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;
307 #define TASK_COMM_LEN 16
308 struct wakeup_entry {
309 struct trace_entry te;
310 char comm[TASK_COMM_LEN];
317 * trace_flag_type is an enumeration that holds different
318 * states when a trace occurs. These are:
319 * IRQS_OFF - interrupts were disabled
320 * IRQS_NOSUPPORT - arch does not support irqs_disabled_flags
321 * NEED_RESCED - reschedule is requested
322 * HARDIRQ - inside an interrupt handler
323 * SOFTIRQ - inside a softirq handler
325 enum trace_flag_type {
326 TRACE_FLAG_IRQS_OFF = 0x01,
327 TRACE_FLAG_IRQS_NOSUPPORT = 0x02,
328 TRACE_FLAG_NEED_RESCHED = 0x04,
329 TRACE_FLAG_HARDIRQ = 0x08,
330 TRACE_FLAG_SOFTIRQ = 0x10,
335 struct sched_switch {
336 struct trace_entry te;
337 char prev_comm[TASK_COMM_LEN];
340 long prev_state; /* Arjan weeps. */
341 char next_comm[TASK_COMM_LEN];
346 static void c_state_start(int cpu, u64 timestamp, int state)
348 cpus_cstate_start_times[cpu] = timestamp;
349 cpus_cstate_state[cpu] = state;
352 static void c_state_end(int cpu, u64 timestamp)
354 struct power_event *pwr;
355 pwr = malloc(sizeof(struct power_event));
358 memset(pwr, 0, sizeof(struct power_event));
360 pwr->state = cpus_cstate_state[cpu];
361 pwr->start_time = cpus_cstate_start_times[cpu];
362 pwr->end_time = timestamp;
365 pwr->next = power_events;
370 static void p_state_change(int cpu, u64 timestamp, u64 new_freq)
372 struct power_event *pwr;
373 pwr = malloc(sizeof(struct power_event));
375 if (new_freq > 8000000) /* detect invalid data */
380 memset(pwr, 0, sizeof(struct power_event));
382 pwr->state = cpus_pstate_state[cpu];
383 pwr->start_time = cpus_pstate_start_times[cpu];
384 pwr->end_time = timestamp;
387 pwr->next = power_events;
389 if (!pwr->start_time)
390 pwr->start_time = first_time;
394 cpus_pstate_state[cpu] = new_freq;
395 cpus_pstate_start_times[cpu] = timestamp;
397 if ((u64)new_freq > max_freq)
400 if (new_freq < min_freq || min_freq == 0)
403 if (new_freq == max_freq - 1000)
404 turbo_frequency = max_freq;
408 sched_wakeup(int cpu, u64 timestamp, int pid, struct trace_entry *te)
410 struct wake_event *we;
412 struct wakeup_entry *wake = (void *)te;
414 we = malloc(sizeof(struct wake_event));
418 memset(we, 0, sizeof(struct wake_event));
419 we->time = timestamp;
422 if ((te->flags & TRACE_FLAG_HARDIRQ) || (te->flags & TRACE_FLAG_SOFTIRQ))
425 we->wakee = wake->pid;
426 we->next = wake_events;
428 p = find_create_pid(we->wakee);
430 if (p && p->current && p->current->state == TYPE_NONE) {
431 p->current->state_since = timestamp;
432 p->current->state = TYPE_WAITING;
434 if (p && p->current && p->current->state == TYPE_BLOCKED) {
435 pid_put_sample(p->pid, p->current->state, cpu, p->current->state_since, timestamp);
436 p->current->state_since = timestamp;
437 p->current->state = TYPE_WAITING;
441 static void sched_switch(int cpu, u64 timestamp, struct trace_entry *te)
443 struct per_pid *p = NULL, *prev_p;
444 struct sched_switch *sw = (void *)te;
447 prev_p = find_create_pid(sw->prev_pid);
449 p = find_create_pid(sw->next_pid);
451 if (prev_p->current && prev_p->current->state != TYPE_NONE)
452 pid_put_sample(sw->prev_pid, TYPE_RUNNING, cpu, prev_p->current->state_since, timestamp);
453 if (p && p->current) {
454 if (p->current->state != TYPE_NONE)
455 pid_put_sample(sw->next_pid, p->current->state, cpu, p->current->state_since, timestamp);
457 p->current->state_since = timestamp;
458 p->current->state = TYPE_RUNNING;
461 if (prev_p->current) {
462 prev_p->current->state = TYPE_NONE;
463 prev_p->current->state_since = timestamp;
464 if (sw->prev_state & 2)
465 prev_p->current->state = TYPE_BLOCKED;
466 if (sw->prev_state == 0)
467 prev_p->current->state = TYPE_WAITING;
472 static int process_sample_event(event_t *event, struct perf_session *session)
474 struct sample_data data;
475 struct trace_entry *te;
477 memset(&data, 0, sizeof(data));
479 event__parse_sample(event, session->sample_type, &data);
481 if (session->sample_type & PERF_SAMPLE_TIME) {
482 if (!first_time || first_time > data.time)
483 first_time = data.time;
484 if (last_time < data.time)
485 last_time = data.time;
488 te = (void *)data.raw_data;
489 if (session->sample_type & PERF_SAMPLE_RAW && data.raw_size > 0) {
491 struct power_entry *pe;
495 event_str = perf_header__find_event(te->type);
500 if (strcmp(event_str, "power:power_start") == 0)
501 c_state_start(data.cpu, data.time, pe->value);
503 if (strcmp(event_str, "power:power_end") == 0)
504 c_state_end(data.cpu, data.time);
506 if (strcmp(event_str, "power:power_frequency") == 0)
507 p_state_change(data.cpu, data.time, pe->value);
509 if (strcmp(event_str, "sched:sched_wakeup") == 0)
510 sched_wakeup(data.cpu, data.time, data.pid, te);
512 if (strcmp(event_str, "sched:sched_switch") == 0)
513 sched_switch(data.cpu, data.time, te);
519 * After the last sample we need to wrap up the current C/P state
520 * and close out each CPU for these.
522 static void end_sample_processing(void)
525 struct power_event *pwr;
527 for (cpu = 0; cpu <= numcpus; cpu++) {
528 pwr = malloc(sizeof(struct power_event));
531 memset(pwr, 0, sizeof(struct power_event));
535 pwr->state = cpus_cstate_state[cpu];
536 pwr->start_time = cpus_cstate_start_times[cpu];
537 pwr->end_time = last_time;
540 pwr->next = power_events;
546 pwr = malloc(sizeof(struct power_event));
549 memset(pwr, 0, sizeof(struct power_event));
551 pwr->state = cpus_pstate_state[cpu];
552 pwr->start_time = cpus_pstate_start_times[cpu];
553 pwr->end_time = last_time;
556 pwr->next = power_events;
558 if (!pwr->start_time)
559 pwr->start_time = first_time;
561 pwr->state = min_freq;
567 * Sort the pid datastructure
569 static void sort_pids(void)
571 struct per_pid *new_list, *p, *cursor, *prev;
572 /* sort by ppid first, then by pid, lowest to highest */
581 if (new_list == NULL) {
589 if (cursor->ppid > p->ppid ||
590 (cursor->ppid == p->ppid && cursor->pid > p->pid)) {
591 /* must insert before */
593 p->next = prev->next;
606 cursor = cursor->next;
615 static void draw_c_p_states(void)
617 struct power_event *pwr;
621 * two pass drawing so that the P state bars are on top of the C state blocks
624 if (pwr->type == CSTATE)
625 svg_cstate(pwr->cpu, pwr->start_time, pwr->end_time, pwr->state);
631 if (pwr->type == PSTATE) {
633 pwr->state = min_freq;
634 svg_pstate(pwr->cpu, pwr->start_time, pwr->end_time, pwr->state);
640 static void draw_wakeups(void)
642 struct wake_event *we;
644 struct per_pidcomm *c;
648 int from = 0, to = 0;
649 char *task_from = NULL, *task_to = NULL;
651 /* locate the column of the waker and wakee */
654 if (p->pid == we->waker || p->pid == we->wakee) {
657 if (c->Y && c->start_time <= we->time && c->end_time >= we->time) {
658 if (p->pid == we->waker && !from) {
660 task_from = strdup(c->comm);
662 if (p->pid == we->wakee && !to) {
664 task_to = strdup(c->comm);
671 if (p->pid == we->waker && !from) {
673 task_from = strdup(c->comm);
675 if (p->pid == we->wakee && !to) {
677 task_to = strdup(c->comm);
686 task_from = malloc(40);
687 sprintf(task_from, "[%i]", we->waker);
690 task_to = malloc(40);
691 sprintf(task_to, "[%i]", we->wakee);
695 svg_interrupt(we->time, to);
696 else if (from && to && abs(from - to) == 1)
697 svg_wakeline(we->time, from, to);
699 svg_partial_wakeline(we->time, from, task_from, to, task_to);
707 static void draw_cpu_usage(void)
710 struct per_pidcomm *c;
711 struct cpu_sample *sample;
718 if (sample->type == TYPE_RUNNING)
719 svg_process(sample->cpu, sample->start_time, sample->end_time, "sample", c->comm);
721 sample = sample->next;
729 static void draw_process_bars(void)
732 struct per_pidcomm *c;
733 struct cpu_sample *sample;
748 svg_box(Y, c->start_time, c->end_time, "process");
751 if (sample->type == TYPE_RUNNING)
752 svg_sample(Y, sample->cpu, sample->start_time, sample->end_time);
753 if (sample->type == TYPE_BLOCKED)
754 svg_box(Y, sample->start_time, sample->end_time, "blocked");
755 if (sample->type == TYPE_WAITING)
756 svg_waiting(Y, sample->start_time, sample->end_time);
757 sample = sample->next;
762 if (c->total_time > 5000000000) /* 5 seconds */
763 sprintf(comm, "%s:%i (%2.2fs)", c->comm, p->pid, c->total_time / 1000000000.0);
765 sprintf(comm, "%s:%i (%3.1fms)", c->comm, p->pid, c->total_time / 1000000.0);
767 svg_text(Y, c->start_time, comm);
777 static void add_process_filter(const char *string)
779 struct process_filter *filt;
782 pid = strtoull(string, NULL, 10);
783 filt = malloc(sizeof(struct process_filter));
787 filt->name = strdup(string);
789 filt->next = process_filter;
791 process_filter = filt;
794 static int passes_filter(struct per_pid *p, struct per_pidcomm *c)
796 struct process_filter *filt;
800 filt = process_filter;
802 if (filt->pid && p->pid == filt->pid)
804 if (strcmp(filt->name, c->comm) == 0)
811 static int determine_display_tasks_filtered(void)
814 struct per_pidcomm *c;
820 if (p->start_time == 1)
821 p->start_time = first_time;
823 /* no exit marker, task kept running to the end */
824 if (p->end_time == 0)
825 p->end_time = last_time;
832 if (c->start_time == 1)
833 c->start_time = first_time;
835 if (passes_filter(p, c)) {
841 if (c->end_time == 0)
842 c->end_time = last_time;
851 static int determine_display_tasks(u64 threshold)
854 struct per_pidcomm *c;
858 return determine_display_tasks_filtered();
863 if (p->start_time == 1)
864 p->start_time = first_time;
866 /* no exit marker, task kept running to the end */
867 if (p->end_time == 0)
868 p->end_time = last_time;
869 if (p->total_time >= threshold && !power_only)
877 if (c->start_time == 1)
878 c->start_time = first_time;
880 if (c->total_time >= threshold && !power_only) {
885 if (c->end_time == 0)
886 c->end_time = last_time;
897 #define TIME_THRESH 10000000
899 static void write_svg_file(const char *filename)
907 count = determine_display_tasks(TIME_THRESH);
909 /* We'd like to show at least 15 tasks; be less picky if we have fewer */
911 count = determine_display_tasks(TIME_THRESH / 10);
913 open_svg(filename, numcpus, count, first_time, last_time);
918 for (i = 0; i < numcpus; i++)
919 svg_cpu_box(i, max_freq, turbo_frequency);
929 static struct perf_event_ops event_ops = {
930 .comm = process_comm_event,
931 .fork = process_fork_event,
932 .exit = process_exit_event,
933 .sample = process_sample_event,
934 .ordered_samples = true,
937 static int __cmd_timechart(void)
939 struct perf_session *session = perf_session__new(input_name, O_RDONLY, 0, false);
945 if (!perf_session__has_traces(session, "timechart record"))
948 ret = perf_session__process_events(session, &event_ops);
952 end_sample_processing();
956 write_svg_file(output_name);
958 pr_info("Written %2.1f seconds of trace to %s.\n",
959 (last_time - first_time) / 1000000000.0, output_name);
961 perf_session__delete(session);
965 static const char * const timechart_usage[] = {
966 "perf timechart [<options>] {record}",
970 static const char *record_args[] = {
976 "-e", "power:power_start",
977 "-e", "power:power_end",
978 "-e", "power:power_frequency",
979 "-e", "sched:sched_wakeup",
980 "-e", "sched:sched_switch",
983 static int __cmd_record(int argc, const char **argv)
985 unsigned int rec_argc, i, j;
986 const char **rec_argv;
988 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
989 rec_argv = calloc(rec_argc + 1, sizeof(char *));
991 for (i = 0; i < ARRAY_SIZE(record_args); i++)
992 rec_argv[i] = strdup(record_args[i]);
994 for (j = 1; j < (unsigned int)argc; j++, i++)
995 rec_argv[i] = argv[j];
997 return cmd_record(i, rec_argv, NULL);
1001 parse_process(const struct option *opt __used, const char *arg, int __used unset)
1004 add_process_filter(arg);
1008 static const struct option options[] = {
1009 OPT_STRING('i', "input", &input_name, "file",
1011 OPT_STRING('o', "output", &output_name, "file",
1012 "output file name"),
1013 OPT_INTEGER('w', "width", &svg_page_width,
1015 OPT_BOOLEAN('P', "power-only", &power_only,
1016 "output power data only"),
1017 OPT_CALLBACK('p', "process", NULL, "process",
1018 "process selector. Pass a pid or process name.",
1024 int cmd_timechart(int argc, const char **argv, const char *prefix __used)
1026 argc = parse_options(argc, argv, options, timechart_usage,
1027 PARSE_OPT_STOP_AT_NON_OPTION);
1031 if (argc && !strncmp(argv[0], "rec", 3))
1032 return __cmd_record(argc, argv);
1034 usage_with_options(timechart_usage, options);
1038 return __cmd_timechart();