4 #include "util/cache.h"
5 #include "util/symbol.h"
6 #include "util/thread.h"
7 #include "util/header.h"
9 #include "util/parse-options.h"
12 #include "util/debug.h"
14 #include "util/trace-event.h"
15 #include <sys/types.h>
20 static char const *input_name = "perf.data";
22 static unsigned long page_size;
23 static unsigned long mmap_window = 32;
25 static unsigned long total_comm = 0;
27 static struct rb_root threads;
28 static struct thread *last_match;
30 static struct perf_header *header;
31 static u64 sample_type;
33 static int replay_mode;
38 * Scheduler benchmarks
40 #include <sys/resource.h>
41 #include <sys/types.h>
44 #include <sys/prctl.h>
46 #include <linux/unistd.h>
48 #include <semaphore.h>
62 #define PR_SET_NAME 15 /* Set process name */
64 #define BUG_ON(x) assert(!(x))
68 typedef unsigned long long nsec_t;
70 static nsec_t run_measurement_overhead;
71 static nsec_t sleep_measurement_overhead;
73 static nsec_t get_nsecs(void)
77 clock_gettime(CLOCK_MONOTONIC, &ts);
79 return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
82 static void burn_nsecs(nsec_t nsecs)
84 nsec_t T0 = get_nsecs(), T1;
88 } while (T1 + run_measurement_overhead < T0 + nsecs);
91 static void sleep_nsecs(nsec_t nsecs)
95 ts.tv_nsec = nsecs % 999999999;
96 ts.tv_sec = nsecs / 999999999;
101 static void calibrate_run_measurement_overhead(void)
103 nsec_t T0, T1, delta, min_delta = 1000000000ULL;
106 for (i = 0; i < 10; i++) {
111 min_delta = min(min_delta, delta);
113 run_measurement_overhead = min_delta;
115 printf("run measurement overhead: %Ld nsecs\n", min_delta);
118 static void calibrate_sleep_measurement_overhead(void)
120 nsec_t T0, T1, delta, min_delta = 1000000000ULL;
123 for (i = 0; i < 10; i++) {
128 min_delta = min(min_delta, delta);
131 sleep_measurement_overhead = min_delta;
133 printf("sleep measurement overhead: %Ld nsecs\n", min_delta);
139 #define MAX_PID 65536
141 static unsigned long nr_tasks;
150 unsigned long nr_events;
151 unsigned long curr_event;
152 struct sched_event **events;
157 sem_t ready_for_work;
163 enum sched_event_type {
170 enum sched_event_type type;
176 struct task_desc *wakee;
179 static struct task_desc *pid_to_task[MAX_PID];
181 static struct task_desc **tasks;
183 static pthread_mutex_t start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
184 static nsec_t start_time;
186 static pthread_mutex_t work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
188 static unsigned long nr_run_events;
189 static unsigned long nr_sleep_events;
190 static unsigned long nr_wakeup_events;
192 static unsigned long nr_sleep_corrections;
193 static unsigned long nr_run_events_optimized;
195 static struct sched_event *
196 get_new_event(struct task_desc *task, nsec_t timestamp)
198 struct sched_event *event = calloc(1, sizeof(*event));
199 unsigned long idx = task->nr_events;
202 event->timestamp = timestamp;
206 size = sizeof(struct sched_event *) * task->nr_events;
207 task->events = realloc(task->events, size);
208 BUG_ON(!task->events);
210 task->events[idx] = event;
215 static struct sched_event *last_event(struct task_desc *task)
217 if (!task->nr_events)
220 return task->events[task->nr_events - 1];
224 add_sched_event_run(struct task_desc *task, nsec_t timestamp, u64 duration)
226 struct sched_event *event, *curr_event = last_event(task);
229 * optimize an existing RUN event by merging this one
232 if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
233 nr_run_events_optimized++;
234 curr_event->duration += duration;
238 event = get_new_event(task, timestamp);
240 event->type = SCHED_EVENT_RUN;
241 event->duration = duration;
246 static unsigned long targetless_wakeups;
247 static unsigned long multitarget_wakeups;
250 add_sched_event_wakeup(struct task_desc *task, nsec_t timestamp,
251 struct task_desc *wakee)
253 struct sched_event *event, *wakee_event;
255 event = get_new_event(task, timestamp);
256 event->type = SCHED_EVENT_WAKEUP;
257 event->wakee = wakee;
259 wakee_event = last_event(wakee);
260 if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
261 targetless_wakeups++;
264 if (wakee_event->wait_sem) {
265 multitarget_wakeups++;
269 wakee_event->wait_sem = calloc(1, sizeof(*wakee_event->wait_sem));
270 sem_init(wakee_event->wait_sem, 0, 0);
271 wakee_event->specific_wait = 1;
272 event->wait_sem = wakee_event->wait_sem;
278 add_sched_event_sleep(struct task_desc *task, nsec_t timestamp,
279 u64 task_state __used)
281 struct sched_event *event = get_new_event(task, timestamp);
283 event->type = SCHED_EVENT_SLEEP;
288 static struct task_desc *register_pid(unsigned long pid, const char *comm)
290 struct task_desc *task;
292 BUG_ON(pid >= MAX_PID);
294 task = pid_to_task[pid];
299 task = calloc(1, sizeof(*task));
302 strcpy(task->comm, comm);
304 * every task starts in sleeping state - this gets ignored
305 * if there's no wakeup pointing to this sleep state:
307 add_sched_event_sleep(task, 0, 0);
309 pid_to_task[pid] = task;
311 tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
313 tasks[task->nr] = task;
316 printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
322 static void print_task_traces(void)
324 struct task_desc *task;
327 for (i = 0; i < nr_tasks; i++) {
329 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
330 task->nr, task->comm, task->pid, task->nr_events);
334 static void add_cross_task_wakeups(void)
336 struct task_desc *task1, *task2;
339 for (i = 0; i < nr_tasks; i++) {
345 add_sched_event_wakeup(task1, 0, task2);
350 process_sched_event(struct task_desc *this_task __used, struct sched_event *event)
357 delta = start_time + event->timestamp - now;
359 switch (event->type) {
360 case SCHED_EVENT_RUN:
361 burn_nsecs(event->duration);
363 case SCHED_EVENT_SLEEP:
365 ret = sem_wait(event->wait_sem);
368 case SCHED_EVENT_WAKEUP:
370 ret = sem_post(event->wait_sem);
378 static nsec_t get_cpu_usage_nsec_parent(void)
384 err = getrusage(RUSAGE_SELF, &ru);
387 sum = ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
388 sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
393 static nsec_t get_cpu_usage_nsec_self(void)
395 char filename [] = "/proc/1234567890/sched";
396 unsigned long msecs, nsecs;
404 sprintf(filename, "/proc/%d/sched", getpid());
405 file = fopen(filename, "r");
408 while ((chars = getline(&line, &len, file)) != -1) {
409 ret = sscanf(line, "se.sum_exec_runtime : %ld.%06ld\n",
412 total = msecs*1e6 + nsecs;
423 static void *thread_func(void *ctx)
425 struct task_desc *this_task = ctx;
426 nsec_t cpu_usage_0, cpu_usage_1;
427 unsigned long i, ret;
430 sprintf(comm2, ":%s", this_task->comm);
431 prctl(PR_SET_NAME, comm2);
434 ret = sem_post(&this_task->ready_for_work);
436 ret = pthread_mutex_lock(&start_work_mutex);
438 ret = pthread_mutex_unlock(&start_work_mutex);
441 cpu_usage_0 = get_cpu_usage_nsec_self();
443 for (i = 0; i < this_task->nr_events; i++) {
444 this_task->curr_event = i;
445 process_sched_event(this_task, this_task->events[i]);
448 cpu_usage_1 = get_cpu_usage_nsec_self();
449 this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
451 ret = sem_post(&this_task->work_done_sem);
454 ret = pthread_mutex_lock(&work_done_wait_mutex);
456 ret = pthread_mutex_unlock(&work_done_wait_mutex);
462 static void create_tasks(void)
464 struct task_desc *task;
469 err = pthread_attr_init(&attr);
471 err = pthread_attr_setstacksize(&attr, (size_t)(16*1024));
473 err = pthread_mutex_lock(&start_work_mutex);
475 err = pthread_mutex_lock(&work_done_wait_mutex);
477 for (i = 0; i < nr_tasks; i++) {
479 sem_init(&task->sleep_sem, 0, 0);
480 sem_init(&task->ready_for_work, 0, 0);
481 sem_init(&task->work_done_sem, 0, 0);
482 task->curr_event = 0;
483 err = pthread_create(&task->thread, &attr, thread_func, task);
488 static nsec_t cpu_usage;
489 static nsec_t runavg_cpu_usage;
490 static nsec_t parent_cpu_usage;
491 static nsec_t runavg_parent_cpu_usage;
493 static void wait_for_tasks(void)
495 nsec_t cpu_usage_0, cpu_usage_1;
496 struct task_desc *task;
497 unsigned long i, ret;
499 start_time = get_nsecs();
501 pthread_mutex_unlock(&work_done_wait_mutex);
503 for (i = 0; i < nr_tasks; i++) {
505 ret = sem_wait(&task->ready_for_work);
507 sem_init(&task->ready_for_work, 0, 0);
509 ret = pthread_mutex_lock(&work_done_wait_mutex);
512 cpu_usage_0 = get_cpu_usage_nsec_parent();
514 pthread_mutex_unlock(&start_work_mutex);
516 for (i = 0; i < nr_tasks; i++) {
518 ret = sem_wait(&task->work_done_sem);
520 sem_init(&task->work_done_sem, 0, 0);
521 cpu_usage += task->cpu_usage;
525 cpu_usage_1 = get_cpu_usage_nsec_parent();
526 if (!runavg_cpu_usage)
527 runavg_cpu_usage = cpu_usage;
528 runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
530 parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
531 if (!runavg_parent_cpu_usage)
532 runavg_parent_cpu_usage = parent_cpu_usage;
533 runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
534 parent_cpu_usage)/10;
536 ret = pthread_mutex_lock(&start_work_mutex);
539 for (i = 0; i < nr_tasks; i++) {
541 sem_init(&task->sleep_sem, 0, 0);
542 task->curr_event = 0;
546 static int __cmd_sched(void);
548 static void parse_trace(void)
552 printf("nr_run_events: %ld\n", nr_run_events);
553 printf("nr_sleep_events: %ld\n", nr_sleep_events);
554 printf("nr_wakeup_events: %ld\n", nr_wakeup_events);
556 if (targetless_wakeups)
557 printf("target-less wakeups: %ld\n", targetless_wakeups);
558 if (multitarget_wakeups)
559 printf("multi-target wakeups: %ld\n", multitarget_wakeups);
560 if (nr_run_events_optimized)
561 printf("run events optimized: %ld\n",
562 nr_run_events_optimized);
565 static unsigned long nr_runs;
566 static nsec_t sum_runtime;
567 static nsec_t sum_fluct;
568 static nsec_t run_avg;
570 static void run_one_test(void)
572 nsec_t T0, T1, delta, avg_delta, fluct, std_dev;
579 sum_runtime += delta;
582 avg_delta = sum_runtime / nr_runs;
583 if (delta < avg_delta)
584 fluct = avg_delta - delta;
586 fluct = delta - avg_delta;
588 std_dev = sum_fluct / nr_runs / sqrt(nr_runs);
591 run_avg = (run_avg*9 + delta)/10;
593 printf("#%-3ld: %0.3f, ",
594 nr_runs, (double)delta/1000000.0);
597 printf("%0.2f +- %0.2f, ",
598 (double)avg_delta/1e6, (double)std_dev/1e6);
600 printf("ravg: %0.2f, ",
601 (double)run_avg/1e6);
603 printf("cpu: %0.2f / %0.2f",
604 (double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6);
608 * rusage statistics done by the parent, these are less
609 * accurate than the sum_exec_runtime based statistics:
611 printf(" [%0.2f / %0.2f]",
612 (double)parent_cpu_usage/1e6,
613 (double)runavg_parent_cpu_usage/1e6);
618 if (nr_sleep_corrections)
619 printf(" (%ld sleep corrections)\n", nr_sleep_corrections);
620 nr_sleep_corrections = 0;
623 static void test_calibrations(void)
631 printf("the run test took %Ld nsecs\n", T1-T0);
637 printf("the sleep test took %Ld nsecs\n", T1-T0);
641 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
643 struct thread *thread;
645 thread = threads__findnew(event->comm.pid, &threads, &last_match);
647 dump_printf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
648 (void *)(offset + head),
649 (void *)(long)(event->header.size),
650 event->comm.comm, event->comm.pid);
652 if (thread == NULL ||
653 thread__set_comm(thread, event->comm.comm)) {
654 dump_printf("problem processing PERF_EVENT_COMM, skipping event.\n");
663 struct raw_event_sample {
668 #define FILL_FIELD(ptr, field, event, data) \
669 ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
671 #define FILL_ARRAY(ptr, array, event, data) \
673 void *__array = raw_field_ptr(event, #array, data); \
674 memcpy(ptr.array, __array, sizeof(ptr.array)); \
677 #define FILL_COMMON_FIELDS(ptr, event, data) \
679 FILL_FIELD(ptr, common_type, event, data); \
680 FILL_FIELD(ptr, common_flags, event, data); \
681 FILL_FIELD(ptr, common_preempt_count, event, data); \
682 FILL_FIELD(ptr, common_pid, event, data); \
683 FILL_FIELD(ptr, common_tgid, event, data); \
688 struct trace_switch_event {
693 u8 common_preempt_count;
707 struct trace_wakeup_event {
712 u8 common_preempt_count;
724 struct trace_fork_event {
729 u8 common_preempt_count;
733 char parent_comm[16];
739 struct trace_sched_handler {
740 void (*switch_event)(struct trace_switch_event *,
744 struct thread *thread);
746 void (*wakeup_event)(struct trace_wakeup_event *,
750 struct thread *thread);
752 void (*fork_event)(struct trace_fork_event *,
756 struct thread *thread);
761 replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
764 u64 timestamp __used,
765 struct thread *thread __used)
767 struct task_desc *waker, *wakee;
770 printf("sched_wakeup event %p\n", event);
772 printf(" ... pid %d woke up %s/%d\n",
773 wakeup_event->common_pid,
778 waker = register_pid(wakeup_event->common_pid, "<unknown>");
779 wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
781 add_sched_event_wakeup(waker, timestamp, wakee);
784 static unsigned long cpu_last_switched[MAX_CPUS];
787 replay_switch_event(struct trace_switch_event *switch_event,
791 struct thread *thread __used)
793 struct task_desc *prev, *next;
798 printf("sched_switch event %p\n", event);
800 if (cpu >= MAX_CPUS || cpu < 0)
803 timestamp0 = cpu_last_switched[cpu];
805 delta = timestamp - timestamp0;
810 die("hm, delta: %Ld < 0 ?\n", delta);
813 printf(" ... switch from %s/%d to %s/%d [ran %Ld nsecs]\n",
814 switch_event->prev_comm, switch_event->prev_pid,
815 switch_event->next_comm, switch_event->next_pid,
819 prev = register_pid(switch_event->prev_pid, switch_event->prev_comm);
820 next = register_pid(switch_event->next_pid, switch_event->next_comm);
822 cpu_last_switched[cpu] = timestamp;
824 add_sched_event_run(prev, timestamp, delta);
825 add_sched_event_sleep(prev, timestamp, switch_event->prev_state);
830 replay_fork_event(struct trace_fork_event *fork_event,
833 u64 timestamp __used,
834 struct thread *thread __used)
837 printf("sched_fork event %p\n", event);
838 printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid);
839 printf("... child: %s/%d\n", fork_event->child_comm, fork_event->child_pid);
841 register_pid(fork_event->parent_pid, fork_event->parent_comm);
842 register_pid(fork_event->child_pid, fork_event->child_comm);
845 static struct trace_sched_handler replay_ops = {
846 .wakeup_event = replay_wakeup_event,
847 .switch_event = replay_switch_event,
848 .fork_event = replay_fork_event,
851 #define TASK_STATE_TO_CHAR_STR "RSDTtZX"
860 struct lat_snapshot {
861 struct list_head list;
862 enum thread_state state;
867 struct thread_latency {
868 struct list_head snapshot_list;
869 struct thread *thread;
873 static struct rb_root lat_snapshot_root;
875 static struct thread_latency *
876 thread_latency_search(struct rb_root *root, struct thread *thread)
878 struct rb_node *node = root->rb_node;
881 struct thread_latency *lat;
883 lat = container_of(node, struct thread_latency, node);
884 if (thread->pid < lat->thread->pid)
885 node = node->rb_left;
886 else if (thread->pid > lat->thread->pid)
887 node = node->rb_right;
896 __thread_latency_insert(struct rb_root *root, struct thread_latency *data)
898 struct rb_node **new = &(root->rb_node), *parent = NULL;
901 struct thread_latency *this;
903 this = container_of(*new, struct thread_latency, node);
905 if (data->thread->pid < this->thread->pid)
906 new = &((*new)->rb_left);
907 else if (data->thread->pid > this->thread->pid)
908 new = &((*new)->rb_right);
910 die("Double thread insertion\n");
913 rb_link_node(&data->node, parent, new);
914 rb_insert_color(&data->node, root);
917 static void thread_latency_insert(struct thread *thread)
919 struct thread_latency *lat;
920 lat = calloc(sizeof(*lat), 1);
924 lat->thread = thread;
925 INIT_LIST_HEAD(&lat->snapshot_list);
926 __thread_latency_insert(&lat_snapshot_root, lat);
930 latency_fork_event(struct trace_fork_event *fork_event __used,
931 struct event *event __used,
933 u64 timestamp __used,
934 struct thread *thread __used)
936 /* should insert the newcomer */
939 static char sched_out_state(struct trace_switch_event *switch_event)
941 const char *str = TASK_STATE_TO_CHAR_STR;
943 return str[switch_event->prev_state];
947 lat_sched_out(struct thread_latency *lat,
948 struct trace_switch_event *switch_event)
950 struct lat_snapshot *snapshot;
952 if (sched_out_state(switch_event) == 'R')
955 snapshot = calloc(sizeof(*snapshot), 1);
959 list_add_tail(&snapshot->list, &lat->snapshot_list);
963 lat_sched_in(struct thread_latency *lat, u64 timestamp)
965 struct lat_snapshot *snapshot;
967 if (list_empty(&lat->snapshot_list))
970 snapshot = list_entry(lat->snapshot_list.prev, struct lat_snapshot,
973 if (snapshot->state != THREAD_WAKED_UP)
976 if (timestamp < snapshot->wake_up_time) {
977 snapshot->state = THREAD_IGNORE;
981 snapshot->state = THREAD_SCHED_IN;
982 snapshot->sched_in_time = timestamp;
987 latency_switch_event(struct trace_switch_event *switch_event,
988 struct event *event __used,
991 struct thread *thread __used)
993 struct thread_latency *out_lat, *in_lat;
994 struct thread *sched_out, *sched_in;
996 sched_out = threads__findnew(switch_event->prev_pid, &threads, &last_match);
997 sched_in = threads__findnew(switch_event->next_pid, &threads, &last_match);
999 in_lat = thread_latency_search(&lat_snapshot_root, sched_in);
1001 thread_latency_insert(sched_in);
1002 in_lat = thread_latency_search(&lat_snapshot_root, sched_in);
1004 die("Internal latency tree error");
1007 out_lat = thread_latency_search(&lat_snapshot_root, sched_out);
1009 thread_latency_insert(sched_out);
1010 out_lat = thread_latency_search(&lat_snapshot_root, sched_out);
1012 die("Internal latency tree error");
1015 lat_sched_in(in_lat, timestamp);
1016 lat_sched_out(out_lat, switch_event);
1020 latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
1021 struct event *event __used,
1024 struct thread *thread __used)
1026 struct thread_latency *lat;
1027 struct lat_snapshot *snapshot;
1028 struct thread *wakee;
1030 /* Note for later, it may be interesting to observe the failing cases */
1031 if (!wakeup_event->success)
1034 wakee = threads__findnew(wakeup_event->pid, &threads, &last_match);
1035 lat = thread_latency_search(&lat_snapshot_root, wakee);
1037 thread_latency_insert(wakee);
1041 if (list_empty(&lat->snapshot_list))
1044 snapshot = list_entry(lat->snapshot_list.prev, struct lat_snapshot,
1047 if (snapshot->state != THREAD_SLEEPING)
1050 snapshot->state = THREAD_WAKED_UP;
1051 snapshot->wake_up_time = timestamp;
1054 static struct trace_sched_handler lat_ops = {
1055 .wakeup_event = latency_wakeup_event,
1056 .switch_event = latency_switch_event,
1057 .fork_event = latency_fork_event,
1060 static void output_lat_thread(struct thread_latency *lat)
1062 struct lat_snapshot *shot;
1067 u64 total = 0, delta;
1069 list_for_each_entry(shot, &lat->snapshot_list, list) {
1070 if (shot->state != THREAD_SCHED_IN)
1075 delta = shot->sched_in_time - shot->wake_up_time;
1084 ret = printf("%s", lat->thread->comm);
1086 for (i = 0; i < 25 - ret; i++)
1089 avg = total / count;
1091 printf("%5d %10llu %10llu %10llu\n", count, total, avg, max);
1094 static void output_lat_results(void)
1096 struct rb_node *next;
1104 next = rb_first(&lat_snapshot_root);
1107 struct thread_latency *lat;
1109 lat = rb_entry(next, struct thread_latency, node);
1110 output_lat_thread(lat);
1111 next = rb_next(next);
1115 static struct trace_sched_handler *trace_handler;
1118 process_sched_wakeup_event(struct raw_event_sample *raw,
1119 struct event *event,
1121 u64 timestamp __used,
1122 struct thread *thread __used)
1124 struct trace_wakeup_event wakeup_event;
1126 FILL_COMMON_FIELDS(wakeup_event, event, raw->data);
1128 FILL_ARRAY(wakeup_event, comm, event, raw->data);
1129 FILL_FIELD(wakeup_event, pid, event, raw->data);
1130 FILL_FIELD(wakeup_event, prio, event, raw->data);
1131 FILL_FIELD(wakeup_event, success, event, raw->data);
1132 FILL_FIELD(wakeup_event, cpu, event, raw->data);
1134 trace_handler->wakeup_event(&wakeup_event, event, cpu, timestamp, thread);
1138 process_sched_switch_event(struct raw_event_sample *raw,
1139 struct event *event,
1141 u64 timestamp __used,
1142 struct thread *thread __used)
1144 struct trace_switch_event switch_event;
1146 FILL_COMMON_FIELDS(switch_event, event, raw->data);
1148 FILL_ARRAY(switch_event, prev_comm, event, raw->data);
1149 FILL_FIELD(switch_event, prev_pid, event, raw->data);
1150 FILL_FIELD(switch_event, prev_prio, event, raw->data);
1151 FILL_FIELD(switch_event, prev_state, event, raw->data);
1152 FILL_ARRAY(switch_event, next_comm, event, raw->data);
1153 FILL_FIELD(switch_event, next_pid, event, raw->data);
1154 FILL_FIELD(switch_event, next_prio, event, raw->data);
1156 trace_handler->switch_event(&switch_event, event, cpu, timestamp, thread);
1160 process_sched_fork_event(struct raw_event_sample *raw,
1161 struct event *event,
1163 u64 timestamp __used,
1164 struct thread *thread __used)
1166 struct trace_fork_event fork_event;
1168 FILL_COMMON_FIELDS(fork_event, event, raw->data);
1170 FILL_ARRAY(fork_event, parent_comm, event, raw->data);
1171 FILL_FIELD(fork_event, parent_pid, event, raw->data);
1172 FILL_ARRAY(fork_event, child_comm, event, raw->data);
1173 FILL_FIELD(fork_event, child_pid, event, raw->data);
1175 trace_handler->fork_event(&fork_event, event, cpu, timestamp, thread);
1179 process_sched_exit_event(struct event *event,
1181 u64 timestamp __used,
1182 struct thread *thread __used)
1185 printf("sched_exit event %p\n", event);
1189 process_raw_event(event_t *raw_event __used, void *more_data,
1190 int cpu, u64 timestamp, struct thread *thread)
1192 struct raw_event_sample *raw = more_data;
1193 struct event *event;
1196 type = trace_parse_common_type(raw->data);
1197 event = trace_find_event(type);
1199 if (!strcmp(event->name, "sched_switch"))
1200 process_sched_switch_event(raw, event, cpu, timestamp, thread);
1201 if (!strcmp(event->name, "sched_wakeup"))
1202 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1203 if (!strcmp(event->name, "sched_wakeup_new"))
1204 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1205 if (!strcmp(event->name, "sched_process_fork"))
1206 process_sched_fork_event(raw, event, cpu, timestamp, thread);
1207 if (!strcmp(event->name, "sched_process_exit"))
1208 process_sched_exit_event(event, cpu, timestamp, thread);
1212 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
1216 struct dso *dso = NULL;
1217 struct thread *thread;
1218 u64 ip = event->ip.ip;
1222 void *more_data = event->ip.__more_data;
1225 thread = threads__findnew(event->ip.pid, &threads, &last_match);
1227 if (sample_type & PERF_SAMPLE_TIME) {
1228 timestamp = *(u64 *)more_data;
1229 more_data += sizeof(u64);
1232 if (sample_type & PERF_SAMPLE_CPU) {
1233 cpu = *(u32 *)more_data;
1234 more_data += sizeof(u32);
1235 more_data += sizeof(u32); /* reserved */
1238 if (sample_type & PERF_SAMPLE_PERIOD) {
1239 period = *(u64 *)more_data;
1240 more_data += sizeof(u64);
1243 dump_printf("%p [%p]: PERF_EVENT_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n",
1244 (void *)(offset + head),
1245 (void *)(long)(event->header.size),
1247 event->ip.pid, event->ip.tid,
1251 dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1253 if (thread == NULL) {
1254 eprintf("problem processing %d event, skipping it.\n",
1255 event->header.type);
1259 cpumode = event->header.misc & PERF_EVENT_MISC_CPUMODE_MASK;
1261 if (cpumode == PERF_EVENT_MISC_KERNEL) {
1267 dump_printf(" ...... dso: %s\n", dso->name);
1269 } else if (cpumode == PERF_EVENT_MISC_USER) {
1278 dso = hypervisor_dso;
1280 dump_printf(" ...... dso: [hypervisor]\n");
1283 if (sample_type & PERF_SAMPLE_RAW)
1284 process_raw_event(event, more_data, cpu, timestamp, thread);
1290 process_event(event_t *event, unsigned long offset, unsigned long head)
1294 switch (event->header.type) {
1295 case PERF_EVENT_MMAP ... PERF_EVENT_LOST:
1298 case PERF_EVENT_COMM:
1299 return process_comm_event(event, offset, head);
1301 case PERF_EVENT_EXIT ... PERF_EVENT_READ:
1304 case PERF_EVENT_SAMPLE:
1305 return process_sample_event(event, offset, head);
1307 case PERF_EVENT_MAX:
1315 static int __cmd_sched(void)
1317 int ret, rc = EXIT_FAILURE;
1318 unsigned long offset = 0;
1319 unsigned long head = 0;
1320 struct stat perf_stat;
1326 register_idle_thread(&threads, &last_match);
1328 input = open(input_name, O_RDONLY);
1330 perror("failed to open file");
1334 ret = fstat(input, &perf_stat);
1336 perror("failed to stat file");
1340 if (!perf_stat.st_size) {
1341 fprintf(stderr, "zero-sized file, nothing to do!\n");
1344 header = perf_header__read(input);
1345 head = header->data_offset;
1346 sample_type = perf_header__sample_type(header);
1348 if (!(sample_type & PERF_SAMPLE_RAW))
1349 die("No trace sample to read. Did you call perf record "
1352 if (load_kernel() < 0) {
1353 perror("failed to load kernel symbols");
1354 return EXIT_FAILURE;
1358 buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1359 MAP_SHARED, input, offset);
1360 if (buf == MAP_FAILED) {
1361 perror("failed to mmap file");
1366 event = (event_t *)(buf + head);
1368 size = event->header.size;
1372 if (head + event->header.size >= page_size * mmap_window) {
1373 unsigned long shift = page_size * (head / page_size);
1376 res = munmap(buf, page_size * mmap_window);
1384 size = event->header.size;
1387 if (!size || process_event(event, offset, head) < 0) {
1390 * assume we lost track of the stream, check alignment, and
1391 * increment a single u64 in the hope to catch on again 'soon'.
1394 if (unlikely(head & 7))
1402 if (offset + head < (unsigned long)perf_stat.st_size)
1411 static const char * const annotate_usage[] = {
1412 "perf trace [<options>] <command>",
1416 static const struct option options[] = {
1417 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1418 "dump raw trace in ASCII"),
1419 OPT_BOOLEAN('r', "replay", &replay_mode,
1420 "replay sched behaviour from traces"),
1421 OPT_BOOLEAN('l', "latency", &lat_mode,
1422 "measure various latencies"),
1423 OPT_BOOLEAN('v', "verbose", &verbose,
1424 "be more verbose (show symbol address, etc)"),
1428 int cmd_sched(int argc, const char **argv, const char *prefix __used)
1430 long nr_iterations = 10, i;
1433 page_size = getpagesize();
1435 argc = parse_options(argc, argv, options, annotate_usage, 0);
1438 * Special case: if there's an argument left then assume tha
1439 * it's a symbol filter:
1442 usage_with_options(annotate_usage, options);
1448 trace_handler = &replay_ops;
1450 trace_handler = &lat_ops;
1451 else /* We may need a default subcommand (perf trace?) */
1452 die("Please select a sub command (-r)\n");
1455 calibrate_run_measurement_overhead();
1456 calibrate_sleep_measurement_overhead();
1458 test_calibrations();
1461 print_task_traces();
1462 add_cross_task_wakeups();
1465 printf("------------------------------------------------------------\n");
1466 for (i = 0; i < nr_iterations; i++)
1468 } else if (lat_mode) {
1471 output_lat_results();