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 int 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;
86 struct per_pidcomm *next;
100 struct cpu_sample *samples;
103 struct sample_wrapper {
104 struct sample_wrapper *next;
107 unsigned char data[0];
111 #define TYPE_RUNNING 1
112 #define TYPE_WAITING 2
113 #define TYPE_BLOCKED 3
116 struct cpu_sample *next;
124 static struct per_pid *all_data;
130 struct power_event *next;
139 struct wake_event *next;
145 static struct power_event *power_events;
146 static struct wake_event *wake_events;
148 struct sample_wrapper *all_samples;
151 struct process_filter;
152 struct process_filter {
155 struct process_filter *next;
158 static struct process_filter *process_filter;
161 static struct per_pid *find_create_pid(int pid)
163 struct per_pid *cursor = all_data;
166 if (cursor->pid == pid)
168 cursor = cursor->next;
170 cursor = malloc(sizeof(struct per_pid));
171 assert(cursor != NULL);
172 memset(cursor, 0, sizeof(struct per_pid));
174 cursor->next = all_data;
179 static void pid_set_comm(int pid, char *comm)
182 struct per_pidcomm *c;
183 p = find_create_pid(pid);
186 if (c->comm && strcmp(c->comm, comm) == 0) {
191 c->comm = strdup(comm);
197 c = malloc(sizeof(struct per_pidcomm));
199 memset(c, 0, sizeof(struct per_pidcomm));
200 c->comm = strdup(comm);
206 static void pid_fork(int pid, int ppid, u64 timestamp)
208 struct per_pid *p, *pp;
209 p = find_create_pid(pid);
210 pp = find_create_pid(ppid);
212 if (pp->current && pp->current->comm && !p->current)
213 pid_set_comm(pid, pp->current->comm);
215 p->start_time = timestamp;
217 p->current->start_time = timestamp;
218 p->current->state_since = timestamp;
222 static void pid_exit(int pid, u64 timestamp)
225 p = find_create_pid(pid);
226 p->end_time = timestamp;
228 p->current->end_time = timestamp;
232 pid_put_sample(int pid, int type, unsigned int cpu, u64 start, u64 end)
235 struct per_pidcomm *c;
236 struct cpu_sample *sample;
238 p = find_create_pid(pid);
241 c = malloc(sizeof(struct per_pidcomm));
243 memset(c, 0, sizeof(struct per_pidcomm));
249 sample = malloc(sizeof(struct cpu_sample));
250 assert(sample != NULL);
251 memset(sample, 0, sizeof(struct cpu_sample));
252 sample->start_time = start;
253 sample->end_time = end;
255 sample->next = c->samples;
259 if (sample->type == TYPE_RUNNING && end > start && start > 0) {
260 c->total_time += (end-start);
261 p->total_time += (end-start);
264 if (c->start_time == 0 || c->start_time > start)
265 c->start_time = start;
266 if (p->start_time == 0 || p->start_time > start)
267 p->start_time = start;
273 #define MAX_CPUS 4096
275 static u64 cpus_cstate_start_times[MAX_CPUS];
276 static int cpus_cstate_state[MAX_CPUS];
277 static u64 cpus_pstate_start_times[MAX_CPUS];
278 static u64 cpus_pstate_state[MAX_CPUS];
280 static int process_comm_event(event_t *event, struct perf_session *session __used)
282 pid_set_comm(event->comm.tid, event->comm.comm);
286 static int process_fork_event(event_t *event, struct perf_session *session __used)
288 pid_fork(event->fork.pid, event->fork.ppid, event->fork.time);
292 static int process_exit_event(event_t *event, struct perf_session *session __used)
294 pid_exit(event->fork.pid, event->fork.time);
301 unsigned char preempt_count;
307 struct trace_entry te;
312 #define TASK_COMM_LEN 16
313 struct wakeup_entry {
314 struct trace_entry te;
315 char comm[TASK_COMM_LEN];
322 * trace_flag_type is an enumeration that holds different
323 * states when a trace occurs. These are:
324 * IRQS_OFF - interrupts were disabled
325 * IRQS_NOSUPPORT - arch does not support irqs_disabled_flags
326 * NEED_RESCED - reschedule is requested
327 * HARDIRQ - inside an interrupt handler
328 * SOFTIRQ - inside a softirq handler
330 enum trace_flag_type {
331 TRACE_FLAG_IRQS_OFF = 0x01,
332 TRACE_FLAG_IRQS_NOSUPPORT = 0x02,
333 TRACE_FLAG_NEED_RESCHED = 0x04,
334 TRACE_FLAG_HARDIRQ = 0x08,
335 TRACE_FLAG_SOFTIRQ = 0x10,
340 struct sched_switch {
341 struct trace_entry te;
342 char prev_comm[TASK_COMM_LEN];
345 long prev_state; /* Arjan weeps. */
346 char next_comm[TASK_COMM_LEN];
351 static void c_state_start(int cpu, u64 timestamp, int state)
353 cpus_cstate_start_times[cpu] = timestamp;
354 cpus_cstate_state[cpu] = state;
357 static void c_state_end(int cpu, u64 timestamp)
359 struct power_event *pwr;
360 pwr = malloc(sizeof(struct power_event));
363 memset(pwr, 0, sizeof(struct power_event));
365 pwr->state = cpus_cstate_state[cpu];
366 pwr->start_time = cpus_cstate_start_times[cpu];
367 pwr->end_time = timestamp;
370 pwr->next = power_events;
375 static void p_state_change(int cpu, u64 timestamp, u64 new_freq)
377 struct power_event *pwr;
378 pwr = malloc(sizeof(struct power_event));
380 if (new_freq > 8000000) /* detect invalid data */
385 memset(pwr, 0, sizeof(struct power_event));
387 pwr->state = cpus_pstate_state[cpu];
388 pwr->start_time = cpus_pstate_start_times[cpu];
389 pwr->end_time = timestamp;
392 pwr->next = power_events;
394 if (!pwr->start_time)
395 pwr->start_time = first_time;
399 cpus_pstate_state[cpu] = new_freq;
400 cpus_pstate_start_times[cpu] = timestamp;
402 if ((u64)new_freq > max_freq)
405 if (new_freq < min_freq || min_freq == 0)
408 if (new_freq == max_freq - 1000)
409 turbo_frequency = max_freq;
413 sched_wakeup(int cpu, u64 timestamp, int pid, struct trace_entry *te)
415 struct wake_event *we;
417 struct wakeup_entry *wake = (void *)te;
419 we = malloc(sizeof(struct wake_event));
423 memset(we, 0, sizeof(struct wake_event));
424 we->time = timestamp;
427 if ((te->flags & TRACE_FLAG_HARDIRQ) || (te->flags & TRACE_FLAG_SOFTIRQ))
430 we->wakee = wake->pid;
431 we->next = wake_events;
433 p = find_create_pid(we->wakee);
435 if (p && p->current && p->current->state == TYPE_NONE) {
436 p->current->state_since = timestamp;
437 p->current->state = TYPE_WAITING;
439 if (p && p->current && p->current->state == TYPE_BLOCKED) {
440 pid_put_sample(p->pid, p->current->state, cpu, p->current->state_since, timestamp);
441 p->current->state_since = timestamp;
442 p->current->state = TYPE_WAITING;
446 static void sched_switch(int cpu, u64 timestamp, struct trace_entry *te)
448 struct per_pid *p = NULL, *prev_p;
449 struct sched_switch *sw = (void *)te;
452 prev_p = find_create_pid(sw->prev_pid);
454 p = find_create_pid(sw->next_pid);
456 if (prev_p->current && prev_p->current->state != TYPE_NONE)
457 pid_put_sample(sw->prev_pid, TYPE_RUNNING, cpu, prev_p->current->state_since, timestamp);
458 if (p && p->current) {
459 if (p->current->state != TYPE_NONE)
460 pid_put_sample(sw->next_pid, p->current->state, cpu, p->current->state_since, timestamp);
462 p->current->state_since = timestamp;
463 p->current->state = TYPE_RUNNING;
466 if (prev_p->current) {
467 prev_p->current->state = TYPE_NONE;
468 prev_p->current->state_since = timestamp;
469 if (sw->prev_state & 2)
470 prev_p->current->state = TYPE_BLOCKED;
471 if (sw->prev_state == 0)
472 prev_p->current->state = TYPE_WAITING;
477 static int process_sample_event(event_t *event, struct perf_session *session)
479 struct sample_data data;
480 struct trace_entry *te;
482 memset(&data, 0, sizeof(data));
484 event__parse_sample(event, session->sample_type, &data);
486 if (session->sample_type & PERF_SAMPLE_TIME) {
487 if (!first_time || first_time > data.time)
488 first_time = data.time;
489 if (last_time < data.time)
490 last_time = data.time;
493 te = (void *)data.raw_data;
494 if (session->sample_type & PERF_SAMPLE_RAW && data.raw_size > 0) {
496 struct power_entry *pe;
500 event_str = perf_header__find_event(te->type);
505 if (strcmp(event_str, "power:power_start") == 0)
506 c_state_start(data.cpu, data.time, pe->value);
508 if (strcmp(event_str, "power:power_end") == 0)
509 c_state_end(data.cpu, data.time);
511 if (strcmp(event_str, "power:power_frequency") == 0)
512 p_state_change(data.cpu, data.time, pe->value);
514 if (strcmp(event_str, "sched:sched_wakeup") == 0)
515 sched_wakeup(data.cpu, data.time, data.pid, te);
517 if (strcmp(event_str, "sched:sched_switch") == 0)
518 sched_switch(data.cpu, data.time, te);
524 * After the last sample we need to wrap up the current C/P state
525 * and close out each CPU for these.
527 static void end_sample_processing(void)
530 struct power_event *pwr;
532 for (cpu = 0; cpu <= numcpus; cpu++) {
533 pwr = malloc(sizeof(struct power_event));
536 memset(pwr, 0, sizeof(struct power_event));
540 pwr->state = cpus_cstate_state[cpu];
541 pwr->start_time = cpus_cstate_start_times[cpu];
542 pwr->end_time = last_time;
545 pwr->next = power_events;
551 pwr = malloc(sizeof(struct power_event));
554 memset(pwr, 0, sizeof(struct power_event));
556 pwr->state = cpus_pstate_state[cpu];
557 pwr->start_time = cpus_pstate_start_times[cpu];
558 pwr->end_time = last_time;
561 pwr->next = power_events;
563 if (!pwr->start_time)
564 pwr->start_time = first_time;
566 pwr->state = min_freq;
571 static u64 sample_time(event_t *event, const struct perf_session *session)
576 if (session->sample_type & PERF_SAMPLE_IP)
578 if (session->sample_type & PERF_SAMPLE_TID)
580 if (session->sample_type & PERF_SAMPLE_TIME)
581 return event->sample.array[cursor];
587 * We first queue all events, sorted backwards by insertion.
588 * The order will get flipped later.
590 static int queue_sample_event(event_t *event, struct perf_session *session)
592 struct sample_wrapper *copy, *prev;
595 size = event->sample.header.size + sizeof(struct sample_wrapper) + 8;
601 memset(copy, 0, size);
604 copy->timestamp = sample_time(event, session);
606 memcpy(©->data, event, event->sample.header.size);
608 /* insert in the right place in the list */
611 /* first sample ever */
616 if (all_samples->timestamp < copy->timestamp) {
617 /* insert at the head of the list */
618 copy->next = all_samples;
625 if (prev->next->timestamp < copy->timestamp) {
626 copy->next = prev->next;
632 /* insert at the end of the list */
638 static void sort_queued_samples(void)
640 struct sample_wrapper *cursor, *next;
642 cursor = all_samples;
647 cursor->next = all_samples;
648 all_samples = cursor;
654 * Sort the pid datastructure
656 static void sort_pids(void)
658 struct per_pid *new_list, *p, *cursor, *prev;
659 /* sort by ppid first, then by pid, lowest to highest */
668 if (new_list == NULL) {
676 if (cursor->ppid > p->ppid ||
677 (cursor->ppid == p->ppid && cursor->pid > p->pid)) {
678 /* must insert before */
680 p->next = prev->next;
693 cursor = cursor->next;
702 static void draw_c_p_states(void)
704 struct power_event *pwr;
708 * two pass drawing so that the P state bars are on top of the C state blocks
711 if (pwr->type == CSTATE)
712 svg_cstate(pwr->cpu, pwr->start_time, pwr->end_time, pwr->state);
718 if (pwr->type == PSTATE) {
720 pwr->state = min_freq;
721 svg_pstate(pwr->cpu, pwr->start_time, pwr->end_time, pwr->state);
727 static void draw_wakeups(void)
729 struct wake_event *we;
731 struct per_pidcomm *c;
735 int from = 0, to = 0;
736 char *task_from = NULL, *task_to = NULL;
738 /* locate the column of the waker and wakee */
741 if (p->pid == we->waker || p->pid == we->wakee) {
744 if (c->Y && c->start_time <= we->time && c->end_time >= we->time) {
745 if (p->pid == we->waker && !from) {
747 task_from = strdup(c->comm);
749 if (p->pid == we->wakee && !to) {
751 task_to = strdup(c->comm);
758 if (p->pid == we->waker && !from) {
760 task_from = strdup(c->comm);
762 if (p->pid == we->wakee && !to) {
764 task_to = strdup(c->comm);
773 task_from = malloc(40);
774 sprintf(task_from, "[%i]", we->waker);
777 task_to = malloc(40);
778 sprintf(task_to, "[%i]", we->wakee);
782 svg_interrupt(we->time, to);
783 else if (from && to && abs(from - to) == 1)
784 svg_wakeline(we->time, from, to);
786 svg_partial_wakeline(we->time, from, task_from, to, task_to);
794 static void draw_cpu_usage(void)
797 struct per_pidcomm *c;
798 struct cpu_sample *sample;
805 if (sample->type == TYPE_RUNNING)
806 svg_process(sample->cpu, sample->start_time, sample->end_time, "sample", c->comm);
808 sample = sample->next;
816 static void draw_process_bars(void)
819 struct per_pidcomm *c;
820 struct cpu_sample *sample;
835 svg_box(Y, c->start_time, c->end_time, "process");
838 if (sample->type == TYPE_RUNNING)
839 svg_sample(Y, sample->cpu, sample->start_time, sample->end_time);
840 if (sample->type == TYPE_BLOCKED)
841 svg_box(Y, sample->start_time, sample->end_time, "blocked");
842 if (sample->type == TYPE_WAITING)
843 svg_waiting(Y, sample->start_time, sample->end_time);
844 sample = sample->next;
849 if (c->total_time > 5000000000) /* 5 seconds */
850 sprintf(comm, "%s:%i (%2.2fs)", c->comm, p->pid, c->total_time / 1000000000.0);
852 sprintf(comm, "%s:%i (%3.1fms)", c->comm, p->pid, c->total_time / 1000000.0);
854 svg_text(Y, c->start_time, comm);
864 static void add_process_filter(const char *string)
866 struct process_filter *filt;
869 pid = strtoull(string, NULL, 10);
870 filt = malloc(sizeof(struct process_filter));
874 filt->name = strdup(string);
876 filt->next = process_filter;
878 process_filter = filt;
881 static int passes_filter(struct per_pid *p, struct per_pidcomm *c)
883 struct process_filter *filt;
887 filt = process_filter;
889 if (filt->pid && p->pid == filt->pid)
891 if (strcmp(filt->name, c->comm) == 0)
898 static int determine_display_tasks_filtered(void)
901 struct per_pidcomm *c;
907 if (p->start_time == 1)
908 p->start_time = first_time;
910 /* no exit marker, task kept running to the end */
911 if (p->end_time == 0)
912 p->end_time = last_time;
919 if (c->start_time == 1)
920 c->start_time = first_time;
922 if (passes_filter(p, c)) {
928 if (c->end_time == 0)
929 c->end_time = last_time;
938 static int determine_display_tasks(u64 threshold)
941 struct per_pidcomm *c;
945 return determine_display_tasks_filtered();
950 if (p->start_time == 1)
951 p->start_time = first_time;
953 /* no exit marker, task kept running to the end */
954 if (p->end_time == 0)
955 p->end_time = last_time;
956 if (p->total_time >= threshold && !power_only)
964 if (c->start_time == 1)
965 c->start_time = first_time;
967 if (c->total_time >= threshold && !power_only) {
972 if (c->end_time == 0)
973 c->end_time = last_time;
984 #define TIME_THRESH 10000000
986 static void write_svg_file(const char *filename)
994 count = determine_display_tasks(TIME_THRESH);
996 /* We'd like to show at least 15 tasks; be less picky if we have fewer */
998 count = determine_display_tasks(TIME_THRESH / 10);
1000 open_svg(filename, numcpus, count, first_time, last_time);
1005 for (i = 0; i < numcpus; i++)
1006 svg_cpu_box(i, max_freq, turbo_frequency);
1009 draw_process_bars();
1016 static void process_samples(struct perf_session *session)
1018 struct sample_wrapper *cursor;
1021 sort_queued_samples();
1023 cursor = all_samples;
1025 event = (void *)&cursor->data;
1026 cursor = cursor->next;
1027 process_sample_event(event, session);
1031 static struct perf_event_ops event_ops = {
1032 .comm = process_comm_event,
1033 .fork = process_fork_event,
1034 .exit = process_exit_event,
1035 .sample = queue_sample_event,
1038 static int __cmd_timechart(void)
1040 struct perf_session *session = perf_session__new(input_name, O_RDONLY, 0);
1043 if (session == NULL)
1046 if (!perf_session__has_traces(session, "timechart record"))
1049 ret = perf_session__process_events(session, &event_ops);
1053 process_samples(session);
1055 end_sample_processing();
1059 write_svg_file(output_name);
1061 pr_info("Written %2.1f seconds of trace to %s.\n",
1062 (last_time - first_time) / 1000000000.0, output_name);
1064 perf_session__delete(session);
1068 static const char * const timechart_usage[] = {
1069 "perf timechart [<options>] {record}",
1073 static const char *record_args[] = {
1080 "-e", "power:power_start",
1081 "-e", "power:power_end",
1082 "-e", "power:power_frequency",
1083 "-e", "sched:sched_wakeup",
1084 "-e", "sched:sched_switch",
1087 static int __cmd_record(int argc, const char **argv)
1089 unsigned int rec_argc, i, j;
1090 const char **rec_argv;
1092 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1093 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1095 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1096 rec_argv[i] = strdup(record_args[i]);
1098 for (j = 1; j < (unsigned int)argc; j++, i++)
1099 rec_argv[i] = argv[j];
1101 return cmd_record(i, rec_argv, NULL);
1105 parse_process(const struct option *opt __used, const char *arg, int __used unset)
1108 add_process_filter(arg);
1112 static const struct option options[] = {
1113 OPT_STRING('i', "input", &input_name, "file",
1115 OPT_STRING('o', "output", &output_name, "file",
1116 "output file name"),
1117 OPT_INTEGER('w', "width", &svg_page_width,
1119 OPT_BOOLEAN('P', "power-only", &power_only,
1120 "output power data only"),
1121 OPT_CALLBACK('p', "process", NULL, "process",
1122 "process selector. Pass a pid or process name.",
1128 int cmd_timechart(int argc, const char **argv, const char *prefix __used)
1130 argc = parse_options(argc, argv, options, timechart_usage,
1131 PARSE_OPT_STOP_AT_NON_OPTION);
1135 if (argc && !strncmp(argv[0], "rec", 3))
1136 return __cmd_record(argc, argv);
1138 usage_with_options(timechart_usage, options);
1142 return __cmd_timechart();