peo;
506 event_str = perf_header__find_event(te->type);
511 if (sample->cpu > numcpus)
512 numcpus = sample->cpu;
514 if (strcmp(event_str, "power:cpu_idle") == 0) {
515 struct power_processor_entry *ppe = (void *)te;
516 if (ppe->state == (u32)PWR_EVENT_EXIT)
517 c_state_end(ppe->cpu_id, sample->time);
519 c_state_start(ppe->cpu_id, sample->time,
522 else if (strcmp(event_str, "power:cpu_frequency") == 0) {
523 struct power_processor_entry *ppe = (void *)te;
524 p_state_change(ppe->cpu_id, sample->time, ppe->state);
527 else if (strcmp(event_str, "sched:sched_wakeup") == 0)
528 sched_wakeup(sample->cpu, sample->time, sample->pid, te);
530 else if (strcmp(event_str, "sched:sched_switch") == 0)
531 sched_switch(sample->cpu, sample->time, te);
533 #ifdef SUPPORT_OLD_POWER_EVENTS
534 if (use_old_power_events) {
535 if (strcmp(event_str, "power:power_start") == 0)
536 c_state_start(peo->cpu_id, sample->time,
539 else if (strcmp(event_str, "power:power_end") == 0)
540 c_state_end(sample->cpu, sample->time);
542 else if (strcmp(event_str,
543 "power:power_frequency") == 0)
544 p_state_change(peo->cpu_id, sample->time,
553 * After the last sample we need to wrap up the current C/P state
554 * and close out each CPU for these.
556 static void end_sample_processing(void)
559 struct power_event *pwr;
561 for (cpu = 0; cpu <= numcpus; cpu++) {
562 pwr = malloc(sizeof(struct power_event));
565 memset(pwr, 0, sizeof(struct power_event));
569 pwr->state = cpus_cstate_state[cpu];
570 pwr->start_time = cpus_cstate_start_times[cpu];
571 pwr->end_time = last_time;
574 pwr->next = power_events;
580 pwr = malloc(sizeof(struct power_event));
583 memset(pwr, 0, sizeof(struct power_event));
585 pwr->state = cpus_pstate_state[cpu];
586 pwr->start_time = cpus_pstate_start_times[cpu];
587 pwr->end_time = last_time;
590 pwr->next = power_events;
592 if (!pwr->start_time)
593 pwr->start_time = first_time;
595 pwr->state = min_freq;
601 * Sort the pid datastructure
603 static void sort_pids(void)
605 struct per_pid *new_list, *p, *cursor, *prev;
606 /* sort by ppid first, then by pid, lowest to highest */
615 if (new_list == NULL) {
623 if (cursor->ppid > p->ppid ||
624 (cursor->ppid == p->ppid && cursor->pid > p->pid)) {
625 /* must insert before */
627 p->next = prev->next;
640 cursor = cursor->next;
649 static void draw_c_p_states(void)
651 struct power_event *pwr;
655 * two pass drawing so that the P state bars are on top of the C state blocks
658 if (pwr->type == CSTATE)
659 svg_cstate(pwr->cpu, pwr->start_time, pwr->end_time, pwr->state);
665 if (pwr->type == PSTATE) {
667 pwr->state = min_freq;
668 svg_pstate(pwr->cpu, pwr->start_time, pwr->end_time, pwr->state);
674 static void draw_wakeups(void)
676 struct wake_event *we;
678 struct per_pidcomm *c;
682 int from = 0, to = 0;
683 char *task_from = NULL, *task_to = NULL;
685 /* locate the column of the waker and wakee */
688 if (p->pid == we->waker || p->pid == we->wakee) {
691 if (c->Y && c->start_time <= we->time && c->end_time >= we->time) {
692 if (p->pid == we->waker && !from) {
694 task_from = strdup(c->comm);
696 if (p->pid == we->wakee && !to) {
698 task_to = strdup(c->comm);
705 if (p->pid == we->waker && !from) {
707 task_from = strdup(c->comm);
709 if (p->pid == we->wakee && !to) {
711 task_to = strdup(c->comm);
720 task_from = malloc(40);
721 sprintf(task_from, "[%i]", we->waker);
724 task_to = malloc(40);
725 sprintf(task_to, "[%i]", we->wakee);
729 svg_interrupt(we->time, to);
730 else if (from && to && abs(from - to) == 1)
731 svg_wakeline(we->time, from, to);
733 svg_partial_wakeline(we->time, from, task_from, to, task_to);
741 static void draw_cpu_usage(void)
744 struct per_pidcomm *c;
745 struct cpu_sample *sample;
752 if (sample->type == TYPE_RUNNING)
753 svg_process(sample->cpu, sample->start_time, sample->end_time, "sample", c->comm);
755 sample = sample->next;
763 static void draw_process_bars(void)
766 struct per_pidcomm *c;
767 struct cpu_sample *sample;
782 svg_box(Y, c->start_time, c->end_time, "process");
785 if (sample->type == TYPE_RUNNING)
786 svg_sample(Y, sample->cpu, sample->start_time, sample->end_time);
787 if (sample->type == TYPE_BLOCKED)
788 svg_box(Y, sample->start_time, sample->end_time, "blocked");
789 if (sample->type == TYPE_WAITING)
790 svg_waiting(Y, sample->start_time, sample->end_time);
791 sample = sample->next;
796 if (c->total_time > 5000000000) /* 5 seconds */
797 sprintf(comm, "%s:%i (%2.2fs)", c->comm, p->pid, c->total_time / 1000000000.0);
799 sprintf(comm, "%s:%i (%3.1fms)", c->comm, p->pid, c->total_time / 1000000.0);
801 svg_text(Y, c->start_time, comm);
811 static void add_process_filter(const char *string)
813 struct process_filter *filt;
816 pid = strtoull(string, NULL, 10);
817 filt = malloc(sizeof(struct process_filter));
821 filt->name = strdup(string);
823 filt->next = process_filter;
825 process_filter = filt;
828 static int passes_filter(struct per_pid *p, struct per_pidcomm *c)
830 struct process_filter *filt;
834 filt = process_filter;
836 if (filt->pid && p->pid == filt->pid)
838 if (strcmp(filt->name, c->comm) == 0)
845 static int determine_display_tasks_filtered(void)
848 struct per_pidcomm *c;
854 if (p->start_time == 1)
855 p->start_time = first_time;
857 /* no exit marker, task kept running to the end */
858 if (p->end_time == 0)
859 p->end_time = last_time;
866 if (c->start_time == 1)
867 c->start_time = first_time;
869 if (passes_filter(p, c)) {
875 if (c->end_time == 0)
876 c->end_time = last_time;
885 static int determine_display_tasks(u64 threshold)
888 struct per_pidcomm *c;
892 return determine_display_tasks_filtered();
897 if (p->start_time == 1)
898 p->start_time = first_time;
900 /* no exit marker, task kept running to the end */
901 if (p->end_time == 0)
902 p->end_time = last_time;
903 if (p->total_time >= threshold && !power_only)
911 if (c->start_time == 1)
912 c->start_time = first_time;
914 if (c->total_time >= threshold && !power_only) {
919 if (c->end_time == 0)
920 c->end_time = last_time;
931 #define TIME_THRESH 10000000
933 static void write_svg_file(const char *filename)
941 count = determine_display_tasks(TIME_THRESH);
943 /* We'd like to show at least 15 tasks; be less picky if we have fewer */
945 count = determine_display_tasks(TIME_THRESH / 10);
947 open_svg(filename, numcpus, count, first_time, last_time);
952 for (i = 0; i < numcpus; i++)
953 svg_cpu_box(i, max_freq, turbo_frequency);
963 static struct perf_event_ops event_ops = {
964 .comm = process_comm_event,
965 .fork = process_fork_event,
966 .exit = process_exit_event,
967 .sample = process_sample_event,
968 .ordered_samples = true,
971 static int __cmd_timechart(void)
973 struct perf_session *session = perf_session__new(input_name, O_RDONLY,
974 0, false, &event_ops);
980 if (!perf_session__has_traces(session, "timechart record"))
983 ret = perf_session__process_events(session, &event_ops);
987 end_sample_processing();
991 write_svg_file(output_name);
993 pr_info("Written %2.1f seconds of trace to %s.\n",
994 (last_time - first_time) / 1000000000.0, output_name);
996 perf_session__delete(session);
1000 static const char * const timechart_usage[] = {
1001 "perf timechart [<options>] {record}",
1005 #ifdef SUPPORT_OLD_POWER_EVENTS
1006 static const char * const record_old_args[] = {
1012 "-e", "power:power_start",
1013 "-e", "power:power_end",
1014 "-e", "power:power_frequency",
1015 "-e", "sched:sched_wakeup",
1016 "-e", "sched:sched_switch",
1020 static const char * const record_new_args[] = {
1026 "-e", "power:cpu_frequency",
1027 "-e", "power:cpu_idle",
1028 "-e", "sched:sched_wakeup",
1029 "-e", "sched:sched_switch",
1032 static int __cmd_record(int argc, const char **argv)
1034 unsigned int rec_argc, i, j;
1035 const char **rec_argv;
1036 const char * const *record_args = record_new_args;
1037 unsigned int record_elems = ARRAY_SIZE(record_new_args);
1039 #ifdef SUPPORT_OLD_POWER_EVENTS
1040 if (!is_valid_tracepoint("power:cpu_idle") &&
1041 is_valid_tracepoint("power:power_start")) {
1042 use_old_power_events = 1;
1043 record_args = record_old_args;
1044 record_elems = ARRAY_SIZE(record_old_args);
1048 rec_argc = record_elems + argc - 1;
1049 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1051 if (rec_argv == NULL)
1054 for (i = 0; i < record_elems; i++)
1055 rec_argv[i] = strdup(record_args[i]);
1057 for (j = 1; j < (unsigned int)argc; j++, i++)
1058 rec_argv[i] = argv[j];
1060 return cmd_record(i, rec_argv, NULL);
1064 parse_process(const struct option *opt __used, const char *arg, int __used unset)
1067 add_process_filter(arg);
1071 static const struct option options[] = {
1072 OPT_STRING('i', "input", &input_name, "file",
1074 OPT_STRING('o', "output", &output_name, "file",
1075 "output file name"),
1076 OPT_INTEGER('w', "width", &svg_page_width,
1078 OPT_BOOLEAN('P', "power-only", &power_only,
1079 "output power data only"),
1080 OPT_CALLBACK('p', "process", NULL, "process",
1081 "process selector. Pass a pid or process name.",
1083 OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
1084 "Look for files with symbols relative to this directory"),
1089 int cmd_timechart(int argc, const char **argv, const char *prefix __used)
1091 argc = parse_options(argc, argv, options, timechart_usage,
1092 PARSE_OPT_STOP_AT_NON_OPTION);
1096 if (argc && !strncmp(argv[0], "rec", 3))
1097 return __cmd_record(argc, argv);
1099 usage_with_options(timechart_usage, options);
1103 return __cmd_timechart();