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);
308 unsigned char preempt_count;
314 struct trace_entry te;
319 #define TASK_COMM_LEN 16
320 struct wakeup_entry {
321 struct trace_entry te;
322 char comm[TASK_COMM_LEN];
329 * trace_flag_type is an enumeration that holds different
330 * states when a trace occurs. These are:
331 * IRQS_OFF - interrupts were disabled
332 * IRQS_NOSUPPORT - arch does not support irqs_disabled_flags
333 * NEED_RESCED - reschedule is requested
334 * HARDIRQ - inside an interrupt handler
335 * SOFTIRQ - inside a softirq handler
337 enum trace_flag_type {
338 TRACE_FLAG_IRQS_OFF = 0x01,
339 TRACE_FLAG_IRQS_NOSUPPORT = 0x02,
340 TRACE_FLAG_NEED_RESCHED = 0x04,
341 TRACE_FLAG_HARDIRQ = 0x08,
342 TRACE_FLAG_SOFTIRQ = 0x10,
347 struct sched_switch {
348 struct trace_entry te;
349 char prev_comm[TASK_COMM_LEN];
352 long prev_state; /* Arjan weeps. */
353 char next_comm[TASK_COMM_LEN];
358 static void c_state_start(int cpu, u64 timestamp, int state)
360 cpus_cstate_start_times[cpu] = timestamp;
361 cpus_cstate_state[cpu] = state;
364 static void c_state_end(int cpu, u64 timestamp)
366 struct power_event *pwr;
367 pwr = malloc(sizeof(struct power_event));
370 memset(pwr, 0, sizeof(struct power_event));
372 pwr->state = cpus_cstate_state[cpu];
373 pwr->start_time = cpus_cstate_start_times[cpu];
374 pwr->end_time = timestamp;
377 pwr->next = power_events;
382 static void p_state_change(int cpu, u64 timestamp, u64 new_freq)
384 struct power_event *pwr;
385 pwr = malloc(sizeof(struct power_event));
387 if (new_freq > 8000000) /* detect invalid data */
392 memset(pwr, 0, sizeof(struct power_event));
394 pwr->state = cpus_pstate_state[cpu];
395 pwr->start_time = cpus_pstate_start_times[cpu];
396 pwr->end_time = timestamp;
399 pwr->next = power_events;
401 if (!pwr->start_time)
402 pwr->start_time = first_time;
406 cpus_pstate_state[cpu] = new_freq;
407 cpus_pstate_start_times[cpu] = timestamp;
409 if ((u64)new_freq > max_freq)
412 if (new_freq < min_freq || min_freq == 0)
415 if (new_freq == max_freq - 1000)
416 turbo_frequency = max_freq;
420 sched_wakeup(int cpu, u64 timestamp, int pid, struct trace_entry *te)
422 struct wake_event *we;
424 struct wakeup_entry *wake = (void *)te;
426 we = malloc(sizeof(struct wake_event));
430 memset(we, 0, sizeof(struct wake_event));
431 we->time = timestamp;
434 if ((te->flags & TRACE_FLAG_HARDIRQ) || (te->flags & TRACE_FLAG_SOFTIRQ))
437 we->wakee = wake->pid;
438 we->next = wake_events;
440 p = find_create_pid(we->wakee);
442 if (p && p->current && p->current->state == TYPE_NONE) {
443 p->current->state_since = timestamp;
444 p->current->state = TYPE_WAITING;
446 if (p && p->current && p->current->state == TYPE_BLOCKED) {
447 pid_put_sample(p->pid, p->current->state, cpu, p->current->state_since, timestamp);
448 p->current->state_since = timestamp;
449 p->current->state = TYPE_WAITING;
453 static void sched_switch(int cpu, u64 timestamp, struct trace_entry *te)
455 struct per_pid *p = NULL, *prev_p;
456 struct sched_switch *sw = (void *)te;
459 prev_p = find_create_pid(sw->prev_pid);
461 p = find_create_pid(sw->next_pid);
463 if (prev_p->current && prev_p->current->state != TYPE_NONE)
464 pid_put_sample(sw->prev_pid, TYPE_RUNNING, cpu, prev_p->current->state_since, timestamp);
465 if (p && p->current) {
466 if (p->current->state != TYPE_NONE)
467 pid_put_sample(sw->next_pid, p->current->state, cpu, p->current->state_since, timestamp);
469 p->current->state_since = timestamp;
470 p->current->state = TYPE_RUNNING;
473 if (prev_p->current) {
474 prev_p->current->state = TYPE_NONE;
475 prev_p->current->state_since = timestamp;
476 if (sw->prev_state & 2)
477 prev_p->current->state = TYPE_BLOCKED;
478 if (sw->prev_state == 0)
479 prev_p->current->state = TYPE_WAITING;
485 process_sample_event(event_t *event)
492 struct trace_entry *te;
494 if (sample_type & PERF_SAMPLE_IP)
497 if (sample_type & PERF_SAMPLE_TID) {
498 pid = event->sample.array[cursor]>>32;
501 if (sample_type & PERF_SAMPLE_TIME) {
502 stamp = event->sample.array[cursor++];
504 if (!first_time || first_time > stamp)
506 if (last_time < stamp)
510 if (sample_type & PERF_SAMPLE_ADDR)
511 addr = event->sample.array[cursor++];
512 if (sample_type & PERF_SAMPLE_ID)
514 if (sample_type & PERF_SAMPLE_STREAM_ID)
516 if (sample_type & PERF_SAMPLE_CPU)
517 cpu = event->sample.array[cursor++] & 0xFFFFFFFF;
518 if (sample_type & PERF_SAMPLE_PERIOD)
521 te = (void *)&event->sample.array[cursor];
523 if (sample_type & PERF_SAMPLE_RAW && te->size > 0) {
525 struct power_entry *pe;
529 event_str = perf_header__find_event(te->type);
534 if (strcmp(event_str, "power:power_start") == 0)
535 c_state_start(cpu, stamp, pe->value);
537 if (strcmp(event_str, "power:power_end") == 0)
538 c_state_end(cpu, stamp);
540 if (strcmp(event_str, "power:power_frequency") == 0)
541 p_state_change(cpu, stamp, pe->value);
543 if (strcmp(event_str, "sched:sched_wakeup") == 0)
544 sched_wakeup(cpu, stamp, pid, te);
546 if (strcmp(event_str, "sched:sched_switch") == 0)
547 sched_switch(cpu, stamp, te);
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;
600 static u64 sample_time(event_t *event)
605 if (sample_type & PERF_SAMPLE_IP)
607 if (sample_type & PERF_SAMPLE_TID)
609 if (sample_type & PERF_SAMPLE_TIME)
610 return event->sample.array[cursor];
616 * We first queue all events, sorted backwards by insertion.
617 * The order will get flipped later.
620 queue_sample_event(event_t *event)
622 struct sample_wrapper *copy, *prev;
625 size = event->sample.header.size + sizeof(struct sample_wrapper) + 8;
631 memset(copy, 0, size);
634 copy->timestamp = sample_time(event);
636 memcpy(©->data, event, event->sample.header.size);
638 /* insert in the right place in the list */
641 /* first sample ever */
646 if (all_samples->timestamp < copy->timestamp) {
647 /* insert at the head of the list */
648 copy->next = all_samples;
655 if (prev->next->timestamp < copy->timestamp) {
656 copy->next = prev->next;
662 /* insert at the end of the list */
668 static void sort_queued_samples(void)
670 struct sample_wrapper *cursor, *next;
672 cursor = all_samples;
677 cursor->next = all_samples;
678 all_samples = cursor;
684 * Sort the pid datastructure
686 static void sort_pids(void)
688 struct per_pid *new_list, *p, *cursor, *prev;
689 /* sort by ppid first, then by pid, lowest to highest */
698 if (new_list == NULL) {
706 if (cursor->ppid > p->ppid ||
707 (cursor->ppid == p->ppid && cursor->pid > p->pid)) {
708 /* must insert before */
710 p->next = prev->next;
723 cursor = cursor->next;
732 static void draw_c_p_states(void)
734 struct power_event *pwr;
738 * two pass drawing so that the P state bars are on top of the C state blocks
741 if (pwr->type == CSTATE)
742 svg_cstate(pwr->cpu, pwr->start_time, pwr->end_time, pwr->state);
748 if (pwr->type == PSTATE) {
750 pwr->state = min_freq;
751 svg_pstate(pwr->cpu, pwr->start_time, pwr->end_time, pwr->state);
757 static void draw_wakeups(void)
759 struct wake_event *we;
761 struct per_pidcomm *c;
765 int from = 0, to = 0;
766 char *task_from = NULL, *task_to = NULL;
768 /* locate the column of the waker and wakee */
771 if (p->pid == we->waker || p->pid == we->wakee) {
774 if (c->Y && c->start_time <= we->time && c->end_time >= we->time) {
775 if (p->pid == we->waker && !from) {
777 task_from = strdup(c->comm);
779 if (p->pid == we->wakee && !to) {
781 task_to = strdup(c->comm);
788 if (p->pid == we->waker && !from) {
790 task_from = strdup(c->comm);
792 if (p->pid == we->wakee && !to) {
794 task_to = strdup(c->comm);
803 task_from = malloc(40);
804 sprintf(task_from, "[%i]", we->waker);
807 task_to = malloc(40);
808 sprintf(task_to, "[%i]", we->wakee);
812 svg_interrupt(we->time, to);
813 else if (from && to && abs(from - to) == 1)
814 svg_wakeline(we->time, from, to);
816 svg_partial_wakeline(we->time, from, task_from, to, task_to);
824 static void draw_cpu_usage(void)
827 struct per_pidcomm *c;
828 struct cpu_sample *sample;
835 if (sample->type == TYPE_RUNNING)
836 svg_process(sample->cpu, sample->start_time, sample->end_time, "sample", c->comm);
838 sample = sample->next;
846 static void draw_process_bars(void)
849 struct per_pidcomm *c;
850 struct cpu_sample *sample;
865 svg_box(Y, c->start_time, c->end_time, "process");
868 if (sample->type == TYPE_RUNNING)
869 svg_sample(Y, sample->cpu, sample->start_time, sample->end_time);
870 if (sample->type == TYPE_BLOCKED)
871 svg_box(Y, sample->start_time, sample->end_time, "blocked");
872 if (sample->type == TYPE_WAITING)
873 svg_waiting(Y, sample->start_time, sample->end_time);
874 sample = sample->next;
879 if (c->total_time > 5000000000) /* 5 seconds */
880 sprintf(comm, "%s:%i (%2.2fs)", c->comm, p->pid, c->total_time / 1000000000.0);
882 sprintf(comm, "%s:%i (%3.1fms)", c->comm, p->pid, c->total_time / 1000000.0);
884 svg_text(Y, c->start_time, comm);
894 static void add_process_filter(const char *string)
896 struct process_filter *filt;
899 pid = strtoull(string, NULL, 10);
900 filt = malloc(sizeof(struct process_filter));
904 filt->name = strdup(string);
906 filt->next = process_filter;
908 process_filter = filt;
911 static int passes_filter(struct per_pid *p, struct per_pidcomm *c)
913 struct process_filter *filt;
917 filt = process_filter;
919 if (filt->pid && p->pid == filt->pid)
921 if (strcmp(filt->name, c->comm) == 0)
928 static int determine_display_tasks_filtered(void)
931 struct per_pidcomm *c;
937 if (p->start_time == 1)
938 p->start_time = first_time;
940 /* no exit marker, task kept running to the end */
941 if (p->end_time == 0)
942 p->end_time = last_time;
949 if (c->start_time == 1)
950 c->start_time = first_time;
952 if (passes_filter(p, c)) {
958 if (c->end_time == 0)
959 c->end_time = last_time;
968 static int determine_display_tasks(u64 threshold)
971 struct per_pidcomm *c;
975 return determine_display_tasks_filtered();
980 if (p->start_time == 1)
981 p->start_time = first_time;
983 /* no exit marker, task kept running to the end */
984 if (p->end_time == 0)
985 p->end_time = last_time;
986 if (p->total_time >= threshold && !power_only)
994 if (c->start_time == 1)
995 c->start_time = first_time;
997 if (c->total_time >= threshold && !power_only) {
1002 if (c->end_time == 0)
1003 c->end_time = last_time;
1014 #define TIME_THRESH 10000000
1016 static void write_svg_file(const char *filename)
1024 count = determine_display_tasks(TIME_THRESH);
1026 /* We'd like to show at least 15 tasks; be less picky if we have fewer */
1028 count = determine_display_tasks(TIME_THRESH / 10);
1030 open_svg(filename, numcpus, count, first_time, last_time);
1035 for (i = 0; i < numcpus; i++)
1036 svg_cpu_box(i, max_freq, turbo_frequency);
1039 draw_process_bars();
1046 static void process_samples(void)
1048 struct sample_wrapper *cursor;
1051 sort_queued_samples();
1053 cursor = all_samples;
1055 event = (void *)&cursor->data;
1056 cursor = cursor->next;
1057 process_sample_event(event);
1061 static int sample_type_check(u64 type)
1065 if (!(sample_type & PERF_SAMPLE_RAW)) {
1066 fprintf(stderr, "No trace samples found in the file.\n"
1067 "Have you used 'perf timechart record' to record it?\n");
1074 static struct perf_file_handler file_handler = {
1075 .process_comm_event = process_comm_event,
1076 .process_fork_event = process_fork_event,
1077 .process_exit_event = process_exit_event,
1078 .process_sample_event = queue_sample_event,
1079 .sample_type_check = sample_type_check,
1082 static int __cmd_timechart(void)
1084 struct perf_header *header;
1087 register_perf_file_handler(&file_handler);
1089 ret = mmap_dispatch_perf_file(&header, input_name, 0, 0,
1090 &event__cwdlen, &event__cwd);
1092 return EXIT_FAILURE;
1096 end_sample_processing();
1100 write_svg_file(output_name);
1102 pr_info("Written %2.1f seconds of trace to %s.\n",
1103 (last_time - first_time) / 1000000000.0, output_name);
1105 return EXIT_SUCCESS;
1108 static const char * const timechart_usage[] = {
1109 "perf timechart [<options>] {record}",
1113 static const char *record_args[] = {
1120 "-e", "power:power_start",
1121 "-e", "power:power_end",
1122 "-e", "power:power_frequency",
1123 "-e", "sched:sched_wakeup",
1124 "-e", "sched:sched_switch",
1127 static int __cmd_record(int argc, const char **argv)
1129 unsigned int rec_argc, i, j;
1130 const char **rec_argv;
1132 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1133 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1135 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1136 rec_argv[i] = strdup(record_args[i]);
1138 for (j = 1; j < (unsigned int)argc; j++, i++)
1139 rec_argv[i] = argv[j];
1141 return cmd_record(i, rec_argv, NULL);
1145 parse_process(const struct option *opt __used, const char *arg, int __used unset)
1148 add_process_filter(arg);
1152 static const struct option options[] = {
1153 OPT_STRING('i', "input", &input_name, "file",
1155 OPT_STRING('o', "output", &output_name, "file",
1156 "output file name"),
1157 OPT_INTEGER('w', "width", &svg_page_width,
1159 OPT_BOOLEAN('P', "power-only", &power_only,
1160 "output power data only"),
1161 OPT_CALLBACK('p', "process", NULL, "process",
1162 "process selector. Pass a pid or process name.",
1168 int cmd_timechart(int argc, const char **argv, const char *prefix __used)
1172 argc = parse_options(argc, argv, options, timechart_usage,
1173 PARSE_OPT_STOP_AT_NON_OPTION);
1175 if (argc && !strncmp(argv[0], "rec", 3))
1176 return __cmd_record(argc, argv);
1178 usage_with_options(timechart_usage, options);
1182 return __cmd_timechart();