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/svghelper.h"
34 static char const *input_name = "perf.data";
35 static char const *output_name = "output.svg";
38 static unsigned long page_size;
39 static unsigned long mmap_window = 32;
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;
52 static struct perf_header *header;
61 struct sample_wrapper;
64 * Datastructure layout:
65 * We keep an list of "pid"s, matching the kernels notion of a task struct.
66 * Each "pid" entry, has a list of "comm"s.
67 * this is because we want to track different programs different, while
68 * exec will reuse the original pid (by design).
69 * Each comm has a list of samples that will be used to draw
84 struct per_pidcomm *all;
85 struct per_pidcomm *current;
92 struct per_pidcomm *next;
106 struct cpu_sample *samples;
109 struct sample_wrapper {
110 struct sample_wrapper *next;
113 unsigned char data[0];
117 #define TYPE_RUNNING 1
118 #define TYPE_WAITING 2
119 #define TYPE_BLOCKED 3
122 struct cpu_sample *next;
130 static struct per_pid *all_data;
136 struct power_event *next;
145 struct wake_event *next;
151 static struct power_event *power_events;
152 static struct wake_event *wake_events;
154 struct sample_wrapper *all_samples;
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];
276 process_comm_event(event_t *event)
278 pid_set_comm(event->comm.pid, event->comm.comm);
282 process_fork_event(event_t *event)
284 pid_fork(event->fork.pid, event->fork.ppid, event->fork.time);
289 process_exit_event(event_t *event)
291 pid_exit(event->fork.pid, event->fork.time);
299 unsigned char preempt_count;
305 struct trace_entry te;
310 #define TASK_COMM_LEN 16
311 struct wakeup_entry {
312 struct trace_entry te;
313 char comm[TASK_COMM_LEN];
320 * trace_flag_type is an enumeration that holds different
321 * states when a trace occurs. These are:
322 * IRQS_OFF - interrupts were disabled
323 * IRQS_NOSUPPORT - arch does not support irqs_disabled_flags
324 * NEED_RESCED - reschedule is requested
325 * HARDIRQ - inside an interrupt handler
326 * SOFTIRQ - inside a softirq handler
328 enum trace_flag_type {
329 TRACE_FLAG_IRQS_OFF = 0x01,
330 TRACE_FLAG_IRQS_NOSUPPORT = 0x02,
331 TRACE_FLAG_NEED_RESCHED = 0x04,
332 TRACE_FLAG_HARDIRQ = 0x08,
333 TRACE_FLAG_SOFTIRQ = 0x10,
338 struct sched_switch {
339 struct trace_entry te;
340 char prev_comm[TASK_COMM_LEN];
343 long prev_state; /* Arjan weeps. */
344 char next_comm[TASK_COMM_LEN];
349 static void c_state_start(int cpu, u64 timestamp, int state)
351 cpus_cstate_start_times[cpu] = timestamp;
352 cpus_cstate_state[cpu] = state;
355 static void c_state_end(int cpu, u64 timestamp)
357 struct power_event *pwr;
358 pwr = malloc(sizeof(struct power_event));
361 memset(pwr, 0, sizeof(struct power_event));
363 pwr->state = cpus_cstate_state[cpu];
364 pwr->start_time = cpus_cstate_start_times[cpu];
365 pwr->end_time = timestamp;
368 pwr->next = power_events;
373 static void p_state_change(int cpu, u64 timestamp, u64 new_freq)
375 struct power_event *pwr;
376 pwr = malloc(sizeof(struct power_event));
378 if (new_freq > 8000000) /* detect invalid data */
383 memset(pwr, 0, sizeof(struct power_event));
385 pwr->state = cpus_pstate_state[cpu];
386 pwr->start_time = cpus_pstate_start_times[cpu];
387 pwr->end_time = timestamp;
390 pwr->next = power_events;
392 if (!pwr->start_time)
393 pwr->start_time = first_time;
397 cpus_pstate_state[cpu] = new_freq;
398 cpus_pstate_start_times[cpu] = timestamp;
400 if ((u64)new_freq > max_freq)
403 if (new_freq < min_freq || min_freq == 0)
406 if (new_freq == max_freq - 1000)
407 turbo_frequency = max_freq;
411 sched_wakeup(int cpu, u64 timestamp, int pid, struct trace_entry *te)
413 struct wake_event *we;
415 struct wakeup_entry *wake = (void *)te;
417 we = malloc(sizeof(struct wake_event));
421 memset(we, 0, sizeof(struct wake_event));
422 we->time = timestamp;
425 if ((te->flags & TRACE_FLAG_HARDIRQ) || (te->flags & TRACE_FLAG_SOFTIRQ))
428 we->wakee = wake->pid;
429 we->next = wake_events;
431 p = find_create_pid(we->wakee);
433 if (p && p->current && p->current->state == TYPE_NONE) {
434 p->current->state_since = timestamp;
435 p->current->state = TYPE_WAITING;
437 if (p && p->current && p->current->state == TYPE_BLOCKED) {
438 pid_put_sample(p->pid, p->current->state, cpu, p->current->state_since, timestamp);
439 p->current->state_since = timestamp;
440 p->current->state = TYPE_WAITING;
444 static void sched_switch(int cpu, u64 timestamp, struct trace_entry *te)
446 struct per_pid *p = NULL, *prev_p;
447 struct sched_switch *sw = (void *)te;
450 prev_p = find_create_pid(sw->prev_pid);
452 p = find_create_pid(sw->next_pid);
454 if (prev_p->current && prev_p->current->state != TYPE_NONE)
455 pid_put_sample(sw->prev_pid, TYPE_RUNNING, cpu, prev_p->current->state_since, timestamp);
456 if (p && p->current) {
457 if (p->current->state != TYPE_NONE)
458 pid_put_sample(sw->next_pid, p->current->state, cpu, p->current->state_since, timestamp);
460 p->current->state_since = timestamp;
461 p->current->state = TYPE_RUNNING;
464 if (prev_p->current) {
465 prev_p->current->state = TYPE_NONE;
466 prev_p->current->state_since = timestamp;
467 if (sw->prev_state & 2)
468 prev_p->current->state = TYPE_BLOCKED;
469 if (sw->prev_state == 0)
470 prev_p->current->state = TYPE_WAITING;
476 process_sample_event(event_t *event)
483 struct trace_entry *te;
485 if (sample_type & PERF_SAMPLE_IP)
488 if (sample_type & PERF_SAMPLE_TID) {
489 pid = event->sample.array[cursor]>>32;
492 if (sample_type & PERF_SAMPLE_TIME) {
493 stamp = event->sample.array[cursor++];
495 if (!first_time || first_time > stamp)
497 if (last_time < stamp)
501 if (sample_type & PERF_SAMPLE_ADDR)
502 addr = event->sample.array[cursor++];
503 if (sample_type & PERF_SAMPLE_ID)
505 if (sample_type & PERF_SAMPLE_STREAM_ID)
507 if (sample_type & PERF_SAMPLE_CPU)
508 cpu = event->sample.array[cursor++] & 0xFFFFFFFF;
509 if (sample_type & PERF_SAMPLE_PERIOD)
512 te = (void *)&event->sample.array[cursor];
514 if (sample_type & PERF_SAMPLE_RAW && te->size > 0) {
516 struct power_entry *pe;
520 event_str = perf_header__find_event(te->type);
525 if (strcmp(event_str, "power:power_start") == 0)
526 c_state_start(cpu, stamp, pe->value);
528 if (strcmp(event_str, "power:power_end") == 0)
529 c_state_end(cpu, stamp);
531 if (strcmp(event_str, "power:power_frequency") == 0)
532 p_state_change(cpu, stamp, pe->value);
534 if (strcmp(event_str, "sched:sched_wakeup") == 0)
535 sched_wakeup(cpu, stamp, pid, te);
537 if (strcmp(event_str, "sched:sched_switch") == 0)
538 sched_switch(cpu, stamp, te);
544 * After the last sample we need to wrap up the current C/P state
545 * and close out each CPU for these.
547 static void end_sample_processing(void)
550 struct power_event *pwr;
552 for (cpu = 0; cpu <= numcpus; cpu++) {
553 pwr = malloc(sizeof(struct power_event));
556 memset(pwr, 0, sizeof(struct power_event));
560 pwr->state = cpus_cstate_state[cpu];
561 pwr->start_time = cpus_cstate_start_times[cpu];
562 pwr->end_time = last_time;
565 pwr->next = power_events;
571 pwr = malloc(sizeof(struct power_event));
574 memset(pwr, 0, sizeof(struct power_event));
576 pwr->state = cpus_pstate_state[cpu];
577 pwr->start_time = cpus_pstate_start_times[cpu];
578 pwr->end_time = last_time;
581 pwr->next = power_events;
583 if (!pwr->start_time)
584 pwr->start_time = first_time;
586 pwr->state = min_freq;
591 static u64 sample_time(event_t *event)
596 if (sample_type & PERF_SAMPLE_IP)
598 if (sample_type & PERF_SAMPLE_TID)
600 if (sample_type & PERF_SAMPLE_TIME)
601 return event->sample.array[cursor];
607 * We first queue all events, sorted backwards by insertion.
608 * The order will get flipped later.
611 queue_sample_event(event_t *event)
613 struct sample_wrapper *copy, *prev;
616 size = event->sample.header.size + sizeof(struct sample_wrapper) + 8;
622 memset(copy, 0, size);
625 copy->timestamp = sample_time(event);
627 memcpy(©->data, event, event->sample.header.size);
629 /* insert in the right place in the list */
632 /* first sample ever */
637 if (all_samples->timestamp < copy->timestamp) {
638 /* insert at the head of the list */
639 copy->next = all_samples;
646 if (prev->next->timestamp < copy->timestamp) {
647 copy->next = prev->next;
653 /* insert at the end of the list */
659 static void sort_queued_samples(void)
661 struct sample_wrapper *cursor, *next;
663 cursor = all_samples;
668 cursor->next = all_samples;
669 all_samples = cursor;
675 * Sort the pid datastructure
677 static void sort_pids(void)
679 struct per_pid *new_list, *p, *cursor, *prev;
680 /* sort by ppid first, then by pid, lowest to highest */
689 if (new_list == NULL) {
697 if (cursor->ppid > p->ppid ||
698 (cursor->ppid == p->ppid && cursor->pid > p->pid)) {
699 /* must insert before */
701 p->next = prev->next;
714 cursor = cursor->next;
723 static void draw_c_p_states(void)
725 struct power_event *pwr;
729 * two pass drawing so that the P state bars are on top of the C state blocks
732 if (pwr->type == CSTATE)
733 svg_cstate(pwr->cpu, pwr->start_time, pwr->end_time, pwr->state);
739 if (pwr->type == PSTATE) {
741 pwr->state = min_freq;
742 svg_pstate(pwr->cpu, pwr->start_time, pwr->end_time, pwr->state);
748 static void draw_wakeups(void)
750 struct wake_event *we;
752 struct per_pidcomm *c;
756 int from = 0, to = 0;
757 char *task_from = NULL, *task_to = NULL;
759 /* locate the column of the waker and wakee */
762 if (p->pid == we->waker || p->pid == we->wakee) {
765 if (c->Y && c->start_time <= we->time && c->end_time >= we->time) {
766 if (p->pid == we->waker) {
770 if (p->pid == 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);
791 static void draw_cpu_usage(void)
794 struct per_pidcomm *c;
795 struct cpu_sample *sample;
802 if (sample->type == TYPE_RUNNING)
803 svg_process(sample->cpu, sample->start_time, sample->end_time, "sample", c->comm);
805 sample = sample->next;
813 static void draw_process_bars(void)
816 struct per_pidcomm *c;
817 struct cpu_sample *sample;
832 svg_box(Y, c->start_time, c->end_time, "process");
835 if (sample->type == TYPE_RUNNING)
836 svg_sample(Y, sample->cpu, sample->start_time, sample->end_time);
837 if (sample->type == TYPE_BLOCKED)
838 svg_box(Y, sample->start_time, sample->end_time, "blocked");
839 if (sample->type == TYPE_WAITING)
840 svg_waiting(Y, sample->start_time, sample->end_time);
841 sample = sample->next;
846 if (c->total_time > 5000000000) /* 5 seconds */
847 sprintf(comm, "%s:%i (%2.2fs)", c->comm, p->pid, c->total_time / 1000000000.0);
849 sprintf(comm, "%s:%i (%3.1fms)", c->comm, p->pid, c->total_time / 1000000.0);
851 svg_text(Y, c->start_time, comm);
861 static int determine_display_tasks(u64 threshold)
864 struct per_pidcomm *c;
870 if (p->start_time == 1)
871 p->start_time = first_time;
873 /* no exit marker, task kept running to the end */
874 if (p->end_time == 0)
875 p->end_time = last_time;
876 if (p->total_time >= threshold && !power_only)
884 if (c->start_time == 1)
885 c->start_time = first_time;
887 if (c->total_time >= threshold && !power_only) {
892 if (c->end_time == 0)
893 c->end_time = last_time;
904 #define TIME_THRESH 10000000
906 static void write_svg_file(const char *filename)
914 count = determine_display_tasks(TIME_THRESH);
916 /* We'd like to show at least 15 tasks; be less picky if we have fewer */
918 count = determine_display_tasks(TIME_THRESH / 10);
920 open_svg(filename, numcpus, count, first_time, last_time);
925 for (i = 0; i < numcpus; i++)
926 svg_cpu_box(i, max_freq, turbo_frequency);
937 process_event(event_t *event)
940 switch (event->header.type) {
942 case PERF_RECORD_COMM:
943 return process_comm_event(event);
944 case PERF_RECORD_FORK:
945 return process_fork_event(event);
946 case PERF_RECORD_EXIT:
947 return process_exit_event(event);
948 case PERF_RECORD_SAMPLE:
949 return queue_sample_event(event);
952 * We dont process them right now but they are fine:
954 case PERF_RECORD_MMAP:
955 case PERF_RECORD_THROTTLE:
956 case PERF_RECORD_UNTHROTTLE:
966 static void process_samples(void)
968 struct sample_wrapper *cursor;
971 sort_queued_samples();
973 cursor = all_samples;
975 event = (void *)&cursor->data;
976 cursor = cursor->next;
977 process_sample_event(event);
982 static int __cmd_timechart(void)
984 int ret, rc = EXIT_FAILURE;
985 unsigned long offset = 0;
986 unsigned long head, shift;
993 input = open(input_name, O_RDONLY);
995 fprintf(stderr, " failed to open file: %s", input_name);
996 if (!strcmp(input_name, "perf.data"))
997 fprintf(stderr, " (try 'perf record' first)");
998 fprintf(stderr, "\n");
1002 ret = fstat(input, &statbuf);
1004 perror("failed to stat file");
1008 if (!statbuf.st_size) {
1009 fprintf(stderr, "zero-sized file, nothing to do!\n");
1013 header = perf_header__read(input);
1014 head = header->data_offset;
1016 sample_type = perf_header__sample_type(header);
1018 shift = page_size * (head / page_size);
1023 buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1024 MAP_SHARED, input, offset);
1025 if (buf == MAP_FAILED) {
1026 perror("failed to mmap file");
1031 event = (event_t *)(buf + head);
1033 size = event->header.size;
1037 if (head + event->header.size >= page_size * mmap_window) {
1040 shift = page_size * (head / page_size);
1042 ret2 = munmap(buf, page_size * mmap_window);
1050 size = event->header.size;
1052 if (!size || process_event(event) < 0) {
1054 printf("%p [%p]: skipping unknown header type: %d\n",
1055 (void *)(offset + head),
1056 (void *)(long)(event->header.size),
1057 event->header.type);
1060 * assume we lost track of the stream, check alignment, and
1061 * increment a single u64 in the hope to catch on again 'soon'.
1064 if (unlikely(head & 7))
1072 if (offset + head >= header->data_offset + header->data_size)
1075 if (offset + head < (unsigned long)statbuf.st_size)
1085 end_sample_processing();
1089 write_svg_file(output_name);
1091 printf("Written %2.1f seconds of trace to %s.\n", (last_time - first_time) / 1000000000.0, output_name);
1096 static const char * const timechart_usage[] = {
1097 "perf timechart [<options>] {record}",
1101 static const char *record_args[] = {
1108 "-e", "power:power_start",
1109 "-e", "power:power_end",
1110 "-e", "power:power_frequency",
1111 "-e", "sched:sched_wakeup",
1112 "-e", "sched:sched_switch",
1115 static int __cmd_record(int argc, const char **argv)
1117 unsigned int rec_argc, i, j;
1118 const char **rec_argv;
1120 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1121 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1123 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1124 rec_argv[i] = strdup(record_args[i]);
1126 for (j = 1; j < (unsigned int)argc; j++, i++)
1127 rec_argv[i] = argv[j];
1129 return cmd_record(i, rec_argv, NULL);
1132 static const struct option options[] = {
1133 OPT_STRING('i', "input", &input_name, "file",
1135 OPT_STRING('o', "output", &output_name, "file",
1136 "output file name"),
1137 OPT_INTEGER('w', "width", &svg_page_width,
1139 OPT_BOOLEAN('p', "power-only", &power_only,
1140 "output power data only"),
1145 int cmd_timechart(int argc, const char **argv, const char *prefix __used)
1149 page_size = getpagesize();
1151 argc = parse_options(argc, argv, options, timechart_usage,
1152 PARSE_OPT_STOP_AT_NON_OPTION);
1154 if (argc && !strncmp(argv[0], "rec", 3))
1155 return __cmd_record(argc, argv);
1157 usage_with_options(timechart_usage, options);
1161 return __cmd_timechart();