5 #include "util/evlist.h"
6 #include "util/cache.h"
7 #include "util/evsel.h"
8 #include "util/symbol.h"
9 #include "util/thread.h"
10 #include "util/header.h"
11 #include "util/session.h"
12 #include "util/tool.h"
13 #include "util/cloexec.h"
15 #include "util/parse-options.h"
16 #include "util/trace-event.h"
18 #include "util/debug.h"
20 #include <sys/prctl.h>
21 #include <sys/resource.h>
23 #include <semaphore.h>
26 #include <api/fs/fs.h>
28 #define PR_SET_NAME 15 /* Set process name */
32 #define MAX_PID 1024000
41 unsigned long nr_events;
42 unsigned long curr_event;
43 struct sched_atom **atoms;
54 enum sched_event_type {
58 SCHED_EVENT_MIGRATION,
62 enum sched_event_type type;
68 struct task_desc *wakee;
71 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
81 struct list_head list;
82 enum thread_state state;
90 struct list_head work_list;
91 struct thread *thread;
100 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
104 struct trace_sched_handler {
105 int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
106 struct perf_sample *sample, struct machine *machine);
108 int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
109 struct perf_sample *sample, struct machine *machine);
111 int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
112 struct perf_sample *sample, struct machine *machine);
114 /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
115 int (*fork_event)(struct perf_sched *sched, union perf_event *event,
116 struct machine *machine);
118 int (*migrate_task_event)(struct perf_sched *sched,
119 struct perf_evsel *evsel,
120 struct perf_sample *sample,
121 struct machine *machine);
125 struct perf_tool tool;
126 const char *sort_order;
127 unsigned long nr_tasks;
128 struct task_desc **pid_to_task;
129 struct task_desc **tasks;
130 const struct trace_sched_handler *tp_handler;
131 pthread_mutex_t start_work_mutex;
132 pthread_mutex_t work_done_wait_mutex;
135 * Track the current task - that way we can know whether there's any
136 * weird events, such as a task being switched away that is not current.
139 u32 curr_pid[MAX_CPUS];
140 struct thread *curr_thread[MAX_CPUS];
141 char next_shortname1;
142 char next_shortname2;
143 unsigned int replay_repeat;
144 unsigned long nr_run_events;
145 unsigned long nr_sleep_events;
146 unsigned long nr_wakeup_events;
147 unsigned long nr_sleep_corrections;
148 unsigned long nr_run_events_optimized;
149 unsigned long targetless_wakeups;
150 unsigned long multitarget_wakeups;
151 unsigned long nr_runs;
152 unsigned long nr_timestamps;
153 unsigned long nr_unordered_timestamps;
154 unsigned long nr_context_switch_bugs;
155 unsigned long nr_events;
156 unsigned long nr_lost_chunks;
157 unsigned long nr_lost_events;
158 u64 run_measurement_overhead;
159 u64 sleep_measurement_overhead;
162 u64 runavg_cpu_usage;
163 u64 parent_cpu_usage;
164 u64 runavg_parent_cpu_usage;
170 u64 cpu_last_switched[MAX_CPUS];
171 struct rb_root atom_root, sorted_atom_root;
172 struct list_head sort_list, cmp_pid;
176 static u64 get_nsecs(void)
180 clock_gettime(CLOCK_MONOTONIC, &ts);
182 return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
185 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
187 u64 T0 = get_nsecs(), T1;
191 } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
194 static void sleep_nsecs(u64 nsecs)
198 ts.tv_nsec = nsecs % 999999999;
199 ts.tv_sec = nsecs / 999999999;
201 nanosleep(&ts, NULL);
204 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
206 u64 T0, T1, delta, min_delta = 1000000000ULL;
209 for (i = 0; i < 10; i++) {
211 burn_nsecs(sched, 0);
214 min_delta = min(min_delta, delta);
216 sched->run_measurement_overhead = min_delta;
218 printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
221 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
223 u64 T0, T1, delta, min_delta = 1000000000ULL;
226 for (i = 0; i < 10; i++) {
231 min_delta = min(min_delta, delta);
234 sched->sleep_measurement_overhead = min_delta;
236 printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
239 static struct sched_atom *
240 get_new_event(struct task_desc *task, u64 timestamp)
242 struct sched_atom *event = zalloc(sizeof(*event));
243 unsigned long idx = task->nr_events;
246 event->timestamp = timestamp;
250 size = sizeof(struct sched_atom *) * task->nr_events;
251 task->atoms = realloc(task->atoms, size);
252 BUG_ON(!task->atoms);
254 task->atoms[idx] = event;
259 static struct sched_atom *last_event(struct task_desc *task)
261 if (!task->nr_events)
264 return task->atoms[task->nr_events - 1];
267 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
268 u64 timestamp, u64 duration)
270 struct sched_atom *event, *curr_event = last_event(task);
273 * optimize an existing RUN event by merging this one
276 if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
277 sched->nr_run_events_optimized++;
278 curr_event->duration += duration;
282 event = get_new_event(task, timestamp);
284 event->type = SCHED_EVENT_RUN;
285 event->duration = duration;
287 sched->nr_run_events++;
290 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
291 u64 timestamp, struct task_desc *wakee)
293 struct sched_atom *event, *wakee_event;
295 event = get_new_event(task, timestamp);
296 event->type = SCHED_EVENT_WAKEUP;
297 event->wakee = wakee;
299 wakee_event = last_event(wakee);
300 if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
301 sched->targetless_wakeups++;
304 if (wakee_event->wait_sem) {
305 sched->multitarget_wakeups++;
309 wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
310 sem_init(wakee_event->wait_sem, 0, 0);
311 wakee_event->specific_wait = 1;
312 event->wait_sem = wakee_event->wait_sem;
314 sched->nr_wakeup_events++;
317 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
318 u64 timestamp, u64 task_state __maybe_unused)
320 struct sched_atom *event = get_new_event(task, timestamp);
322 event->type = SCHED_EVENT_SLEEP;
324 sched->nr_sleep_events++;
327 static struct task_desc *register_pid(struct perf_sched *sched,
328 unsigned long pid, const char *comm)
330 struct task_desc *task;
333 if (sched->pid_to_task == NULL) {
334 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
336 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
338 if (pid >= (unsigned long)pid_max) {
339 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
340 sizeof(struct task_desc *))) == NULL);
341 while (pid >= (unsigned long)pid_max)
342 sched->pid_to_task[pid_max++] = NULL;
345 task = sched->pid_to_task[pid];
350 task = zalloc(sizeof(*task));
352 task->nr = sched->nr_tasks;
353 strcpy(task->comm, comm);
355 * every task starts in sleeping state - this gets ignored
356 * if there's no wakeup pointing to this sleep state:
358 add_sched_event_sleep(sched, task, 0, 0);
360 sched->pid_to_task[pid] = task;
362 sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
363 BUG_ON(!sched->tasks);
364 sched->tasks[task->nr] = task;
367 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
373 static void print_task_traces(struct perf_sched *sched)
375 struct task_desc *task;
378 for (i = 0; i < sched->nr_tasks; i++) {
379 task = sched->tasks[i];
380 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
381 task->nr, task->comm, task->pid, task->nr_events);
385 static void add_cross_task_wakeups(struct perf_sched *sched)
387 struct task_desc *task1, *task2;
390 for (i = 0; i < sched->nr_tasks; i++) {
391 task1 = sched->tasks[i];
393 if (j == sched->nr_tasks)
395 task2 = sched->tasks[j];
396 add_sched_event_wakeup(sched, task1, 0, task2);
400 static void perf_sched__process_event(struct perf_sched *sched,
401 struct sched_atom *atom)
405 switch (atom->type) {
406 case SCHED_EVENT_RUN:
407 burn_nsecs(sched, atom->duration);
409 case SCHED_EVENT_SLEEP:
411 ret = sem_wait(atom->wait_sem);
414 case SCHED_EVENT_WAKEUP:
416 ret = sem_post(atom->wait_sem);
419 case SCHED_EVENT_MIGRATION:
426 static u64 get_cpu_usage_nsec_parent(void)
432 err = getrusage(RUSAGE_SELF, &ru);
435 sum = ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
436 sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
441 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
443 struct perf_event_attr attr;
444 char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
447 bool need_privilege = false;
449 memset(&attr, 0, sizeof(attr));
451 attr.type = PERF_TYPE_SOFTWARE;
452 attr.config = PERF_COUNT_SW_TASK_CLOCK;
455 fd = sys_perf_event_open(&attr, 0, -1, -1,
456 perf_event_open_cloexec_flag());
459 if (errno == EMFILE) {
461 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
462 limit.rlim_cur += sched->nr_tasks - cur_task;
463 if (limit.rlim_cur > limit.rlim_max) {
464 limit.rlim_max = limit.rlim_cur;
465 need_privilege = true;
467 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
468 if (need_privilege && errno == EPERM)
469 strcpy(info, "Need privilege\n");
473 strcpy(info, "Have a try with -f option\n");
475 pr_err("Error: sys_perf_event_open() syscall returned "
476 "with %d (%s)\n%s", fd,
477 strerror_r(errno, sbuf, sizeof(sbuf)), info);
483 static u64 get_cpu_usage_nsec_self(int fd)
488 ret = read(fd, &runtime, sizeof(runtime));
489 BUG_ON(ret != sizeof(runtime));
494 struct sched_thread_parms {
495 struct task_desc *task;
496 struct perf_sched *sched;
500 static void *thread_func(void *ctx)
502 struct sched_thread_parms *parms = ctx;
503 struct task_desc *this_task = parms->task;
504 struct perf_sched *sched = parms->sched;
505 u64 cpu_usage_0, cpu_usage_1;
506 unsigned long i, ret;
512 sprintf(comm2, ":%s", this_task->comm);
513 prctl(PR_SET_NAME, comm2);
517 ret = sem_post(&this_task->ready_for_work);
519 ret = pthread_mutex_lock(&sched->start_work_mutex);
521 ret = pthread_mutex_unlock(&sched->start_work_mutex);
524 cpu_usage_0 = get_cpu_usage_nsec_self(fd);
526 for (i = 0; i < this_task->nr_events; i++) {
527 this_task->curr_event = i;
528 perf_sched__process_event(sched, this_task->atoms[i]);
531 cpu_usage_1 = get_cpu_usage_nsec_self(fd);
532 this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
533 ret = sem_post(&this_task->work_done_sem);
536 ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
538 ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
544 static void create_tasks(struct perf_sched *sched)
546 struct task_desc *task;
551 err = pthread_attr_init(&attr);
553 err = pthread_attr_setstacksize(&attr,
554 (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
556 err = pthread_mutex_lock(&sched->start_work_mutex);
558 err = pthread_mutex_lock(&sched->work_done_wait_mutex);
560 for (i = 0; i < sched->nr_tasks; i++) {
561 struct sched_thread_parms *parms = malloc(sizeof(*parms));
562 BUG_ON(parms == NULL);
563 parms->task = task = sched->tasks[i];
564 parms->sched = sched;
565 parms->fd = self_open_counters(sched, i);
566 sem_init(&task->sleep_sem, 0, 0);
567 sem_init(&task->ready_for_work, 0, 0);
568 sem_init(&task->work_done_sem, 0, 0);
569 task->curr_event = 0;
570 err = pthread_create(&task->thread, &attr, thread_func, parms);
575 static void wait_for_tasks(struct perf_sched *sched)
577 u64 cpu_usage_0, cpu_usage_1;
578 struct task_desc *task;
579 unsigned long i, ret;
581 sched->start_time = get_nsecs();
582 sched->cpu_usage = 0;
583 pthread_mutex_unlock(&sched->work_done_wait_mutex);
585 for (i = 0; i < sched->nr_tasks; i++) {
586 task = sched->tasks[i];
587 ret = sem_wait(&task->ready_for_work);
589 sem_init(&task->ready_for_work, 0, 0);
591 ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
594 cpu_usage_0 = get_cpu_usage_nsec_parent();
596 pthread_mutex_unlock(&sched->start_work_mutex);
598 for (i = 0; i < sched->nr_tasks; i++) {
599 task = sched->tasks[i];
600 ret = sem_wait(&task->work_done_sem);
602 sem_init(&task->work_done_sem, 0, 0);
603 sched->cpu_usage += task->cpu_usage;
607 cpu_usage_1 = get_cpu_usage_nsec_parent();
608 if (!sched->runavg_cpu_usage)
609 sched->runavg_cpu_usage = sched->cpu_usage;
610 sched->runavg_cpu_usage = (sched->runavg_cpu_usage * 9 + sched->cpu_usage) / 10;
612 sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
613 if (!sched->runavg_parent_cpu_usage)
614 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
615 sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * 9 +
616 sched->parent_cpu_usage)/10;
618 ret = pthread_mutex_lock(&sched->start_work_mutex);
621 for (i = 0; i < sched->nr_tasks; i++) {
622 task = sched->tasks[i];
623 sem_init(&task->sleep_sem, 0, 0);
624 task->curr_event = 0;
628 static void run_one_test(struct perf_sched *sched)
630 u64 T0, T1, delta, avg_delta, fluct;
633 wait_for_tasks(sched);
637 sched->sum_runtime += delta;
640 avg_delta = sched->sum_runtime / sched->nr_runs;
641 if (delta < avg_delta)
642 fluct = avg_delta - delta;
644 fluct = delta - avg_delta;
645 sched->sum_fluct += fluct;
647 sched->run_avg = delta;
648 sched->run_avg = (sched->run_avg * 9 + delta) / 10;
650 printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / 1000000.0);
652 printf("ravg: %0.2f, ", (double)sched->run_avg / 1e6);
654 printf("cpu: %0.2f / %0.2f",
655 (double)sched->cpu_usage / 1e6, (double)sched->runavg_cpu_usage / 1e6);
659 * rusage statistics done by the parent, these are less
660 * accurate than the sched->sum_exec_runtime based statistics:
662 printf(" [%0.2f / %0.2f]",
663 (double)sched->parent_cpu_usage/1e6,
664 (double)sched->runavg_parent_cpu_usage/1e6);
669 if (sched->nr_sleep_corrections)
670 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
671 sched->nr_sleep_corrections = 0;
674 static void test_calibrations(struct perf_sched *sched)
679 burn_nsecs(sched, 1e6);
682 printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
688 printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
692 replay_wakeup_event(struct perf_sched *sched,
693 struct perf_evsel *evsel, struct perf_sample *sample,
694 struct machine *machine __maybe_unused)
696 const char *comm = perf_evsel__strval(evsel, sample, "comm");
697 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
698 struct task_desc *waker, *wakee;
701 printf("sched_wakeup event %p\n", evsel);
703 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
706 waker = register_pid(sched, sample->tid, "<unknown>");
707 wakee = register_pid(sched, pid, comm);
709 add_sched_event_wakeup(sched, waker, sample->time, wakee);
713 static int replay_switch_event(struct perf_sched *sched,
714 struct perf_evsel *evsel,
715 struct perf_sample *sample,
716 struct machine *machine __maybe_unused)
718 const char *prev_comm = perf_evsel__strval(evsel, sample, "prev_comm"),
719 *next_comm = perf_evsel__strval(evsel, sample, "next_comm");
720 const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
721 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
722 const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
723 struct task_desc *prev, __maybe_unused *next;
724 u64 timestamp0, timestamp = sample->time;
725 int cpu = sample->cpu;
729 printf("sched_switch event %p\n", evsel);
731 if (cpu >= MAX_CPUS || cpu < 0)
734 timestamp0 = sched->cpu_last_switched[cpu];
736 delta = timestamp - timestamp0;
741 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
745 pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
746 prev_comm, prev_pid, next_comm, next_pid, delta);
748 prev = register_pid(sched, prev_pid, prev_comm);
749 next = register_pid(sched, next_pid, next_comm);
751 sched->cpu_last_switched[cpu] = timestamp;
753 add_sched_event_run(sched, prev, timestamp, delta);
754 add_sched_event_sleep(sched, prev, timestamp, prev_state);
759 static int replay_fork_event(struct perf_sched *sched,
760 union perf_event *event,
761 struct machine *machine)
763 struct thread *child, *parent;
765 child = machine__findnew_thread(machine, event->fork.pid,
767 parent = machine__findnew_thread(machine, event->fork.ppid,
770 if (child == NULL || parent == NULL) {
771 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
777 printf("fork event\n");
778 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
779 printf("... child: %s/%d\n", thread__comm_str(child), child->tid);
782 register_pid(sched, parent->tid, thread__comm_str(parent));
783 register_pid(sched, child->tid, thread__comm_str(child));
787 struct sort_dimension {
790 struct list_head list;
794 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
796 struct sort_dimension *sort;
799 BUG_ON(list_empty(list));
801 list_for_each_entry(sort, list, list) {
802 ret = sort->cmp(l, r);
810 static struct work_atoms *
811 thread_atoms_search(struct rb_root *root, struct thread *thread,
812 struct list_head *sort_list)
814 struct rb_node *node = root->rb_node;
815 struct work_atoms key = { .thread = thread };
818 struct work_atoms *atoms;
821 atoms = container_of(node, struct work_atoms, node);
823 cmp = thread_lat_cmp(sort_list, &key, atoms);
825 node = node->rb_left;
827 node = node->rb_right;
829 BUG_ON(thread != atoms->thread);
837 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
838 struct list_head *sort_list)
840 struct rb_node **new = &(root->rb_node), *parent = NULL;
843 struct work_atoms *this;
846 this = container_of(*new, struct work_atoms, node);
849 cmp = thread_lat_cmp(sort_list, data, this);
852 new = &((*new)->rb_left);
854 new = &((*new)->rb_right);
857 rb_link_node(&data->node, parent, new);
858 rb_insert_color(&data->node, root);
861 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
863 struct work_atoms *atoms = zalloc(sizeof(*atoms));
865 pr_err("No memory at %s\n", __func__);
869 atoms->thread = thread__get(thread);
870 INIT_LIST_HEAD(&atoms->work_list);
871 __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
875 static char sched_out_state(u64 prev_state)
877 const char *str = TASK_STATE_TO_CHAR_STR;
879 return str[prev_state];
883 add_sched_out_event(struct work_atoms *atoms,
887 struct work_atom *atom = zalloc(sizeof(*atom));
889 pr_err("Non memory at %s", __func__);
893 atom->sched_out_time = timestamp;
895 if (run_state == 'R') {
896 atom->state = THREAD_WAIT_CPU;
897 atom->wake_up_time = atom->sched_out_time;
900 list_add_tail(&atom->list, &atoms->work_list);
905 add_runtime_event(struct work_atoms *atoms, u64 delta,
906 u64 timestamp __maybe_unused)
908 struct work_atom *atom;
910 BUG_ON(list_empty(&atoms->work_list));
912 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
914 atom->runtime += delta;
915 atoms->total_runtime += delta;
919 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
921 struct work_atom *atom;
924 if (list_empty(&atoms->work_list))
927 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
929 if (atom->state != THREAD_WAIT_CPU)
932 if (timestamp < atom->wake_up_time) {
933 atom->state = THREAD_IGNORE;
937 atom->state = THREAD_SCHED_IN;
938 atom->sched_in_time = timestamp;
940 delta = atom->sched_in_time - atom->wake_up_time;
941 atoms->total_lat += delta;
942 if (delta > atoms->max_lat) {
943 atoms->max_lat = delta;
944 atoms->max_lat_at = timestamp;
949 static int latency_switch_event(struct perf_sched *sched,
950 struct perf_evsel *evsel,
951 struct perf_sample *sample,
952 struct machine *machine)
954 const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
955 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
956 const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
957 struct work_atoms *out_events, *in_events;
958 struct thread *sched_out, *sched_in;
959 u64 timestamp0, timestamp = sample->time;
960 int cpu = sample->cpu;
963 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
965 timestamp0 = sched->cpu_last_switched[cpu];
966 sched->cpu_last_switched[cpu] = timestamp;
968 delta = timestamp - timestamp0;
973 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
977 sched_out = machine__findnew_thread(machine, -1, prev_pid);
978 sched_in = machine__findnew_thread(machine, -1, next_pid);
980 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
982 if (thread_atoms_insert(sched, sched_out))
984 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
986 pr_err("out-event: Internal tree error");
990 if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
993 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
995 if (thread_atoms_insert(sched, sched_in))
997 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
999 pr_err("in-event: Internal tree error");
1003 * Take came in we have not heard about yet,
1004 * add in an initial atom in runnable state:
1006 if (add_sched_out_event(in_events, 'R', timestamp))
1009 add_sched_in_event(in_events, timestamp);
1014 static int latency_runtime_event(struct perf_sched *sched,
1015 struct perf_evsel *evsel,
1016 struct perf_sample *sample,
1017 struct machine *machine)
1019 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1020 const u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
1021 struct thread *thread = machine__findnew_thread(machine, -1, pid);
1022 struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1023 u64 timestamp = sample->time;
1024 int cpu = sample->cpu;
1026 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1028 if (thread_atoms_insert(sched, thread))
1030 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1032 pr_err("in-event: Internal tree error");
1035 if (add_sched_out_event(atoms, 'R', timestamp))
1039 add_runtime_event(atoms, runtime, timestamp);
1043 static int latency_wakeup_event(struct perf_sched *sched,
1044 struct perf_evsel *evsel,
1045 struct perf_sample *sample,
1046 struct machine *machine)
1048 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1049 struct work_atoms *atoms;
1050 struct work_atom *atom;
1051 struct thread *wakee;
1052 u64 timestamp = sample->time;
1054 wakee = machine__findnew_thread(machine, -1, pid);
1055 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1057 if (thread_atoms_insert(sched, wakee))
1059 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1061 pr_err("wakeup-event: Internal tree error");
1064 if (add_sched_out_event(atoms, 'S', timestamp))
1068 BUG_ON(list_empty(&atoms->work_list));
1070 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1073 * As we do not guarantee the wakeup event happens when
1074 * task is out of run queue, also may happen when task is
1075 * on run queue and wakeup only change ->state to TASK_RUNNING,
1076 * then we should not set the ->wake_up_time when wake up a
1077 * task which is on run queue.
1079 * You WILL be missing events if you've recorded only
1080 * one CPU, or are only looking at only one, so don't
1081 * skip in this case.
1083 if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1086 sched->nr_timestamps++;
1087 if (atom->sched_out_time > timestamp) {
1088 sched->nr_unordered_timestamps++;
1092 atom->state = THREAD_WAIT_CPU;
1093 atom->wake_up_time = timestamp;
1097 static int latency_migrate_task_event(struct perf_sched *sched,
1098 struct perf_evsel *evsel,
1099 struct perf_sample *sample,
1100 struct machine *machine)
1102 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1103 u64 timestamp = sample->time;
1104 struct work_atoms *atoms;
1105 struct work_atom *atom;
1106 struct thread *migrant;
1109 * Only need to worry about migration when profiling one CPU.
1111 if (sched->profile_cpu == -1)
1114 migrant = machine__findnew_thread(machine, -1, pid);
1115 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1117 if (thread_atoms_insert(sched, migrant))
1119 register_pid(sched, migrant->tid, thread__comm_str(migrant));
1120 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1122 pr_err("migration-event: Internal tree error");
1125 if (add_sched_out_event(atoms, 'R', timestamp))
1129 BUG_ON(list_empty(&atoms->work_list));
1131 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1132 atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1134 sched->nr_timestamps++;
1136 if (atom->sched_out_time > timestamp)
1137 sched->nr_unordered_timestamps++;
1142 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1148 if (!work_list->nb_atoms)
1151 * Ignore idle threads:
1153 if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1156 sched->all_runtime += work_list->total_runtime;
1157 sched->all_count += work_list->nb_atoms;
1159 ret = printf(" %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1161 for (i = 0; i < 24 - ret; i++)
1164 avg = work_list->total_lat / work_list->nb_atoms;
1166 printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13.6f s\n",
1167 (double)work_list->total_runtime / 1e6,
1168 work_list->nb_atoms, (double)avg / 1e6,
1169 (double)work_list->max_lat / 1e6,
1170 (double)work_list->max_lat_at / 1e9);
1173 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1175 if (l->thread->tid < r->thread->tid)
1177 if (l->thread->tid > r->thread->tid)
1183 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1193 avgl = l->total_lat / l->nb_atoms;
1194 avgr = r->total_lat / r->nb_atoms;
1204 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1206 if (l->max_lat < r->max_lat)
1208 if (l->max_lat > r->max_lat)
1214 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1216 if (l->nb_atoms < r->nb_atoms)
1218 if (l->nb_atoms > r->nb_atoms)
1224 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1226 if (l->total_runtime < r->total_runtime)
1228 if (l->total_runtime > r->total_runtime)
1234 static int sort_dimension__add(const char *tok, struct list_head *list)
1237 static struct sort_dimension avg_sort_dimension = {
1241 static struct sort_dimension max_sort_dimension = {
1245 static struct sort_dimension pid_sort_dimension = {
1249 static struct sort_dimension runtime_sort_dimension = {
1253 static struct sort_dimension switch_sort_dimension = {
1257 struct sort_dimension *available_sorts[] = {
1258 &pid_sort_dimension,
1259 &avg_sort_dimension,
1260 &max_sort_dimension,
1261 &switch_sort_dimension,
1262 &runtime_sort_dimension,
1265 for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1266 if (!strcmp(available_sorts[i]->name, tok)) {
1267 list_add_tail(&available_sorts[i]->list, list);
1276 static void perf_sched__sort_lat(struct perf_sched *sched)
1278 struct rb_node *node;
1281 struct work_atoms *data;
1282 node = rb_first(&sched->atom_root);
1286 rb_erase(node, &sched->atom_root);
1287 data = rb_entry(node, struct work_atoms, node);
1288 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1292 static int process_sched_wakeup_event(struct perf_tool *tool,
1293 struct perf_evsel *evsel,
1294 struct perf_sample *sample,
1295 struct machine *machine)
1297 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1299 if (sched->tp_handler->wakeup_event)
1300 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1305 static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1306 struct perf_sample *sample, struct machine *machine)
1308 const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1309 struct thread *sched_in;
1311 u64 timestamp0, timestamp = sample->time;
1313 int cpu, this_cpu = sample->cpu;
1315 BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1317 if (this_cpu > sched->max_cpu)
1318 sched->max_cpu = this_cpu;
1320 timestamp0 = sched->cpu_last_switched[this_cpu];
1321 sched->cpu_last_switched[this_cpu] = timestamp;
1323 delta = timestamp - timestamp0;
1328 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1332 sched_in = machine__findnew_thread(machine, -1, next_pid);
1334 sched->curr_thread[this_cpu] = sched_in;
1339 if (!sched_in->shortname[0]) {
1340 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1342 * Don't allocate a letter-number for swapper:0
1343 * as a shortname. Instead, we use '.' for it.
1345 sched_in->shortname[0] = '.';
1346 sched_in->shortname[1] = ' ';
1348 sched_in->shortname[0] = sched->next_shortname1;
1349 sched_in->shortname[1] = sched->next_shortname2;
1351 if (sched->next_shortname1 < 'Z') {
1352 sched->next_shortname1++;
1354 sched->next_shortname1 = 'A';
1355 if (sched->next_shortname2 < '9')
1356 sched->next_shortname2++;
1358 sched->next_shortname2 = '0';
1364 for (cpu = 0; cpu <= sched->max_cpu; cpu++) {
1365 if (cpu != this_cpu)
1370 if (sched->curr_thread[cpu])
1371 printf("%2s ", sched->curr_thread[cpu]->shortname);
1376 printf(" %12.6f secs ", (double)timestamp/1e9);
1377 if (new_shortname) {
1378 printf("%s => %s:%d\n",
1379 sched_in->shortname, thread__comm_str(sched_in), sched_in->tid);
1387 static int process_sched_switch_event(struct perf_tool *tool,
1388 struct perf_evsel *evsel,
1389 struct perf_sample *sample,
1390 struct machine *machine)
1392 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1393 int this_cpu = sample->cpu, err = 0;
1394 u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1395 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1397 if (sched->curr_pid[this_cpu] != (u32)-1) {
1399 * Are we trying to switch away a PID that is
1402 if (sched->curr_pid[this_cpu] != prev_pid)
1403 sched->nr_context_switch_bugs++;
1406 if (sched->tp_handler->switch_event)
1407 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1409 sched->curr_pid[this_cpu] = next_pid;
1413 static int process_sched_runtime_event(struct perf_tool *tool,
1414 struct perf_evsel *evsel,
1415 struct perf_sample *sample,
1416 struct machine *machine)
1418 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1420 if (sched->tp_handler->runtime_event)
1421 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1426 static int perf_sched__process_fork_event(struct perf_tool *tool,
1427 union perf_event *event,
1428 struct perf_sample *sample,
1429 struct machine *machine)
1431 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1433 /* run the fork event through the perf machineruy */
1434 perf_event__process_fork(tool, event, sample, machine);
1436 /* and then run additional processing needed for this command */
1437 if (sched->tp_handler->fork_event)
1438 return sched->tp_handler->fork_event(sched, event, machine);
1443 static int process_sched_migrate_task_event(struct perf_tool *tool,
1444 struct perf_evsel *evsel,
1445 struct perf_sample *sample,
1446 struct machine *machine)
1448 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1450 if (sched->tp_handler->migrate_task_event)
1451 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1456 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1457 struct perf_evsel *evsel,
1458 struct perf_sample *sample,
1459 struct machine *machine);
1461 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1462 union perf_event *event __maybe_unused,
1463 struct perf_sample *sample,
1464 struct perf_evsel *evsel,
1465 struct machine *machine)
1469 if (evsel->handler != NULL) {
1470 tracepoint_handler f = evsel->handler;
1471 err = f(tool, evsel, sample, machine);
1477 static int perf_sched__read_events(struct perf_sched *sched)
1479 const struct perf_evsel_str_handler handlers[] = {
1480 { "sched:sched_switch", process_sched_switch_event, },
1481 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1482 { "sched:sched_wakeup", process_sched_wakeup_event, },
1483 { "sched:sched_wakeup_new", process_sched_wakeup_event, },
1484 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1486 struct perf_session *session;
1487 struct perf_data_file file = {
1489 .mode = PERF_DATA_MODE_READ,
1493 session = perf_session__new(&file, false, &sched->tool);
1494 if (session == NULL) {
1495 pr_debug("No Memory for session\n");
1499 symbol__init(&session->header.env);
1501 if (perf_session__set_tracepoints_handlers(session, handlers))
1504 if (perf_session__has_traces(session, "record -R")) {
1505 int err = perf_session__process_events(session);
1507 pr_err("Failed to process events, error %d", err);
1511 sched->nr_events = session->evlist->stats.nr_events[0];
1512 sched->nr_lost_events = session->evlist->stats.total_lost;
1513 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1518 perf_session__delete(session);
1522 static void print_bad_events(struct perf_sched *sched)
1524 if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
1525 printf(" INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1526 (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
1527 sched->nr_unordered_timestamps, sched->nr_timestamps);
1529 if (sched->nr_lost_events && sched->nr_events) {
1530 printf(" INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1531 (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
1532 sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
1534 if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
1535 printf(" INFO: %.3f%% context switch bugs (%ld out of %ld)",
1536 (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
1537 sched->nr_context_switch_bugs, sched->nr_timestamps);
1538 if (sched->nr_lost_events)
1539 printf(" (due to lost events?)");
1544 static int perf_sched__lat(struct perf_sched *sched)
1546 struct rb_node *next;
1550 if (perf_sched__read_events(sched))
1553 perf_sched__sort_lat(sched);
1555 printf("\n -----------------------------------------------------------------------------------------------------------------\n");
1556 printf(" Task | Runtime ms | Switches | Average delay ms | Maximum delay ms | Maximum delay at |\n");
1557 printf(" -----------------------------------------------------------------------------------------------------------------\n");
1559 next = rb_first(&sched->sorted_atom_root);
1562 struct work_atoms *work_list;
1564 work_list = rb_entry(next, struct work_atoms, node);
1565 output_lat_thread(sched, work_list);
1566 next = rb_next(next);
1567 thread__zput(work_list->thread);
1570 printf(" -----------------------------------------------------------------------------------------------------------------\n");
1571 printf(" TOTAL: |%11.3f ms |%9" PRIu64 " |\n",
1572 (double)sched->all_runtime / 1e6, sched->all_count);
1574 printf(" ---------------------------------------------------\n");
1576 print_bad_events(sched);
1582 static int perf_sched__map(struct perf_sched *sched)
1584 sched->max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1587 if (perf_sched__read_events(sched))
1589 print_bad_events(sched);
1593 static int perf_sched__replay(struct perf_sched *sched)
1597 calibrate_run_measurement_overhead(sched);
1598 calibrate_sleep_measurement_overhead(sched);
1600 test_calibrations(sched);
1602 if (perf_sched__read_events(sched))
1605 printf("nr_run_events: %ld\n", sched->nr_run_events);
1606 printf("nr_sleep_events: %ld\n", sched->nr_sleep_events);
1607 printf("nr_wakeup_events: %ld\n", sched->nr_wakeup_events);
1609 if (sched->targetless_wakeups)
1610 printf("target-less wakeups: %ld\n", sched->targetless_wakeups);
1611 if (sched->multitarget_wakeups)
1612 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
1613 if (sched->nr_run_events_optimized)
1614 printf("run atoms optimized: %ld\n",
1615 sched->nr_run_events_optimized);
1617 print_task_traces(sched);
1618 add_cross_task_wakeups(sched);
1620 create_tasks(sched);
1621 printf("------------------------------------------------------------\n");
1622 for (i = 0; i < sched->replay_repeat; i++)
1623 run_one_test(sched);
1628 static void setup_sorting(struct perf_sched *sched, const struct option *options,
1629 const char * const usage_msg[])
1631 char *tmp, *tok, *str = strdup(sched->sort_order);
1633 for (tok = strtok_r(str, ", ", &tmp);
1634 tok; tok = strtok_r(NULL, ", ", &tmp)) {
1635 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
1636 error("Unknown --sort key: `%s'", tok);
1637 usage_with_options(usage_msg, options);
1643 sort_dimension__add("pid", &sched->cmp_pid);
1646 static int __cmd_record(int argc, const char **argv)
1648 unsigned int rec_argc, i, j;
1649 const char **rec_argv;
1650 const char * const record_args[] = {
1656 "-e", "sched:sched_switch",
1657 "-e", "sched:sched_stat_wait",
1658 "-e", "sched:sched_stat_sleep",
1659 "-e", "sched:sched_stat_iowait",
1660 "-e", "sched:sched_stat_runtime",
1661 "-e", "sched:sched_process_fork",
1662 "-e", "sched:sched_wakeup",
1663 "-e", "sched:sched_wakeup_new",
1664 "-e", "sched:sched_migrate_task",
1667 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1668 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1670 if (rec_argv == NULL)
1673 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1674 rec_argv[i] = strdup(record_args[i]);
1676 for (j = 1; j < (unsigned int)argc; j++, i++)
1677 rec_argv[i] = argv[j];
1679 BUG_ON(i != rec_argc);
1681 return cmd_record(i, rec_argv, NULL);
1684 int cmd_sched(int argc, const char **argv, const char *prefix __maybe_unused)
1686 const char default_sort_order[] = "avg, max, switch, runtime";
1687 struct perf_sched sched = {
1689 .sample = perf_sched__process_tracepoint_sample,
1690 .comm = perf_event__process_comm,
1691 .lost = perf_event__process_lost,
1692 .fork = perf_sched__process_fork_event,
1693 .ordered_events = true,
1695 .cmp_pid = LIST_HEAD_INIT(sched.cmp_pid),
1696 .sort_list = LIST_HEAD_INIT(sched.sort_list),
1697 .start_work_mutex = PTHREAD_MUTEX_INITIALIZER,
1698 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
1699 .sort_order = default_sort_order,
1700 .replay_repeat = 10,
1702 .next_shortname1 = 'A',
1703 .next_shortname2 = '0',
1705 const struct option latency_options[] = {
1706 OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
1707 "sort by key(s): runtime, switch, avg, max"),
1708 OPT_INCR('v', "verbose", &verbose,
1709 "be more verbose (show symbol address, etc)"),
1710 OPT_INTEGER('C', "CPU", &sched.profile_cpu,
1711 "CPU to profile on"),
1712 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1713 "dump raw trace in ASCII"),
1716 const struct option replay_options[] = {
1717 OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
1718 "repeat the workload replay N times (-1: infinite)"),
1719 OPT_INCR('v', "verbose", &verbose,
1720 "be more verbose (show symbol address, etc)"),
1721 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1722 "dump raw trace in ASCII"),
1723 OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
1726 const struct option sched_options[] = {
1727 OPT_STRING('i', "input", &input_name, "file",
1729 OPT_INCR('v', "verbose", &verbose,
1730 "be more verbose (show symbol address, etc)"),
1731 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1732 "dump raw trace in ASCII"),
1735 const char * const latency_usage[] = {
1736 "perf sched latency [<options>]",
1739 const char * const replay_usage[] = {
1740 "perf sched replay [<options>]",
1743 const char *const sched_subcommands[] = { "record", "latency", "map",
1744 "replay", "script", NULL };
1745 const char *sched_usage[] = {
1749 struct trace_sched_handler lat_ops = {
1750 .wakeup_event = latency_wakeup_event,
1751 .switch_event = latency_switch_event,
1752 .runtime_event = latency_runtime_event,
1753 .migrate_task_event = latency_migrate_task_event,
1755 struct trace_sched_handler map_ops = {
1756 .switch_event = map_switch_event,
1758 struct trace_sched_handler replay_ops = {
1759 .wakeup_event = replay_wakeup_event,
1760 .switch_event = replay_switch_event,
1761 .fork_event = replay_fork_event,
1765 for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
1766 sched.curr_pid[i] = -1;
1768 argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
1769 sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
1771 usage_with_options(sched_usage, sched_options);
1774 * Aliased to 'perf script' for now:
1776 if (!strcmp(argv[0], "script"))
1777 return cmd_script(argc, argv, prefix);
1779 if (!strncmp(argv[0], "rec", 3)) {
1780 return __cmd_record(argc, argv);
1781 } else if (!strncmp(argv[0], "lat", 3)) {
1782 sched.tp_handler = &lat_ops;
1784 argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1786 usage_with_options(latency_usage, latency_options);
1788 setup_sorting(&sched, latency_options, latency_usage);
1789 return perf_sched__lat(&sched);
1790 } else if (!strcmp(argv[0], "map")) {
1791 sched.tp_handler = &map_ops;
1792 setup_sorting(&sched, latency_options, latency_usage);
1793 return perf_sched__map(&sched);
1794 } else if (!strncmp(argv[0], "rep", 3)) {
1795 sched.tp_handler = &replay_ops;
1797 argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1799 usage_with_options(replay_usage, replay_options);
1801 return perf_sched__replay(&sched);
1803 usage_with_options(sched_usage, sched_options);