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;
140 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
144 struct trace_sched_handler {
145 int (*switch_event)(struct perf_sched *sched, struct evsel *evsel,
146 struct perf_sample *sample, struct machine *machine);
148 int (*runtime_event)(struct perf_sched *sched, struct evsel *evsel,
149 struct perf_sample *sample, struct machine *machine);
151 int (*wakeup_event)(struct perf_sched *sched, struct evsel *evsel,
152 struct perf_sample *sample, struct machine *machine);
154 /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
155 int (*fork_event)(struct perf_sched *sched, union perf_event *event,
156 struct machine *machine);
158 int (*migrate_task_event)(struct perf_sched *sched,
160 struct perf_sample *sample,
161 struct machine *machine);
164 #define COLOR_PIDS PERF_COLOR_BLUE
165 #define COLOR_CPUS PERF_COLOR_BG_RED
167 struct perf_sched_map {
168 DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
171 struct perf_thread_map *color_pids;
172 const char *color_pids_str;
173 struct perf_cpu_map *color_cpus;
174 const char *color_cpus_str;
175 struct perf_cpu_map *cpus;
176 const char *cpus_str;
180 struct perf_tool tool;
181 const char *sort_order;
182 unsigned long nr_tasks;
183 struct task_desc **pid_to_task;
184 struct task_desc **tasks;
185 const struct trace_sched_handler *tp_handler;
186 pthread_mutex_t start_work_mutex;
187 pthread_mutex_t work_done_wait_mutex;
190 * Track the current task - that way we can know whether there's any
191 * weird events, such as a task being switched away that is not current.
194 u32 curr_pid[MAX_CPUS];
195 struct thread *curr_thread[MAX_CPUS];
196 char next_shortname1;
197 char next_shortname2;
198 unsigned int replay_repeat;
199 unsigned long nr_run_events;
200 unsigned long nr_sleep_events;
201 unsigned long nr_wakeup_events;
202 unsigned long nr_sleep_corrections;
203 unsigned long nr_run_events_optimized;
204 unsigned long targetless_wakeups;
205 unsigned long multitarget_wakeups;
206 unsigned long nr_runs;
207 unsigned long nr_timestamps;
208 unsigned long nr_unordered_timestamps;
209 unsigned long nr_context_switch_bugs;
210 unsigned long nr_events;
211 unsigned long nr_lost_chunks;
212 unsigned long nr_lost_events;
213 u64 run_measurement_overhead;
214 u64 sleep_measurement_overhead;
217 u64 runavg_cpu_usage;
218 u64 parent_cpu_usage;
219 u64 runavg_parent_cpu_usage;
225 u64 cpu_last_switched[MAX_CPUS];
226 struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
227 struct list_head sort_list, cmp_pid;
230 struct perf_sched_map map;
232 /* options for timehist command */
237 unsigned int max_stack;
238 bool show_cpu_visual;
241 bool show_migrations;
244 const char *time_str;
245 struct perf_time_interval ptime;
246 struct perf_time_interval hist_time;
249 /* per thread run time data */
250 struct thread_runtime {
251 u64 last_time; /* time of previous sched in/out event */
252 u64 dt_run; /* run time */
253 u64 dt_sleep; /* time between CPU access by sleep (off cpu) */
254 u64 dt_iowait; /* time between CPU access by iowait (off cpu) */
255 u64 dt_preempt; /* time between CPU access by preempt (off cpu) */
256 u64 dt_delay; /* time between wakeup and sched-in */
257 u64 ready_to_run; /* time of wakeup */
259 struct stats run_stats;
261 u64 total_sleep_time;
262 u64 total_iowait_time;
263 u64 total_preempt_time;
264 u64 total_delay_time;
274 /* per event run time data */
275 struct evsel_runtime {
276 u64 *last_time; /* time this event was last seen per cpu */
277 u32 ncpu; /* highest cpu slot allocated */
280 /* per cpu idle time data */
281 struct idle_thread_runtime {
282 struct thread_runtime tr;
283 struct thread *last_thread;
284 struct rb_root_cached sorted_root;
285 struct callchain_root callchain;
286 struct callchain_cursor cursor;
289 /* track idle times per cpu */
290 static struct thread **idle_threads;
291 static int idle_max_cpu;
292 static char idle_comm[] = "<idle>";
294 static u64 get_nsecs(void)
298 clock_gettime(CLOCK_MONOTONIC, &ts);
300 return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
303 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
305 u64 T0 = get_nsecs(), T1;
309 } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
312 static void sleep_nsecs(u64 nsecs)
316 ts.tv_nsec = nsecs % 999999999;
317 ts.tv_sec = nsecs / 999999999;
319 nanosleep(&ts, NULL);
322 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
324 u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
327 for (i = 0; i < 10; i++) {
329 burn_nsecs(sched, 0);
332 min_delta = min(min_delta, delta);
334 sched->run_measurement_overhead = min_delta;
336 printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
339 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
341 u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
344 for (i = 0; i < 10; i++) {
349 min_delta = min(min_delta, delta);
352 sched->sleep_measurement_overhead = min_delta;
354 printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
357 static struct sched_atom *
358 get_new_event(struct task_desc *task, u64 timestamp)
360 struct sched_atom *event = zalloc(sizeof(*event));
361 unsigned long idx = task->nr_events;
364 event->timestamp = timestamp;
368 size = sizeof(struct sched_atom *) * task->nr_events;
369 task->atoms = realloc(task->atoms, size);
370 BUG_ON(!task->atoms);
372 task->atoms[idx] = event;
377 static struct sched_atom *last_event(struct task_desc *task)
379 if (!task->nr_events)
382 return task->atoms[task->nr_events - 1];
385 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
386 u64 timestamp, u64 duration)
388 struct sched_atom *event, *curr_event = last_event(task);
391 * optimize an existing RUN event by merging this one
394 if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
395 sched->nr_run_events_optimized++;
396 curr_event->duration += duration;
400 event = get_new_event(task, timestamp);
402 event->type = SCHED_EVENT_RUN;
403 event->duration = duration;
405 sched->nr_run_events++;
408 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
409 u64 timestamp, struct task_desc *wakee)
411 struct sched_atom *event, *wakee_event;
413 event = get_new_event(task, timestamp);
414 event->type = SCHED_EVENT_WAKEUP;
415 event->wakee = wakee;
417 wakee_event = last_event(wakee);
418 if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
419 sched->targetless_wakeups++;
422 if (wakee_event->wait_sem) {
423 sched->multitarget_wakeups++;
427 wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
428 sem_init(wakee_event->wait_sem, 0, 0);
429 wakee_event->specific_wait = 1;
430 event->wait_sem = wakee_event->wait_sem;
432 sched->nr_wakeup_events++;
435 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
436 u64 timestamp, u64 task_state __maybe_unused)
438 struct sched_atom *event = get_new_event(task, timestamp);
440 event->type = SCHED_EVENT_SLEEP;
442 sched->nr_sleep_events++;
445 static struct task_desc *register_pid(struct perf_sched *sched,
446 unsigned long pid, const char *comm)
448 struct task_desc *task;
451 if (sched->pid_to_task == NULL) {
452 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
454 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
456 if (pid >= (unsigned long)pid_max) {
457 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
458 sizeof(struct task_desc *))) == NULL);
459 while (pid >= (unsigned long)pid_max)
460 sched->pid_to_task[pid_max++] = NULL;
463 task = sched->pid_to_task[pid];
468 task = zalloc(sizeof(*task));
470 task->nr = sched->nr_tasks;
471 strcpy(task->comm, comm);
473 * every task starts in sleeping state - this gets ignored
474 * if there's no wakeup pointing to this sleep state:
476 add_sched_event_sleep(sched, task, 0, 0);
478 sched->pid_to_task[pid] = task;
480 sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
481 BUG_ON(!sched->tasks);
482 sched->tasks[task->nr] = task;
485 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
491 static void print_task_traces(struct perf_sched *sched)
493 struct task_desc *task;
496 for (i = 0; i < sched->nr_tasks; i++) {
497 task = sched->tasks[i];
498 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
499 task->nr, task->comm, task->pid, task->nr_events);
503 static void add_cross_task_wakeups(struct perf_sched *sched)
505 struct task_desc *task1, *task2;
508 for (i = 0; i < sched->nr_tasks; i++) {
509 task1 = sched->tasks[i];
511 if (j == sched->nr_tasks)
513 task2 = sched->tasks[j];
514 add_sched_event_wakeup(sched, task1, 0, task2);
518 static void perf_sched__process_event(struct perf_sched *sched,
519 struct sched_atom *atom)
523 switch (atom->type) {
524 case SCHED_EVENT_RUN:
525 burn_nsecs(sched, atom->duration);
527 case SCHED_EVENT_SLEEP:
529 ret = sem_wait(atom->wait_sem);
532 case SCHED_EVENT_WAKEUP:
534 ret = sem_post(atom->wait_sem);
537 case SCHED_EVENT_MIGRATION:
544 static u64 get_cpu_usage_nsec_parent(void)
550 err = getrusage(RUSAGE_SELF, &ru);
553 sum = ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
554 sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
559 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
561 struct perf_event_attr attr;
562 char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
565 bool need_privilege = false;
567 memset(&attr, 0, sizeof(attr));
569 attr.type = PERF_TYPE_SOFTWARE;
570 attr.config = PERF_COUNT_SW_TASK_CLOCK;
573 fd = sys_perf_event_open(&attr, 0, -1, -1,
574 perf_event_open_cloexec_flag());
577 if (errno == EMFILE) {
579 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
580 limit.rlim_cur += sched->nr_tasks - cur_task;
581 if (limit.rlim_cur > limit.rlim_max) {
582 limit.rlim_max = limit.rlim_cur;
583 need_privilege = true;
585 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
586 if (need_privilege && errno == EPERM)
587 strcpy(info, "Need privilege\n");
591 strcpy(info, "Have a try with -f option\n");
593 pr_err("Error: sys_perf_event_open() syscall returned "
594 "with %d (%s)\n%s", fd,
595 str_error_r(errno, sbuf, sizeof(sbuf)), info);
601 static u64 get_cpu_usage_nsec_self(int fd)
606 ret = read(fd, &runtime, sizeof(runtime));
607 BUG_ON(ret != sizeof(runtime));
612 struct sched_thread_parms {
613 struct task_desc *task;
614 struct perf_sched *sched;
618 static void *thread_func(void *ctx)
620 struct sched_thread_parms *parms = ctx;
621 struct task_desc *this_task = parms->task;
622 struct perf_sched *sched = parms->sched;
623 u64 cpu_usage_0, cpu_usage_1;
624 unsigned long i, ret;
630 sprintf(comm2, ":%s", this_task->comm);
631 prctl(PR_SET_NAME, comm2);
635 ret = sem_post(&this_task->ready_for_work);
637 ret = pthread_mutex_lock(&sched->start_work_mutex);
639 ret = pthread_mutex_unlock(&sched->start_work_mutex);
642 cpu_usage_0 = get_cpu_usage_nsec_self(fd);
644 for (i = 0; i < this_task->nr_events; i++) {
645 this_task->curr_event = i;
646 perf_sched__process_event(sched, this_task->atoms[i]);
649 cpu_usage_1 = get_cpu_usage_nsec_self(fd);
650 this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
651 ret = sem_post(&this_task->work_done_sem);
654 ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
656 ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
662 static void create_tasks(struct perf_sched *sched)
664 struct task_desc *task;
669 err = pthread_attr_init(&attr);
671 err = pthread_attr_setstacksize(&attr,
672 (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
674 err = pthread_mutex_lock(&sched->start_work_mutex);
676 err = pthread_mutex_lock(&sched->work_done_wait_mutex);
678 for (i = 0; i < sched->nr_tasks; i++) {
679 struct sched_thread_parms *parms = malloc(sizeof(*parms));
680 BUG_ON(parms == NULL);
681 parms->task = task = sched->tasks[i];
682 parms->sched = sched;
683 parms->fd = self_open_counters(sched, i);
684 sem_init(&task->sleep_sem, 0, 0);
685 sem_init(&task->ready_for_work, 0, 0);
686 sem_init(&task->work_done_sem, 0, 0);
687 task->curr_event = 0;
688 err = pthread_create(&task->thread, &attr, thread_func, parms);
693 static void wait_for_tasks(struct perf_sched *sched)
695 u64 cpu_usage_0, cpu_usage_1;
696 struct task_desc *task;
697 unsigned long i, ret;
699 sched->start_time = get_nsecs();
700 sched->cpu_usage = 0;
701 pthread_mutex_unlock(&sched->work_done_wait_mutex);
703 for (i = 0; i < sched->nr_tasks; i++) {
704 task = sched->tasks[i];
705 ret = sem_wait(&task->ready_for_work);
707 sem_init(&task->ready_for_work, 0, 0);
709 ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
712 cpu_usage_0 = get_cpu_usage_nsec_parent();
714 pthread_mutex_unlock(&sched->start_work_mutex);
716 for (i = 0; i < sched->nr_tasks; i++) {
717 task = sched->tasks[i];
718 ret = sem_wait(&task->work_done_sem);
720 sem_init(&task->work_done_sem, 0, 0);
721 sched->cpu_usage += task->cpu_usage;
725 cpu_usage_1 = get_cpu_usage_nsec_parent();
726 if (!sched->runavg_cpu_usage)
727 sched->runavg_cpu_usage = sched->cpu_usage;
728 sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
730 sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
731 if (!sched->runavg_parent_cpu_usage)
732 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
733 sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
734 sched->parent_cpu_usage)/sched->replay_repeat;
736 ret = pthread_mutex_lock(&sched->start_work_mutex);
739 for (i = 0; i < sched->nr_tasks; i++) {
740 task = sched->tasks[i];
741 sem_init(&task->sleep_sem, 0, 0);
742 task->curr_event = 0;
746 static void run_one_test(struct perf_sched *sched)
748 u64 T0, T1, delta, avg_delta, fluct;
751 wait_for_tasks(sched);
755 sched->sum_runtime += delta;
758 avg_delta = sched->sum_runtime / sched->nr_runs;
759 if (delta < avg_delta)
760 fluct = avg_delta - delta;
762 fluct = delta - avg_delta;
763 sched->sum_fluct += fluct;
765 sched->run_avg = delta;
766 sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
768 printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
770 printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
772 printf("cpu: %0.2f / %0.2f",
773 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
777 * rusage statistics done by the parent, these are less
778 * accurate than the sched->sum_exec_runtime based statistics:
780 printf(" [%0.2f / %0.2f]",
781 (double)sched->parent_cpu_usage / NSEC_PER_MSEC,
782 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
787 if (sched->nr_sleep_corrections)
788 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
789 sched->nr_sleep_corrections = 0;
792 static void test_calibrations(struct perf_sched *sched)
797 burn_nsecs(sched, NSEC_PER_MSEC);
800 printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
803 sleep_nsecs(NSEC_PER_MSEC);
806 printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
810 replay_wakeup_event(struct perf_sched *sched,
811 struct evsel *evsel, struct perf_sample *sample,
812 struct machine *machine __maybe_unused)
814 const char *comm = evsel__strval(evsel, sample, "comm");
815 const u32 pid = evsel__intval(evsel, sample, "pid");
816 struct task_desc *waker, *wakee;
819 printf("sched_wakeup event %p\n", evsel);
821 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
824 waker = register_pid(sched, sample->tid, "<unknown>");
825 wakee = register_pid(sched, pid, comm);
827 add_sched_event_wakeup(sched, waker, sample->time, wakee);
831 static int replay_switch_event(struct perf_sched *sched,
833 struct perf_sample *sample,
834 struct machine *machine __maybe_unused)
836 const char *prev_comm = evsel__strval(evsel, sample, "prev_comm"),
837 *next_comm = evsel__strval(evsel, sample, "next_comm");
838 const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
839 next_pid = evsel__intval(evsel, sample, "next_pid");
840 const u64 prev_state = evsel__intval(evsel, sample, "prev_state");
841 struct task_desc *prev, __maybe_unused *next;
842 u64 timestamp0, timestamp = sample->time;
843 int cpu = sample->cpu;
847 printf("sched_switch event %p\n", evsel);
849 if (cpu >= MAX_CPUS || cpu < 0)
852 timestamp0 = sched->cpu_last_switched[cpu];
854 delta = timestamp - timestamp0;
859 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
863 pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
864 prev_comm, prev_pid, next_comm, next_pid, delta);
866 prev = register_pid(sched, prev_pid, prev_comm);
867 next = register_pid(sched, next_pid, next_comm);
869 sched->cpu_last_switched[cpu] = timestamp;
871 add_sched_event_run(sched, prev, timestamp, delta);
872 add_sched_event_sleep(sched, prev, timestamp, prev_state);
877 static int replay_fork_event(struct perf_sched *sched,
878 union perf_event *event,
879 struct machine *machine)
881 struct thread *child, *parent;
883 child = machine__findnew_thread(machine, event->fork.pid,
885 parent = machine__findnew_thread(machine, event->fork.ppid,
888 if (child == NULL || parent == NULL) {
889 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
895 printf("fork event\n");
896 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
897 printf("... child: %s/%d\n", thread__comm_str(child), child->tid);
900 register_pid(sched, parent->tid, thread__comm_str(parent));
901 register_pid(sched, child->tid, thread__comm_str(child));
908 struct sort_dimension {
911 struct list_head list;
915 * handle runtime stats saved per thread
917 static struct thread_runtime *thread__init_runtime(struct thread *thread)
919 struct thread_runtime *r;
921 r = zalloc(sizeof(struct thread_runtime));
925 init_stats(&r->run_stats);
926 thread__set_priv(thread, r);
931 static struct thread_runtime *thread__get_runtime(struct thread *thread)
933 struct thread_runtime *tr;
935 tr = thread__priv(thread);
937 tr = thread__init_runtime(thread);
939 pr_debug("Failed to malloc memory for runtime data.\n");
946 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
948 struct sort_dimension *sort;
951 BUG_ON(list_empty(list));
953 list_for_each_entry(sort, list, list) {
954 ret = sort->cmp(l, r);
962 static struct work_atoms *
963 thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
964 struct list_head *sort_list)
966 struct rb_node *node = root->rb_root.rb_node;
967 struct work_atoms key = { .thread = thread };
970 struct work_atoms *atoms;
973 atoms = container_of(node, struct work_atoms, node);
975 cmp = thread_lat_cmp(sort_list, &key, atoms);
977 node = node->rb_left;
979 node = node->rb_right;
981 BUG_ON(thread != atoms->thread);
989 __thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
990 struct list_head *sort_list)
992 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
993 bool leftmost = true;
996 struct work_atoms *this;
999 this = container_of(*new, struct work_atoms, node);
1002 cmp = thread_lat_cmp(sort_list, data, this);
1005 new = &((*new)->rb_left);
1007 new = &((*new)->rb_right);
1012 rb_link_node(&data->node, parent, new);
1013 rb_insert_color_cached(&data->node, root, leftmost);
1016 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1018 struct work_atoms *atoms = zalloc(sizeof(*atoms));
1020 pr_err("No memory at %s\n", __func__);
1024 atoms->thread = thread__get(thread);
1025 INIT_LIST_HEAD(&atoms->work_list);
1026 __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1030 static char sched_out_state(u64 prev_state)
1032 const char *str = TASK_STATE_TO_CHAR_STR;
1034 return str[prev_state];
1038 add_sched_out_event(struct work_atoms *atoms,
1042 struct work_atom *atom = zalloc(sizeof(*atom));
1044 pr_err("Non memory at %s", __func__);
1048 atom->sched_out_time = timestamp;
1050 if (run_state == 'R') {
1051 atom->state = THREAD_WAIT_CPU;
1052 atom->wake_up_time = atom->sched_out_time;
1055 list_add_tail(&atom->list, &atoms->work_list);
1060 add_runtime_event(struct work_atoms *atoms, u64 delta,
1061 u64 timestamp __maybe_unused)
1063 struct work_atom *atom;
1065 BUG_ON(list_empty(&atoms->work_list));
1067 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1069 atom->runtime += delta;
1070 atoms->total_runtime += delta;
1074 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1076 struct work_atom *atom;
1079 if (list_empty(&atoms->work_list))
1082 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1084 if (atom->state != THREAD_WAIT_CPU)
1087 if (timestamp < atom->wake_up_time) {
1088 atom->state = THREAD_IGNORE;
1092 atom->state = THREAD_SCHED_IN;
1093 atom->sched_in_time = timestamp;
1095 delta = atom->sched_in_time - atom->wake_up_time;
1096 atoms->total_lat += delta;
1097 if (delta > atoms->max_lat) {
1098 atoms->max_lat = delta;
1099 atoms->max_lat_at = timestamp;
1104 static int latency_switch_event(struct perf_sched *sched,
1105 struct evsel *evsel,
1106 struct perf_sample *sample,
1107 struct machine *machine)
1109 const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1110 next_pid = evsel__intval(evsel, sample, "next_pid");
1111 const u64 prev_state = evsel__intval(evsel, sample, "prev_state");
1112 struct work_atoms *out_events, *in_events;
1113 struct thread *sched_out, *sched_in;
1114 u64 timestamp0, timestamp = sample->time;
1115 int cpu = sample->cpu, err = -1;
1118 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1120 timestamp0 = sched->cpu_last_switched[cpu];
1121 sched->cpu_last_switched[cpu] = timestamp;
1123 delta = timestamp - timestamp0;
1128 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1132 sched_out = machine__findnew_thread(machine, -1, prev_pid);
1133 sched_in = machine__findnew_thread(machine, -1, next_pid);
1134 if (sched_out == NULL || sched_in == NULL)
1137 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1139 if (thread_atoms_insert(sched, sched_out))
1141 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1143 pr_err("out-event: Internal tree error");
1147 if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1150 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1152 if (thread_atoms_insert(sched, sched_in))
1154 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1156 pr_err("in-event: Internal tree error");
1160 * Take came in we have not heard about yet,
1161 * add in an initial atom in runnable state:
1163 if (add_sched_out_event(in_events, 'R', timestamp))
1166 add_sched_in_event(in_events, timestamp);
1169 thread__put(sched_out);
1170 thread__put(sched_in);
1174 static int latency_runtime_event(struct perf_sched *sched,
1175 struct evsel *evsel,
1176 struct perf_sample *sample,
1177 struct machine *machine)
1179 const u32 pid = evsel__intval(evsel, sample, "pid");
1180 const u64 runtime = evsel__intval(evsel, sample, "runtime");
1181 struct thread *thread = machine__findnew_thread(machine, -1, pid);
1182 struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1183 u64 timestamp = sample->time;
1184 int cpu = sample->cpu, err = -1;
1189 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1191 if (thread_atoms_insert(sched, thread))
1193 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1195 pr_err("in-event: Internal tree error");
1198 if (add_sched_out_event(atoms, 'R', timestamp))
1202 add_runtime_event(atoms, runtime, timestamp);
1205 thread__put(thread);
1209 static int latency_wakeup_event(struct perf_sched *sched,
1210 struct evsel *evsel,
1211 struct perf_sample *sample,
1212 struct machine *machine)
1214 const u32 pid = evsel__intval(evsel, sample, "pid");
1215 struct work_atoms *atoms;
1216 struct work_atom *atom;
1217 struct thread *wakee;
1218 u64 timestamp = sample->time;
1221 wakee = machine__findnew_thread(machine, -1, pid);
1224 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1226 if (thread_atoms_insert(sched, wakee))
1228 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1230 pr_err("wakeup-event: Internal tree error");
1233 if (add_sched_out_event(atoms, 'S', timestamp))
1237 BUG_ON(list_empty(&atoms->work_list));
1239 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1242 * As we do not guarantee the wakeup event happens when
1243 * task is out of run queue, also may happen when task is
1244 * on run queue and wakeup only change ->state to TASK_RUNNING,
1245 * then we should not set the ->wake_up_time when wake up a
1246 * task which is on run queue.
1248 * You WILL be missing events if you've recorded only
1249 * one CPU, or are only looking at only one, so don't
1250 * skip in this case.
1252 if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1255 sched->nr_timestamps++;
1256 if (atom->sched_out_time > timestamp) {
1257 sched->nr_unordered_timestamps++;
1261 atom->state = THREAD_WAIT_CPU;
1262 atom->wake_up_time = timestamp;
1270 static int latency_migrate_task_event(struct perf_sched *sched,
1271 struct evsel *evsel,
1272 struct perf_sample *sample,
1273 struct machine *machine)
1275 const u32 pid = evsel__intval(evsel, sample, "pid");
1276 u64 timestamp = sample->time;
1277 struct work_atoms *atoms;
1278 struct work_atom *atom;
1279 struct thread *migrant;
1283 * Only need to worry about migration when profiling one CPU.
1285 if (sched->profile_cpu == -1)
1288 migrant = machine__findnew_thread(machine, -1, pid);
1289 if (migrant == NULL)
1291 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1293 if (thread_atoms_insert(sched, migrant))
1295 register_pid(sched, migrant->tid, thread__comm_str(migrant));
1296 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1298 pr_err("migration-event: Internal tree error");
1301 if (add_sched_out_event(atoms, 'R', timestamp))
1305 BUG_ON(list_empty(&atoms->work_list));
1307 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1308 atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1310 sched->nr_timestamps++;
1312 if (atom->sched_out_time > timestamp)
1313 sched->nr_unordered_timestamps++;
1316 thread__put(migrant);
1320 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1325 char max_lat_at[32];
1327 if (!work_list->nb_atoms)
1330 * Ignore idle threads:
1332 if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1335 sched->all_runtime += work_list->total_runtime;
1336 sched->all_count += work_list->nb_atoms;
1338 if (work_list->num_merged > 1)
1339 ret = printf(" %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1341 ret = printf(" %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1343 for (i = 0; i < 24 - ret; i++)
1346 avg = work_list->total_lat / work_list->nb_atoms;
1347 timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at));
1349 printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n",
1350 (double)work_list->total_runtime / NSEC_PER_MSEC,
1351 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1352 (double)work_list->max_lat / NSEC_PER_MSEC,
1356 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1358 if (l->thread == r->thread)
1360 if (l->thread->tid < r->thread->tid)
1362 if (l->thread->tid > r->thread->tid)
1364 return (int)(l->thread - r->thread);
1367 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1377 avgl = l->total_lat / l->nb_atoms;
1378 avgr = r->total_lat / r->nb_atoms;
1388 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1390 if (l->max_lat < r->max_lat)
1392 if (l->max_lat > r->max_lat)
1398 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1400 if (l->nb_atoms < r->nb_atoms)
1402 if (l->nb_atoms > r->nb_atoms)
1408 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1410 if (l->total_runtime < r->total_runtime)
1412 if (l->total_runtime > r->total_runtime)
1418 static int sort_dimension__add(const char *tok, struct list_head *list)
1421 static struct sort_dimension avg_sort_dimension = {
1425 static struct sort_dimension max_sort_dimension = {
1429 static struct sort_dimension pid_sort_dimension = {
1433 static struct sort_dimension runtime_sort_dimension = {
1437 static struct sort_dimension switch_sort_dimension = {
1441 struct sort_dimension *available_sorts[] = {
1442 &pid_sort_dimension,
1443 &avg_sort_dimension,
1444 &max_sort_dimension,
1445 &switch_sort_dimension,
1446 &runtime_sort_dimension,
1449 for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1450 if (!strcmp(available_sorts[i]->name, tok)) {
1451 list_add_tail(&available_sorts[i]->list, list);
1460 static void perf_sched__sort_lat(struct perf_sched *sched)
1462 struct rb_node *node;
1463 struct rb_root_cached *root = &sched->atom_root;
1466 struct work_atoms *data;
1467 node = rb_first_cached(root);
1471 rb_erase_cached(node, root);
1472 data = rb_entry(node, struct work_atoms, node);
1473 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1475 if (root == &sched->atom_root) {
1476 root = &sched->merged_atom_root;
1481 static int process_sched_wakeup_event(struct perf_tool *tool,
1482 struct evsel *evsel,
1483 struct perf_sample *sample,
1484 struct machine *machine)
1486 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1488 if (sched->tp_handler->wakeup_event)
1489 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1499 static bool thread__has_color(struct thread *thread)
1501 union map_priv priv = {
1502 .ptr = thread__priv(thread),
1508 static struct thread*
1509 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1511 struct thread *thread = machine__findnew_thread(machine, pid, tid);
1512 union map_priv priv = {
1516 if (!sched->map.color_pids || !thread || thread__priv(thread))
1519 if (thread_map__has(sched->map.color_pids, tid))
1522 thread__set_priv(thread, priv.ptr);
1526 static int map_switch_event(struct perf_sched *sched, struct evsel *evsel,
1527 struct perf_sample *sample, struct machine *machine)
1529 const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
1530 struct thread *sched_in;
1531 struct thread_runtime *tr;
1533 u64 timestamp0, timestamp = sample->time;
1535 int i, this_cpu = sample->cpu;
1537 bool new_cpu = false;
1538 const char *color = PERF_COLOR_NORMAL;
1539 char stimestamp[32];
1541 BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1543 if (this_cpu > sched->max_cpu)
1544 sched->max_cpu = this_cpu;
1546 if (sched->map.comp) {
1547 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1548 if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1549 sched->map.comp_cpus[cpus_nr++] = this_cpu;
1553 cpus_nr = sched->max_cpu;
1555 timestamp0 = sched->cpu_last_switched[this_cpu];
1556 sched->cpu_last_switched[this_cpu] = timestamp;
1558 delta = timestamp - timestamp0;
1563 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1567 sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1568 if (sched_in == NULL)
1571 tr = thread__get_runtime(sched_in);
1573 thread__put(sched_in);
1577 sched->curr_thread[this_cpu] = thread__get(sched_in);
1582 if (!tr->shortname[0]) {
1583 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1585 * Don't allocate a letter-number for swapper:0
1586 * as a shortname. Instead, we use '.' for it.
1588 tr->shortname[0] = '.';
1589 tr->shortname[1] = ' ';
1591 tr->shortname[0] = sched->next_shortname1;
1592 tr->shortname[1] = sched->next_shortname2;
1594 if (sched->next_shortname1 < 'Z') {
1595 sched->next_shortname1++;
1597 sched->next_shortname1 = 'A';
1598 if (sched->next_shortname2 < '9')
1599 sched->next_shortname2++;
1601 sched->next_shortname2 = '0';
1607 for (i = 0; i < cpus_nr; i++) {
1608 int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
1609 struct thread *curr_thread = sched->curr_thread[cpu];
1610 struct thread_runtime *curr_tr;
1611 const char *pid_color = color;
1612 const char *cpu_color = color;
1614 if (curr_thread && thread__has_color(curr_thread))
1615 pid_color = COLOR_PIDS;
1617 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1620 if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1621 cpu_color = COLOR_CPUS;
1623 if (cpu != this_cpu)
1624 color_fprintf(stdout, color, " ");
1626 color_fprintf(stdout, cpu_color, "*");
1628 if (sched->curr_thread[cpu]) {
1629 curr_tr = thread__get_runtime(sched->curr_thread[cpu]);
1630 if (curr_tr == NULL) {
1631 thread__put(sched_in);
1634 color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1636 color_fprintf(stdout, color, " ");
1639 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1642 timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1643 color_fprintf(stdout, color, " %12s secs ", stimestamp);
1644 if (new_shortname || tr->comm_changed || (verbose > 0 && sched_in->tid)) {
1645 const char *pid_color = color;
1647 if (thread__has_color(sched_in))
1648 pid_color = COLOR_PIDS;
1650 color_fprintf(stdout, pid_color, "%s => %s:%d",
1651 tr->shortname, thread__comm_str(sched_in), sched_in->tid);
1652 tr->comm_changed = false;
1655 if (sched->map.comp && new_cpu)
1656 color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1659 color_fprintf(stdout, color, "\n");
1661 thread__put(sched_in);
1666 static int process_sched_switch_event(struct perf_tool *tool,
1667 struct evsel *evsel,
1668 struct perf_sample *sample,
1669 struct machine *machine)
1671 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1672 int this_cpu = sample->cpu, err = 0;
1673 u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1674 next_pid = evsel__intval(evsel, sample, "next_pid");
1676 if (sched->curr_pid[this_cpu] != (u32)-1) {
1678 * Are we trying to switch away a PID that is
1681 if (sched->curr_pid[this_cpu] != prev_pid)
1682 sched->nr_context_switch_bugs++;
1685 if (sched->tp_handler->switch_event)
1686 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1688 sched->curr_pid[this_cpu] = next_pid;
1692 static int process_sched_runtime_event(struct perf_tool *tool,
1693 struct evsel *evsel,
1694 struct perf_sample *sample,
1695 struct machine *machine)
1697 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1699 if (sched->tp_handler->runtime_event)
1700 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1705 static int perf_sched__process_fork_event(struct perf_tool *tool,
1706 union perf_event *event,
1707 struct perf_sample *sample,
1708 struct machine *machine)
1710 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1712 /* run the fork event through the perf machineruy */
1713 perf_event__process_fork(tool, event, sample, machine);
1715 /* and then run additional processing needed for this command */
1716 if (sched->tp_handler->fork_event)
1717 return sched->tp_handler->fork_event(sched, event, machine);
1722 static int process_sched_migrate_task_event(struct perf_tool *tool,
1723 struct evsel *evsel,
1724 struct perf_sample *sample,
1725 struct machine *machine)
1727 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1729 if (sched->tp_handler->migrate_task_event)
1730 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1735 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1736 struct evsel *evsel,
1737 struct perf_sample *sample,
1738 struct machine *machine);
1740 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1741 union perf_event *event __maybe_unused,
1742 struct perf_sample *sample,
1743 struct evsel *evsel,
1744 struct machine *machine)
1748 if (evsel->handler != NULL) {
1749 tracepoint_handler f = evsel->handler;
1750 err = f(tool, evsel, sample, machine);
1756 static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1757 union perf_event *event,
1758 struct perf_sample *sample,
1759 struct machine *machine)
1761 struct thread *thread;
1762 struct thread_runtime *tr;
1765 err = perf_event__process_comm(tool, event, sample, machine);
1769 thread = machine__find_thread(machine, sample->pid, sample->tid);
1771 pr_err("Internal error: can't find thread\n");
1775 tr = thread__get_runtime(thread);
1777 thread__put(thread);
1781 tr->comm_changed = true;
1782 thread__put(thread);
1787 static int perf_sched__read_events(struct perf_sched *sched)
1789 const struct evsel_str_handler handlers[] = {
1790 { "sched:sched_switch", process_sched_switch_event, },
1791 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1792 { "sched:sched_wakeup", process_sched_wakeup_event, },
1793 { "sched:sched_wakeup_new", process_sched_wakeup_event, },
1794 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1796 struct perf_session *session;
1797 struct perf_data data = {
1799 .mode = PERF_DATA_MODE_READ,
1800 .force = sched->force,
1804 session = perf_session__new(&data, false, &sched->tool);
1805 if (IS_ERR(session)) {
1806 pr_debug("Error creating perf session");
1807 return PTR_ERR(session);
1810 symbol__init(&session->header.env);
1812 if (perf_session__set_tracepoints_handlers(session, handlers))
1815 if (perf_session__has_traces(session, "record -R")) {
1816 int err = perf_session__process_events(session);
1818 pr_err("Failed to process events, error %d", err);
1822 sched->nr_events = session->evlist->stats.nr_events[0];
1823 sched->nr_lost_events = session->evlist->stats.total_lost;
1824 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1829 perf_session__delete(session);
1834 * scheduling times are printed as msec.usec
1836 static inline void print_sched_time(unsigned long long nsecs, int width)
1838 unsigned long msecs;
1839 unsigned long usecs;
1841 msecs = nsecs / NSEC_PER_MSEC;
1842 nsecs -= msecs * NSEC_PER_MSEC;
1843 usecs = nsecs / NSEC_PER_USEC;
1844 printf("%*lu.%03lu ", width, msecs, usecs);
1848 * returns runtime data for event, allocating memory for it the
1849 * first time it is used.
1851 static struct evsel_runtime *evsel__get_runtime(struct evsel *evsel)
1853 struct evsel_runtime *r = evsel->priv;
1856 r = zalloc(sizeof(struct evsel_runtime));
1864 * save last time event was seen per cpu
1866 static void evsel__save_time(struct evsel *evsel, u64 timestamp, u32 cpu)
1868 struct evsel_runtime *r = evsel__get_runtime(evsel);
1873 if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1874 int i, n = __roundup_pow_of_two(cpu+1);
1875 void *p = r->last_time;
1877 p = realloc(r->last_time, n * sizeof(u64));
1882 for (i = r->ncpu; i < n; ++i)
1883 r->last_time[i] = (u64) 0;
1888 r->last_time[cpu] = timestamp;
1891 /* returns last time this event was seen on the given cpu */
1892 static u64 evsel__get_time(struct evsel *evsel, u32 cpu)
1894 struct evsel_runtime *r = evsel__get_runtime(evsel);
1896 if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1899 return r->last_time[cpu];
1902 static int comm_width = 30;
1904 static char *timehist_get_commstr(struct thread *thread)
1906 static char str[32];
1907 const char *comm = thread__comm_str(thread);
1908 pid_t tid = thread->tid;
1909 pid_t pid = thread->pid_;
1913 n = scnprintf(str, sizeof(str), "%s", comm);
1915 else if (tid != pid)
1916 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1919 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1927 static void timehist_header(struct perf_sched *sched)
1929 u32 ncpus = sched->max_cpu + 1;
1932 printf("%15s %6s ", "time", "cpu");
1934 if (sched->show_cpu_visual) {
1936 for (i = 0, j = 0; i < ncpus; ++i) {
1944 printf(" %-*s %9s %9s %9s", comm_width,
1945 "task name", "wait time", "sch delay", "run time");
1947 if (sched->show_state)
1948 printf(" %s", "state");
1955 printf("%15s %-6s ", "", "");
1957 if (sched->show_cpu_visual)
1958 printf(" %*s ", ncpus, "");
1960 printf(" %-*s %9s %9s %9s", comm_width,
1961 "[tid/pid]", "(msec)", "(msec)", "(msec)");
1963 if (sched->show_state)
1971 printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1973 if (sched->show_cpu_visual)
1974 printf(" %.*s ", ncpus, graph_dotted_line);
1976 printf(" %.*s %.9s %.9s %.9s", comm_width,
1977 graph_dotted_line, graph_dotted_line, graph_dotted_line,
1980 if (sched->show_state)
1981 printf(" %.5s", graph_dotted_line);
1986 static char task_state_char(struct thread *thread, int state)
1988 static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
1989 unsigned bit = state ? ffs(state) : 0;
1992 if (thread->tid == 0)
1995 return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
1998 static void timehist_print_sample(struct perf_sched *sched,
1999 struct evsel *evsel,
2000 struct perf_sample *sample,
2001 struct addr_location *al,
2002 struct thread *thread,
2005 struct thread_runtime *tr = thread__priv(thread);
2006 const char *next_comm = evsel__strval(evsel, sample, "next_comm");
2007 const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
2008 u32 max_cpus = sched->max_cpu + 1;
2013 if (cpu_list && !test_bit(sample->cpu, cpu_bitmap))
2016 timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2017 printf("%15s [%04d] ", tstr, sample->cpu);
2019 if (sched->show_cpu_visual) {
2024 for (i = 0; i < max_cpus; ++i) {
2025 /* flag idle times with 'i'; others are sched events */
2026 if (i == sample->cpu)
2027 c = (thread->tid == 0) ? 'i' : 's';
2035 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2037 wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2038 print_sched_time(wait_time, 6);
2040 print_sched_time(tr->dt_delay, 6);
2041 print_sched_time(tr->dt_run, 6);
2043 if (sched->show_state)
2044 printf(" %5c ", task_state_char(thread, state));
2046 if (sched->show_next) {
2047 snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2048 printf(" %-*s", comm_width, nstr);
2051 if (sched->show_wakeups && !sched->show_next)
2052 printf(" %-*s", comm_width, "");
2054 if (thread->tid == 0)
2057 if (sched->show_callchain)
2060 sample__fprintf_sym(sample, al, 0,
2061 EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2062 EVSEL__PRINT_CALLCHAIN_ARROW |
2063 EVSEL__PRINT_SKIP_IGNORED,
2064 &callchain_cursor, symbol_conf.bt_stop_list, stdout);
2071 * Explanation of delta-time stats:
2073 * t = time of current schedule out event
2074 * tprev = time of previous sched out event
2075 * also time of schedule-in event for current task
2076 * last_time = time of last sched change event for current task
2077 * (i.e, time process was last scheduled out)
2078 * ready_to_run = time of wakeup for current task
2080 * -----|------------|------------|------------|------
2081 * last ready tprev t
2084 * |-------- dt_wait --------|
2085 * |- dt_delay -|-- dt_run --|
2087 * dt_run = run time of current task
2088 * dt_wait = time between last schedule out event for task and tprev
2089 * represents time spent off the cpu
2090 * dt_delay = time between wakeup and schedule-in of task
2093 static void timehist_update_runtime_stats(struct thread_runtime *r,
2103 r->dt_run = t - tprev;
2104 if (r->ready_to_run) {
2105 if (r->ready_to_run > tprev)
2106 pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2108 r->dt_delay = tprev - r->ready_to_run;
2111 if (r->last_time > tprev)
2112 pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2113 else if (r->last_time) {
2114 u64 dt_wait = tprev - r->last_time;
2116 if (r->last_state == TASK_RUNNING)
2117 r->dt_preempt = dt_wait;
2118 else if (r->last_state == TASK_UNINTERRUPTIBLE)
2119 r->dt_iowait = dt_wait;
2121 r->dt_sleep = dt_wait;
2125 update_stats(&r->run_stats, r->dt_run);
2127 r->total_run_time += r->dt_run;
2128 r->total_delay_time += r->dt_delay;
2129 r->total_sleep_time += r->dt_sleep;
2130 r->total_iowait_time += r->dt_iowait;
2131 r->total_preempt_time += r->dt_preempt;
2134 static bool is_idle_sample(struct perf_sample *sample,
2135 struct evsel *evsel)
2137 /* pid 0 == swapper == idle task */
2138 if (strcmp(evsel__name(evsel), "sched:sched_switch") == 0)
2139 return evsel__intval(evsel, sample, "prev_pid") == 0;
2141 return sample->pid == 0;
2144 static void save_task_callchain(struct perf_sched *sched,
2145 struct perf_sample *sample,
2146 struct evsel *evsel,
2147 struct machine *machine)
2149 struct callchain_cursor *cursor = &callchain_cursor;
2150 struct thread *thread;
2152 /* want main thread for process - has maps */
2153 thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2154 if (thread == NULL) {
2155 pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2159 if (!sched->show_callchain || sample->callchain == NULL)
2162 if (thread__resolve_callchain(thread, cursor, evsel, sample,
2163 NULL, NULL, sched->max_stack + 2) != 0) {
2165 pr_err("Failed to resolve callchain. Skipping\n");
2170 callchain_cursor_commit(cursor);
2173 struct callchain_cursor_node *node;
2176 node = callchain_cursor_current(cursor);
2182 if (!strcmp(sym->name, "schedule") ||
2183 !strcmp(sym->name, "__schedule") ||
2184 !strcmp(sym->name, "preempt_schedule"))
2188 callchain_cursor_advance(cursor);
2192 static int init_idle_thread(struct thread *thread)
2194 struct idle_thread_runtime *itr;
2196 thread__set_comm(thread, idle_comm, 0);
2198 itr = zalloc(sizeof(*itr));
2202 init_stats(&itr->tr.run_stats);
2203 callchain_init(&itr->callchain);
2204 callchain_cursor_reset(&itr->cursor);
2205 thread__set_priv(thread, itr);
2211 * Track idle stats per cpu by maintaining a local thread
2212 * struct for the idle task on each cpu.
2214 static int init_idle_threads(int ncpu)
2218 idle_threads = zalloc(ncpu * sizeof(struct thread *));
2222 idle_max_cpu = ncpu;
2224 /* allocate the actual thread struct if needed */
2225 for (i = 0; i < ncpu; ++i) {
2226 idle_threads[i] = thread__new(0, 0);
2227 if (idle_threads[i] == NULL)
2230 ret = init_idle_thread(idle_threads[i]);
2238 static void free_idle_threads(void)
2242 if (idle_threads == NULL)
2245 for (i = 0; i < idle_max_cpu; ++i) {
2246 if ((idle_threads[i]))
2247 thread__delete(idle_threads[i]);
2253 static struct thread *get_idle_thread(int cpu)
2256 * expand/allocate array of pointers to local thread
2259 if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2260 int i, j = __roundup_pow_of_two(cpu+1);
2263 p = realloc(idle_threads, j * sizeof(struct thread *));
2267 idle_threads = (struct thread **) p;
2268 for (i = idle_max_cpu; i < j; ++i)
2269 idle_threads[i] = NULL;
2274 /* allocate a new thread struct if needed */
2275 if (idle_threads[cpu] == NULL) {
2276 idle_threads[cpu] = thread__new(0, 0);
2278 if (idle_threads[cpu]) {
2279 if (init_idle_thread(idle_threads[cpu]) < 0)
2284 return idle_threads[cpu];
2287 static void save_idle_callchain(struct perf_sched *sched,
2288 struct idle_thread_runtime *itr,
2289 struct perf_sample *sample)
2291 if (!sched->show_callchain || sample->callchain == NULL)
2294 callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2297 static struct thread *timehist_get_thread(struct perf_sched *sched,
2298 struct perf_sample *sample,
2299 struct machine *machine,
2300 struct evsel *evsel)
2302 struct thread *thread;
2304 if (is_idle_sample(sample, evsel)) {
2305 thread = get_idle_thread(sample->cpu);
2307 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2310 /* there were samples with tid 0 but non-zero pid */
2311 thread = machine__findnew_thread(machine, sample->pid,
2312 sample->tid ?: sample->pid);
2313 if (thread == NULL) {
2314 pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2318 save_task_callchain(sched, sample, evsel, machine);
2319 if (sched->idle_hist) {
2320 struct thread *idle;
2321 struct idle_thread_runtime *itr;
2323 idle = get_idle_thread(sample->cpu);
2325 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2329 itr = thread__priv(idle);
2333 itr->last_thread = thread;
2335 /* copy task callchain when entering to idle */
2336 if (evsel__intval(evsel, sample, "next_pid") == 0)
2337 save_idle_callchain(sched, itr, sample);
2344 static bool timehist_skip_sample(struct perf_sched *sched,
2345 struct thread *thread,
2346 struct evsel *evsel,
2347 struct perf_sample *sample)
2351 if (thread__is_filtered(thread)) {
2353 sched->skipped_samples++;
2356 if (sched->idle_hist) {
2357 if (strcmp(evsel__name(evsel), "sched:sched_switch"))
2359 else if (evsel__intval(evsel, sample, "prev_pid") != 0 &&
2360 evsel__intval(evsel, sample, "next_pid") != 0)
2367 static void timehist_print_wakeup_event(struct perf_sched *sched,
2368 struct evsel *evsel,
2369 struct perf_sample *sample,
2370 struct machine *machine,
2371 struct thread *awakened)
2373 struct thread *thread;
2376 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2380 /* show wakeup unless both awakee and awaker are filtered */
2381 if (timehist_skip_sample(sched, thread, evsel, sample) &&
2382 timehist_skip_sample(sched, awakened, evsel, sample)) {
2386 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2387 printf("%15s [%04d] ", tstr, sample->cpu);
2388 if (sched->show_cpu_visual)
2389 printf(" %*s ", sched->max_cpu + 1, "");
2391 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2394 printf(" %9s %9s %9s ", "", "", "");
2396 printf("awakened: %s", timehist_get_commstr(awakened));
2401 static int timehist_sched_wakeup_ignore(struct perf_tool *tool __maybe_unused,
2402 union perf_event *event __maybe_unused,
2403 struct evsel *evsel __maybe_unused,
2404 struct perf_sample *sample __maybe_unused,
2405 struct machine *machine __maybe_unused)
2410 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2411 union perf_event *event __maybe_unused,
2412 struct evsel *evsel,
2413 struct perf_sample *sample,
2414 struct machine *machine)
2416 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2417 struct thread *thread;
2418 struct thread_runtime *tr = NULL;
2419 /* want pid of awakened task not pid in sample */
2420 const u32 pid = evsel__intval(evsel, sample, "pid");
2422 thread = machine__findnew_thread(machine, 0, pid);
2426 tr = thread__get_runtime(thread);
2430 if (tr->ready_to_run == 0)
2431 tr->ready_to_run = sample->time;
2433 /* show wakeups if requested */
2434 if (sched->show_wakeups &&
2435 !perf_time__skip_sample(&sched->ptime, sample->time))
2436 timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2441 static void timehist_print_migration_event(struct perf_sched *sched,
2442 struct evsel *evsel,
2443 struct perf_sample *sample,
2444 struct machine *machine,
2445 struct thread *migrated)
2447 struct thread *thread;
2449 u32 max_cpus = sched->max_cpu + 1;
2452 if (sched->summary_only)
2455 max_cpus = sched->max_cpu + 1;
2456 ocpu = evsel__intval(evsel, sample, "orig_cpu");
2457 dcpu = evsel__intval(evsel, sample, "dest_cpu");
2459 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2463 if (timehist_skip_sample(sched, thread, evsel, sample) &&
2464 timehist_skip_sample(sched, migrated, evsel, sample)) {
2468 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2469 printf("%15s [%04d] ", tstr, sample->cpu);
2471 if (sched->show_cpu_visual) {
2476 for (i = 0; i < max_cpus; ++i) {
2477 c = (i == sample->cpu) ? 'm' : ' ';
2483 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2486 printf(" %9s %9s %9s ", "", "", "");
2488 printf("migrated: %s", timehist_get_commstr(migrated));
2489 printf(" cpu %d => %d", ocpu, dcpu);
2494 static int timehist_migrate_task_event(struct perf_tool *tool,
2495 union perf_event *event __maybe_unused,
2496 struct evsel *evsel,
2497 struct perf_sample *sample,
2498 struct machine *machine)
2500 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2501 struct thread *thread;
2502 struct thread_runtime *tr = NULL;
2503 /* want pid of migrated task not pid in sample */
2504 const u32 pid = evsel__intval(evsel, sample, "pid");
2506 thread = machine__findnew_thread(machine, 0, pid);
2510 tr = thread__get_runtime(thread);
2516 /* show migrations if requested */
2517 timehist_print_migration_event(sched, evsel, sample, machine, thread);
2522 static int timehist_sched_change_event(struct perf_tool *tool,
2523 union perf_event *event,
2524 struct evsel *evsel,
2525 struct perf_sample *sample,
2526 struct machine *machine)
2528 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2529 struct perf_time_interval *ptime = &sched->ptime;
2530 struct addr_location al;
2531 struct thread *thread;
2532 struct thread_runtime *tr = NULL;
2533 u64 tprev, t = sample->time;
2535 int state = evsel__intval(evsel, sample, "prev_state");
2537 if (machine__resolve(machine, &al, sample) < 0) {
2538 pr_err("problem processing %d event. skipping it\n",
2539 event->header.type);
2544 thread = timehist_get_thread(sched, sample, machine, evsel);
2545 if (thread == NULL) {
2550 if (timehist_skip_sample(sched, thread, evsel, sample))
2553 tr = thread__get_runtime(thread);
2559 tprev = evsel__get_time(evsel, sample->cpu);
2562 * If start time given:
2563 * - sample time is under window user cares about - skip sample
2564 * - tprev is under window user cares about - reset to start of window
2566 if (ptime->start && ptime->start > t)
2569 if (tprev && ptime->start > tprev)
2570 tprev = ptime->start;
2573 * If end time given:
2574 * - previous sched event is out of window - we are done
2575 * - sample time is beyond window user cares about - reset it
2576 * to close out stats for time window interest
2579 if (tprev > ptime->end)
2586 if (!sched->idle_hist || thread->tid == 0) {
2587 timehist_update_runtime_stats(tr, t, tprev);
2589 if (sched->idle_hist) {
2590 struct idle_thread_runtime *itr = (void *)tr;
2591 struct thread_runtime *last_tr;
2593 BUG_ON(thread->tid != 0);
2595 if (itr->last_thread == NULL)
2598 /* add current idle time as last thread's runtime */
2599 last_tr = thread__get_runtime(itr->last_thread);
2600 if (last_tr == NULL)
2603 timehist_update_runtime_stats(last_tr, t, tprev);
2605 * remove delta time of last thread as it's not updated
2606 * and otherwise it will show an invalid value next
2607 * time. we only care total run time and run stat.
2609 last_tr->dt_run = 0;
2610 last_tr->dt_delay = 0;
2611 last_tr->dt_sleep = 0;
2612 last_tr->dt_iowait = 0;
2613 last_tr->dt_preempt = 0;
2616 callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2618 itr->last_thread = NULL;
2622 if (!sched->summary_only)
2623 timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2626 if (sched->hist_time.start == 0 && t >= ptime->start)
2627 sched->hist_time.start = t;
2628 if (ptime->end == 0 || t <= ptime->end)
2629 sched->hist_time.end = t;
2632 /* time of this sched_switch event becomes last time task seen */
2633 tr->last_time = sample->time;
2635 /* last state is used to determine where to account wait time */
2636 tr->last_state = state;
2638 /* sched out event for task so reset ready to run time */
2639 tr->ready_to_run = 0;
2642 evsel__save_time(evsel, sample->time, sample->cpu);
2647 static int timehist_sched_switch_event(struct perf_tool *tool,
2648 union perf_event *event,
2649 struct evsel *evsel,
2650 struct perf_sample *sample,
2651 struct machine *machine __maybe_unused)
2653 return timehist_sched_change_event(tool, event, evsel, sample, machine);
2656 static int process_lost(struct perf_tool *tool __maybe_unused,
2657 union perf_event *event,
2658 struct perf_sample *sample,
2659 struct machine *machine __maybe_unused)
2663 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2664 printf("%15s ", tstr);
2665 printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2671 static void print_thread_runtime(struct thread *t,
2672 struct thread_runtime *r)
2674 double mean = avg_stats(&r->run_stats);
2677 printf("%*s %5d %9" PRIu64 " ",
2678 comm_width, timehist_get_commstr(t), t->ppid,
2679 (u64) r->run_stats.n);
2681 print_sched_time(r->total_run_time, 8);
2682 stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2683 print_sched_time(r->run_stats.min, 6);
2685 print_sched_time((u64) mean, 6);
2687 print_sched_time(r->run_stats.max, 6);
2689 printf("%5.2f", stddev);
2690 printf(" %5" PRIu64, r->migrations);
2694 static void print_thread_waittime(struct thread *t,
2695 struct thread_runtime *r)
2697 printf("%*s %5d %9" PRIu64 " ",
2698 comm_width, timehist_get_commstr(t), t->ppid,
2699 (u64) r->run_stats.n);
2701 print_sched_time(r->total_run_time, 8);
2702 print_sched_time(r->total_sleep_time, 6);
2704 print_sched_time(r->total_iowait_time, 6);
2706 print_sched_time(r->total_preempt_time, 6);
2708 print_sched_time(r->total_delay_time, 6);
2712 struct total_run_stats {
2713 struct perf_sched *sched;
2719 static int __show_thread_runtime(struct thread *t, void *priv)
2721 struct total_run_stats *stats = priv;
2722 struct thread_runtime *r;
2724 if (thread__is_filtered(t))
2727 r = thread__priv(t);
2728 if (r && r->run_stats.n) {
2729 stats->task_count++;
2730 stats->sched_count += r->run_stats.n;
2731 stats->total_run_time += r->total_run_time;
2733 if (stats->sched->show_state)
2734 print_thread_waittime(t, r);
2736 print_thread_runtime(t, r);
2742 static int show_thread_runtime(struct thread *t, void *priv)
2747 return __show_thread_runtime(t, priv);
2750 static int show_deadthread_runtime(struct thread *t, void *priv)
2755 return __show_thread_runtime(t, priv);
2758 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2760 const char *sep = " <- ";
2761 struct callchain_list *chain;
2769 ret = callchain__fprintf_folded(fp, node->parent);
2772 list_for_each_entry(chain, &node->val, list) {
2773 if (chain->ip >= PERF_CONTEXT_MAX)
2775 if (chain->ms.sym && chain->ms.sym->ignore)
2777 ret += fprintf(fp, "%s%s", first ? "" : sep,
2778 callchain_list__sym_name(chain, bf, sizeof(bf),
2786 static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2790 struct callchain_node *chain;
2791 struct rb_node *rb_node = rb_first_cached(root);
2793 printf(" %16s %8s %s\n", "Idle time (msec)", "Count", "Callchains");
2794 printf(" %.16s %.8s %.50s\n", graph_dotted_line, graph_dotted_line,
2798 chain = rb_entry(rb_node, struct callchain_node, rb_node);
2799 rb_node = rb_next(rb_node);
2801 ret += fprintf(fp, " ");
2802 print_sched_time(chain->hit, 12);
2803 ret += 16; /* print_sched_time returns 2nd arg + 4 */
2804 ret += fprintf(fp, " %8d ", chain->count);
2805 ret += callchain__fprintf_folded(fp, chain);
2806 ret += fprintf(fp, "\n");
2812 static void timehist_print_summary(struct perf_sched *sched,
2813 struct perf_session *session)
2815 struct machine *m = &session->machines.host;
2816 struct total_run_stats totals;
2819 struct thread_runtime *r;
2821 u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2823 memset(&totals, 0, sizeof(totals));
2824 totals.sched = sched;
2826 if (sched->idle_hist) {
2827 printf("\nIdle-time summary\n");
2828 printf("%*s parent sched-out ", comm_width, "comm");
2829 printf(" idle-time min-idle avg-idle max-idle stddev migrations\n");
2830 } else if (sched->show_state) {
2831 printf("\nWait-time summary\n");
2832 printf("%*s parent sched-in ", comm_width, "comm");
2833 printf(" run-time sleep iowait preempt delay\n");
2835 printf("\nRuntime summary\n");
2836 printf("%*s parent sched-in ", comm_width, "comm");
2837 printf(" run-time min-run avg-run max-run stddev migrations\n");
2839 printf("%*s (count) ", comm_width, "");
2840 printf(" (msec) (msec) (msec) (msec) %s\n",
2841 sched->show_state ? "(msec)" : "%");
2842 printf("%.117s\n", graph_dotted_line);
2844 machine__for_each_thread(m, show_thread_runtime, &totals);
2845 task_count = totals.task_count;
2847 printf("<no still running tasks>\n");
2849 printf("\nTerminated tasks:\n");
2850 machine__for_each_thread(m, show_deadthread_runtime, &totals);
2851 if (task_count == totals.task_count)
2852 printf("<no terminated tasks>\n");
2854 /* CPU idle stats not tracked when samples were skipped */
2855 if (sched->skipped_samples && !sched->idle_hist)
2858 printf("\nIdle stats:\n");
2859 for (i = 0; i < idle_max_cpu; ++i) {
2860 t = idle_threads[i];
2864 r = thread__priv(t);
2865 if (r && r->run_stats.n) {
2866 totals.sched_count += r->run_stats.n;
2867 printf(" CPU %2d idle for ", i);
2868 print_sched_time(r->total_run_time, 6);
2869 printf(" msec (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2871 printf(" CPU %2d idle entire time window\n", i);
2874 if (sched->idle_hist && sched->show_callchain) {
2875 callchain_param.mode = CHAIN_FOLDED;
2876 callchain_param.value = CCVAL_PERIOD;
2878 callchain_register_param(&callchain_param);
2880 printf("\nIdle stats by callchain:\n");
2881 for (i = 0; i < idle_max_cpu; ++i) {
2882 struct idle_thread_runtime *itr;
2884 t = idle_threads[i];
2888 itr = thread__priv(t);
2892 callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
2893 0, &callchain_param);
2895 printf(" CPU %2d:", i);
2896 print_sched_time(itr->tr.total_run_time, 6);
2898 timehist_print_idlehist_callchain(&itr->sorted_root);
2904 " Total number of unique tasks: %" PRIu64 "\n"
2905 "Total number of context switches: %" PRIu64 "\n",
2906 totals.task_count, totals.sched_count);
2908 printf(" Total run time (msec): ");
2909 print_sched_time(totals.total_run_time, 2);
2912 printf(" Total scheduling time (msec): ");
2913 print_sched_time(hist_time, 2);
2914 printf(" (x %d)\n", sched->max_cpu);
2917 typedef int (*sched_handler)(struct perf_tool *tool,
2918 union perf_event *event,
2919 struct evsel *evsel,
2920 struct perf_sample *sample,
2921 struct machine *machine);
2923 static int perf_timehist__process_sample(struct perf_tool *tool,
2924 union perf_event *event,
2925 struct perf_sample *sample,
2926 struct evsel *evsel,
2927 struct machine *machine)
2929 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2931 int this_cpu = sample->cpu;
2933 if (this_cpu > sched->max_cpu)
2934 sched->max_cpu = this_cpu;
2936 if (evsel->handler != NULL) {
2937 sched_handler f = evsel->handler;
2939 err = f(tool, event, evsel, sample, machine);
2945 static int timehist_check_attr(struct perf_sched *sched,
2946 struct evlist *evlist)
2948 struct evsel *evsel;
2949 struct evsel_runtime *er;
2951 list_for_each_entry(evsel, &evlist->core.entries, core.node) {
2952 er = evsel__get_runtime(evsel);
2954 pr_err("Failed to allocate memory for evsel runtime data\n");
2958 if (sched->show_callchain && !evsel__has_callchain(evsel)) {
2959 pr_info("Samples do not have callchains.\n");
2960 sched->show_callchain = 0;
2961 symbol_conf.use_callchain = 0;
2968 static int perf_sched__timehist(struct perf_sched *sched)
2970 struct evsel_str_handler handlers[] = {
2971 { "sched:sched_switch", timehist_sched_switch_event, },
2972 { "sched:sched_wakeup", timehist_sched_wakeup_event, },
2973 { "sched:sched_waking", timehist_sched_wakeup_event, },
2974 { "sched:sched_wakeup_new", timehist_sched_wakeup_event, },
2976 const struct evsel_str_handler migrate_handlers[] = {
2977 { "sched:sched_migrate_task", timehist_migrate_task_event, },
2979 struct perf_data data = {
2981 .mode = PERF_DATA_MODE_READ,
2982 .force = sched->force,
2985 struct perf_session *session;
2986 struct evlist *evlist;
2990 * event handlers for timehist option
2992 sched->tool.sample = perf_timehist__process_sample;
2993 sched->tool.mmap = perf_event__process_mmap;
2994 sched->tool.comm = perf_event__process_comm;
2995 sched->tool.exit = perf_event__process_exit;
2996 sched->tool.fork = perf_event__process_fork;
2997 sched->tool.lost = process_lost;
2998 sched->tool.attr = perf_event__process_attr;
2999 sched->tool.tracing_data = perf_event__process_tracing_data;
3000 sched->tool.build_id = perf_event__process_build_id;
3002 sched->tool.ordered_events = true;
3003 sched->tool.ordering_requires_timestamps = true;
3005 symbol_conf.use_callchain = sched->show_callchain;
3007 session = perf_session__new(&data, false, &sched->tool);
3008 if (IS_ERR(session))
3009 return PTR_ERR(session);
3012 err = perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap);
3017 evlist = session->evlist;
3019 symbol__init(&session->header.env);
3021 if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
3022 pr_err("Invalid time string\n");
3026 if (timehist_check_attr(sched, evlist) != 0)
3031 /* prefer sched_waking if it is captured */
3032 if (perf_evlist__find_tracepoint_by_name(session->evlist,
3033 "sched:sched_waking"))
3034 handlers[1].handler = timehist_sched_wakeup_ignore;
3036 /* setup per-evsel handlers */
3037 if (perf_session__set_tracepoints_handlers(session, handlers))
3040 /* sched_switch event at a minimum needs to exist */
3041 if (!perf_evlist__find_tracepoint_by_name(session->evlist,
3042 "sched:sched_switch")) {
3043 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3047 if (sched->show_migrations &&
3048 perf_session__set_tracepoints_handlers(session, migrate_handlers))
3051 /* pre-allocate struct for per-CPU idle stats */
3052 sched->max_cpu = session->header.env.nr_cpus_online;
3053 if (sched->max_cpu == 0)
3055 if (init_idle_threads(sched->max_cpu))
3058 /* summary_only implies summary option, but don't overwrite summary if set */
3059 if (sched->summary_only)
3060 sched->summary = sched->summary_only;
3062 if (!sched->summary_only)
3063 timehist_header(sched);
3065 err = perf_session__process_events(session);
3067 pr_err("Failed to process events, error %d", err);
3071 sched->nr_events = evlist->stats.nr_events[0];
3072 sched->nr_lost_events = evlist->stats.total_lost;
3073 sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3076 timehist_print_summary(sched, session);
3079 free_idle_threads();
3080 perf_session__delete(session);
3086 static void print_bad_events(struct perf_sched *sched)
3088 if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3089 printf(" INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3090 (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3091 sched->nr_unordered_timestamps, sched->nr_timestamps);
3093 if (sched->nr_lost_events && sched->nr_events) {
3094 printf(" INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3095 (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3096 sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3098 if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3099 printf(" INFO: %.3f%% context switch bugs (%ld out of %ld)",
3100 (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3101 sched->nr_context_switch_bugs, sched->nr_timestamps);
3102 if (sched->nr_lost_events)
3103 printf(" (due to lost events?)");
3108 static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3110 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3111 struct work_atoms *this;
3112 const char *comm = thread__comm_str(data->thread), *this_comm;
3113 bool leftmost = true;
3118 this = container_of(*new, struct work_atoms, node);
3121 this_comm = thread__comm_str(this->thread);
3122 cmp = strcmp(comm, this_comm);
3124 new = &((*new)->rb_left);
3125 } else if (cmp < 0) {
3126 new = &((*new)->rb_right);
3130 this->total_runtime += data->total_runtime;
3131 this->nb_atoms += data->nb_atoms;
3132 this->total_lat += data->total_lat;
3133 list_splice(&data->work_list, &this->work_list);
3134 if (this->max_lat < data->max_lat) {
3135 this->max_lat = data->max_lat;
3136 this->max_lat_at = data->max_lat_at;
3144 rb_link_node(&data->node, parent, new);
3145 rb_insert_color_cached(&data->node, root, leftmost);
3148 static void perf_sched__merge_lat(struct perf_sched *sched)
3150 struct work_atoms *data;
3151 struct rb_node *node;
3153 if (sched->skip_merge)
3156 while ((node = rb_first_cached(&sched->atom_root))) {
3157 rb_erase_cached(node, &sched->atom_root);
3158 data = rb_entry(node, struct work_atoms, node);
3159 __merge_work_atoms(&sched->merged_atom_root, data);
3163 static int perf_sched__lat(struct perf_sched *sched)
3165 struct rb_node *next;
3169 if (perf_sched__read_events(sched))
3172 perf_sched__merge_lat(sched);
3173 perf_sched__sort_lat(sched);
3175 printf("\n -----------------------------------------------------------------------------------------------------------------\n");
3176 printf(" Task | Runtime ms | Switches | Average delay ms | Maximum delay ms | Maximum delay at |\n");
3177 printf(" -----------------------------------------------------------------------------------------------------------------\n");
3179 next = rb_first_cached(&sched->sorted_atom_root);
3182 struct work_atoms *work_list;
3184 work_list = rb_entry(next, struct work_atoms, node);
3185 output_lat_thread(sched, work_list);
3186 next = rb_next(next);
3187 thread__zput(work_list->thread);
3190 printf(" -----------------------------------------------------------------------------------------------------------------\n");
3191 printf(" TOTAL: |%11.3f ms |%9" PRIu64 " |\n",
3192 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3194 printf(" ---------------------------------------------------\n");
3196 print_bad_events(sched);
3202 static int setup_map_cpus(struct perf_sched *sched)
3204 struct perf_cpu_map *map;
3206 sched->max_cpu = sysconf(_SC_NPROCESSORS_CONF);
3208 if (sched->map.comp) {
3209 sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
3210 if (!sched->map.comp_cpus)
3214 if (!sched->map.cpus_str)
3217 map = perf_cpu_map__new(sched->map.cpus_str);
3219 pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3223 sched->map.cpus = map;
3227 static int setup_color_pids(struct perf_sched *sched)
3229 struct perf_thread_map *map;
3231 if (!sched->map.color_pids_str)
3234 map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3236 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3240 sched->map.color_pids = map;
3244 static int setup_color_cpus(struct perf_sched *sched)
3246 struct perf_cpu_map *map;
3248 if (!sched->map.color_cpus_str)
3251 map = perf_cpu_map__new(sched->map.color_cpus_str);
3253 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3257 sched->map.color_cpus = map;
3261 static int perf_sched__map(struct perf_sched *sched)
3263 if (setup_map_cpus(sched))
3266 if (setup_color_pids(sched))
3269 if (setup_color_cpus(sched))
3273 if (perf_sched__read_events(sched))
3275 print_bad_events(sched);
3279 static int perf_sched__replay(struct perf_sched *sched)
3283 calibrate_run_measurement_overhead(sched);
3284 calibrate_sleep_measurement_overhead(sched);
3286 test_calibrations(sched);
3288 if (perf_sched__read_events(sched))
3291 printf("nr_run_events: %ld\n", sched->nr_run_events);
3292 printf("nr_sleep_events: %ld\n", sched->nr_sleep_events);
3293 printf("nr_wakeup_events: %ld\n", sched->nr_wakeup_events);
3295 if (sched->targetless_wakeups)
3296 printf("target-less wakeups: %ld\n", sched->targetless_wakeups);
3297 if (sched->multitarget_wakeups)
3298 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3299 if (sched->nr_run_events_optimized)
3300 printf("run atoms optimized: %ld\n",
3301 sched->nr_run_events_optimized);
3303 print_task_traces(sched);
3304 add_cross_task_wakeups(sched);
3306 create_tasks(sched);
3307 printf("------------------------------------------------------------\n");
3308 for (i = 0; i < sched->replay_repeat; i++)
3309 run_one_test(sched);
3314 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3315 const char * const usage_msg[])
3317 char *tmp, *tok, *str = strdup(sched->sort_order);
3319 for (tok = strtok_r(str, ", ", &tmp);
3320 tok; tok = strtok_r(NULL, ", ", &tmp)) {
3321 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3322 usage_with_options_msg(usage_msg, options,
3323 "Unknown --sort key: `%s'", tok);
3329 sort_dimension__add("pid", &sched->cmp_pid);
3332 static int __cmd_record(int argc, const char **argv)
3334 unsigned int rec_argc, i, j;
3335 const char **rec_argv;
3336 const char * const record_args[] = {
3342 "-e", "sched:sched_switch",
3343 "-e", "sched:sched_stat_wait",
3344 "-e", "sched:sched_stat_sleep",
3345 "-e", "sched:sched_stat_iowait",
3346 "-e", "sched:sched_stat_runtime",
3347 "-e", "sched:sched_process_fork",
3348 "-e", "sched:sched_wakeup_new",
3349 "-e", "sched:sched_migrate_task",
3351 struct tep_event *waking_event;
3354 * +2 for either "-e", "sched:sched_wakeup" or
3355 * "-e", "sched:sched_waking"
3357 rec_argc = ARRAY_SIZE(record_args) + 2 + argc - 1;
3358 rec_argv = calloc(rec_argc + 1, sizeof(char *));
3360 if (rec_argv == NULL)
3363 for (i = 0; i < ARRAY_SIZE(record_args); i++)
3364 rec_argv[i] = strdup(record_args[i]);
3366 rec_argv[i++] = "-e";
3367 waking_event = trace_event__tp_format("sched", "sched_waking");
3368 if (!IS_ERR(waking_event))
3369 rec_argv[i++] = strdup("sched:sched_waking");
3371 rec_argv[i++] = strdup("sched:sched_wakeup");
3373 for (j = 1; j < (unsigned int)argc; j++, i++)
3374 rec_argv[i] = argv[j];
3376 BUG_ON(i != rec_argc);
3378 return cmd_record(i, rec_argv);
3381 int cmd_sched(int argc, const char **argv)
3383 static const char default_sort_order[] = "avg, max, switch, runtime";
3384 struct perf_sched sched = {
3386 .sample = perf_sched__process_tracepoint_sample,
3387 .comm = perf_sched__process_comm,
3388 .namespaces = perf_event__process_namespaces,
3389 .lost = perf_event__process_lost,
3390 .fork = perf_sched__process_fork_event,
3391 .ordered_events = true,
3393 .cmp_pid = LIST_HEAD_INIT(sched.cmp_pid),
3394 .sort_list = LIST_HEAD_INIT(sched.sort_list),
3395 .start_work_mutex = PTHREAD_MUTEX_INITIALIZER,
3396 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
3397 .sort_order = default_sort_order,
3398 .replay_repeat = 10,
3400 .next_shortname1 = 'A',
3401 .next_shortname2 = '0',
3403 .show_callchain = 1,
3406 const struct option sched_options[] = {
3407 OPT_STRING('i', "input", &input_name, "file",
3409 OPT_INCR('v', "verbose", &verbose,
3410 "be more verbose (show symbol address, etc)"),
3411 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3412 "dump raw trace in ASCII"),
3413 OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3416 const struct option latency_options[] = {
3417 OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3418 "sort by key(s): runtime, switch, avg, max"),
3419 OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3420 "CPU to profile on"),
3421 OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3422 "latency stats per pid instead of per comm"),
3423 OPT_PARENT(sched_options)
3425 const struct option replay_options[] = {
3426 OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3427 "repeat the workload replay N times (-1: infinite)"),
3428 OPT_PARENT(sched_options)
3430 const struct option map_options[] = {
3431 OPT_BOOLEAN(0, "compact", &sched.map.comp,
3432 "map output in compact mode"),
3433 OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3434 "highlight given pids in map"),
3435 OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3436 "highlight given CPUs in map"),
3437 OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3438 "display given CPUs in map"),
3439 OPT_PARENT(sched_options)
3441 const struct option timehist_options[] = {
3442 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3443 "file", "vmlinux pathname"),
3444 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3445 "file", "kallsyms pathname"),
3446 OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3447 "Display call chains if present (default on)"),
3448 OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3449 "Maximum number of functions to display backtrace."),
3450 OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3451 "Look for files with symbols relative to this directory"),
3452 OPT_BOOLEAN('s', "summary", &sched.summary_only,
3453 "Show only syscall summary with statistics"),
3454 OPT_BOOLEAN('S', "with-summary", &sched.summary,
3455 "Show all syscalls and summary with statistics"),
3456 OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3457 OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3458 OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3459 OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3460 OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3461 OPT_STRING(0, "time", &sched.time_str, "str",
3462 "Time span for analysis (start,stop)"),
3463 OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3464 OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3465 "analyze events only for given process id(s)"),
3466 OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3467 "analyze events only for given thread id(s)"),
3468 OPT_STRING('C', "cpu", &cpu_list, "cpu", "list of cpus to profile"),
3469 OPT_PARENT(sched_options)
3472 const char * const latency_usage[] = {
3473 "perf sched latency [<options>]",
3476 const char * const replay_usage[] = {
3477 "perf sched replay [<options>]",
3480 const char * const map_usage[] = {
3481 "perf sched map [<options>]",
3484 const char * const timehist_usage[] = {
3485 "perf sched timehist [<options>]",
3488 const char *const sched_subcommands[] = { "record", "latency", "map",
3491 const char *sched_usage[] = {
3495 struct trace_sched_handler lat_ops = {
3496 .wakeup_event = latency_wakeup_event,
3497 .switch_event = latency_switch_event,
3498 .runtime_event = latency_runtime_event,
3499 .migrate_task_event = latency_migrate_task_event,
3501 struct trace_sched_handler map_ops = {
3502 .switch_event = map_switch_event,
3504 struct trace_sched_handler replay_ops = {
3505 .wakeup_event = replay_wakeup_event,
3506 .switch_event = replay_switch_event,
3507 .fork_event = replay_fork_event,
3511 for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3512 sched.curr_pid[i] = -1;
3514 argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3515 sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3517 usage_with_options(sched_usage, sched_options);
3520 * Aliased to 'perf script' for now:
3522 if (!strcmp(argv[0], "script"))
3523 return cmd_script(argc, argv);
3525 if (!strncmp(argv[0], "rec", 3)) {
3526 return __cmd_record(argc, argv);
3527 } else if (!strncmp(argv[0], "lat", 3)) {
3528 sched.tp_handler = &lat_ops;
3530 argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3532 usage_with_options(latency_usage, latency_options);
3534 setup_sorting(&sched, latency_options, latency_usage);
3535 return perf_sched__lat(&sched);
3536 } else if (!strcmp(argv[0], "map")) {
3538 argc = parse_options(argc, argv, map_options, map_usage, 0);
3540 usage_with_options(map_usage, map_options);
3542 sched.tp_handler = &map_ops;
3543 setup_sorting(&sched, latency_options, latency_usage);
3544 return perf_sched__map(&sched);
3545 } else if (!strncmp(argv[0], "rep", 3)) {
3546 sched.tp_handler = &replay_ops;
3548 argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3550 usage_with_options(replay_usage, replay_options);
3552 return perf_sched__replay(&sched);
3553 } else if (!strcmp(argv[0], "timehist")) {
3555 argc = parse_options(argc, argv, timehist_options,
3558 usage_with_options(timehist_usage, timehist_options);
3560 if ((sched.show_wakeups || sched.show_next) &&
3561 sched.summary_only) {
3562 pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3563 parse_options_usage(timehist_usage, timehist_options, "s", true);
3564 if (sched.show_wakeups)
3565 parse_options_usage(NULL, timehist_options, "w", true);
3566 if (sched.show_next)
3567 parse_options_usage(NULL, timehist_options, "n", true);
3571 return perf_sched__timehist(&sched);
3573 usage_with_options(sched_usage, sched_options);