1 // SPDX-License-Identifier: GPL-2.0
6 #include "util/cpumap.h"
7 #include "util/evlist.h"
8 #include "util/evsel.h"
9 #include "util/evsel_fprintf.h"
10 #include "util/symbol.h"
11 #include "util/thread.h"
12 #include "util/header.h"
13 #include "util/session.h"
14 #include "util/tool.h"
15 #include "util/cloexec.h"
16 #include "util/thread_map.h"
17 #include "util/color.h"
18 #include "util/stat.h"
19 #include "util/string2.h"
20 #include "util/callchain.h"
21 #include "util/time-utils.h"
23 #include <subcmd/pager.h>
24 #include <subcmd/parse-options.h>
25 #include "util/trace-event.h"
27 #include "util/debug.h"
28 #include "util/event.h"
30 #include <linux/kernel.h>
31 #include <linux/log2.h>
32 #include <linux/zalloc.h>
33 #include <sys/prctl.h>
34 #include <sys/resource.h>
38 #include <semaphore.h>
41 #include <api/fs/fs.h>
42 #include <perf/cpumap.h>
43 #include <linux/time64.h>
44 #include <linux/err.h>
46 #include <linux/ctype.h>
48 #define PR_SET_NAME 15 /* Set process name */
52 #define MAX_PID 1024000
54 static const char *cpu_list;
55 static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
64 unsigned long nr_events;
65 unsigned long curr_event;
66 struct sched_atom **atoms;
77 enum sched_event_type {
81 SCHED_EVENT_MIGRATION,
85 enum sched_event_type type;
91 struct task_desc *wakee;
94 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
96 /* task state bitmask, copied from include/linux/sched.h */
97 #define TASK_RUNNING 0
98 #define TASK_INTERRUPTIBLE 1
99 #define TASK_UNINTERRUPTIBLE 2
100 #define __TASK_STOPPED 4
101 #define __TASK_TRACED 8
102 /* in tsk->exit_state */
104 #define EXIT_ZOMBIE 32
105 #define EXIT_TRACE (EXIT_ZOMBIE | EXIT_DEAD)
106 /* in tsk->state again */
108 #define TASK_WAKEKILL 128
109 #define TASK_WAKING 256
110 #define TASK_PARKED 512
120 struct list_head list;
121 enum thread_state state;
129 struct list_head work_list;
130 struct thread *thread;
141 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
145 struct trace_sched_handler {
146 int (*switch_event)(struct perf_sched *sched, struct evsel *evsel,
147 struct perf_sample *sample, struct machine *machine);
149 int (*runtime_event)(struct perf_sched *sched, struct evsel *evsel,
150 struct perf_sample *sample, struct machine *machine);
152 int (*wakeup_event)(struct perf_sched *sched, struct evsel *evsel,
153 struct perf_sample *sample, struct machine *machine);
155 /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
156 int (*fork_event)(struct perf_sched *sched, union perf_event *event,
157 struct machine *machine);
159 int (*migrate_task_event)(struct perf_sched *sched,
161 struct perf_sample *sample,
162 struct machine *machine);
165 #define COLOR_PIDS PERF_COLOR_BLUE
166 #define COLOR_CPUS PERF_COLOR_BG_RED
168 struct perf_sched_map {
169 DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
170 struct perf_cpu *comp_cpus;
172 struct perf_thread_map *color_pids;
173 const char *color_pids_str;
174 struct perf_cpu_map *color_cpus;
175 const char *color_cpus_str;
176 struct perf_cpu_map *cpus;
177 const char *cpus_str;
181 struct perf_tool tool;
182 const char *sort_order;
183 unsigned long nr_tasks;
184 struct task_desc **pid_to_task;
185 struct task_desc **tasks;
186 const struct trace_sched_handler *tp_handler;
187 pthread_mutex_t start_work_mutex;
188 pthread_mutex_t work_done_wait_mutex;
191 * Track the current task - that way we can know whether there's any
192 * weird events, such as a task being switched away that is not current.
194 struct perf_cpu max_cpu;
195 u32 curr_pid[MAX_CPUS];
196 struct thread *curr_thread[MAX_CPUS];
197 char next_shortname1;
198 char next_shortname2;
199 unsigned int replay_repeat;
200 unsigned long nr_run_events;
201 unsigned long nr_sleep_events;
202 unsigned long nr_wakeup_events;
203 unsigned long nr_sleep_corrections;
204 unsigned long nr_run_events_optimized;
205 unsigned long targetless_wakeups;
206 unsigned long multitarget_wakeups;
207 unsigned long nr_runs;
208 unsigned long nr_timestamps;
209 unsigned long nr_unordered_timestamps;
210 unsigned long nr_context_switch_bugs;
211 unsigned long nr_events;
212 unsigned long nr_lost_chunks;
213 unsigned long nr_lost_events;
214 u64 run_measurement_overhead;
215 u64 sleep_measurement_overhead;
218 u64 runavg_cpu_usage;
219 u64 parent_cpu_usage;
220 u64 runavg_parent_cpu_usage;
226 u64 cpu_last_switched[MAX_CPUS];
227 struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
228 struct list_head sort_list, cmp_pid;
231 struct perf_sched_map map;
233 /* options for timehist command */
238 unsigned int max_stack;
239 bool show_cpu_visual;
242 bool show_migrations;
245 const char *time_str;
246 struct perf_time_interval ptime;
247 struct perf_time_interval hist_time;
250 /* per thread run time data */
251 struct thread_runtime {
252 u64 last_time; /* time of previous sched in/out event */
253 u64 dt_run; /* run time */
254 u64 dt_sleep; /* time between CPU access by sleep (off cpu) */
255 u64 dt_iowait; /* time between CPU access by iowait (off cpu) */
256 u64 dt_preempt; /* time between CPU access by preempt (off cpu) */
257 u64 dt_delay; /* time between wakeup and sched-in */
258 u64 ready_to_run; /* time of wakeup */
260 struct stats run_stats;
262 u64 total_sleep_time;
263 u64 total_iowait_time;
264 u64 total_preempt_time;
265 u64 total_delay_time;
275 /* per event run time data */
276 struct evsel_runtime {
277 u64 *last_time; /* time this event was last seen per cpu */
278 u32 ncpu; /* highest cpu slot allocated */
281 /* per cpu idle time data */
282 struct idle_thread_runtime {
283 struct thread_runtime tr;
284 struct thread *last_thread;
285 struct rb_root_cached sorted_root;
286 struct callchain_root callchain;
287 struct callchain_cursor cursor;
290 /* track idle times per cpu */
291 static struct thread **idle_threads;
292 static int idle_max_cpu;
293 static char idle_comm[] = "<idle>";
295 static u64 get_nsecs(void)
299 clock_gettime(CLOCK_MONOTONIC, &ts);
301 return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
304 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
306 u64 T0 = get_nsecs(), T1;
310 } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
313 static void sleep_nsecs(u64 nsecs)
317 ts.tv_nsec = nsecs % 999999999;
318 ts.tv_sec = nsecs / 999999999;
320 nanosleep(&ts, NULL);
323 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
325 u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
328 for (i = 0; i < 10; i++) {
330 burn_nsecs(sched, 0);
333 min_delta = min(min_delta, delta);
335 sched->run_measurement_overhead = min_delta;
337 printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
340 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
342 u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
345 for (i = 0; i < 10; i++) {
350 min_delta = min(min_delta, delta);
353 sched->sleep_measurement_overhead = min_delta;
355 printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
358 static struct sched_atom *
359 get_new_event(struct task_desc *task, u64 timestamp)
361 struct sched_atom *event = zalloc(sizeof(*event));
362 unsigned long idx = task->nr_events;
365 event->timestamp = timestamp;
369 size = sizeof(struct sched_atom *) * task->nr_events;
370 task->atoms = realloc(task->atoms, size);
371 BUG_ON(!task->atoms);
373 task->atoms[idx] = event;
378 static struct sched_atom *last_event(struct task_desc *task)
380 if (!task->nr_events)
383 return task->atoms[task->nr_events - 1];
386 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
387 u64 timestamp, u64 duration)
389 struct sched_atom *event, *curr_event = last_event(task);
392 * optimize an existing RUN event by merging this one
395 if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
396 sched->nr_run_events_optimized++;
397 curr_event->duration += duration;
401 event = get_new_event(task, timestamp);
403 event->type = SCHED_EVENT_RUN;
404 event->duration = duration;
406 sched->nr_run_events++;
409 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
410 u64 timestamp, struct task_desc *wakee)
412 struct sched_atom *event, *wakee_event;
414 event = get_new_event(task, timestamp);
415 event->type = SCHED_EVENT_WAKEUP;
416 event->wakee = wakee;
418 wakee_event = last_event(wakee);
419 if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
420 sched->targetless_wakeups++;
423 if (wakee_event->wait_sem) {
424 sched->multitarget_wakeups++;
428 wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
429 sem_init(wakee_event->wait_sem, 0, 0);
430 wakee_event->specific_wait = 1;
431 event->wait_sem = wakee_event->wait_sem;
433 sched->nr_wakeup_events++;
436 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
437 u64 timestamp, u64 task_state __maybe_unused)
439 struct sched_atom *event = get_new_event(task, timestamp);
441 event->type = SCHED_EVENT_SLEEP;
443 sched->nr_sleep_events++;
446 static struct task_desc *register_pid(struct perf_sched *sched,
447 unsigned long pid, const char *comm)
449 struct task_desc *task;
452 if (sched->pid_to_task == NULL) {
453 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
455 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
457 if (pid >= (unsigned long)pid_max) {
458 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
459 sizeof(struct task_desc *))) == NULL);
460 while (pid >= (unsigned long)pid_max)
461 sched->pid_to_task[pid_max++] = NULL;
464 task = sched->pid_to_task[pid];
469 task = zalloc(sizeof(*task));
471 task->nr = sched->nr_tasks;
472 strcpy(task->comm, comm);
474 * every task starts in sleeping state - this gets ignored
475 * if there's no wakeup pointing to this sleep state:
477 add_sched_event_sleep(sched, task, 0, 0);
479 sched->pid_to_task[pid] = task;
481 sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
482 BUG_ON(!sched->tasks);
483 sched->tasks[task->nr] = task;
486 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
492 static void print_task_traces(struct perf_sched *sched)
494 struct task_desc *task;
497 for (i = 0; i < sched->nr_tasks; i++) {
498 task = sched->tasks[i];
499 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
500 task->nr, task->comm, task->pid, task->nr_events);
504 static void add_cross_task_wakeups(struct perf_sched *sched)
506 struct task_desc *task1, *task2;
509 for (i = 0; i < sched->nr_tasks; i++) {
510 task1 = sched->tasks[i];
512 if (j == sched->nr_tasks)
514 task2 = sched->tasks[j];
515 add_sched_event_wakeup(sched, task1, 0, task2);
519 static void perf_sched__process_event(struct perf_sched *sched,
520 struct sched_atom *atom)
524 switch (atom->type) {
525 case SCHED_EVENT_RUN:
526 burn_nsecs(sched, atom->duration);
528 case SCHED_EVENT_SLEEP:
530 ret = sem_wait(atom->wait_sem);
533 case SCHED_EVENT_WAKEUP:
535 ret = sem_post(atom->wait_sem);
538 case SCHED_EVENT_MIGRATION:
545 static u64 get_cpu_usage_nsec_parent(void)
551 err = getrusage(RUSAGE_SELF, &ru);
554 sum = ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
555 sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
560 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
562 struct perf_event_attr attr;
563 char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
566 bool need_privilege = false;
568 memset(&attr, 0, sizeof(attr));
570 attr.type = PERF_TYPE_SOFTWARE;
571 attr.config = PERF_COUNT_SW_TASK_CLOCK;
574 fd = sys_perf_event_open(&attr, 0, -1, -1,
575 perf_event_open_cloexec_flag());
578 if (errno == EMFILE) {
580 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
581 limit.rlim_cur += sched->nr_tasks - cur_task;
582 if (limit.rlim_cur > limit.rlim_max) {
583 limit.rlim_max = limit.rlim_cur;
584 need_privilege = true;
586 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
587 if (need_privilege && errno == EPERM)
588 strcpy(info, "Need privilege\n");
592 strcpy(info, "Have a try with -f option\n");
594 pr_err("Error: sys_perf_event_open() syscall returned "
595 "with %d (%s)\n%s", fd,
596 str_error_r(errno, sbuf, sizeof(sbuf)), info);
602 static u64 get_cpu_usage_nsec_self(int fd)
607 ret = read(fd, &runtime, sizeof(runtime));
608 BUG_ON(ret != sizeof(runtime));
613 struct sched_thread_parms {
614 struct task_desc *task;
615 struct perf_sched *sched;
619 static void *thread_func(void *ctx)
621 struct sched_thread_parms *parms = ctx;
622 struct task_desc *this_task = parms->task;
623 struct perf_sched *sched = parms->sched;
624 u64 cpu_usage_0, cpu_usage_1;
625 unsigned long i, ret;
631 sprintf(comm2, ":%s", this_task->comm);
632 prctl(PR_SET_NAME, comm2);
636 ret = sem_post(&this_task->ready_for_work);
638 ret = pthread_mutex_lock(&sched->start_work_mutex);
640 ret = pthread_mutex_unlock(&sched->start_work_mutex);
643 cpu_usage_0 = get_cpu_usage_nsec_self(fd);
645 for (i = 0; i < this_task->nr_events; i++) {
646 this_task->curr_event = i;
647 perf_sched__process_event(sched, this_task->atoms[i]);
650 cpu_usage_1 = get_cpu_usage_nsec_self(fd);
651 this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
652 ret = sem_post(&this_task->work_done_sem);
655 ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
657 ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
663 static void create_tasks(struct perf_sched *sched)
665 struct task_desc *task;
670 err = pthread_attr_init(&attr);
672 err = pthread_attr_setstacksize(&attr,
673 (size_t) max(16 * 1024, (int)PTHREAD_STACK_MIN));
675 err = pthread_mutex_lock(&sched->start_work_mutex);
677 err = pthread_mutex_lock(&sched->work_done_wait_mutex);
679 for (i = 0; i < sched->nr_tasks; i++) {
680 struct sched_thread_parms *parms = malloc(sizeof(*parms));
681 BUG_ON(parms == NULL);
682 parms->task = task = sched->tasks[i];
683 parms->sched = sched;
684 parms->fd = self_open_counters(sched, i);
685 sem_init(&task->sleep_sem, 0, 0);
686 sem_init(&task->ready_for_work, 0, 0);
687 sem_init(&task->work_done_sem, 0, 0);
688 task->curr_event = 0;
689 err = pthread_create(&task->thread, &attr, thread_func, parms);
694 static void wait_for_tasks(struct perf_sched *sched)
696 u64 cpu_usage_0, cpu_usage_1;
697 struct task_desc *task;
698 unsigned long i, ret;
700 sched->start_time = get_nsecs();
701 sched->cpu_usage = 0;
702 pthread_mutex_unlock(&sched->work_done_wait_mutex);
704 for (i = 0; i < sched->nr_tasks; i++) {
705 task = sched->tasks[i];
706 ret = sem_wait(&task->ready_for_work);
708 sem_init(&task->ready_for_work, 0, 0);
710 ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
713 cpu_usage_0 = get_cpu_usage_nsec_parent();
715 pthread_mutex_unlock(&sched->start_work_mutex);
717 for (i = 0; i < sched->nr_tasks; i++) {
718 task = sched->tasks[i];
719 ret = sem_wait(&task->work_done_sem);
721 sem_init(&task->work_done_sem, 0, 0);
722 sched->cpu_usage += task->cpu_usage;
726 cpu_usage_1 = get_cpu_usage_nsec_parent();
727 if (!sched->runavg_cpu_usage)
728 sched->runavg_cpu_usage = sched->cpu_usage;
729 sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
731 sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
732 if (!sched->runavg_parent_cpu_usage)
733 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
734 sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
735 sched->parent_cpu_usage)/sched->replay_repeat;
737 ret = pthread_mutex_lock(&sched->start_work_mutex);
740 for (i = 0; i < sched->nr_tasks; i++) {
741 task = sched->tasks[i];
742 sem_init(&task->sleep_sem, 0, 0);
743 task->curr_event = 0;
747 static void run_one_test(struct perf_sched *sched)
749 u64 T0, T1, delta, avg_delta, fluct;
752 wait_for_tasks(sched);
756 sched->sum_runtime += delta;
759 avg_delta = sched->sum_runtime / sched->nr_runs;
760 if (delta < avg_delta)
761 fluct = avg_delta - delta;
763 fluct = delta - avg_delta;
764 sched->sum_fluct += fluct;
766 sched->run_avg = delta;
767 sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
769 printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
771 printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
773 printf("cpu: %0.2f / %0.2f",
774 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
778 * rusage statistics done by the parent, these are less
779 * accurate than the sched->sum_exec_runtime based statistics:
781 printf(" [%0.2f / %0.2f]",
782 (double)sched->parent_cpu_usage / NSEC_PER_MSEC,
783 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
788 if (sched->nr_sleep_corrections)
789 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
790 sched->nr_sleep_corrections = 0;
793 static void test_calibrations(struct perf_sched *sched)
798 burn_nsecs(sched, NSEC_PER_MSEC);
801 printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
804 sleep_nsecs(NSEC_PER_MSEC);
807 printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
811 replay_wakeup_event(struct perf_sched *sched,
812 struct evsel *evsel, struct perf_sample *sample,
813 struct machine *machine __maybe_unused)
815 const char *comm = evsel__strval(evsel, sample, "comm");
816 const u32 pid = evsel__intval(evsel, sample, "pid");
817 struct task_desc *waker, *wakee;
820 printf("sched_wakeup event %p\n", evsel);
822 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
825 waker = register_pid(sched, sample->tid, "<unknown>");
826 wakee = register_pid(sched, pid, comm);
828 add_sched_event_wakeup(sched, waker, sample->time, wakee);
832 static int replay_switch_event(struct perf_sched *sched,
834 struct perf_sample *sample,
835 struct machine *machine __maybe_unused)
837 const char *prev_comm = evsel__strval(evsel, sample, "prev_comm"),
838 *next_comm = evsel__strval(evsel, sample, "next_comm");
839 const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
840 next_pid = evsel__intval(evsel, sample, "next_pid");
841 const u64 prev_state = evsel__intval(evsel, sample, "prev_state");
842 struct task_desc *prev, __maybe_unused *next;
843 u64 timestamp0, timestamp = sample->time;
844 int cpu = sample->cpu;
848 printf("sched_switch event %p\n", evsel);
850 if (cpu >= MAX_CPUS || cpu < 0)
853 timestamp0 = sched->cpu_last_switched[cpu];
855 delta = timestamp - timestamp0;
860 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
864 pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
865 prev_comm, prev_pid, next_comm, next_pid, delta);
867 prev = register_pid(sched, prev_pid, prev_comm);
868 next = register_pid(sched, next_pid, next_comm);
870 sched->cpu_last_switched[cpu] = timestamp;
872 add_sched_event_run(sched, prev, timestamp, delta);
873 add_sched_event_sleep(sched, prev, timestamp, prev_state);
878 static int replay_fork_event(struct perf_sched *sched,
879 union perf_event *event,
880 struct machine *machine)
882 struct thread *child, *parent;
884 child = machine__findnew_thread(machine, event->fork.pid,
886 parent = machine__findnew_thread(machine, event->fork.ppid,
889 if (child == NULL || parent == NULL) {
890 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
896 printf("fork event\n");
897 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
898 printf("... child: %s/%d\n", thread__comm_str(child), child->tid);
901 register_pid(sched, parent->tid, thread__comm_str(parent));
902 register_pid(sched, child->tid, thread__comm_str(child));
909 struct sort_dimension {
912 struct list_head list;
916 * handle runtime stats saved per thread
918 static struct thread_runtime *thread__init_runtime(struct thread *thread)
920 struct thread_runtime *r;
922 r = zalloc(sizeof(struct thread_runtime));
926 init_stats(&r->run_stats);
927 thread__set_priv(thread, r);
932 static struct thread_runtime *thread__get_runtime(struct thread *thread)
934 struct thread_runtime *tr;
936 tr = thread__priv(thread);
938 tr = thread__init_runtime(thread);
940 pr_debug("Failed to malloc memory for runtime data.\n");
947 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
949 struct sort_dimension *sort;
952 BUG_ON(list_empty(list));
954 list_for_each_entry(sort, list, list) {
955 ret = sort->cmp(l, r);
963 static struct work_atoms *
964 thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
965 struct list_head *sort_list)
967 struct rb_node *node = root->rb_root.rb_node;
968 struct work_atoms key = { .thread = thread };
971 struct work_atoms *atoms;
974 atoms = container_of(node, struct work_atoms, node);
976 cmp = thread_lat_cmp(sort_list, &key, atoms);
978 node = node->rb_left;
980 node = node->rb_right;
982 BUG_ON(thread != atoms->thread);
990 __thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
991 struct list_head *sort_list)
993 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
994 bool leftmost = true;
997 struct work_atoms *this;
1000 this = container_of(*new, struct work_atoms, node);
1003 cmp = thread_lat_cmp(sort_list, data, this);
1006 new = &((*new)->rb_left);
1008 new = &((*new)->rb_right);
1013 rb_link_node(&data->node, parent, new);
1014 rb_insert_color_cached(&data->node, root, leftmost);
1017 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1019 struct work_atoms *atoms = zalloc(sizeof(*atoms));
1021 pr_err("No memory at %s\n", __func__);
1025 atoms->thread = thread__get(thread);
1026 INIT_LIST_HEAD(&atoms->work_list);
1027 __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1031 static char sched_out_state(u64 prev_state)
1033 const char *str = TASK_STATE_TO_CHAR_STR;
1035 return str[prev_state];
1039 add_sched_out_event(struct work_atoms *atoms,
1043 struct work_atom *atom = zalloc(sizeof(*atom));
1045 pr_err("Non memory at %s", __func__);
1049 atom->sched_out_time = timestamp;
1051 if (run_state == 'R') {
1052 atom->state = THREAD_WAIT_CPU;
1053 atom->wake_up_time = atom->sched_out_time;
1056 list_add_tail(&atom->list, &atoms->work_list);
1061 add_runtime_event(struct work_atoms *atoms, u64 delta,
1062 u64 timestamp __maybe_unused)
1064 struct work_atom *atom;
1066 BUG_ON(list_empty(&atoms->work_list));
1068 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1070 atom->runtime += delta;
1071 atoms->total_runtime += delta;
1075 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1077 struct work_atom *atom;
1080 if (list_empty(&atoms->work_list))
1083 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1085 if (atom->state != THREAD_WAIT_CPU)
1088 if (timestamp < atom->wake_up_time) {
1089 atom->state = THREAD_IGNORE;
1093 atom->state = THREAD_SCHED_IN;
1094 atom->sched_in_time = timestamp;
1096 delta = atom->sched_in_time - atom->wake_up_time;
1097 atoms->total_lat += delta;
1098 if (delta > atoms->max_lat) {
1099 atoms->max_lat = delta;
1100 atoms->max_lat_start = atom->wake_up_time;
1101 atoms->max_lat_end = timestamp;
1106 static int latency_switch_event(struct perf_sched *sched,
1107 struct evsel *evsel,
1108 struct perf_sample *sample,
1109 struct machine *machine)
1111 const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1112 next_pid = evsel__intval(evsel, sample, "next_pid");
1113 const u64 prev_state = evsel__intval(evsel, sample, "prev_state");
1114 struct work_atoms *out_events, *in_events;
1115 struct thread *sched_out, *sched_in;
1116 u64 timestamp0, timestamp = sample->time;
1117 int cpu = sample->cpu, err = -1;
1120 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1122 timestamp0 = sched->cpu_last_switched[cpu];
1123 sched->cpu_last_switched[cpu] = timestamp;
1125 delta = timestamp - timestamp0;
1130 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1134 sched_out = machine__findnew_thread(machine, -1, prev_pid);
1135 sched_in = machine__findnew_thread(machine, -1, next_pid);
1136 if (sched_out == NULL || sched_in == NULL)
1139 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1141 if (thread_atoms_insert(sched, sched_out))
1143 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1145 pr_err("out-event: Internal tree error");
1149 if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1152 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1154 if (thread_atoms_insert(sched, sched_in))
1156 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1158 pr_err("in-event: Internal tree error");
1162 * Take came in we have not heard about yet,
1163 * add in an initial atom in runnable state:
1165 if (add_sched_out_event(in_events, 'R', timestamp))
1168 add_sched_in_event(in_events, timestamp);
1171 thread__put(sched_out);
1172 thread__put(sched_in);
1176 static int latency_runtime_event(struct perf_sched *sched,
1177 struct evsel *evsel,
1178 struct perf_sample *sample,
1179 struct machine *machine)
1181 const u32 pid = evsel__intval(evsel, sample, "pid");
1182 const u64 runtime = evsel__intval(evsel, sample, "runtime");
1183 struct thread *thread = machine__findnew_thread(machine, -1, pid);
1184 struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1185 u64 timestamp = sample->time;
1186 int cpu = sample->cpu, err = -1;
1191 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1193 if (thread_atoms_insert(sched, thread))
1195 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1197 pr_err("in-event: Internal tree error");
1200 if (add_sched_out_event(atoms, 'R', timestamp))
1204 add_runtime_event(atoms, runtime, timestamp);
1207 thread__put(thread);
1211 static int latency_wakeup_event(struct perf_sched *sched,
1212 struct evsel *evsel,
1213 struct perf_sample *sample,
1214 struct machine *machine)
1216 const u32 pid = evsel__intval(evsel, sample, "pid");
1217 struct work_atoms *atoms;
1218 struct work_atom *atom;
1219 struct thread *wakee;
1220 u64 timestamp = sample->time;
1223 wakee = machine__findnew_thread(machine, -1, pid);
1226 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1228 if (thread_atoms_insert(sched, wakee))
1230 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1232 pr_err("wakeup-event: Internal tree error");
1235 if (add_sched_out_event(atoms, 'S', timestamp))
1239 BUG_ON(list_empty(&atoms->work_list));
1241 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1244 * As we do not guarantee the wakeup event happens when
1245 * task is out of run queue, also may happen when task is
1246 * on run queue and wakeup only change ->state to TASK_RUNNING,
1247 * then we should not set the ->wake_up_time when wake up a
1248 * task which is on run queue.
1250 * You WILL be missing events if you've recorded only
1251 * one CPU, or are only looking at only one, so don't
1252 * skip in this case.
1254 if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1257 sched->nr_timestamps++;
1258 if (atom->sched_out_time > timestamp) {
1259 sched->nr_unordered_timestamps++;
1263 atom->state = THREAD_WAIT_CPU;
1264 atom->wake_up_time = timestamp;
1272 static int latency_migrate_task_event(struct perf_sched *sched,
1273 struct evsel *evsel,
1274 struct perf_sample *sample,
1275 struct machine *machine)
1277 const u32 pid = evsel__intval(evsel, sample, "pid");
1278 u64 timestamp = sample->time;
1279 struct work_atoms *atoms;
1280 struct work_atom *atom;
1281 struct thread *migrant;
1285 * Only need to worry about migration when profiling one CPU.
1287 if (sched->profile_cpu == -1)
1290 migrant = machine__findnew_thread(machine, -1, pid);
1291 if (migrant == NULL)
1293 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1295 if (thread_atoms_insert(sched, migrant))
1297 register_pid(sched, migrant->tid, thread__comm_str(migrant));
1298 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1300 pr_err("migration-event: Internal tree error");
1303 if (add_sched_out_event(atoms, 'R', timestamp))
1307 BUG_ON(list_empty(&atoms->work_list));
1309 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1310 atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1312 sched->nr_timestamps++;
1314 if (atom->sched_out_time > timestamp)
1315 sched->nr_unordered_timestamps++;
1318 thread__put(migrant);
1322 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1327 char max_lat_start[32], max_lat_end[32];
1329 if (!work_list->nb_atoms)
1332 * Ignore idle threads:
1334 if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1337 sched->all_runtime += work_list->total_runtime;
1338 sched->all_count += work_list->nb_atoms;
1340 if (work_list->num_merged > 1)
1341 ret = printf(" %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1343 ret = printf(" %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1345 for (i = 0; i < 24 - ret; i++)
1348 avg = work_list->total_lat / work_list->nb_atoms;
1349 timestamp__scnprintf_usec(work_list->max_lat_start, max_lat_start, sizeof(max_lat_start));
1350 timestamp__scnprintf_usec(work_list->max_lat_end, max_lat_end, sizeof(max_lat_end));
1352 printf("|%11.3f ms |%9" PRIu64 " | avg:%8.3f ms | max:%8.3f ms | max start: %12s s | max end: %12s s\n",
1353 (double)work_list->total_runtime / NSEC_PER_MSEC,
1354 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1355 (double)work_list->max_lat / NSEC_PER_MSEC,
1356 max_lat_start, max_lat_end);
1359 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1361 if (l->thread == r->thread)
1363 if (l->thread->tid < r->thread->tid)
1365 if (l->thread->tid > r->thread->tid)
1367 return (int)(l->thread - r->thread);
1370 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1380 avgl = l->total_lat / l->nb_atoms;
1381 avgr = r->total_lat / r->nb_atoms;
1391 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1393 if (l->max_lat < r->max_lat)
1395 if (l->max_lat > r->max_lat)
1401 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1403 if (l->nb_atoms < r->nb_atoms)
1405 if (l->nb_atoms > r->nb_atoms)
1411 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1413 if (l->total_runtime < r->total_runtime)
1415 if (l->total_runtime > r->total_runtime)
1421 static int sort_dimension__add(const char *tok, struct list_head *list)
1424 static struct sort_dimension avg_sort_dimension = {
1428 static struct sort_dimension max_sort_dimension = {
1432 static struct sort_dimension pid_sort_dimension = {
1436 static struct sort_dimension runtime_sort_dimension = {
1440 static struct sort_dimension switch_sort_dimension = {
1444 struct sort_dimension *available_sorts[] = {
1445 &pid_sort_dimension,
1446 &avg_sort_dimension,
1447 &max_sort_dimension,
1448 &switch_sort_dimension,
1449 &runtime_sort_dimension,
1452 for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1453 if (!strcmp(available_sorts[i]->name, tok)) {
1454 list_add_tail(&available_sorts[i]->list, list);
1463 static void perf_sched__sort_lat(struct perf_sched *sched)
1465 struct rb_node *node;
1466 struct rb_root_cached *root = &sched->atom_root;
1469 struct work_atoms *data;
1470 node = rb_first_cached(root);
1474 rb_erase_cached(node, root);
1475 data = rb_entry(node, struct work_atoms, node);
1476 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1478 if (root == &sched->atom_root) {
1479 root = &sched->merged_atom_root;
1484 static int process_sched_wakeup_event(struct perf_tool *tool,
1485 struct evsel *evsel,
1486 struct perf_sample *sample,
1487 struct machine *machine)
1489 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1491 if (sched->tp_handler->wakeup_event)
1492 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1502 static bool thread__has_color(struct thread *thread)
1504 union map_priv priv = {
1505 .ptr = thread__priv(thread),
1511 static struct thread*
1512 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1514 struct thread *thread = machine__findnew_thread(machine, pid, tid);
1515 union map_priv priv = {
1519 if (!sched->map.color_pids || !thread || thread__priv(thread))
1522 if (thread_map__has(sched->map.color_pids, tid))
1525 thread__set_priv(thread, priv.ptr);
1529 static int map_switch_event(struct perf_sched *sched, struct evsel *evsel,
1530 struct perf_sample *sample, struct machine *machine)
1532 const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
1533 struct thread *sched_in;
1534 struct thread_runtime *tr;
1536 u64 timestamp0, timestamp = sample->time;
1539 struct perf_cpu this_cpu = {
1543 bool new_cpu = false;
1544 const char *color = PERF_COLOR_NORMAL;
1545 char stimestamp[32];
1547 BUG_ON(this_cpu.cpu >= MAX_CPUS || this_cpu.cpu < 0);
1549 if (this_cpu.cpu > sched->max_cpu.cpu)
1550 sched->max_cpu = this_cpu;
1552 if (sched->map.comp) {
1553 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1554 if (!test_and_set_bit(this_cpu.cpu, sched->map.comp_cpus_mask)) {
1555 sched->map.comp_cpus[cpus_nr++] = this_cpu;
1559 cpus_nr = sched->max_cpu.cpu;
1561 timestamp0 = sched->cpu_last_switched[this_cpu.cpu];
1562 sched->cpu_last_switched[this_cpu.cpu] = timestamp;
1564 delta = timestamp - timestamp0;
1569 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1573 sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1574 if (sched_in == NULL)
1577 tr = thread__get_runtime(sched_in);
1579 thread__put(sched_in);
1583 sched->curr_thread[this_cpu.cpu] = thread__get(sched_in);
1588 if (!tr->shortname[0]) {
1589 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1591 * Don't allocate a letter-number for swapper:0
1592 * as a shortname. Instead, we use '.' for it.
1594 tr->shortname[0] = '.';
1595 tr->shortname[1] = ' ';
1597 tr->shortname[0] = sched->next_shortname1;
1598 tr->shortname[1] = sched->next_shortname2;
1600 if (sched->next_shortname1 < 'Z') {
1601 sched->next_shortname1++;
1603 sched->next_shortname1 = 'A';
1604 if (sched->next_shortname2 < '9')
1605 sched->next_shortname2++;
1607 sched->next_shortname2 = '0';
1613 for (i = 0; i < cpus_nr; i++) {
1614 struct perf_cpu cpu = {
1615 .cpu = sched->map.comp ? sched->map.comp_cpus[i].cpu : i,
1617 struct thread *curr_thread = sched->curr_thread[cpu.cpu];
1618 struct thread_runtime *curr_tr;
1619 const char *pid_color = color;
1620 const char *cpu_color = color;
1622 if (curr_thread && thread__has_color(curr_thread))
1623 pid_color = COLOR_PIDS;
1625 if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, cpu))
1628 if (sched->map.color_cpus && perf_cpu_map__has(sched->map.color_cpus, cpu))
1629 cpu_color = COLOR_CPUS;
1631 if (cpu.cpu != this_cpu.cpu)
1632 color_fprintf(stdout, color, " ");
1634 color_fprintf(stdout, cpu_color, "*");
1636 if (sched->curr_thread[cpu.cpu]) {
1637 curr_tr = thread__get_runtime(sched->curr_thread[cpu.cpu]);
1638 if (curr_tr == NULL) {
1639 thread__put(sched_in);
1642 color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1644 color_fprintf(stdout, color, " ");
1647 if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, this_cpu))
1650 timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1651 color_fprintf(stdout, color, " %12s secs ", stimestamp);
1652 if (new_shortname || tr->comm_changed || (verbose > 0 && sched_in->tid)) {
1653 const char *pid_color = color;
1655 if (thread__has_color(sched_in))
1656 pid_color = COLOR_PIDS;
1658 color_fprintf(stdout, pid_color, "%s => %s:%d",
1659 tr->shortname, thread__comm_str(sched_in), sched_in->tid);
1660 tr->comm_changed = false;
1663 if (sched->map.comp && new_cpu)
1664 color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1667 color_fprintf(stdout, color, "\n");
1669 thread__put(sched_in);
1674 static int process_sched_switch_event(struct perf_tool *tool,
1675 struct evsel *evsel,
1676 struct perf_sample *sample,
1677 struct machine *machine)
1679 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1680 int this_cpu = sample->cpu, err = 0;
1681 u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1682 next_pid = evsel__intval(evsel, sample, "next_pid");
1684 if (sched->curr_pid[this_cpu] != (u32)-1) {
1686 * Are we trying to switch away a PID that is
1689 if (sched->curr_pid[this_cpu] != prev_pid)
1690 sched->nr_context_switch_bugs++;
1693 if (sched->tp_handler->switch_event)
1694 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1696 sched->curr_pid[this_cpu] = next_pid;
1700 static int process_sched_runtime_event(struct perf_tool *tool,
1701 struct evsel *evsel,
1702 struct perf_sample *sample,
1703 struct machine *machine)
1705 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1707 if (sched->tp_handler->runtime_event)
1708 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1713 static int perf_sched__process_fork_event(struct perf_tool *tool,
1714 union perf_event *event,
1715 struct perf_sample *sample,
1716 struct machine *machine)
1718 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1720 /* run the fork event through the perf machinery */
1721 perf_event__process_fork(tool, event, sample, machine);
1723 /* and then run additional processing needed for this command */
1724 if (sched->tp_handler->fork_event)
1725 return sched->tp_handler->fork_event(sched, event, machine);
1730 static int process_sched_migrate_task_event(struct perf_tool *tool,
1731 struct evsel *evsel,
1732 struct perf_sample *sample,
1733 struct machine *machine)
1735 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1737 if (sched->tp_handler->migrate_task_event)
1738 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1743 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1744 struct evsel *evsel,
1745 struct perf_sample *sample,
1746 struct machine *machine);
1748 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1749 union perf_event *event __maybe_unused,
1750 struct perf_sample *sample,
1751 struct evsel *evsel,
1752 struct machine *machine)
1756 if (evsel->handler != NULL) {
1757 tracepoint_handler f = evsel->handler;
1758 err = f(tool, evsel, sample, machine);
1764 static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1765 union perf_event *event,
1766 struct perf_sample *sample,
1767 struct machine *machine)
1769 struct thread *thread;
1770 struct thread_runtime *tr;
1773 err = perf_event__process_comm(tool, event, sample, machine);
1777 thread = machine__find_thread(machine, sample->pid, sample->tid);
1779 pr_err("Internal error: can't find thread\n");
1783 tr = thread__get_runtime(thread);
1785 thread__put(thread);
1789 tr->comm_changed = true;
1790 thread__put(thread);
1795 static int perf_sched__read_events(struct perf_sched *sched)
1797 const struct evsel_str_handler handlers[] = {
1798 { "sched:sched_switch", process_sched_switch_event, },
1799 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1800 { "sched:sched_wakeup", process_sched_wakeup_event, },
1801 { "sched:sched_wakeup_new", process_sched_wakeup_event, },
1802 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1804 struct perf_session *session;
1805 struct perf_data data = {
1807 .mode = PERF_DATA_MODE_READ,
1808 .force = sched->force,
1812 session = perf_session__new(&data, &sched->tool);
1813 if (IS_ERR(session)) {
1814 pr_debug("Error creating perf session");
1815 return PTR_ERR(session);
1818 symbol__init(&session->header.env);
1820 if (perf_session__set_tracepoints_handlers(session, handlers))
1823 if (perf_session__has_traces(session, "record -R")) {
1824 int err = perf_session__process_events(session);
1826 pr_err("Failed to process events, error %d", err);
1830 sched->nr_events = session->evlist->stats.nr_events[0];
1831 sched->nr_lost_events = session->evlist->stats.total_lost;
1832 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1837 perf_session__delete(session);
1842 * scheduling times are printed as msec.usec
1844 static inline void print_sched_time(unsigned long long nsecs, int width)
1846 unsigned long msecs;
1847 unsigned long usecs;
1849 msecs = nsecs / NSEC_PER_MSEC;
1850 nsecs -= msecs * NSEC_PER_MSEC;
1851 usecs = nsecs / NSEC_PER_USEC;
1852 printf("%*lu.%03lu ", width, msecs, usecs);
1856 * returns runtime data for event, allocating memory for it the
1857 * first time it is used.
1859 static struct evsel_runtime *evsel__get_runtime(struct evsel *evsel)
1861 struct evsel_runtime *r = evsel->priv;
1864 r = zalloc(sizeof(struct evsel_runtime));
1872 * save last time event was seen per cpu
1874 static void evsel__save_time(struct evsel *evsel, u64 timestamp, u32 cpu)
1876 struct evsel_runtime *r = evsel__get_runtime(evsel);
1881 if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1882 int i, n = __roundup_pow_of_two(cpu+1);
1883 void *p = r->last_time;
1885 p = realloc(r->last_time, n * sizeof(u64));
1890 for (i = r->ncpu; i < n; ++i)
1891 r->last_time[i] = (u64) 0;
1896 r->last_time[cpu] = timestamp;
1899 /* returns last time this event was seen on the given cpu */
1900 static u64 evsel__get_time(struct evsel *evsel, u32 cpu)
1902 struct evsel_runtime *r = evsel__get_runtime(evsel);
1904 if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1907 return r->last_time[cpu];
1910 static int comm_width = 30;
1912 static char *timehist_get_commstr(struct thread *thread)
1914 static char str[32];
1915 const char *comm = thread__comm_str(thread);
1916 pid_t tid = thread->tid;
1917 pid_t pid = thread->pid_;
1921 n = scnprintf(str, sizeof(str), "%s", comm);
1923 else if (tid != pid)
1924 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1927 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1935 static void timehist_header(struct perf_sched *sched)
1937 u32 ncpus = sched->max_cpu.cpu + 1;
1940 printf("%15s %6s ", "time", "cpu");
1942 if (sched->show_cpu_visual) {
1944 for (i = 0, j = 0; i < ncpus; ++i) {
1952 printf(" %-*s %9s %9s %9s", comm_width,
1953 "task name", "wait time", "sch delay", "run time");
1955 if (sched->show_state)
1956 printf(" %s", "state");
1963 printf("%15s %-6s ", "", "");
1965 if (sched->show_cpu_visual)
1966 printf(" %*s ", ncpus, "");
1968 printf(" %-*s %9s %9s %9s", comm_width,
1969 "[tid/pid]", "(msec)", "(msec)", "(msec)");
1971 if (sched->show_state)
1979 printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1981 if (sched->show_cpu_visual)
1982 printf(" %.*s ", ncpus, graph_dotted_line);
1984 printf(" %.*s %.9s %.9s %.9s", comm_width,
1985 graph_dotted_line, graph_dotted_line, graph_dotted_line,
1988 if (sched->show_state)
1989 printf(" %.5s", graph_dotted_line);
1994 static char task_state_char(struct thread *thread, int state)
1996 static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
1997 unsigned bit = state ? ffs(state) : 0;
2000 if (thread->tid == 0)
2003 return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
2006 static void timehist_print_sample(struct perf_sched *sched,
2007 struct evsel *evsel,
2008 struct perf_sample *sample,
2009 struct addr_location *al,
2010 struct thread *thread,
2013 struct thread_runtime *tr = thread__priv(thread);
2014 const char *next_comm = evsel__strval(evsel, sample, "next_comm");
2015 const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
2016 u32 max_cpus = sched->max_cpu.cpu + 1;
2021 if (cpu_list && !test_bit(sample->cpu, cpu_bitmap))
2024 timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2025 printf("%15s [%04d] ", tstr, sample->cpu);
2027 if (sched->show_cpu_visual) {
2032 for (i = 0; i < max_cpus; ++i) {
2033 /* flag idle times with 'i'; others are sched events */
2034 if (i == sample->cpu)
2035 c = (thread->tid == 0) ? 'i' : 's';
2043 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2045 wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2046 print_sched_time(wait_time, 6);
2048 print_sched_time(tr->dt_delay, 6);
2049 print_sched_time(tr->dt_run, 6);
2051 if (sched->show_state)
2052 printf(" %5c ", task_state_char(thread, state));
2054 if (sched->show_next) {
2055 snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2056 printf(" %-*s", comm_width, nstr);
2059 if (sched->show_wakeups && !sched->show_next)
2060 printf(" %-*s", comm_width, "");
2062 if (thread->tid == 0)
2065 if (sched->show_callchain)
2068 sample__fprintf_sym(sample, al, 0,
2069 EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2070 EVSEL__PRINT_CALLCHAIN_ARROW |
2071 EVSEL__PRINT_SKIP_IGNORED,
2072 &callchain_cursor, symbol_conf.bt_stop_list, stdout);
2079 * Explanation of delta-time stats:
2081 * t = time of current schedule out event
2082 * tprev = time of previous sched out event
2083 * also time of schedule-in event for current task
2084 * last_time = time of last sched change event for current task
2085 * (i.e, time process was last scheduled out)
2086 * ready_to_run = time of wakeup for current task
2088 * -----|------------|------------|------------|------
2089 * last ready tprev t
2092 * |-------- dt_wait --------|
2093 * |- dt_delay -|-- dt_run --|
2095 * dt_run = run time of current task
2096 * dt_wait = time between last schedule out event for task and tprev
2097 * represents time spent off the cpu
2098 * dt_delay = time between wakeup and schedule-in of task
2101 static void timehist_update_runtime_stats(struct thread_runtime *r,
2111 r->dt_run = t - tprev;
2112 if (r->ready_to_run) {
2113 if (r->ready_to_run > tprev)
2114 pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2116 r->dt_delay = tprev - r->ready_to_run;
2119 if (r->last_time > tprev)
2120 pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2121 else if (r->last_time) {
2122 u64 dt_wait = tprev - r->last_time;
2124 if (r->last_state == TASK_RUNNING)
2125 r->dt_preempt = dt_wait;
2126 else if (r->last_state == TASK_UNINTERRUPTIBLE)
2127 r->dt_iowait = dt_wait;
2129 r->dt_sleep = dt_wait;
2133 update_stats(&r->run_stats, r->dt_run);
2135 r->total_run_time += r->dt_run;
2136 r->total_delay_time += r->dt_delay;
2137 r->total_sleep_time += r->dt_sleep;
2138 r->total_iowait_time += r->dt_iowait;
2139 r->total_preempt_time += r->dt_preempt;
2142 static bool is_idle_sample(struct perf_sample *sample,
2143 struct evsel *evsel)
2145 /* pid 0 == swapper == idle task */
2146 if (strcmp(evsel__name(evsel), "sched:sched_switch") == 0)
2147 return evsel__intval(evsel, sample, "prev_pid") == 0;
2149 return sample->pid == 0;
2152 static void save_task_callchain(struct perf_sched *sched,
2153 struct perf_sample *sample,
2154 struct evsel *evsel,
2155 struct machine *machine)
2157 struct callchain_cursor *cursor = &callchain_cursor;
2158 struct thread *thread;
2160 /* want main thread for process - has maps */
2161 thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2162 if (thread == NULL) {
2163 pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2167 if (!sched->show_callchain || sample->callchain == NULL)
2170 if (thread__resolve_callchain(thread, cursor, evsel, sample,
2171 NULL, NULL, sched->max_stack + 2) != 0) {
2173 pr_err("Failed to resolve callchain. Skipping\n");
2178 callchain_cursor_commit(cursor);
2181 struct callchain_cursor_node *node;
2184 node = callchain_cursor_current(cursor);
2190 if (!strcmp(sym->name, "schedule") ||
2191 !strcmp(sym->name, "__schedule") ||
2192 !strcmp(sym->name, "preempt_schedule"))
2196 callchain_cursor_advance(cursor);
2200 static int init_idle_thread(struct thread *thread)
2202 struct idle_thread_runtime *itr;
2204 thread__set_comm(thread, idle_comm, 0);
2206 itr = zalloc(sizeof(*itr));
2210 init_stats(&itr->tr.run_stats);
2211 callchain_init(&itr->callchain);
2212 callchain_cursor_reset(&itr->cursor);
2213 thread__set_priv(thread, itr);
2219 * Track idle stats per cpu by maintaining a local thread
2220 * struct for the idle task on each cpu.
2222 static int init_idle_threads(int ncpu)
2226 idle_threads = zalloc(ncpu * sizeof(struct thread *));
2230 idle_max_cpu = ncpu;
2232 /* allocate the actual thread struct if needed */
2233 for (i = 0; i < ncpu; ++i) {
2234 idle_threads[i] = thread__new(0, 0);
2235 if (idle_threads[i] == NULL)
2238 ret = init_idle_thread(idle_threads[i]);
2246 static void free_idle_threads(void)
2250 if (idle_threads == NULL)
2253 for (i = 0; i < idle_max_cpu; ++i) {
2254 if ((idle_threads[i]))
2255 thread__delete(idle_threads[i]);
2261 static struct thread *get_idle_thread(int cpu)
2264 * expand/allocate array of pointers to local thread
2267 if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2268 int i, j = __roundup_pow_of_two(cpu+1);
2271 p = realloc(idle_threads, j * sizeof(struct thread *));
2275 idle_threads = (struct thread **) p;
2276 for (i = idle_max_cpu; i < j; ++i)
2277 idle_threads[i] = NULL;
2282 /* allocate a new thread struct if needed */
2283 if (idle_threads[cpu] == NULL) {
2284 idle_threads[cpu] = thread__new(0, 0);
2286 if (idle_threads[cpu]) {
2287 if (init_idle_thread(idle_threads[cpu]) < 0)
2292 return idle_threads[cpu];
2295 static void save_idle_callchain(struct perf_sched *sched,
2296 struct idle_thread_runtime *itr,
2297 struct perf_sample *sample)
2299 if (!sched->show_callchain || sample->callchain == NULL)
2302 callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2305 static struct thread *timehist_get_thread(struct perf_sched *sched,
2306 struct perf_sample *sample,
2307 struct machine *machine,
2308 struct evsel *evsel)
2310 struct thread *thread;
2312 if (is_idle_sample(sample, evsel)) {
2313 thread = get_idle_thread(sample->cpu);
2315 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2318 /* there were samples with tid 0 but non-zero pid */
2319 thread = machine__findnew_thread(machine, sample->pid,
2320 sample->tid ?: sample->pid);
2321 if (thread == NULL) {
2322 pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2326 save_task_callchain(sched, sample, evsel, machine);
2327 if (sched->idle_hist) {
2328 struct thread *idle;
2329 struct idle_thread_runtime *itr;
2331 idle = get_idle_thread(sample->cpu);
2333 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2337 itr = thread__priv(idle);
2341 itr->last_thread = thread;
2343 /* copy task callchain when entering to idle */
2344 if (evsel__intval(evsel, sample, "next_pid") == 0)
2345 save_idle_callchain(sched, itr, sample);
2352 static bool timehist_skip_sample(struct perf_sched *sched,
2353 struct thread *thread,
2354 struct evsel *evsel,
2355 struct perf_sample *sample)
2359 if (thread__is_filtered(thread)) {
2361 sched->skipped_samples++;
2364 if (sched->idle_hist) {
2365 if (strcmp(evsel__name(evsel), "sched:sched_switch"))
2367 else if (evsel__intval(evsel, sample, "prev_pid") != 0 &&
2368 evsel__intval(evsel, sample, "next_pid") != 0)
2375 static void timehist_print_wakeup_event(struct perf_sched *sched,
2376 struct evsel *evsel,
2377 struct perf_sample *sample,
2378 struct machine *machine,
2379 struct thread *awakened)
2381 struct thread *thread;
2384 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2388 /* show wakeup unless both awakee and awaker are filtered */
2389 if (timehist_skip_sample(sched, thread, evsel, sample) &&
2390 timehist_skip_sample(sched, awakened, evsel, sample)) {
2394 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2395 printf("%15s [%04d] ", tstr, sample->cpu);
2396 if (sched->show_cpu_visual)
2397 printf(" %*s ", sched->max_cpu.cpu + 1, "");
2399 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2402 printf(" %9s %9s %9s ", "", "", "");
2404 printf("awakened: %s", timehist_get_commstr(awakened));
2409 static int timehist_sched_wakeup_ignore(struct perf_tool *tool __maybe_unused,
2410 union perf_event *event __maybe_unused,
2411 struct evsel *evsel __maybe_unused,
2412 struct perf_sample *sample __maybe_unused,
2413 struct machine *machine __maybe_unused)
2418 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2419 union perf_event *event __maybe_unused,
2420 struct evsel *evsel,
2421 struct perf_sample *sample,
2422 struct machine *machine)
2424 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2425 struct thread *thread;
2426 struct thread_runtime *tr = NULL;
2427 /* want pid of awakened task not pid in sample */
2428 const u32 pid = evsel__intval(evsel, sample, "pid");
2430 thread = machine__findnew_thread(machine, 0, pid);
2434 tr = thread__get_runtime(thread);
2438 if (tr->ready_to_run == 0)
2439 tr->ready_to_run = sample->time;
2441 /* show wakeups if requested */
2442 if (sched->show_wakeups &&
2443 !perf_time__skip_sample(&sched->ptime, sample->time))
2444 timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2449 static void timehist_print_migration_event(struct perf_sched *sched,
2450 struct evsel *evsel,
2451 struct perf_sample *sample,
2452 struct machine *machine,
2453 struct thread *migrated)
2455 struct thread *thread;
2460 if (sched->summary_only)
2463 max_cpus = sched->max_cpu.cpu + 1;
2464 ocpu = evsel__intval(evsel, sample, "orig_cpu");
2465 dcpu = evsel__intval(evsel, sample, "dest_cpu");
2467 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2471 if (timehist_skip_sample(sched, thread, evsel, sample) &&
2472 timehist_skip_sample(sched, migrated, evsel, sample)) {
2476 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2477 printf("%15s [%04d] ", tstr, sample->cpu);
2479 if (sched->show_cpu_visual) {
2484 for (i = 0; i < max_cpus; ++i) {
2485 c = (i == sample->cpu) ? 'm' : ' ';
2491 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2494 printf(" %9s %9s %9s ", "", "", "");
2496 printf("migrated: %s", timehist_get_commstr(migrated));
2497 printf(" cpu %d => %d", ocpu, dcpu);
2502 static int timehist_migrate_task_event(struct perf_tool *tool,
2503 union perf_event *event __maybe_unused,
2504 struct evsel *evsel,
2505 struct perf_sample *sample,
2506 struct machine *machine)
2508 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2509 struct thread *thread;
2510 struct thread_runtime *tr = NULL;
2511 /* want pid of migrated task not pid in sample */
2512 const u32 pid = evsel__intval(evsel, sample, "pid");
2514 thread = machine__findnew_thread(machine, 0, pid);
2518 tr = thread__get_runtime(thread);
2524 /* show migrations if requested */
2525 timehist_print_migration_event(sched, evsel, sample, machine, thread);
2530 static int timehist_sched_change_event(struct perf_tool *tool,
2531 union perf_event *event,
2532 struct evsel *evsel,
2533 struct perf_sample *sample,
2534 struct machine *machine)
2536 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2537 struct perf_time_interval *ptime = &sched->ptime;
2538 struct addr_location al;
2539 struct thread *thread;
2540 struct thread_runtime *tr = NULL;
2541 u64 tprev, t = sample->time;
2543 int state = evsel__intval(evsel, sample, "prev_state");
2545 if (machine__resolve(machine, &al, sample) < 0) {
2546 pr_err("problem processing %d event. skipping it\n",
2547 event->header.type);
2552 thread = timehist_get_thread(sched, sample, machine, evsel);
2553 if (thread == NULL) {
2558 if (timehist_skip_sample(sched, thread, evsel, sample))
2561 tr = thread__get_runtime(thread);
2567 tprev = evsel__get_time(evsel, sample->cpu);
2570 * If start time given:
2571 * - sample time is under window user cares about - skip sample
2572 * - tprev is under window user cares about - reset to start of window
2574 if (ptime->start && ptime->start > t)
2577 if (tprev && ptime->start > tprev)
2578 tprev = ptime->start;
2581 * If end time given:
2582 * - previous sched event is out of window - we are done
2583 * - sample time is beyond window user cares about - reset it
2584 * to close out stats for time window interest
2587 if (tprev > ptime->end)
2594 if (!sched->idle_hist || thread->tid == 0) {
2595 if (!cpu_list || test_bit(sample->cpu, cpu_bitmap))
2596 timehist_update_runtime_stats(tr, t, tprev);
2598 if (sched->idle_hist) {
2599 struct idle_thread_runtime *itr = (void *)tr;
2600 struct thread_runtime *last_tr;
2602 BUG_ON(thread->tid != 0);
2604 if (itr->last_thread == NULL)
2607 /* add current idle time as last thread's runtime */
2608 last_tr = thread__get_runtime(itr->last_thread);
2609 if (last_tr == NULL)
2612 timehist_update_runtime_stats(last_tr, t, tprev);
2614 * remove delta time of last thread as it's not updated
2615 * and otherwise it will show an invalid value next
2616 * time. we only care total run time and run stat.
2618 last_tr->dt_run = 0;
2619 last_tr->dt_delay = 0;
2620 last_tr->dt_sleep = 0;
2621 last_tr->dt_iowait = 0;
2622 last_tr->dt_preempt = 0;
2625 callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2627 itr->last_thread = NULL;
2631 if (!sched->summary_only)
2632 timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2635 if (sched->hist_time.start == 0 && t >= ptime->start)
2636 sched->hist_time.start = t;
2637 if (ptime->end == 0 || t <= ptime->end)
2638 sched->hist_time.end = t;
2641 /* time of this sched_switch event becomes last time task seen */
2642 tr->last_time = sample->time;
2644 /* last state is used to determine where to account wait time */
2645 tr->last_state = state;
2647 /* sched out event for task so reset ready to run time */
2648 tr->ready_to_run = 0;
2651 evsel__save_time(evsel, sample->time, sample->cpu);
2656 static int timehist_sched_switch_event(struct perf_tool *tool,
2657 union perf_event *event,
2658 struct evsel *evsel,
2659 struct perf_sample *sample,
2660 struct machine *machine __maybe_unused)
2662 return timehist_sched_change_event(tool, event, evsel, sample, machine);
2665 static int process_lost(struct perf_tool *tool __maybe_unused,
2666 union perf_event *event,
2667 struct perf_sample *sample,
2668 struct machine *machine __maybe_unused)
2672 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2673 printf("%15s ", tstr);
2674 printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2680 static void print_thread_runtime(struct thread *t,
2681 struct thread_runtime *r)
2683 double mean = avg_stats(&r->run_stats);
2686 printf("%*s %5d %9" PRIu64 " ",
2687 comm_width, timehist_get_commstr(t), t->ppid,
2688 (u64) r->run_stats.n);
2690 print_sched_time(r->total_run_time, 8);
2691 stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2692 print_sched_time(r->run_stats.min, 6);
2694 print_sched_time((u64) mean, 6);
2696 print_sched_time(r->run_stats.max, 6);
2698 printf("%5.2f", stddev);
2699 printf(" %5" PRIu64, r->migrations);
2703 static void print_thread_waittime(struct thread *t,
2704 struct thread_runtime *r)
2706 printf("%*s %5d %9" PRIu64 " ",
2707 comm_width, timehist_get_commstr(t), t->ppid,
2708 (u64) r->run_stats.n);
2710 print_sched_time(r->total_run_time, 8);
2711 print_sched_time(r->total_sleep_time, 6);
2713 print_sched_time(r->total_iowait_time, 6);
2715 print_sched_time(r->total_preempt_time, 6);
2717 print_sched_time(r->total_delay_time, 6);
2721 struct total_run_stats {
2722 struct perf_sched *sched;
2728 static int __show_thread_runtime(struct thread *t, void *priv)
2730 struct total_run_stats *stats = priv;
2731 struct thread_runtime *r;
2733 if (thread__is_filtered(t))
2736 r = thread__priv(t);
2737 if (r && r->run_stats.n) {
2738 stats->task_count++;
2739 stats->sched_count += r->run_stats.n;
2740 stats->total_run_time += r->total_run_time;
2742 if (stats->sched->show_state)
2743 print_thread_waittime(t, r);
2745 print_thread_runtime(t, r);
2751 static int show_thread_runtime(struct thread *t, void *priv)
2756 return __show_thread_runtime(t, priv);
2759 static int show_deadthread_runtime(struct thread *t, void *priv)
2764 return __show_thread_runtime(t, priv);
2767 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2769 const char *sep = " <- ";
2770 struct callchain_list *chain;
2778 ret = callchain__fprintf_folded(fp, node->parent);
2781 list_for_each_entry(chain, &node->val, list) {
2782 if (chain->ip >= PERF_CONTEXT_MAX)
2784 if (chain->ms.sym && chain->ms.sym->ignore)
2786 ret += fprintf(fp, "%s%s", first ? "" : sep,
2787 callchain_list__sym_name(chain, bf, sizeof(bf),
2795 static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2799 struct callchain_node *chain;
2800 struct rb_node *rb_node = rb_first_cached(root);
2802 printf(" %16s %8s %s\n", "Idle time (msec)", "Count", "Callchains");
2803 printf(" %.16s %.8s %.50s\n", graph_dotted_line, graph_dotted_line,
2807 chain = rb_entry(rb_node, struct callchain_node, rb_node);
2808 rb_node = rb_next(rb_node);
2810 ret += fprintf(fp, " ");
2811 print_sched_time(chain->hit, 12);
2812 ret += 16; /* print_sched_time returns 2nd arg + 4 */
2813 ret += fprintf(fp, " %8d ", chain->count);
2814 ret += callchain__fprintf_folded(fp, chain);
2815 ret += fprintf(fp, "\n");
2821 static void timehist_print_summary(struct perf_sched *sched,
2822 struct perf_session *session)
2824 struct machine *m = &session->machines.host;
2825 struct total_run_stats totals;
2828 struct thread_runtime *r;
2830 u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2832 memset(&totals, 0, sizeof(totals));
2833 totals.sched = sched;
2835 if (sched->idle_hist) {
2836 printf("\nIdle-time summary\n");
2837 printf("%*s parent sched-out ", comm_width, "comm");
2838 printf(" idle-time min-idle avg-idle max-idle stddev migrations\n");
2839 } else if (sched->show_state) {
2840 printf("\nWait-time summary\n");
2841 printf("%*s parent sched-in ", comm_width, "comm");
2842 printf(" run-time sleep iowait preempt delay\n");
2844 printf("\nRuntime summary\n");
2845 printf("%*s parent sched-in ", comm_width, "comm");
2846 printf(" run-time min-run avg-run max-run stddev migrations\n");
2848 printf("%*s (count) ", comm_width, "");
2849 printf(" (msec) (msec) (msec) (msec) %s\n",
2850 sched->show_state ? "(msec)" : "%");
2851 printf("%.117s\n", graph_dotted_line);
2853 machine__for_each_thread(m, show_thread_runtime, &totals);
2854 task_count = totals.task_count;
2856 printf("<no still running tasks>\n");
2858 printf("\nTerminated tasks:\n");
2859 machine__for_each_thread(m, show_deadthread_runtime, &totals);
2860 if (task_count == totals.task_count)
2861 printf("<no terminated tasks>\n");
2863 /* CPU idle stats not tracked when samples were skipped */
2864 if (sched->skipped_samples && !sched->idle_hist)
2867 printf("\nIdle stats:\n");
2868 for (i = 0; i < idle_max_cpu; ++i) {
2869 if (cpu_list && !test_bit(i, cpu_bitmap))
2872 t = idle_threads[i];
2876 r = thread__priv(t);
2877 if (r && r->run_stats.n) {
2878 totals.sched_count += r->run_stats.n;
2879 printf(" CPU %2d idle for ", i);
2880 print_sched_time(r->total_run_time, 6);
2881 printf(" msec (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2883 printf(" CPU %2d idle entire time window\n", i);
2886 if (sched->idle_hist && sched->show_callchain) {
2887 callchain_param.mode = CHAIN_FOLDED;
2888 callchain_param.value = CCVAL_PERIOD;
2890 callchain_register_param(&callchain_param);
2892 printf("\nIdle stats by callchain:\n");
2893 for (i = 0; i < idle_max_cpu; ++i) {
2894 struct idle_thread_runtime *itr;
2896 t = idle_threads[i];
2900 itr = thread__priv(t);
2904 callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
2905 0, &callchain_param);
2907 printf(" CPU %2d:", i);
2908 print_sched_time(itr->tr.total_run_time, 6);
2910 timehist_print_idlehist_callchain(&itr->sorted_root);
2916 " Total number of unique tasks: %" PRIu64 "\n"
2917 "Total number of context switches: %" PRIu64 "\n",
2918 totals.task_count, totals.sched_count);
2920 printf(" Total run time (msec): ");
2921 print_sched_time(totals.total_run_time, 2);
2924 printf(" Total scheduling time (msec): ");
2925 print_sched_time(hist_time, 2);
2926 printf(" (x %d)\n", sched->max_cpu.cpu);
2929 typedef int (*sched_handler)(struct perf_tool *tool,
2930 union perf_event *event,
2931 struct evsel *evsel,
2932 struct perf_sample *sample,
2933 struct machine *machine);
2935 static int perf_timehist__process_sample(struct perf_tool *tool,
2936 union perf_event *event,
2937 struct perf_sample *sample,
2938 struct evsel *evsel,
2939 struct machine *machine)
2941 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2943 struct perf_cpu this_cpu = {
2947 if (this_cpu.cpu > sched->max_cpu.cpu)
2948 sched->max_cpu = this_cpu;
2950 if (evsel->handler != NULL) {
2951 sched_handler f = evsel->handler;
2953 err = f(tool, event, evsel, sample, machine);
2959 static int timehist_check_attr(struct perf_sched *sched,
2960 struct evlist *evlist)
2962 struct evsel *evsel;
2963 struct evsel_runtime *er;
2965 list_for_each_entry(evsel, &evlist->core.entries, core.node) {
2966 er = evsel__get_runtime(evsel);
2968 pr_err("Failed to allocate memory for evsel runtime data\n");
2972 if (sched->show_callchain && !evsel__has_callchain(evsel)) {
2973 pr_info("Samples do not have callchains.\n");
2974 sched->show_callchain = 0;
2975 symbol_conf.use_callchain = 0;
2982 static int perf_sched__timehist(struct perf_sched *sched)
2984 struct evsel_str_handler handlers[] = {
2985 { "sched:sched_switch", timehist_sched_switch_event, },
2986 { "sched:sched_wakeup", timehist_sched_wakeup_event, },
2987 { "sched:sched_waking", timehist_sched_wakeup_event, },
2988 { "sched:sched_wakeup_new", timehist_sched_wakeup_event, },
2990 const struct evsel_str_handler migrate_handlers[] = {
2991 { "sched:sched_migrate_task", timehist_migrate_task_event, },
2993 struct perf_data data = {
2995 .mode = PERF_DATA_MODE_READ,
2996 .force = sched->force,
2999 struct perf_session *session;
3000 struct evlist *evlist;
3004 * event handlers for timehist option
3006 sched->tool.sample = perf_timehist__process_sample;
3007 sched->tool.mmap = perf_event__process_mmap;
3008 sched->tool.comm = perf_event__process_comm;
3009 sched->tool.exit = perf_event__process_exit;
3010 sched->tool.fork = perf_event__process_fork;
3011 sched->tool.lost = process_lost;
3012 sched->tool.attr = perf_event__process_attr;
3013 sched->tool.tracing_data = perf_event__process_tracing_data;
3014 sched->tool.build_id = perf_event__process_build_id;
3016 sched->tool.ordered_events = true;
3017 sched->tool.ordering_requires_timestamps = true;
3019 symbol_conf.use_callchain = sched->show_callchain;
3021 session = perf_session__new(&data, &sched->tool);
3022 if (IS_ERR(session))
3023 return PTR_ERR(session);
3026 err = perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap);
3031 evlist = session->evlist;
3033 symbol__init(&session->header.env);
3035 if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
3036 pr_err("Invalid time string\n");
3040 if (timehist_check_attr(sched, evlist) != 0)
3045 /* prefer sched_waking if it is captured */
3046 if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
3047 handlers[1].handler = timehist_sched_wakeup_ignore;
3049 /* setup per-evsel handlers */
3050 if (perf_session__set_tracepoints_handlers(session, handlers))
3053 /* sched_switch event at a minimum needs to exist */
3054 if (!evlist__find_tracepoint_by_name(session->evlist, "sched:sched_switch")) {
3055 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3059 if (sched->show_migrations &&
3060 perf_session__set_tracepoints_handlers(session, migrate_handlers))
3063 /* pre-allocate struct for per-CPU idle stats */
3064 sched->max_cpu.cpu = session->header.env.nr_cpus_online;
3065 if (sched->max_cpu.cpu == 0)
3066 sched->max_cpu.cpu = 4;
3067 if (init_idle_threads(sched->max_cpu.cpu))
3070 /* summary_only implies summary option, but don't overwrite summary if set */
3071 if (sched->summary_only)
3072 sched->summary = sched->summary_only;
3074 if (!sched->summary_only)
3075 timehist_header(sched);
3077 err = perf_session__process_events(session);
3079 pr_err("Failed to process events, error %d", err);
3083 sched->nr_events = evlist->stats.nr_events[0];
3084 sched->nr_lost_events = evlist->stats.total_lost;
3085 sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3088 timehist_print_summary(sched, session);
3091 free_idle_threads();
3092 perf_session__delete(session);
3098 static void print_bad_events(struct perf_sched *sched)
3100 if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3101 printf(" INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3102 (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3103 sched->nr_unordered_timestamps, sched->nr_timestamps);
3105 if (sched->nr_lost_events && sched->nr_events) {
3106 printf(" INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3107 (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3108 sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3110 if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3111 printf(" INFO: %.3f%% context switch bugs (%ld out of %ld)",
3112 (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3113 sched->nr_context_switch_bugs, sched->nr_timestamps);
3114 if (sched->nr_lost_events)
3115 printf(" (due to lost events?)");
3120 static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3122 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3123 struct work_atoms *this;
3124 const char *comm = thread__comm_str(data->thread), *this_comm;
3125 bool leftmost = true;
3130 this = container_of(*new, struct work_atoms, node);
3133 this_comm = thread__comm_str(this->thread);
3134 cmp = strcmp(comm, this_comm);
3136 new = &((*new)->rb_left);
3137 } else if (cmp < 0) {
3138 new = &((*new)->rb_right);
3142 this->total_runtime += data->total_runtime;
3143 this->nb_atoms += data->nb_atoms;
3144 this->total_lat += data->total_lat;
3145 list_splice(&data->work_list, &this->work_list);
3146 if (this->max_lat < data->max_lat) {
3147 this->max_lat = data->max_lat;
3148 this->max_lat_start = data->max_lat_start;
3149 this->max_lat_end = data->max_lat_end;
3157 rb_link_node(&data->node, parent, new);
3158 rb_insert_color_cached(&data->node, root, leftmost);
3161 static void perf_sched__merge_lat(struct perf_sched *sched)
3163 struct work_atoms *data;
3164 struct rb_node *node;
3166 if (sched->skip_merge)
3169 while ((node = rb_first_cached(&sched->atom_root))) {
3170 rb_erase_cached(node, &sched->atom_root);
3171 data = rb_entry(node, struct work_atoms, node);
3172 __merge_work_atoms(&sched->merged_atom_root, data);
3176 static int perf_sched__lat(struct perf_sched *sched)
3178 struct rb_node *next;
3182 if (perf_sched__read_events(sched))
3185 perf_sched__merge_lat(sched);
3186 perf_sched__sort_lat(sched);
3188 printf("\n -------------------------------------------------------------------------------------------------------------------------------------------\n");
3189 printf(" Task | Runtime ms | Switches | Avg delay ms | Max delay ms | Max delay start | Max delay end |\n");
3190 printf(" -------------------------------------------------------------------------------------------------------------------------------------------\n");
3192 next = rb_first_cached(&sched->sorted_atom_root);
3195 struct work_atoms *work_list;
3197 work_list = rb_entry(next, struct work_atoms, node);
3198 output_lat_thread(sched, work_list);
3199 next = rb_next(next);
3200 thread__zput(work_list->thread);
3203 printf(" -----------------------------------------------------------------------------------------------------------------\n");
3204 printf(" TOTAL: |%11.3f ms |%9" PRIu64 " |\n",
3205 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3207 printf(" ---------------------------------------------------\n");
3209 print_bad_events(sched);
3215 static int setup_map_cpus(struct perf_sched *sched)
3217 struct perf_cpu_map *map;
3219 sched->max_cpu.cpu = sysconf(_SC_NPROCESSORS_CONF);
3221 if (sched->map.comp) {
3222 sched->map.comp_cpus = zalloc(sched->max_cpu.cpu * sizeof(int));
3223 if (!sched->map.comp_cpus)
3227 if (!sched->map.cpus_str)
3230 map = perf_cpu_map__new(sched->map.cpus_str);
3232 pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3236 sched->map.cpus = map;
3240 static int setup_color_pids(struct perf_sched *sched)
3242 struct perf_thread_map *map;
3244 if (!sched->map.color_pids_str)
3247 map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3249 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3253 sched->map.color_pids = map;
3257 static int setup_color_cpus(struct perf_sched *sched)
3259 struct perf_cpu_map *map;
3261 if (!sched->map.color_cpus_str)
3264 map = perf_cpu_map__new(sched->map.color_cpus_str);
3266 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3270 sched->map.color_cpus = map;
3274 static int perf_sched__map(struct perf_sched *sched)
3276 if (setup_map_cpus(sched))
3279 if (setup_color_pids(sched))
3282 if (setup_color_cpus(sched))
3286 if (perf_sched__read_events(sched))
3288 print_bad_events(sched);
3292 static int perf_sched__replay(struct perf_sched *sched)
3296 calibrate_run_measurement_overhead(sched);
3297 calibrate_sleep_measurement_overhead(sched);
3299 test_calibrations(sched);
3301 if (perf_sched__read_events(sched))
3304 printf("nr_run_events: %ld\n", sched->nr_run_events);
3305 printf("nr_sleep_events: %ld\n", sched->nr_sleep_events);
3306 printf("nr_wakeup_events: %ld\n", sched->nr_wakeup_events);
3308 if (sched->targetless_wakeups)
3309 printf("target-less wakeups: %ld\n", sched->targetless_wakeups);
3310 if (sched->multitarget_wakeups)
3311 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3312 if (sched->nr_run_events_optimized)
3313 printf("run atoms optimized: %ld\n",
3314 sched->nr_run_events_optimized);
3316 print_task_traces(sched);
3317 add_cross_task_wakeups(sched);
3319 create_tasks(sched);
3320 printf("------------------------------------------------------------\n");
3321 for (i = 0; i < sched->replay_repeat; i++)
3322 run_one_test(sched);
3327 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3328 const char * const usage_msg[])
3330 char *tmp, *tok, *str = strdup(sched->sort_order);
3332 for (tok = strtok_r(str, ", ", &tmp);
3333 tok; tok = strtok_r(NULL, ", ", &tmp)) {
3334 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3335 usage_with_options_msg(usage_msg, options,
3336 "Unknown --sort key: `%s'", tok);
3342 sort_dimension__add("pid", &sched->cmp_pid);
3345 static bool schedstat_events_exposed(void)
3348 * Select "sched:sched_stat_wait" event to check
3349 * whether schedstat tracepoints are exposed.
3351 return IS_ERR(trace_event__tp_format("sched", "sched_stat_wait")) ?
3355 static int __cmd_record(int argc, const char **argv)
3357 unsigned int rec_argc, i, j;
3358 const char **rec_argv;
3359 const char * const record_args[] = {
3365 "-e", "sched:sched_switch",
3366 "-e", "sched:sched_stat_runtime",
3367 "-e", "sched:sched_process_fork",
3368 "-e", "sched:sched_wakeup_new",
3369 "-e", "sched:sched_migrate_task",
3373 * The tracepoints trace_sched_stat_{wait, sleep, iowait}
3374 * are not exposed to user if CONFIG_SCHEDSTATS is not set,
3375 * to prevent "perf sched record" execution failure, determine
3376 * whether to record schedstat events according to actual situation.
3378 const char * const schedstat_args[] = {
3379 "-e", "sched:sched_stat_wait",
3380 "-e", "sched:sched_stat_sleep",
3381 "-e", "sched:sched_stat_iowait",
3383 unsigned int schedstat_argc = schedstat_events_exposed() ?
3384 ARRAY_SIZE(schedstat_args) : 0;
3386 struct tep_event *waking_event;
3389 * +2 for either "-e", "sched:sched_wakeup" or
3390 * "-e", "sched:sched_waking"
3392 rec_argc = ARRAY_SIZE(record_args) + 2 + schedstat_argc + argc - 1;
3393 rec_argv = calloc(rec_argc + 1, sizeof(char *));
3395 if (rec_argv == NULL)
3398 for (i = 0; i < ARRAY_SIZE(record_args); i++)
3399 rec_argv[i] = strdup(record_args[i]);
3401 rec_argv[i++] = "-e";
3402 waking_event = trace_event__tp_format("sched", "sched_waking");
3403 if (!IS_ERR(waking_event))
3404 rec_argv[i++] = strdup("sched:sched_waking");
3406 rec_argv[i++] = strdup("sched:sched_wakeup");
3408 for (j = 0; j < schedstat_argc; j++)
3409 rec_argv[i++] = strdup(schedstat_args[j]);
3411 for (j = 1; j < (unsigned int)argc; j++, i++)
3412 rec_argv[i] = argv[j];
3414 BUG_ON(i != rec_argc);
3416 return cmd_record(i, rec_argv);
3419 int cmd_sched(int argc, const char **argv)
3421 static const char default_sort_order[] = "avg, max, switch, runtime";
3422 struct perf_sched sched = {
3424 .sample = perf_sched__process_tracepoint_sample,
3425 .comm = perf_sched__process_comm,
3426 .namespaces = perf_event__process_namespaces,
3427 .lost = perf_event__process_lost,
3428 .fork = perf_sched__process_fork_event,
3429 .ordered_events = true,
3431 .cmp_pid = LIST_HEAD_INIT(sched.cmp_pid),
3432 .sort_list = LIST_HEAD_INIT(sched.sort_list),
3433 .start_work_mutex = PTHREAD_MUTEX_INITIALIZER,
3434 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
3435 .sort_order = default_sort_order,
3436 .replay_repeat = 10,
3438 .next_shortname1 = 'A',
3439 .next_shortname2 = '0',
3441 .show_callchain = 1,
3444 const struct option sched_options[] = {
3445 OPT_STRING('i', "input", &input_name, "file",
3447 OPT_INCR('v', "verbose", &verbose,
3448 "be more verbose (show symbol address, etc)"),
3449 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3450 "dump raw trace in ASCII"),
3451 OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3454 const struct option latency_options[] = {
3455 OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3456 "sort by key(s): runtime, switch, avg, max"),
3457 OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3458 "CPU to profile on"),
3459 OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3460 "latency stats per pid instead of per comm"),
3461 OPT_PARENT(sched_options)
3463 const struct option replay_options[] = {
3464 OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3465 "repeat the workload replay N times (-1: infinite)"),
3466 OPT_PARENT(sched_options)
3468 const struct option map_options[] = {
3469 OPT_BOOLEAN(0, "compact", &sched.map.comp,
3470 "map output in compact mode"),
3471 OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3472 "highlight given pids in map"),
3473 OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3474 "highlight given CPUs in map"),
3475 OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3476 "display given CPUs in map"),
3477 OPT_PARENT(sched_options)
3479 const struct option timehist_options[] = {
3480 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3481 "file", "vmlinux pathname"),
3482 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3483 "file", "kallsyms pathname"),
3484 OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3485 "Display call chains if present (default on)"),
3486 OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3487 "Maximum number of functions to display backtrace."),
3488 OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3489 "Look for files with symbols relative to this directory"),
3490 OPT_BOOLEAN('s', "summary", &sched.summary_only,
3491 "Show only syscall summary with statistics"),
3492 OPT_BOOLEAN('S', "with-summary", &sched.summary,
3493 "Show all syscalls and summary with statistics"),
3494 OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3495 OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3496 OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3497 OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3498 OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3499 OPT_STRING(0, "time", &sched.time_str, "str",
3500 "Time span for analysis (start,stop)"),
3501 OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3502 OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3503 "analyze events only for given process id(s)"),
3504 OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3505 "analyze events only for given thread id(s)"),
3506 OPT_STRING('C', "cpu", &cpu_list, "cpu", "list of cpus to profile"),
3507 OPT_PARENT(sched_options)
3510 const char * const latency_usage[] = {
3511 "perf sched latency [<options>]",
3514 const char * const replay_usage[] = {
3515 "perf sched replay [<options>]",
3518 const char * const map_usage[] = {
3519 "perf sched map [<options>]",
3522 const char * const timehist_usage[] = {
3523 "perf sched timehist [<options>]",
3526 const char *const sched_subcommands[] = { "record", "latency", "map",
3529 const char *sched_usage[] = {
3533 struct trace_sched_handler lat_ops = {
3534 .wakeup_event = latency_wakeup_event,
3535 .switch_event = latency_switch_event,
3536 .runtime_event = latency_runtime_event,
3537 .migrate_task_event = latency_migrate_task_event,
3539 struct trace_sched_handler map_ops = {
3540 .switch_event = map_switch_event,
3542 struct trace_sched_handler replay_ops = {
3543 .wakeup_event = replay_wakeup_event,
3544 .switch_event = replay_switch_event,
3545 .fork_event = replay_fork_event,
3550 for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3551 sched.curr_pid[i] = -1;
3553 argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3554 sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3556 usage_with_options(sched_usage, sched_options);
3559 * Aliased to 'perf script' for now:
3561 if (!strcmp(argv[0], "script"))
3562 return cmd_script(argc, argv);
3564 if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
3565 return __cmd_record(argc, argv);
3566 } else if (strlen(argv[0]) > 2 && strstarts("latency", argv[0])) {
3567 sched.tp_handler = &lat_ops;
3569 argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3571 usage_with_options(latency_usage, latency_options);
3573 setup_sorting(&sched, latency_options, latency_usage);
3574 return perf_sched__lat(&sched);
3575 } else if (!strcmp(argv[0], "map")) {
3577 argc = parse_options(argc, argv, map_options, map_usage, 0);
3579 usage_with_options(map_usage, map_options);
3581 sched.tp_handler = &map_ops;
3582 setup_sorting(&sched, latency_options, latency_usage);
3583 return perf_sched__map(&sched);
3584 } else if (strlen(argv[0]) > 2 && strstarts("replay", argv[0])) {
3585 sched.tp_handler = &replay_ops;
3587 argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3589 usage_with_options(replay_usage, replay_options);
3591 return perf_sched__replay(&sched);
3592 } else if (!strcmp(argv[0], "timehist")) {
3594 argc = parse_options(argc, argv, timehist_options,
3597 usage_with_options(timehist_usage, timehist_options);
3599 if ((sched.show_wakeups || sched.show_next) &&
3600 sched.summary_only) {
3601 pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3602 parse_options_usage(timehist_usage, timehist_options, "s", true);
3603 if (sched.show_wakeups)
3604 parse_options_usage(NULL, timehist_options, "w", true);
3605 if (sched.show_next)
3606 parse_options_usage(NULL, timehist_options, "n", true);
3609 ret = symbol__validate_sym_arguments();
3613 return perf_sched__timehist(&sched);
3615 usage_with_options(sched_usage, sched_options);