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/string.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/data_map.h"
34 #include "util/svghelper.h"
36 static char const *input_name = "perf.data";
37 static char const *output_name = "output.svg";
40 static u64 sample_type;
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 int 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;
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 sample_wrapper *all_samples;
155 struct process_filter;
156 struct process_filter {
159 struct process_filter *next;
162 static struct process_filter *process_filter;
165 static struct per_pid *find_create_pid(int pid)
167 struct per_pid *cursor = all_data;
170 if (cursor->pid == pid)
172 cursor = cursor->next;
174 cursor = malloc(sizeof(struct per_pid));
175 assert(cursor != NULL);
176 memset(cursor, 0, sizeof(struct per_pid));
178 cursor->next = all_data;
183 static void pid_set_comm(int pid, char *comm)
186 struct per_pidcomm *c;
187 p = find_create_pid(pid);
190 if (c->comm && strcmp(c->comm, comm) == 0) {
195 c->comm = strdup(comm);
201 c = malloc(sizeof(struct per_pidcomm));
203 memset(c, 0, sizeof(struct per_pidcomm));
204 c->comm = strdup(comm);
210 static void pid_fork(int pid, int ppid, u64 timestamp)
212 struct per_pid *p, *pp;
213 p = find_create_pid(pid);
214 pp = find_create_pid(ppid);
216 if (pp->current && pp->current->comm && !p->current)
217 pid_set_comm(pid, pp->current->comm);
219 p->start_time = timestamp;
221 p->current->start_time = timestamp;
222 p->current->state_since = timestamp;
226 static void pid_exit(int pid, u64 timestamp)
229 p = find_create_pid(pid);
230 p->end_time = timestamp;
232 p->current->end_time = timestamp;
236 pid_put_sample(int pid, int type, unsigned int cpu, u64 start, u64 end)
239 struct per_pidcomm *c;
240 struct cpu_sample *sample;
242 p = find_create_pid(pid);
245 c = malloc(sizeof(struct per_pidcomm));
247 memset(c, 0, sizeof(struct per_pidcomm));
253 sample = malloc(sizeof(struct cpu_sample));
254 assert(sample != NULL);
255 memset(sample, 0, sizeof(struct cpu_sample));
256 sample->start_time = start;
257 sample->end_time = end;
259 sample->next = c->samples;
263 if (sample->type == TYPE_RUNNING && end > start && start > 0) {
264 c->total_time += (end-start);
265 p->total_time += (end-start);
268 if (c->start_time == 0 || c->start_time > start)
269 c->start_time = start;
270 if (p->start_time == 0 || p->start_time > start)
271 p->start_time = start;
277 #define MAX_CPUS 4096
279 static u64 cpus_cstate_start_times[MAX_CPUS];
280 static int cpus_cstate_state[MAX_CPUS];
281 static u64 cpus_pstate_start_times[MAX_CPUS];
282 static u64 cpus_pstate_state[MAX_CPUS];
285 process_comm_event(event_t *event)
287 pid_set_comm(event->comm.pid, event->comm.comm);
291 process_fork_event(event_t *event)
293 pid_fork(event->fork.pid, event->fork.ppid, event->fork.time);
298 process_exit_event(event_t *event)
300 pid_exit(event->fork.pid, event->fork.time);
307 unsigned char preempt_count;
313 struct trace_entry te;
318 #define TASK_COMM_LEN 16
319 struct wakeup_entry {
320 struct trace_entry te;
321 char comm[TASK_COMM_LEN];
328 * trace_flag_type is an enumeration that holds different
329 * states when a trace occurs. These are:
330 * IRQS_OFF - interrupts were disabled
331 * IRQS_NOSUPPORT - arch does not support irqs_disabled_flags
332 * NEED_RESCED - reschedule is requested
333 * HARDIRQ - inside an interrupt handler
334 * SOFTIRQ - inside a softirq handler
336 enum trace_flag_type {
337 TRACE_FLAG_IRQS_OFF = 0x01,
338 TRACE_FLAG_IRQS_NOSUPPORT = 0x02,
339 TRACE_FLAG_NEED_RESCHED = 0x04,
340 TRACE_FLAG_HARDIRQ = 0x08,
341 TRACE_FLAG_SOFTIRQ = 0x10,
346 struct sched_switch {
347 struct trace_entry te;
348 char prev_comm[TASK_COMM_LEN];
351 long prev_state; /* Arjan weeps. */
352 char next_comm[TASK_COMM_LEN];
357 static void c_state_start(int cpu, u64 timestamp, int state)
359 cpus_cstate_start_times[cpu] = timestamp;
360 cpus_cstate_state[cpu] = state;
363 static void c_state_end(int cpu, u64 timestamp)
365 struct power_event *pwr;
366 pwr = malloc(sizeof(struct power_event));
369 memset(pwr, 0, sizeof(struct power_event));
371 pwr->state = cpus_cstate_state[cpu];
372 pwr->start_time = cpus_cstate_start_times[cpu];
373 pwr->end_time = timestamp;
376 pwr->next = power_events;
381 static void p_state_change(int cpu, u64 timestamp, u64 new_freq)
383 struct power_event *pwr;
384 pwr = malloc(sizeof(struct power_event));
386 if (new_freq > 8000000) /* detect invalid data */
391 memset(pwr, 0, sizeof(struct power_event));
393 pwr->state = cpus_pstate_state[cpu];
394 pwr->start_time = cpus_pstate_start_times[cpu];
395 pwr->end_time = timestamp;
398 pwr->next = power_events;
400 if (!pwr->start_time)
401 pwr->start_time = first_time;
405 cpus_pstate_state[cpu] = new_freq;
406 cpus_pstate_start_times[cpu] = timestamp;
408 if ((u64)new_freq > max_freq)
411 if (new_freq < min_freq || min_freq == 0)
414 if (new_freq == max_freq - 1000)
415 turbo_frequency = max_freq;
419 sched_wakeup(int cpu, u64 timestamp, int pid, struct trace_entry *te)
421 struct wake_event *we;
423 struct wakeup_entry *wake = (void *)te;
425 we = malloc(sizeof(struct wake_event));
429 memset(we, 0, sizeof(struct wake_event));
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;
484 process_sample_event(event_t *event)
486 struct sample_data data;
487 struct trace_entry *te;
489 memset(&data, 0, sizeof(data));
491 event__parse_sample(event, sample_type, &data);
493 if (sample_type & PERF_SAMPLE_TIME) {
494 if (!first_time || first_time > data.time)
495 first_time = data.time;
496 if (last_time < data.time)
497 last_time = data.time;
500 te = (void *)data.raw_data;
501 if (sample_type & PERF_SAMPLE_RAW && data.raw_size > 0) {
503 struct power_entry *pe;
507 event_str = perf_header__find_event(te->type);
512 if (strcmp(event_str, "power:power_start") == 0)
513 c_state_start(data.cpu, data.time, pe->value);
515 if (strcmp(event_str, "power:power_end") == 0)
516 c_state_end(data.cpu, data.time);
518 if (strcmp(event_str, "power:power_frequency") == 0)
519 p_state_change(data.cpu, data.time, pe->value);
521 if (strcmp(event_str, "sched:sched_wakeup") == 0)
522 sched_wakeup(data.cpu, data.time, data.pid, te);
524 if (strcmp(event_str, "sched:sched_switch") == 0)
525 sched_switch(data.cpu, data.time, te);
531 * After the last sample we need to wrap up the current C/P state
532 * and close out each CPU for these.
534 static void end_sample_processing(void)
537 struct power_event *pwr;
539 for (cpu = 0; cpu <= numcpus; cpu++) {
540 pwr = malloc(sizeof(struct power_event));
543 memset(pwr, 0, sizeof(struct power_event));
547 pwr->state = cpus_cstate_state[cpu];
548 pwr->start_time = cpus_cstate_start_times[cpu];
549 pwr->end_time = last_time;
552 pwr->next = power_events;
558 pwr = malloc(sizeof(struct power_event));
561 memset(pwr, 0, sizeof(struct power_event));
563 pwr->state = cpus_pstate_state[cpu];
564 pwr->start_time = cpus_pstate_start_times[cpu];
565 pwr->end_time = last_time;
568 pwr->next = power_events;
570 if (!pwr->start_time)
571 pwr->start_time = first_time;
573 pwr->state = min_freq;
578 static u64 sample_time(event_t *event)
583 if (sample_type & PERF_SAMPLE_IP)
585 if (sample_type & PERF_SAMPLE_TID)
587 if (sample_type & PERF_SAMPLE_TIME)
588 return event->sample.array[cursor];
594 * We first queue all events, sorted backwards by insertion.
595 * The order will get flipped later.
598 queue_sample_event(event_t *event)
600 struct sample_wrapper *copy, *prev;
603 size = event->sample.header.size + sizeof(struct sample_wrapper) + 8;
609 memset(copy, 0, size);
612 copy->timestamp = sample_time(event);
614 memcpy(©->data, event, event->sample.header.size);
616 /* insert in the right place in the list */
619 /* first sample ever */
624 if (all_samples->timestamp < copy->timestamp) {
625 /* insert at the head of the list */
626 copy->next = all_samples;
633 if (prev->next->timestamp < copy->timestamp) {
634 copy->next = prev->next;
640 /* insert at the end of the list */
646 static void sort_queued_samples(void)
648 struct sample_wrapper *cursor, *next;
650 cursor = all_samples;
655 cursor->next = all_samples;
656 all_samples = cursor;
662 * Sort the pid datastructure
664 static void sort_pids(void)
666 struct per_pid *new_list, *p, *cursor, *prev;
667 /* sort by ppid first, then by pid, lowest to highest */
676 if (new_list == NULL) {
684 if (cursor->ppid > p->ppid ||
685 (cursor->ppid == p->ppid && cursor->pid > p->pid)) {
686 /* must insert before */
688 p->next = prev->next;
701 cursor = cursor->next;
710 static void draw_c_p_states(void)
712 struct power_event *pwr;
716 * two pass drawing so that the P state bars are on top of the C state blocks
719 if (pwr->type == CSTATE)
720 svg_cstate(pwr->cpu, pwr->start_time, pwr->end_time, pwr->state);
726 if (pwr->type == PSTATE) {
728 pwr->state = min_freq;
729 svg_pstate(pwr->cpu, pwr->start_time, pwr->end_time, pwr->state);
735 static void draw_wakeups(void)
737 struct wake_event *we;
739 struct per_pidcomm *c;
743 int from = 0, to = 0;
744 char *task_from = NULL, *task_to = NULL;
746 /* locate the column of the waker and wakee */
749 if (p->pid == we->waker || p->pid == we->wakee) {
752 if (c->Y && c->start_time <= we->time && c->end_time >= we->time) {
753 if (p->pid == we->waker && !from) {
755 task_from = strdup(c->comm);
757 if (p->pid == we->wakee && !to) {
759 task_to = strdup(c->comm);
766 if (p->pid == we->waker && !from) {
768 task_from = strdup(c->comm);
770 if (p->pid == we->wakee && !to) {
772 task_to = strdup(c->comm);
781 task_from = malloc(40);
782 sprintf(task_from, "[%i]", we->waker);
785 task_to = malloc(40);
786 sprintf(task_to, "[%i]", we->wakee);
790 svg_interrupt(we->time, to);
791 else if (from && to && abs(from - to) == 1)
792 svg_wakeline(we->time, from, to);
794 svg_partial_wakeline(we->time, from, task_from, to, task_to);
802 static void draw_cpu_usage(void)
805 struct per_pidcomm *c;
806 struct cpu_sample *sample;
813 if (sample->type == TYPE_RUNNING)
814 svg_process(sample->cpu, sample->start_time, sample->end_time, "sample", c->comm);
816 sample = sample->next;
824 static void draw_process_bars(void)
827 struct per_pidcomm *c;
828 struct cpu_sample *sample;
843 svg_box(Y, c->start_time, c->end_time, "process");
846 if (sample->type == TYPE_RUNNING)
847 svg_sample(Y, sample->cpu, sample->start_time, sample->end_time);
848 if (sample->type == TYPE_BLOCKED)
849 svg_box(Y, sample->start_time, sample->end_time, "blocked");
850 if (sample->type == TYPE_WAITING)
851 svg_waiting(Y, sample->start_time, sample->end_time);
852 sample = sample->next;
857 if (c->total_time > 5000000000) /* 5 seconds */
858 sprintf(comm, "%s:%i (%2.2fs)", c->comm, p->pid, c->total_time / 1000000000.0);
860 sprintf(comm, "%s:%i (%3.1fms)", c->comm, p->pid, c->total_time / 1000000.0);
862 svg_text(Y, c->start_time, comm);
872 static void add_process_filter(const char *string)
874 struct process_filter *filt;
877 pid = strtoull(string, NULL, 10);
878 filt = malloc(sizeof(struct process_filter));
882 filt->name = strdup(string);
884 filt->next = process_filter;
886 process_filter = filt;
889 static int passes_filter(struct per_pid *p, struct per_pidcomm *c)
891 struct process_filter *filt;
895 filt = process_filter;
897 if (filt->pid && p->pid == filt->pid)
899 if (strcmp(filt->name, c->comm) == 0)
906 static int determine_display_tasks_filtered(void)
909 struct per_pidcomm *c;
915 if (p->start_time == 1)
916 p->start_time = first_time;
918 /* no exit marker, task kept running to the end */
919 if (p->end_time == 0)
920 p->end_time = last_time;
927 if (c->start_time == 1)
928 c->start_time = first_time;
930 if (passes_filter(p, c)) {
936 if (c->end_time == 0)
937 c->end_time = last_time;
946 static int determine_display_tasks(u64 threshold)
949 struct per_pidcomm *c;
953 return determine_display_tasks_filtered();
958 if (p->start_time == 1)
959 p->start_time = first_time;
961 /* no exit marker, task kept running to the end */
962 if (p->end_time == 0)
963 p->end_time = last_time;
964 if (p->total_time >= threshold && !power_only)
972 if (c->start_time == 1)
973 c->start_time = first_time;
975 if (c->total_time >= threshold && !power_only) {
980 if (c->end_time == 0)
981 c->end_time = last_time;
992 #define TIME_THRESH 10000000
994 static void write_svg_file(const char *filename)
1002 count = determine_display_tasks(TIME_THRESH);
1004 /* We'd like to show at least 15 tasks; be less picky if we have fewer */
1006 count = determine_display_tasks(TIME_THRESH / 10);
1008 open_svg(filename, numcpus, count, first_time, last_time);
1013 for (i = 0; i < numcpus; i++)
1014 svg_cpu_box(i, max_freq, turbo_frequency);
1017 draw_process_bars();
1024 static void process_samples(void)
1026 struct sample_wrapper *cursor;
1029 sort_queued_samples();
1031 cursor = all_samples;
1033 event = (void *)&cursor->data;
1034 cursor = cursor->next;
1035 process_sample_event(event);
1039 static int sample_type_check(u64 type)
1043 if (!(sample_type & PERF_SAMPLE_RAW)) {
1044 fprintf(stderr, "No trace samples found in the file.\n"
1045 "Have you used 'perf timechart record' to record it?\n");
1052 static struct perf_file_handler file_handler = {
1053 .process_comm_event = process_comm_event,
1054 .process_fork_event = process_fork_event,
1055 .process_exit_event = process_exit_event,
1056 .process_sample_event = queue_sample_event,
1057 .sample_type_check = sample_type_check,
1060 static int __cmd_timechart(void)
1062 struct perf_session *session = perf_session__new(input_name, O_RDONLY, 0);
1065 if (session == NULL)
1068 register_perf_file_handler(&file_handler);
1070 ret = perf_session__process_events(session, 0, &event__cwdlen, &event__cwd);
1076 end_sample_processing();
1080 write_svg_file(output_name);
1082 pr_info("Written %2.1f seconds of trace to %s.\n",
1083 (last_time - first_time) / 1000000000.0, output_name);
1085 perf_session__delete(session);
1089 static const char * const timechart_usage[] = {
1090 "perf timechart [<options>] {record}",
1094 static const char *record_args[] = {
1101 "-e", "power:power_start",
1102 "-e", "power:power_end",
1103 "-e", "power:power_frequency",
1104 "-e", "sched:sched_wakeup",
1105 "-e", "sched:sched_switch",
1108 static int __cmd_record(int argc, const char **argv)
1110 unsigned int rec_argc, i, j;
1111 const char **rec_argv;
1113 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1114 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1116 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1117 rec_argv[i] = strdup(record_args[i]);
1119 for (j = 1; j < (unsigned int)argc; j++, i++)
1120 rec_argv[i] = argv[j];
1122 return cmd_record(i, rec_argv, NULL);
1126 parse_process(const struct option *opt __used, const char *arg, int __used unset)
1129 add_process_filter(arg);
1133 static const struct option options[] = {
1134 OPT_STRING('i', "input", &input_name, "file",
1136 OPT_STRING('o', "output", &output_name, "file",
1137 "output file name"),
1138 OPT_INTEGER('w', "width", &svg_page_width,
1140 OPT_BOOLEAN('P', "power-only", &power_only,
1141 "output power data only"),
1142 OPT_CALLBACK('p', "process", NULL, "process",
1143 "process selector. Pass a pid or process name.",
1149 int cmd_timechart(int argc, const char **argv, const char *prefix __used)
1153 argc = parse_options(argc, argv, options, timechart_usage,
1154 PARSE_OPT_STOP_AT_NON_OPTION);
1156 if (argc && !strncmp(argv[0], "rec", 3))
1157 return __cmd_record(argc, argv);
1159 usage_with_options(timechart_usage, options);
1163 return __cmd_timechart();