5 #include "util/cache.h"
6 #include "util/symbol.h"
7 #include "util/thread.h"
8 #include "util/header.h"
10 #include "util/parse-options.h"
11 #include "util/trace-event.h"
13 #include "util/debug.h"
15 #include <sys/types.h>
16 #include <sys/prctl.h>
18 #include <semaphore.h>
22 static char const *input_name = "perf.data";
24 static unsigned long page_size;
25 static unsigned long mmap_window = 32;
27 static unsigned long total_comm = 0;
29 static struct rb_root threads;
30 static struct thread *last_match;
32 static struct perf_header *header;
33 static u64 sample_type;
35 static char default_sort_order[] = "avg, max, switch, runtime";
36 static char *sort_order = default_sort_order;
38 #define PR_SET_NAME 15 /* Set process name */
41 #define BUG_ON(x) assert(!(x))
43 static u64 run_measurement_overhead;
44 static u64 sleep_measurement_overhead;
51 static unsigned long nr_tasks;
60 unsigned long nr_events;
61 unsigned long curr_event;
62 struct sched_event **events;
73 enum sched_event_type {
80 enum sched_event_type type;
86 struct task_desc *wakee;
89 static struct task_desc *pid_to_task[MAX_PID];
91 static struct task_desc **tasks;
93 static pthread_mutex_t start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
94 static u64 start_time;
96 static pthread_mutex_t work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
98 static unsigned long nr_run_events;
99 static unsigned long nr_sleep_events;
100 static unsigned long nr_wakeup_events;
102 static unsigned long nr_sleep_corrections;
103 static unsigned long nr_run_events_optimized;
105 static unsigned long targetless_wakeups;
106 static unsigned long multitarget_wakeups;
108 static u64 cpu_usage;
109 static u64 runavg_cpu_usage;
110 static u64 parent_cpu_usage;
111 static u64 runavg_parent_cpu_usage;
113 static unsigned long nr_runs;
114 static u64 sum_runtime;
115 static u64 sum_fluct;
118 static unsigned long replay_repeat = 10;
120 #define TASK_STATE_TO_CHAR_STR "RSDTtZX"
130 struct list_head list;
131 enum thread_state state;
138 struct list_head atom_list;
139 struct thread *thread;
147 typedef int (*sort_fn_t)(struct task_atoms *, struct task_atoms *);
149 static struct rb_root atom_root, sorted_atom_root;
151 static u64 all_runtime;
152 static u64 all_count;
154 static int read_events(void);
157 static u64 get_nsecs(void)
161 clock_gettime(CLOCK_MONOTONIC, &ts);
163 return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
166 static void burn_nsecs(u64 nsecs)
168 u64 T0 = get_nsecs(), T1;
172 } while (T1 + run_measurement_overhead < T0 + nsecs);
175 static void sleep_nsecs(u64 nsecs)
179 ts.tv_nsec = nsecs % 999999999;
180 ts.tv_sec = nsecs / 999999999;
182 nanosleep(&ts, NULL);
185 static void calibrate_run_measurement_overhead(void)
187 u64 T0, T1, delta, min_delta = 1000000000ULL;
190 for (i = 0; i < 10; i++) {
195 min_delta = min(min_delta, delta);
197 run_measurement_overhead = min_delta;
199 printf("run measurement overhead: %Ld nsecs\n", min_delta);
202 static void calibrate_sleep_measurement_overhead(void)
204 u64 T0, T1, delta, min_delta = 1000000000ULL;
207 for (i = 0; i < 10; i++) {
212 min_delta = min(min_delta, delta);
215 sleep_measurement_overhead = min_delta;
217 printf("sleep measurement overhead: %Ld nsecs\n", min_delta);
220 static struct sched_event *
221 get_new_event(struct task_desc *task, u64 timestamp)
223 struct sched_event *event = calloc(1, sizeof(*event));
224 unsigned long idx = task->nr_events;
227 event->timestamp = timestamp;
231 size = sizeof(struct sched_event *) * task->nr_events;
232 task->events = realloc(task->events, size);
233 BUG_ON(!task->events);
235 task->events[idx] = event;
240 static struct sched_event *last_event(struct task_desc *task)
242 if (!task->nr_events)
245 return task->events[task->nr_events - 1];
249 add_sched_event_run(struct task_desc *task, u64 timestamp, u64 duration)
251 struct sched_event *event, *curr_event = last_event(task);
254 * optimize an existing RUN event by merging this one
257 if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
258 nr_run_events_optimized++;
259 curr_event->duration += duration;
263 event = get_new_event(task, timestamp);
265 event->type = SCHED_EVENT_RUN;
266 event->duration = duration;
272 add_sched_event_wakeup(struct task_desc *task, u64 timestamp,
273 struct task_desc *wakee)
275 struct sched_event *event, *wakee_event;
277 event = get_new_event(task, timestamp);
278 event->type = SCHED_EVENT_WAKEUP;
279 event->wakee = wakee;
281 wakee_event = last_event(wakee);
282 if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
283 targetless_wakeups++;
286 if (wakee_event->wait_sem) {
287 multitarget_wakeups++;
291 wakee_event->wait_sem = calloc(1, sizeof(*wakee_event->wait_sem));
292 sem_init(wakee_event->wait_sem, 0, 0);
293 wakee_event->specific_wait = 1;
294 event->wait_sem = wakee_event->wait_sem;
300 add_sched_event_sleep(struct task_desc *task, u64 timestamp,
301 u64 task_state __used)
303 struct sched_event *event = get_new_event(task, timestamp);
305 event->type = SCHED_EVENT_SLEEP;
310 static struct task_desc *register_pid(unsigned long pid, const char *comm)
312 struct task_desc *task;
314 BUG_ON(pid >= MAX_PID);
316 task = pid_to_task[pid];
321 task = calloc(1, sizeof(*task));
324 strcpy(task->comm, comm);
326 * every task starts in sleeping state - this gets ignored
327 * if there's no wakeup pointing to this sleep state:
329 add_sched_event_sleep(task, 0, 0);
331 pid_to_task[pid] = task;
333 tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
335 tasks[task->nr] = task;
338 printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
344 static void print_task_traces(void)
346 struct task_desc *task;
349 for (i = 0; i < nr_tasks; i++) {
351 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
352 task->nr, task->comm, task->pid, task->nr_events);
356 static void add_cross_task_wakeups(void)
358 struct task_desc *task1, *task2;
361 for (i = 0; i < nr_tasks; i++) {
367 add_sched_event_wakeup(task1, 0, task2);
372 process_sched_event(struct task_desc *this_task __used, struct sched_event *event)
379 delta = start_time + event->timestamp - now;
381 switch (event->type) {
382 case SCHED_EVENT_RUN:
383 burn_nsecs(event->duration);
385 case SCHED_EVENT_SLEEP:
387 ret = sem_wait(event->wait_sem);
390 case SCHED_EVENT_WAKEUP:
392 ret = sem_post(event->wait_sem);
400 static u64 get_cpu_usage_nsec_parent(void)
406 err = getrusage(RUSAGE_SELF, &ru);
409 sum = ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
410 sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
415 static u64 get_cpu_usage_nsec_self(void)
417 char filename [] = "/proc/1234567890/sched";
418 unsigned long msecs, nsecs;
426 sprintf(filename, "/proc/%d/sched", getpid());
427 file = fopen(filename, "r");
430 while ((chars = getline(&line, &len, file)) != -1) {
431 ret = sscanf(line, "se.sum_exec_runtime : %ld.%06ld\n",
434 total = msecs*1e6 + nsecs;
445 static void *thread_func(void *ctx)
447 struct task_desc *this_task = ctx;
448 u64 cpu_usage_0, cpu_usage_1;
449 unsigned long i, ret;
452 sprintf(comm2, ":%s", this_task->comm);
453 prctl(PR_SET_NAME, comm2);
456 ret = sem_post(&this_task->ready_for_work);
458 ret = pthread_mutex_lock(&start_work_mutex);
460 ret = pthread_mutex_unlock(&start_work_mutex);
463 cpu_usage_0 = get_cpu_usage_nsec_self();
465 for (i = 0; i < this_task->nr_events; i++) {
466 this_task->curr_event = i;
467 process_sched_event(this_task, this_task->events[i]);
470 cpu_usage_1 = get_cpu_usage_nsec_self();
471 this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
473 ret = sem_post(&this_task->work_done_sem);
476 ret = pthread_mutex_lock(&work_done_wait_mutex);
478 ret = pthread_mutex_unlock(&work_done_wait_mutex);
484 static void create_tasks(void)
486 struct task_desc *task;
491 err = pthread_attr_init(&attr);
493 err = pthread_attr_setstacksize(&attr, (size_t)(16*1024));
495 err = pthread_mutex_lock(&start_work_mutex);
497 err = pthread_mutex_lock(&work_done_wait_mutex);
499 for (i = 0; i < nr_tasks; i++) {
501 sem_init(&task->sleep_sem, 0, 0);
502 sem_init(&task->ready_for_work, 0, 0);
503 sem_init(&task->work_done_sem, 0, 0);
504 task->curr_event = 0;
505 err = pthread_create(&task->thread, &attr, thread_func, task);
510 static void wait_for_tasks(void)
512 u64 cpu_usage_0, cpu_usage_1;
513 struct task_desc *task;
514 unsigned long i, ret;
516 start_time = get_nsecs();
518 pthread_mutex_unlock(&work_done_wait_mutex);
520 for (i = 0; i < nr_tasks; i++) {
522 ret = sem_wait(&task->ready_for_work);
524 sem_init(&task->ready_for_work, 0, 0);
526 ret = pthread_mutex_lock(&work_done_wait_mutex);
529 cpu_usage_0 = get_cpu_usage_nsec_parent();
531 pthread_mutex_unlock(&start_work_mutex);
533 for (i = 0; i < nr_tasks; i++) {
535 ret = sem_wait(&task->work_done_sem);
537 sem_init(&task->work_done_sem, 0, 0);
538 cpu_usage += task->cpu_usage;
542 cpu_usage_1 = get_cpu_usage_nsec_parent();
543 if (!runavg_cpu_usage)
544 runavg_cpu_usage = cpu_usage;
545 runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
547 parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
548 if (!runavg_parent_cpu_usage)
549 runavg_parent_cpu_usage = parent_cpu_usage;
550 runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
551 parent_cpu_usage)/10;
553 ret = pthread_mutex_lock(&start_work_mutex);
556 for (i = 0; i < nr_tasks; i++) {
558 sem_init(&task->sleep_sem, 0, 0);
559 task->curr_event = 0;
563 static void run_one_test(void)
565 u64 T0, T1, delta, avg_delta, fluct, std_dev;
572 sum_runtime += delta;
575 avg_delta = sum_runtime / nr_runs;
576 if (delta < avg_delta)
577 fluct = avg_delta - delta;
579 fluct = delta - avg_delta;
581 std_dev = sum_fluct / nr_runs / sqrt(nr_runs);
584 run_avg = (run_avg*9 + delta)/10;
586 printf("#%-3ld: %0.3f, ",
587 nr_runs, (double)delta/1000000.0);
589 printf("ravg: %0.2f, ",
590 (double)run_avg/1e6);
592 printf("cpu: %0.2f / %0.2f",
593 (double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6);
597 * rusage statistics done by the parent, these are less
598 * accurate than the sum_exec_runtime based statistics:
600 printf(" [%0.2f / %0.2f]",
601 (double)parent_cpu_usage/1e6,
602 (double)runavg_parent_cpu_usage/1e6);
607 if (nr_sleep_corrections)
608 printf(" (%ld sleep corrections)\n", nr_sleep_corrections);
609 nr_sleep_corrections = 0;
612 static void test_calibrations(void)
620 printf("the run test took %Ld nsecs\n", T1-T0);
626 printf("the sleep test took %Ld nsecs\n", T1-T0);
629 static void __cmd_replay(void)
633 calibrate_run_measurement_overhead();
634 calibrate_sleep_measurement_overhead();
640 printf("nr_run_events: %ld\n", nr_run_events);
641 printf("nr_sleep_events: %ld\n", nr_sleep_events);
642 printf("nr_wakeup_events: %ld\n", nr_wakeup_events);
644 if (targetless_wakeups)
645 printf("target-less wakeups: %ld\n", targetless_wakeups);
646 if (multitarget_wakeups)
647 printf("multi-target wakeups: %ld\n", multitarget_wakeups);
648 if (nr_run_events_optimized)
649 printf("run events optimized: %ld\n",
650 nr_run_events_optimized);
653 add_cross_task_wakeups();
656 printf("------------------------------------------------------------\n");
657 for (i = 0; i < replay_repeat; i++)
662 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
664 struct thread *thread;
666 thread = threads__findnew(event->comm.pid, &threads, &last_match);
668 dump_printf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
669 (void *)(offset + head),
670 (void *)(long)(event->header.size),
671 event->comm.comm, event->comm.pid);
673 if (thread == NULL ||
674 thread__set_comm(thread, event->comm.comm)) {
675 dump_printf("problem processing PERF_EVENT_COMM, skipping event.\n");
684 struct raw_event_sample {
689 #define FILL_FIELD(ptr, field, event, data) \
690 ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
692 #define FILL_ARRAY(ptr, array, event, data) \
694 void *__array = raw_field_ptr(event, #array, data); \
695 memcpy(ptr.array, __array, sizeof(ptr.array)); \
698 #define FILL_COMMON_FIELDS(ptr, event, data) \
700 FILL_FIELD(ptr, common_type, event, data); \
701 FILL_FIELD(ptr, common_flags, event, data); \
702 FILL_FIELD(ptr, common_preempt_count, event, data); \
703 FILL_FIELD(ptr, common_pid, event, data); \
704 FILL_FIELD(ptr, common_tgid, event, data); \
709 struct trace_switch_event {
714 u8 common_preempt_count;
728 struct trace_wakeup_event {
733 u8 common_preempt_count;
745 struct trace_fork_event {
750 u8 common_preempt_count;
754 char parent_comm[16];
760 struct trace_sched_handler {
761 void (*switch_event)(struct trace_switch_event *,
765 struct thread *thread);
767 void (*wakeup_event)(struct trace_wakeup_event *,
771 struct thread *thread);
773 void (*fork_event)(struct trace_fork_event *,
777 struct thread *thread);
782 replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
785 u64 timestamp __used,
786 struct thread *thread __used)
788 struct task_desc *waker, *wakee;
791 printf("sched_wakeup event %p\n", event);
793 printf(" ... pid %d woke up %s/%d\n",
794 wakeup_event->common_pid,
799 waker = register_pid(wakeup_event->common_pid, "<unknown>");
800 wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
802 add_sched_event_wakeup(waker, timestamp, wakee);
805 static unsigned long cpu_last_switched[MAX_CPUS];
808 replay_switch_event(struct trace_switch_event *switch_event,
812 struct thread *thread __used)
814 struct task_desc *prev, *next;
819 printf("sched_switch event %p\n", event);
821 if (cpu >= MAX_CPUS || cpu < 0)
824 timestamp0 = cpu_last_switched[cpu];
826 delta = timestamp - timestamp0;
831 die("hm, delta: %Ld < 0 ?\n", delta);
834 printf(" ... switch from %s/%d to %s/%d [ran %Ld nsecs]\n",
835 switch_event->prev_comm, switch_event->prev_pid,
836 switch_event->next_comm, switch_event->next_pid,
840 prev = register_pid(switch_event->prev_pid, switch_event->prev_comm);
841 next = register_pid(switch_event->next_pid, switch_event->next_comm);
843 cpu_last_switched[cpu] = timestamp;
845 add_sched_event_run(prev, timestamp, delta);
846 add_sched_event_sleep(prev, timestamp, switch_event->prev_state);
851 replay_fork_event(struct trace_fork_event *fork_event,
854 u64 timestamp __used,
855 struct thread *thread __used)
858 printf("sched_fork event %p\n", event);
859 printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid);
860 printf("... child: %s/%d\n", fork_event->child_comm, fork_event->child_pid);
862 register_pid(fork_event->parent_pid, fork_event->parent_comm);
863 register_pid(fork_event->child_pid, fork_event->child_comm);
866 static struct trace_sched_handler replay_ops = {
867 .wakeup_event = replay_wakeup_event,
868 .switch_event = replay_switch_event,
869 .fork_event = replay_fork_event,
872 struct sort_dimension {
875 struct list_head list;
878 static LIST_HEAD(cmp_pid);
881 thread_lat_cmp(struct list_head *list, struct task_atoms *l, struct task_atoms *r)
883 struct sort_dimension *sort;
886 BUG_ON(list_empty(list));
888 list_for_each_entry(sort, list, list) {
889 ret = sort->cmp(l, r);
897 static struct task_atoms *
898 thread_atoms_search(struct rb_root *root, struct thread *thread,
899 struct list_head *sort_list)
901 struct rb_node *node = root->rb_node;
902 struct task_atoms key = { .thread = thread };
905 struct task_atoms *atoms;
908 atoms = container_of(node, struct task_atoms, node);
910 cmp = thread_lat_cmp(sort_list, &key, atoms);
912 node = node->rb_left;
914 node = node->rb_right;
916 BUG_ON(thread != atoms->thread);
924 __thread_latency_insert(struct rb_root *root, struct task_atoms *data,
925 struct list_head *sort_list)
927 struct rb_node **new = &(root->rb_node), *parent = NULL;
930 struct task_atoms *this;
933 this = container_of(*new, struct task_atoms, node);
936 cmp = thread_lat_cmp(sort_list, data, this);
939 new = &((*new)->rb_left);
941 new = &((*new)->rb_right);
944 rb_link_node(&data->node, parent, new);
945 rb_insert_color(&data->node, root);
948 static void thread_atoms_insert(struct thread *thread)
950 struct task_atoms *atoms;
952 atoms = calloc(sizeof(*atoms), 1);
956 atoms->thread = thread;
957 INIT_LIST_HEAD(&atoms->atom_list);
958 __thread_latency_insert(&atom_root, atoms, &cmp_pid);
962 latency_fork_event(struct trace_fork_event *fork_event __used,
963 struct event *event __used,
965 u64 timestamp __used,
966 struct thread *thread __used)
968 /* should insert the newcomer */
972 static char sched_out_state(struct trace_switch_event *switch_event)
974 const char *str = TASK_STATE_TO_CHAR_STR;
976 return str[switch_event->prev_state];
980 lat_sched_out(struct task_atoms *atoms,
981 struct trace_switch_event *switch_event __used,
985 struct work_atom *atom;
987 atom = calloc(sizeof(*atom), 1);
991 if (sched_out_state(switch_event) == 'R') {
992 atom->state = THREAD_WAIT_CPU;
993 atom->wake_up_time = timestamp;
996 atom->runtime = delta;
997 list_add_tail(&atom->list, &atoms->atom_list);
1001 lat_sched_in(struct task_atoms *atoms, u64 timestamp)
1003 struct work_atom *atom;
1006 if (list_empty(&atoms->atom_list))
1009 atom = list_entry(atoms->atom_list.prev, struct work_atom, list);
1011 if (atom->state != THREAD_WAIT_CPU)
1014 if (timestamp < atom->wake_up_time) {
1015 atom->state = THREAD_IGNORE;
1019 atom->state = THREAD_SCHED_IN;
1020 atom->sched_in_time = timestamp;
1022 delta = atom->sched_in_time - atom->wake_up_time;
1023 atoms->total_lat += delta;
1024 if (delta > atoms->max_lat)
1025 atoms->max_lat = delta;
1027 atoms->total_runtime += atom->runtime;
1031 latency_switch_event(struct trace_switch_event *switch_event,
1032 struct event *event __used,
1035 struct thread *thread __used)
1037 struct task_atoms *out_atoms, *in_atoms;
1038 struct thread *sched_out, *sched_in;
1042 if (cpu >= MAX_CPUS || cpu < 0)
1045 timestamp0 = cpu_last_switched[cpu];
1046 cpu_last_switched[cpu] = timestamp;
1048 delta = timestamp - timestamp0;
1053 die("hm, delta: %Ld < 0 ?\n", delta);
1056 sched_out = threads__findnew(switch_event->prev_pid, &threads, &last_match);
1057 sched_in = threads__findnew(switch_event->next_pid, &threads, &last_match);
1059 in_atoms = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1061 thread_atoms_insert(sched_in);
1062 in_atoms = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1064 die("in-atom: Internal tree error");
1067 out_atoms = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1069 thread_atoms_insert(sched_out);
1070 out_atoms = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1072 die("out-atom: Internal tree error");
1075 lat_sched_in(in_atoms, timestamp);
1076 lat_sched_out(out_atoms, switch_event, delta, timestamp);
1080 latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
1081 struct event *event __used,
1084 struct thread *thread __used)
1086 struct task_atoms *atoms;
1087 struct work_atom *atom;
1088 struct thread *wakee;
1090 /* Note for later, it may be interesting to observe the failing cases */
1091 if (!wakeup_event->success)
1094 wakee = threads__findnew(wakeup_event->pid, &threads, &last_match);
1095 atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1097 thread_atoms_insert(wakee);
1101 if (list_empty(&atoms->atom_list))
1104 atom = list_entry(atoms->atom_list.prev, struct work_atom, list);
1106 if (atom->state != THREAD_SLEEPING)
1109 atom->state = THREAD_WAIT_CPU;
1110 atom->wake_up_time = timestamp;
1113 static struct trace_sched_handler lat_ops = {
1114 .wakeup_event = latency_wakeup_event,
1115 .switch_event = latency_switch_event,
1116 .fork_event = latency_fork_event,
1119 static void output_lat_thread(struct task_atoms *atom_list)
1125 if (!atom_list->nb_atoms)
1128 all_runtime += atom_list->total_runtime;
1129 all_count += atom_list->nb_atoms;
1131 ret = printf(" %s ", atom_list->thread->comm);
1133 for (i = 0; i < 19 - ret; i++)
1136 avg = atom_list->total_lat / atom_list->nb_atoms;
1138 printf("|%9.3f ms |%9llu | avg:%9.3f ms | max:%9.3f ms |\n",
1139 (double)atom_list->total_runtime / 1e6,
1140 atom_list->nb_atoms, (double)avg / 1e6,
1141 (double)atom_list->max_lat / 1e6);
1144 static int pid_cmp(struct task_atoms *l, struct task_atoms *r)
1146 if (l->thread->pid < r->thread->pid)
1148 if (l->thread->pid > r->thread->pid)
1154 static struct sort_dimension pid_sort_dimension = {
1159 static int avg_cmp(struct task_atoms *l, struct task_atoms *r)
1169 avgl = l->total_lat / l->nb_atoms;
1170 avgr = r->total_lat / r->nb_atoms;
1180 static struct sort_dimension avg_sort_dimension = {
1185 static int max_cmp(struct task_atoms *l, struct task_atoms *r)
1187 if (l->max_lat < r->max_lat)
1189 if (l->max_lat > r->max_lat)
1195 static struct sort_dimension max_sort_dimension = {
1200 static int switch_cmp(struct task_atoms *l, struct task_atoms *r)
1202 if (l->nb_atoms < r->nb_atoms)
1204 if (l->nb_atoms > r->nb_atoms)
1210 static struct sort_dimension switch_sort_dimension = {
1215 static int runtime_cmp(struct task_atoms *l, struct task_atoms *r)
1217 if (l->total_runtime < r->total_runtime)
1219 if (l->total_runtime > r->total_runtime)
1225 static struct sort_dimension runtime_sort_dimension = {
1230 static struct sort_dimension *available_sorts[] = {
1231 &pid_sort_dimension,
1232 &avg_sort_dimension,
1233 &max_sort_dimension,
1234 &switch_sort_dimension,
1235 &runtime_sort_dimension,
1238 #define NB_AVAILABLE_SORTS (int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
1240 static LIST_HEAD(sort_list);
1242 static int sort_dimension__add(char *tok, struct list_head *list)
1246 for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
1247 if (!strcmp(available_sorts[i]->name, tok)) {
1248 list_add_tail(&available_sorts[i]->list, list);
1257 static void setup_sorting(void);
1259 static void sort_lat(void)
1261 struct rb_node *node;
1264 struct task_atoms *data;
1265 node = rb_first(&atom_root);
1269 rb_erase(node, &atom_root);
1270 data = rb_entry(node, struct task_atoms, node);
1271 __thread_latency_insert(&sorted_atom_root, data, &sort_list);
1275 static void __cmd_lat(void)
1277 struct rb_node *next;
1283 printf("-----------------------------------------------------------------------------------\n");
1284 printf(" Task | Runtime ms | Switches | Average delay ms | Maximum delay ms |\n");
1285 printf("-----------------------------------------------------------------------------------\n");
1287 next = rb_first(&sorted_atom_root);
1290 struct task_atoms *atom_list;
1292 atom_list = rb_entry(next, struct task_atoms, node);
1293 output_lat_thread(atom_list);
1294 next = rb_next(next);
1297 printf("-----------------------------------------------------------------------------------\n");
1298 printf(" TOTAL: |%9.3f ms |%9Ld |\n",
1299 (double)all_runtime/1e6, all_count);
1300 printf("---------------------------------------------\n");
1303 static struct trace_sched_handler *trace_handler;
1306 process_sched_wakeup_event(struct raw_event_sample *raw,
1307 struct event *event,
1309 u64 timestamp __used,
1310 struct thread *thread __used)
1312 struct trace_wakeup_event wakeup_event;
1314 FILL_COMMON_FIELDS(wakeup_event, event, raw->data);
1316 FILL_ARRAY(wakeup_event, comm, event, raw->data);
1317 FILL_FIELD(wakeup_event, pid, event, raw->data);
1318 FILL_FIELD(wakeup_event, prio, event, raw->data);
1319 FILL_FIELD(wakeup_event, success, event, raw->data);
1320 FILL_FIELD(wakeup_event, cpu, event, raw->data);
1322 trace_handler->wakeup_event(&wakeup_event, event, cpu, timestamp, thread);
1326 process_sched_switch_event(struct raw_event_sample *raw,
1327 struct event *event,
1329 u64 timestamp __used,
1330 struct thread *thread __used)
1332 struct trace_switch_event switch_event;
1334 FILL_COMMON_FIELDS(switch_event, event, raw->data);
1336 FILL_ARRAY(switch_event, prev_comm, event, raw->data);
1337 FILL_FIELD(switch_event, prev_pid, event, raw->data);
1338 FILL_FIELD(switch_event, prev_prio, event, raw->data);
1339 FILL_FIELD(switch_event, prev_state, event, raw->data);
1340 FILL_ARRAY(switch_event, next_comm, event, raw->data);
1341 FILL_FIELD(switch_event, next_pid, event, raw->data);
1342 FILL_FIELD(switch_event, next_prio, event, raw->data);
1344 trace_handler->switch_event(&switch_event, event, cpu, timestamp, thread);
1348 process_sched_fork_event(struct raw_event_sample *raw,
1349 struct event *event,
1351 u64 timestamp __used,
1352 struct thread *thread __used)
1354 struct trace_fork_event fork_event;
1356 FILL_COMMON_FIELDS(fork_event, event, raw->data);
1358 FILL_ARRAY(fork_event, parent_comm, event, raw->data);
1359 FILL_FIELD(fork_event, parent_pid, event, raw->data);
1360 FILL_ARRAY(fork_event, child_comm, event, raw->data);
1361 FILL_FIELD(fork_event, child_pid, event, raw->data);
1363 trace_handler->fork_event(&fork_event, event, cpu, timestamp, thread);
1367 process_sched_exit_event(struct event *event,
1369 u64 timestamp __used,
1370 struct thread *thread __used)
1373 printf("sched_exit event %p\n", event);
1377 process_raw_event(event_t *raw_event __used, void *more_data,
1378 int cpu, u64 timestamp, struct thread *thread)
1380 struct raw_event_sample *raw = more_data;
1381 struct event *event;
1384 type = trace_parse_common_type(raw->data);
1385 event = trace_find_event(type);
1387 if (!strcmp(event->name, "sched_switch"))
1388 process_sched_switch_event(raw, event, cpu, timestamp, thread);
1389 if (!strcmp(event->name, "sched_wakeup"))
1390 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1391 if (!strcmp(event->name, "sched_wakeup_new"))
1392 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1393 if (!strcmp(event->name, "sched_process_fork"))
1394 process_sched_fork_event(raw, event, cpu, timestamp, thread);
1395 if (!strcmp(event->name, "sched_process_exit"))
1396 process_sched_exit_event(event, cpu, timestamp, thread);
1400 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
1404 struct dso *dso = NULL;
1405 struct thread *thread;
1406 u64 ip = event->ip.ip;
1410 void *more_data = event->ip.__more_data;
1413 thread = threads__findnew(event->ip.pid, &threads, &last_match);
1415 if (sample_type & PERF_SAMPLE_TIME) {
1416 timestamp = *(u64 *)more_data;
1417 more_data += sizeof(u64);
1420 if (sample_type & PERF_SAMPLE_CPU) {
1421 cpu = *(u32 *)more_data;
1422 more_data += sizeof(u32);
1423 more_data += sizeof(u32); /* reserved */
1426 if (sample_type & PERF_SAMPLE_PERIOD) {
1427 period = *(u64 *)more_data;
1428 more_data += sizeof(u64);
1431 dump_printf("%p [%p]: PERF_EVENT_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n",
1432 (void *)(offset + head),
1433 (void *)(long)(event->header.size),
1435 event->ip.pid, event->ip.tid,
1439 dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1441 if (thread == NULL) {
1442 eprintf("problem processing %d event, skipping it.\n",
1443 event->header.type);
1447 cpumode = event->header.misc & PERF_EVENT_MISC_CPUMODE_MASK;
1449 if (cpumode == PERF_EVENT_MISC_KERNEL) {
1455 dump_printf(" ...... dso: %s\n", dso->name);
1457 } else if (cpumode == PERF_EVENT_MISC_USER) {
1466 dso = hypervisor_dso;
1468 dump_printf(" ...... dso: [hypervisor]\n");
1471 if (sample_type & PERF_SAMPLE_RAW)
1472 process_raw_event(event, more_data, cpu, timestamp, thread);
1478 process_event(event_t *event, unsigned long offset, unsigned long head)
1482 switch (event->header.type) {
1483 case PERF_EVENT_MMAP ... PERF_EVENT_LOST:
1486 case PERF_EVENT_COMM:
1487 return process_comm_event(event, offset, head);
1489 case PERF_EVENT_EXIT ... PERF_EVENT_READ:
1492 case PERF_EVENT_SAMPLE:
1493 return process_sample_event(event, offset, head);
1495 case PERF_EVENT_MAX:
1503 static int read_events(void)
1505 int ret, rc = EXIT_FAILURE;
1506 unsigned long offset = 0;
1507 unsigned long head = 0;
1508 struct stat perf_stat;
1514 register_idle_thread(&threads, &last_match);
1516 input = open(input_name, O_RDONLY);
1518 perror("failed to open file");
1522 ret = fstat(input, &perf_stat);
1524 perror("failed to stat file");
1528 if (!perf_stat.st_size) {
1529 fprintf(stderr, "zero-sized file, nothing to do!\n");
1532 header = perf_header__read(input);
1533 head = header->data_offset;
1534 sample_type = perf_header__sample_type(header);
1536 if (!(sample_type & PERF_SAMPLE_RAW))
1537 die("No trace sample to read. Did you call perf record "
1540 if (load_kernel() < 0) {
1541 perror("failed to load kernel symbols");
1542 return EXIT_FAILURE;
1546 buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1547 MAP_SHARED, input, offset);
1548 if (buf == MAP_FAILED) {
1549 perror("failed to mmap file");
1554 event = (event_t *)(buf + head);
1556 size = event->header.size;
1560 if (head + event->header.size >= page_size * mmap_window) {
1561 unsigned long shift = page_size * (head / page_size);
1564 res = munmap(buf, page_size * mmap_window);
1572 size = event->header.size;
1575 if (!size || process_event(event, offset, head) < 0) {
1578 * assume we lost track of the stream, check alignment, and
1579 * increment a single u64 in the hope to catch on again 'soon'.
1582 if (unlikely(head & 7))
1590 if (offset + head < (unsigned long)perf_stat.st_size)
1599 static const char * const sched_usage[] = {
1600 "perf sched [<options>] {record|latency|replay|trace}",
1604 static const struct option sched_options[] = {
1605 OPT_BOOLEAN('v', "verbose", &verbose,
1606 "be more verbose (show symbol address, etc)"),
1607 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1608 "dump raw trace in ASCII"),
1612 static const char * const latency_usage[] = {
1613 "perf sched latency [<options>]",
1617 static const struct option latency_options[] = {
1618 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1619 "sort by key(s): runtime, switch, avg, max"),
1620 OPT_BOOLEAN('v', "verbose", &verbose,
1621 "be more verbose (show symbol address, etc)"),
1622 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1623 "dump raw trace in ASCII"),
1627 static const char * const replay_usage[] = {
1628 "perf sched replay [<options>]",
1632 static const struct option replay_options[] = {
1633 OPT_INTEGER('r', "repeat", &replay_repeat,
1634 "repeat the workload replay N times (-1: infinite)"),
1635 OPT_BOOLEAN('v', "verbose", &verbose,
1636 "be more verbose (show symbol address, etc)"),
1637 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1638 "dump raw trace in ASCII"),
1642 static void setup_sorting(void)
1644 char *tmp, *tok, *str = strdup(sort_order);
1646 for (tok = strtok_r(str, ", ", &tmp);
1647 tok; tok = strtok_r(NULL, ", ", &tmp)) {
1648 if (sort_dimension__add(tok, &sort_list) < 0) {
1649 error("Unknown --sort key: `%s'", tok);
1650 usage_with_options(latency_usage, latency_options);
1656 sort_dimension__add((char *)"pid", &cmp_pid);
1659 static const char *record_args[] = {
1664 "-e", "sched:sched_switch:r",
1665 "-e", "sched:sched_stat_wait:r",
1666 "-e", "sched:sched_stat_sleep:r",
1667 "-e", "sched:sched_stat_iowait:r",
1668 "-e", "sched:sched_process_exit:r",
1669 "-e", "sched:sched_process_fork:r",
1670 "-e", "sched:sched_wakeup:r",
1671 "-e", "sched:sched_migrate_task:r",
1674 static int __cmd_record(int argc, const char **argv)
1676 unsigned int rec_argc, i, j;
1677 const char **rec_argv;
1679 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1680 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1682 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1683 rec_argv[i] = strdup(record_args[i]);
1685 for (j = 1; j < (unsigned int)argc; j++, i++)
1686 rec_argv[i] = argv[j];
1688 BUG_ON(i != rec_argc);
1690 return cmd_record(i, rec_argv, NULL);
1693 int cmd_sched(int argc, const char **argv, const char *prefix __used)
1696 page_size = getpagesize();
1698 argc = parse_options(argc, argv, sched_options, sched_usage,
1699 PARSE_OPT_STOP_AT_NON_OPTION);
1701 usage_with_options(sched_usage, sched_options);
1703 if (!strncmp(argv[0], "rec", 3)) {
1704 return __cmd_record(argc, argv);
1705 } else if (!strncmp(argv[0], "lat", 3)) {
1706 trace_handler = &lat_ops;
1708 argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1710 usage_with_options(latency_usage, latency_options);
1714 } else if (!strncmp(argv[0], "rep", 3)) {
1715 trace_handler = &replay_ops;
1717 argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1719 usage_with_options(replay_usage, replay_options);
1722 } else if (!strcmp(argv[0], "trace")) {
1724 * Aliased to 'perf trace' for now:
1726 return cmd_trace(argc, argv, prefix);
1728 usage_with_options(sched_usage, sched_options);