1 // SPDX-License-Identifier: GPL-2.0
6 #include "util/evlist.h"
7 #include "util/cache.h"
8 #include "util/evsel.h"
9 #include "util/symbol.h"
10 #include "util/thread.h"
11 #include "util/header.h"
12 #include "util/session.h"
13 #include "util/tool.h"
14 #include "util/cloexec.h"
15 #include "util/thread_map.h"
16 #include "util/color.h"
17 #include "util/stat.h"
18 #include "util/callchain.h"
19 #include "util/time-utils.h"
21 #include <subcmd/parse-options.h>
22 #include "util/trace-event.h"
24 #include "util/debug.h"
26 #include <linux/kernel.h>
27 #include <linux/log2.h>
28 #include <sys/prctl.h>
29 #include <sys/resource.h>
33 #include <semaphore.h>
36 #include <api/fs/fs.h>
37 #include <linux/time64.h>
39 #include "sane_ctype.h"
41 #define PR_SET_NAME 15 /* Set process name */
45 #define MAX_PID 1024000
54 unsigned long nr_events;
55 unsigned long curr_event;
56 struct sched_atom **atoms;
67 enum sched_event_type {
71 SCHED_EVENT_MIGRATION,
75 enum sched_event_type type;
81 struct task_desc *wakee;
84 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
86 /* task state bitmask, copied from include/linux/sched.h */
87 #define TASK_RUNNING 0
88 #define TASK_INTERRUPTIBLE 1
89 #define TASK_UNINTERRUPTIBLE 2
90 #define __TASK_STOPPED 4
91 #define __TASK_TRACED 8
92 /* in tsk->exit_state */
94 #define EXIT_ZOMBIE 32
95 #define EXIT_TRACE (EXIT_ZOMBIE | EXIT_DEAD)
96 /* in tsk->state again */
98 #define TASK_WAKEKILL 128
99 #define TASK_WAKING 256
100 #define TASK_PARKED 512
110 struct list_head list;
111 enum thread_state state;
119 struct list_head work_list;
120 struct thread *thread;
130 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
134 struct trace_sched_handler {
135 int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
136 struct perf_sample *sample, struct machine *machine);
138 int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
139 struct perf_sample *sample, struct machine *machine);
141 int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
142 struct perf_sample *sample, struct machine *machine);
144 /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
145 int (*fork_event)(struct perf_sched *sched, union perf_event *event,
146 struct machine *machine);
148 int (*migrate_task_event)(struct perf_sched *sched,
149 struct perf_evsel *evsel,
150 struct perf_sample *sample,
151 struct machine *machine);
154 #define COLOR_PIDS PERF_COLOR_BLUE
155 #define COLOR_CPUS PERF_COLOR_BG_RED
157 struct perf_sched_map {
158 DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
161 struct thread_map *color_pids;
162 const char *color_pids_str;
163 struct cpu_map *color_cpus;
164 const char *color_cpus_str;
165 struct cpu_map *cpus;
166 const char *cpus_str;
170 struct perf_tool tool;
171 const char *sort_order;
172 unsigned long nr_tasks;
173 struct task_desc **pid_to_task;
174 struct task_desc **tasks;
175 const struct trace_sched_handler *tp_handler;
176 pthread_mutex_t start_work_mutex;
177 pthread_mutex_t work_done_wait_mutex;
180 * Track the current task - that way we can know whether there's any
181 * weird events, such as a task being switched away that is not current.
184 u32 curr_pid[MAX_CPUS];
185 struct thread *curr_thread[MAX_CPUS];
186 char next_shortname1;
187 char next_shortname2;
188 unsigned int replay_repeat;
189 unsigned long nr_run_events;
190 unsigned long nr_sleep_events;
191 unsigned long nr_wakeup_events;
192 unsigned long nr_sleep_corrections;
193 unsigned long nr_run_events_optimized;
194 unsigned long targetless_wakeups;
195 unsigned long multitarget_wakeups;
196 unsigned long nr_runs;
197 unsigned long nr_timestamps;
198 unsigned long nr_unordered_timestamps;
199 unsigned long nr_context_switch_bugs;
200 unsigned long nr_events;
201 unsigned long nr_lost_chunks;
202 unsigned long nr_lost_events;
203 u64 run_measurement_overhead;
204 u64 sleep_measurement_overhead;
207 u64 runavg_cpu_usage;
208 u64 parent_cpu_usage;
209 u64 runavg_parent_cpu_usage;
215 u64 cpu_last_switched[MAX_CPUS];
216 struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
217 struct list_head sort_list, cmp_pid;
220 struct perf_sched_map map;
222 /* options for timehist command */
227 unsigned int max_stack;
228 bool show_cpu_visual;
231 bool show_migrations;
234 const char *time_str;
235 struct perf_time_interval ptime;
236 struct perf_time_interval hist_time;
239 /* per thread run time data */
240 struct thread_runtime {
241 u64 last_time; /* time of previous sched in/out event */
242 u64 dt_run; /* run time */
243 u64 dt_sleep; /* time between CPU access by sleep (off cpu) */
244 u64 dt_iowait; /* time between CPU access by iowait (off cpu) */
245 u64 dt_preempt; /* time between CPU access by preempt (off cpu) */
246 u64 dt_delay; /* time between wakeup and sched-in */
247 u64 ready_to_run; /* time of wakeup */
249 struct stats run_stats;
251 u64 total_sleep_time;
252 u64 total_iowait_time;
253 u64 total_preempt_time;
254 u64 total_delay_time;
264 /* per event run time data */
265 struct evsel_runtime {
266 u64 *last_time; /* time this event was last seen per cpu */
267 u32 ncpu; /* highest cpu slot allocated */
270 /* per cpu idle time data */
271 struct idle_thread_runtime {
272 struct thread_runtime tr;
273 struct thread *last_thread;
274 struct rb_root_cached sorted_root;
275 struct callchain_root callchain;
276 struct callchain_cursor cursor;
279 /* track idle times per cpu */
280 static struct thread **idle_threads;
281 static int idle_max_cpu;
282 static char idle_comm[] = "<idle>";
284 static u64 get_nsecs(void)
288 clock_gettime(CLOCK_MONOTONIC, &ts);
290 return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
293 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
295 u64 T0 = get_nsecs(), T1;
299 } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
302 static void sleep_nsecs(u64 nsecs)
306 ts.tv_nsec = nsecs % 999999999;
307 ts.tv_sec = nsecs / 999999999;
309 nanosleep(&ts, NULL);
312 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
314 u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
317 for (i = 0; i < 10; i++) {
319 burn_nsecs(sched, 0);
322 min_delta = min(min_delta, delta);
324 sched->run_measurement_overhead = min_delta;
326 printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
329 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
331 u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
334 for (i = 0; i < 10; i++) {
339 min_delta = min(min_delta, delta);
342 sched->sleep_measurement_overhead = min_delta;
344 printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
347 static struct sched_atom *
348 get_new_event(struct task_desc *task, u64 timestamp)
350 struct sched_atom *event = zalloc(sizeof(*event));
351 unsigned long idx = task->nr_events;
354 event->timestamp = timestamp;
358 size = sizeof(struct sched_atom *) * task->nr_events;
359 task->atoms = realloc(task->atoms, size);
360 BUG_ON(!task->atoms);
362 task->atoms[idx] = event;
367 static struct sched_atom *last_event(struct task_desc *task)
369 if (!task->nr_events)
372 return task->atoms[task->nr_events - 1];
375 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
376 u64 timestamp, u64 duration)
378 struct sched_atom *event, *curr_event = last_event(task);
381 * optimize an existing RUN event by merging this one
384 if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
385 sched->nr_run_events_optimized++;
386 curr_event->duration += duration;
390 event = get_new_event(task, timestamp);
392 event->type = SCHED_EVENT_RUN;
393 event->duration = duration;
395 sched->nr_run_events++;
398 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
399 u64 timestamp, struct task_desc *wakee)
401 struct sched_atom *event, *wakee_event;
403 event = get_new_event(task, timestamp);
404 event->type = SCHED_EVENT_WAKEUP;
405 event->wakee = wakee;
407 wakee_event = last_event(wakee);
408 if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
409 sched->targetless_wakeups++;
412 if (wakee_event->wait_sem) {
413 sched->multitarget_wakeups++;
417 wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
418 sem_init(wakee_event->wait_sem, 0, 0);
419 wakee_event->specific_wait = 1;
420 event->wait_sem = wakee_event->wait_sem;
422 sched->nr_wakeup_events++;
425 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
426 u64 timestamp, u64 task_state __maybe_unused)
428 struct sched_atom *event = get_new_event(task, timestamp);
430 event->type = SCHED_EVENT_SLEEP;
432 sched->nr_sleep_events++;
435 static struct task_desc *register_pid(struct perf_sched *sched,
436 unsigned long pid, const char *comm)
438 struct task_desc *task;
441 if (sched->pid_to_task == NULL) {
442 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
444 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
446 if (pid >= (unsigned long)pid_max) {
447 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
448 sizeof(struct task_desc *))) == NULL);
449 while (pid >= (unsigned long)pid_max)
450 sched->pid_to_task[pid_max++] = NULL;
453 task = sched->pid_to_task[pid];
458 task = zalloc(sizeof(*task));
460 task->nr = sched->nr_tasks;
461 strcpy(task->comm, comm);
463 * every task starts in sleeping state - this gets ignored
464 * if there's no wakeup pointing to this sleep state:
466 add_sched_event_sleep(sched, task, 0, 0);
468 sched->pid_to_task[pid] = task;
470 sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
471 BUG_ON(!sched->tasks);
472 sched->tasks[task->nr] = task;
475 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
481 static void print_task_traces(struct perf_sched *sched)
483 struct task_desc *task;
486 for (i = 0; i < sched->nr_tasks; i++) {
487 task = sched->tasks[i];
488 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
489 task->nr, task->comm, task->pid, task->nr_events);
493 static void add_cross_task_wakeups(struct perf_sched *sched)
495 struct task_desc *task1, *task2;
498 for (i = 0; i < sched->nr_tasks; i++) {
499 task1 = sched->tasks[i];
501 if (j == sched->nr_tasks)
503 task2 = sched->tasks[j];
504 add_sched_event_wakeup(sched, task1, 0, task2);
508 static void perf_sched__process_event(struct perf_sched *sched,
509 struct sched_atom *atom)
513 switch (atom->type) {
514 case SCHED_EVENT_RUN:
515 burn_nsecs(sched, atom->duration);
517 case SCHED_EVENT_SLEEP:
519 ret = sem_wait(atom->wait_sem);
522 case SCHED_EVENT_WAKEUP:
524 ret = sem_post(atom->wait_sem);
527 case SCHED_EVENT_MIGRATION:
534 static u64 get_cpu_usage_nsec_parent(void)
540 err = getrusage(RUSAGE_SELF, &ru);
543 sum = ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
544 sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
549 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
551 struct perf_event_attr attr;
552 char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
555 bool need_privilege = false;
557 memset(&attr, 0, sizeof(attr));
559 attr.type = PERF_TYPE_SOFTWARE;
560 attr.config = PERF_COUNT_SW_TASK_CLOCK;
563 fd = sys_perf_event_open(&attr, 0, -1, -1,
564 perf_event_open_cloexec_flag());
567 if (errno == EMFILE) {
569 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
570 limit.rlim_cur += sched->nr_tasks - cur_task;
571 if (limit.rlim_cur > limit.rlim_max) {
572 limit.rlim_max = limit.rlim_cur;
573 need_privilege = true;
575 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
576 if (need_privilege && errno == EPERM)
577 strcpy(info, "Need privilege\n");
581 strcpy(info, "Have a try with -f option\n");
583 pr_err("Error: sys_perf_event_open() syscall returned "
584 "with %d (%s)\n%s", fd,
585 str_error_r(errno, sbuf, sizeof(sbuf)), info);
591 static u64 get_cpu_usage_nsec_self(int fd)
596 ret = read(fd, &runtime, sizeof(runtime));
597 BUG_ON(ret != sizeof(runtime));
602 struct sched_thread_parms {
603 struct task_desc *task;
604 struct perf_sched *sched;
608 static void *thread_func(void *ctx)
610 struct sched_thread_parms *parms = ctx;
611 struct task_desc *this_task = parms->task;
612 struct perf_sched *sched = parms->sched;
613 u64 cpu_usage_0, cpu_usage_1;
614 unsigned long i, ret;
620 sprintf(comm2, ":%s", this_task->comm);
621 prctl(PR_SET_NAME, comm2);
625 ret = sem_post(&this_task->ready_for_work);
627 ret = pthread_mutex_lock(&sched->start_work_mutex);
629 ret = pthread_mutex_unlock(&sched->start_work_mutex);
632 cpu_usage_0 = get_cpu_usage_nsec_self(fd);
634 for (i = 0; i < this_task->nr_events; i++) {
635 this_task->curr_event = i;
636 perf_sched__process_event(sched, this_task->atoms[i]);
639 cpu_usage_1 = get_cpu_usage_nsec_self(fd);
640 this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
641 ret = sem_post(&this_task->work_done_sem);
644 ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
646 ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
652 static void create_tasks(struct perf_sched *sched)
654 struct task_desc *task;
659 err = pthread_attr_init(&attr);
661 err = pthread_attr_setstacksize(&attr,
662 (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
664 err = pthread_mutex_lock(&sched->start_work_mutex);
666 err = pthread_mutex_lock(&sched->work_done_wait_mutex);
668 for (i = 0; i < sched->nr_tasks; i++) {
669 struct sched_thread_parms *parms = malloc(sizeof(*parms));
670 BUG_ON(parms == NULL);
671 parms->task = task = sched->tasks[i];
672 parms->sched = sched;
673 parms->fd = self_open_counters(sched, i);
674 sem_init(&task->sleep_sem, 0, 0);
675 sem_init(&task->ready_for_work, 0, 0);
676 sem_init(&task->work_done_sem, 0, 0);
677 task->curr_event = 0;
678 err = pthread_create(&task->thread, &attr, thread_func, parms);
683 static void wait_for_tasks(struct perf_sched *sched)
685 u64 cpu_usage_0, cpu_usage_1;
686 struct task_desc *task;
687 unsigned long i, ret;
689 sched->start_time = get_nsecs();
690 sched->cpu_usage = 0;
691 pthread_mutex_unlock(&sched->work_done_wait_mutex);
693 for (i = 0; i < sched->nr_tasks; i++) {
694 task = sched->tasks[i];
695 ret = sem_wait(&task->ready_for_work);
697 sem_init(&task->ready_for_work, 0, 0);
699 ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
702 cpu_usage_0 = get_cpu_usage_nsec_parent();
704 pthread_mutex_unlock(&sched->start_work_mutex);
706 for (i = 0; i < sched->nr_tasks; i++) {
707 task = sched->tasks[i];
708 ret = sem_wait(&task->work_done_sem);
710 sem_init(&task->work_done_sem, 0, 0);
711 sched->cpu_usage += task->cpu_usage;
715 cpu_usage_1 = get_cpu_usage_nsec_parent();
716 if (!sched->runavg_cpu_usage)
717 sched->runavg_cpu_usage = sched->cpu_usage;
718 sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
720 sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
721 if (!sched->runavg_parent_cpu_usage)
722 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
723 sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
724 sched->parent_cpu_usage)/sched->replay_repeat;
726 ret = pthread_mutex_lock(&sched->start_work_mutex);
729 for (i = 0; i < sched->nr_tasks; i++) {
730 task = sched->tasks[i];
731 sem_init(&task->sleep_sem, 0, 0);
732 task->curr_event = 0;
736 static void run_one_test(struct perf_sched *sched)
738 u64 T0, T1, delta, avg_delta, fluct;
741 wait_for_tasks(sched);
745 sched->sum_runtime += delta;
748 avg_delta = sched->sum_runtime / sched->nr_runs;
749 if (delta < avg_delta)
750 fluct = avg_delta - delta;
752 fluct = delta - avg_delta;
753 sched->sum_fluct += fluct;
755 sched->run_avg = delta;
756 sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
758 printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
760 printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
762 printf("cpu: %0.2f / %0.2f",
763 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
767 * rusage statistics done by the parent, these are less
768 * accurate than the sched->sum_exec_runtime based statistics:
770 printf(" [%0.2f / %0.2f]",
771 (double)sched->parent_cpu_usage / NSEC_PER_MSEC,
772 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
777 if (sched->nr_sleep_corrections)
778 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
779 sched->nr_sleep_corrections = 0;
782 static void test_calibrations(struct perf_sched *sched)
787 burn_nsecs(sched, NSEC_PER_MSEC);
790 printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
793 sleep_nsecs(NSEC_PER_MSEC);
796 printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
800 replay_wakeup_event(struct perf_sched *sched,
801 struct perf_evsel *evsel, struct perf_sample *sample,
802 struct machine *machine __maybe_unused)
804 const char *comm = perf_evsel__strval(evsel, sample, "comm");
805 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
806 struct task_desc *waker, *wakee;
809 printf("sched_wakeup event %p\n", evsel);
811 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
814 waker = register_pid(sched, sample->tid, "<unknown>");
815 wakee = register_pid(sched, pid, comm);
817 add_sched_event_wakeup(sched, waker, sample->time, wakee);
821 static int replay_switch_event(struct perf_sched *sched,
822 struct perf_evsel *evsel,
823 struct perf_sample *sample,
824 struct machine *machine __maybe_unused)
826 const char *prev_comm = perf_evsel__strval(evsel, sample, "prev_comm"),
827 *next_comm = perf_evsel__strval(evsel, sample, "next_comm");
828 const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
829 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
830 const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
831 struct task_desc *prev, __maybe_unused *next;
832 u64 timestamp0, timestamp = sample->time;
833 int cpu = sample->cpu;
837 printf("sched_switch event %p\n", evsel);
839 if (cpu >= MAX_CPUS || cpu < 0)
842 timestamp0 = sched->cpu_last_switched[cpu];
844 delta = timestamp - timestamp0;
849 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
853 pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
854 prev_comm, prev_pid, next_comm, next_pid, delta);
856 prev = register_pid(sched, prev_pid, prev_comm);
857 next = register_pid(sched, next_pid, next_comm);
859 sched->cpu_last_switched[cpu] = timestamp;
861 add_sched_event_run(sched, prev, timestamp, delta);
862 add_sched_event_sleep(sched, prev, timestamp, prev_state);
867 static int replay_fork_event(struct perf_sched *sched,
868 union perf_event *event,
869 struct machine *machine)
871 struct thread *child, *parent;
873 child = machine__findnew_thread(machine, event->fork.pid,
875 parent = machine__findnew_thread(machine, event->fork.ppid,
878 if (child == NULL || parent == NULL) {
879 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
885 printf("fork event\n");
886 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
887 printf("... child: %s/%d\n", thread__comm_str(child), child->tid);
890 register_pid(sched, parent->tid, thread__comm_str(parent));
891 register_pid(sched, child->tid, thread__comm_str(child));
898 struct sort_dimension {
901 struct list_head list;
905 * handle runtime stats saved per thread
907 static struct thread_runtime *thread__init_runtime(struct thread *thread)
909 struct thread_runtime *r;
911 r = zalloc(sizeof(struct thread_runtime));
915 init_stats(&r->run_stats);
916 thread__set_priv(thread, r);
921 static struct thread_runtime *thread__get_runtime(struct thread *thread)
923 struct thread_runtime *tr;
925 tr = thread__priv(thread);
927 tr = thread__init_runtime(thread);
929 pr_debug("Failed to malloc memory for runtime data.\n");
936 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
938 struct sort_dimension *sort;
941 BUG_ON(list_empty(list));
943 list_for_each_entry(sort, list, list) {
944 ret = sort->cmp(l, r);
952 static struct work_atoms *
953 thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
954 struct list_head *sort_list)
956 struct rb_node *node = root->rb_root.rb_node;
957 struct work_atoms key = { .thread = thread };
960 struct work_atoms *atoms;
963 atoms = container_of(node, struct work_atoms, node);
965 cmp = thread_lat_cmp(sort_list, &key, atoms);
967 node = node->rb_left;
969 node = node->rb_right;
971 BUG_ON(thread != atoms->thread);
979 __thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
980 struct list_head *sort_list)
982 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
983 bool leftmost = true;
986 struct work_atoms *this;
989 this = container_of(*new, struct work_atoms, node);
992 cmp = thread_lat_cmp(sort_list, data, this);
995 new = &((*new)->rb_left);
997 new = &((*new)->rb_right);
1002 rb_link_node(&data->node, parent, new);
1003 rb_insert_color_cached(&data->node, root, leftmost);
1006 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1008 struct work_atoms *atoms = zalloc(sizeof(*atoms));
1010 pr_err("No memory at %s\n", __func__);
1014 atoms->thread = thread__get(thread);
1015 INIT_LIST_HEAD(&atoms->work_list);
1016 __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1020 static char sched_out_state(u64 prev_state)
1022 const char *str = TASK_STATE_TO_CHAR_STR;
1024 return str[prev_state];
1028 add_sched_out_event(struct work_atoms *atoms,
1032 struct work_atom *atom = zalloc(sizeof(*atom));
1034 pr_err("Non memory at %s", __func__);
1038 atom->sched_out_time = timestamp;
1040 if (run_state == 'R') {
1041 atom->state = THREAD_WAIT_CPU;
1042 atom->wake_up_time = atom->sched_out_time;
1045 list_add_tail(&atom->list, &atoms->work_list);
1050 add_runtime_event(struct work_atoms *atoms, u64 delta,
1051 u64 timestamp __maybe_unused)
1053 struct work_atom *atom;
1055 BUG_ON(list_empty(&atoms->work_list));
1057 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1059 atom->runtime += delta;
1060 atoms->total_runtime += delta;
1064 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1066 struct work_atom *atom;
1069 if (list_empty(&atoms->work_list))
1072 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1074 if (atom->state != THREAD_WAIT_CPU)
1077 if (timestamp < atom->wake_up_time) {
1078 atom->state = THREAD_IGNORE;
1082 atom->state = THREAD_SCHED_IN;
1083 atom->sched_in_time = timestamp;
1085 delta = atom->sched_in_time - atom->wake_up_time;
1086 atoms->total_lat += delta;
1087 if (delta > atoms->max_lat) {
1088 atoms->max_lat = delta;
1089 atoms->max_lat_at = timestamp;
1094 static int latency_switch_event(struct perf_sched *sched,
1095 struct perf_evsel *evsel,
1096 struct perf_sample *sample,
1097 struct machine *machine)
1099 const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1100 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1101 const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
1102 struct work_atoms *out_events, *in_events;
1103 struct thread *sched_out, *sched_in;
1104 u64 timestamp0, timestamp = sample->time;
1105 int cpu = sample->cpu, err = -1;
1108 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1110 timestamp0 = sched->cpu_last_switched[cpu];
1111 sched->cpu_last_switched[cpu] = timestamp;
1113 delta = timestamp - timestamp0;
1118 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1122 sched_out = machine__findnew_thread(machine, -1, prev_pid);
1123 sched_in = machine__findnew_thread(machine, -1, next_pid);
1124 if (sched_out == NULL || sched_in == NULL)
1127 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1129 if (thread_atoms_insert(sched, sched_out))
1131 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1133 pr_err("out-event: Internal tree error");
1137 if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1140 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1142 if (thread_atoms_insert(sched, sched_in))
1144 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1146 pr_err("in-event: Internal tree error");
1150 * Take came in we have not heard about yet,
1151 * add in an initial atom in runnable state:
1153 if (add_sched_out_event(in_events, 'R', timestamp))
1156 add_sched_in_event(in_events, timestamp);
1159 thread__put(sched_out);
1160 thread__put(sched_in);
1164 static int latency_runtime_event(struct perf_sched *sched,
1165 struct perf_evsel *evsel,
1166 struct perf_sample *sample,
1167 struct machine *machine)
1169 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1170 const u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
1171 struct thread *thread = machine__findnew_thread(machine, -1, pid);
1172 struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1173 u64 timestamp = sample->time;
1174 int cpu = sample->cpu, err = -1;
1179 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1181 if (thread_atoms_insert(sched, thread))
1183 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1185 pr_err("in-event: Internal tree error");
1188 if (add_sched_out_event(atoms, 'R', timestamp))
1192 add_runtime_event(atoms, runtime, timestamp);
1195 thread__put(thread);
1199 static int latency_wakeup_event(struct perf_sched *sched,
1200 struct perf_evsel *evsel,
1201 struct perf_sample *sample,
1202 struct machine *machine)
1204 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1205 struct work_atoms *atoms;
1206 struct work_atom *atom;
1207 struct thread *wakee;
1208 u64 timestamp = sample->time;
1211 wakee = machine__findnew_thread(machine, -1, pid);
1214 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1216 if (thread_atoms_insert(sched, wakee))
1218 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1220 pr_err("wakeup-event: Internal tree error");
1223 if (add_sched_out_event(atoms, 'S', timestamp))
1227 BUG_ON(list_empty(&atoms->work_list));
1229 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1232 * As we do not guarantee the wakeup event happens when
1233 * task is out of run queue, also may happen when task is
1234 * on run queue and wakeup only change ->state to TASK_RUNNING,
1235 * then we should not set the ->wake_up_time when wake up a
1236 * task which is on run queue.
1238 * You WILL be missing events if you've recorded only
1239 * one CPU, or are only looking at only one, so don't
1240 * skip in this case.
1242 if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1245 sched->nr_timestamps++;
1246 if (atom->sched_out_time > timestamp) {
1247 sched->nr_unordered_timestamps++;
1251 atom->state = THREAD_WAIT_CPU;
1252 atom->wake_up_time = timestamp;
1260 static int latency_migrate_task_event(struct perf_sched *sched,
1261 struct perf_evsel *evsel,
1262 struct perf_sample *sample,
1263 struct machine *machine)
1265 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1266 u64 timestamp = sample->time;
1267 struct work_atoms *atoms;
1268 struct work_atom *atom;
1269 struct thread *migrant;
1273 * Only need to worry about migration when profiling one CPU.
1275 if (sched->profile_cpu == -1)
1278 migrant = machine__findnew_thread(machine, -1, pid);
1279 if (migrant == NULL)
1281 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1283 if (thread_atoms_insert(sched, migrant))
1285 register_pid(sched, migrant->tid, thread__comm_str(migrant));
1286 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1288 pr_err("migration-event: Internal tree error");
1291 if (add_sched_out_event(atoms, 'R', timestamp))
1295 BUG_ON(list_empty(&atoms->work_list));
1297 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1298 atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1300 sched->nr_timestamps++;
1302 if (atom->sched_out_time > timestamp)
1303 sched->nr_unordered_timestamps++;
1306 thread__put(migrant);
1310 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1315 char max_lat_at[32];
1317 if (!work_list->nb_atoms)
1320 * Ignore idle threads:
1322 if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1325 sched->all_runtime += work_list->total_runtime;
1326 sched->all_count += work_list->nb_atoms;
1328 if (work_list->num_merged > 1)
1329 ret = printf(" %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1331 ret = printf(" %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1333 for (i = 0; i < 24 - ret; i++)
1336 avg = work_list->total_lat / work_list->nb_atoms;
1337 timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at));
1339 printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n",
1340 (double)work_list->total_runtime / NSEC_PER_MSEC,
1341 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1342 (double)work_list->max_lat / NSEC_PER_MSEC,
1346 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1348 if (l->thread == r->thread)
1350 if (l->thread->tid < r->thread->tid)
1352 if (l->thread->tid > r->thread->tid)
1354 return (int)(l->thread - r->thread);
1357 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1367 avgl = l->total_lat / l->nb_atoms;
1368 avgr = r->total_lat / r->nb_atoms;
1378 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1380 if (l->max_lat < r->max_lat)
1382 if (l->max_lat > r->max_lat)
1388 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1390 if (l->nb_atoms < r->nb_atoms)
1392 if (l->nb_atoms > r->nb_atoms)
1398 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1400 if (l->total_runtime < r->total_runtime)
1402 if (l->total_runtime > r->total_runtime)
1408 static int sort_dimension__add(const char *tok, struct list_head *list)
1411 static struct sort_dimension avg_sort_dimension = {
1415 static struct sort_dimension max_sort_dimension = {
1419 static struct sort_dimension pid_sort_dimension = {
1423 static struct sort_dimension runtime_sort_dimension = {
1427 static struct sort_dimension switch_sort_dimension = {
1431 struct sort_dimension *available_sorts[] = {
1432 &pid_sort_dimension,
1433 &avg_sort_dimension,
1434 &max_sort_dimension,
1435 &switch_sort_dimension,
1436 &runtime_sort_dimension,
1439 for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1440 if (!strcmp(available_sorts[i]->name, tok)) {
1441 list_add_tail(&available_sorts[i]->list, list);
1450 static void perf_sched__sort_lat(struct perf_sched *sched)
1452 struct rb_node *node;
1453 struct rb_root_cached *root = &sched->atom_root;
1456 struct work_atoms *data;
1457 node = rb_first_cached(root);
1461 rb_erase_cached(node, root);
1462 data = rb_entry(node, struct work_atoms, node);
1463 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1465 if (root == &sched->atom_root) {
1466 root = &sched->merged_atom_root;
1471 static int process_sched_wakeup_event(struct perf_tool *tool,
1472 struct perf_evsel *evsel,
1473 struct perf_sample *sample,
1474 struct machine *machine)
1476 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1478 if (sched->tp_handler->wakeup_event)
1479 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1489 static bool thread__has_color(struct thread *thread)
1491 union map_priv priv = {
1492 .ptr = thread__priv(thread),
1498 static struct thread*
1499 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1501 struct thread *thread = machine__findnew_thread(machine, pid, tid);
1502 union map_priv priv = {
1506 if (!sched->map.color_pids || !thread || thread__priv(thread))
1509 if (thread_map__has(sched->map.color_pids, tid))
1512 thread__set_priv(thread, priv.ptr);
1516 static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1517 struct perf_sample *sample, struct machine *machine)
1519 const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1520 struct thread *sched_in;
1521 struct thread_runtime *tr;
1523 u64 timestamp0, timestamp = sample->time;
1525 int i, this_cpu = sample->cpu;
1527 bool new_cpu = false;
1528 const char *color = PERF_COLOR_NORMAL;
1529 char stimestamp[32];
1531 BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1533 if (this_cpu > sched->max_cpu)
1534 sched->max_cpu = this_cpu;
1536 if (sched->map.comp) {
1537 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1538 if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1539 sched->map.comp_cpus[cpus_nr++] = this_cpu;
1543 cpus_nr = sched->max_cpu;
1545 timestamp0 = sched->cpu_last_switched[this_cpu];
1546 sched->cpu_last_switched[this_cpu] = timestamp;
1548 delta = timestamp - timestamp0;
1553 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1557 sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1558 if (sched_in == NULL)
1561 tr = thread__get_runtime(sched_in);
1563 thread__put(sched_in);
1567 sched->curr_thread[this_cpu] = thread__get(sched_in);
1572 if (!tr->shortname[0]) {
1573 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1575 * Don't allocate a letter-number for swapper:0
1576 * as a shortname. Instead, we use '.' for it.
1578 tr->shortname[0] = '.';
1579 tr->shortname[1] = ' ';
1581 tr->shortname[0] = sched->next_shortname1;
1582 tr->shortname[1] = sched->next_shortname2;
1584 if (sched->next_shortname1 < 'Z') {
1585 sched->next_shortname1++;
1587 sched->next_shortname1 = 'A';
1588 if (sched->next_shortname2 < '9')
1589 sched->next_shortname2++;
1591 sched->next_shortname2 = '0';
1597 for (i = 0; i < cpus_nr; i++) {
1598 int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
1599 struct thread *curr_thread = sched->curr_thread[cpu];
1600 struct thread_runtime *curr_tr;
1601 const char *pid_color = color;
1602 const char *cpu_color = color;
1604 if (curr_thread && thread__has_color(curr_thread))
1605 pid_color = COLOR_PIDS;
1607 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1610 if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1611 cpu_color = COLOR_CPUS;
1613 if (cpu != this_cpu)
1614 color_fprintf(stdout, color, " ");
1616 color_fprintf(stdout, cpu_color, "*");
1618 if (sched->curr_thread[cpu]) {
1619 curr_tr = thread__get_runtime(sched->curr_thread[cpu]);
1620 if (curr_tr == NULL) {
1621 thread__put(sched_in);
1624 color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1626 color_fprintf(stdout, color, " ");
1629 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1632 timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1633 color_fprintf(stdout, color, " %12s secs ", stimestamp);
1634 if (new_shortname || tr->comm_changed || (verbose > 0 && sched_in->tid)) {
1635 const char *pid_color = color;
1637 if (thread__has_color(sched_in))
1638 pid_color = COLOR_PIDS;
1640 color_fprintf(stdout, pid_color, "%s => %s:%d",
1641 tr->shortname, thread__comm_str(sched_in), sched_in->tid);
1642 tr->comm_changed = false;
1645 if (sched->map.comp && new_cpu)
1646 color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1649 color_fprintf(stdout, color, "\n");
1651 thread__put(sched_in);
1656 static int process_sched_switch_event(struct perf_tool *tool,
1657 struct perf_evsel *evsel,
1658 struct perf_sample *sample,
1659 struct machine *machine)
1661 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1662 int this_cpu = sample->cpu, err = 0;
1663 u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1664 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1666 if (sched->curr_pid[this_cpu] != (u32)-1) {
1668 * Are we trying to switch away a PID that is
1671 if (sched->curr_pid[this_cpu] != prev_pid)
1672 sched->nr_context_switch_bugs++;
1675 if (sched->tp_handler->switch_event)
1676 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1678 sched->curr_pid[this_cpu] = next_pid;
1682 static int process_sched_runtime_event(struct perf_tool *tool,
1683 struct perf_evsel *evsel,
1684 struct perf_sample *sample,
1685 struct machine *machine)
1687 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1689 if (sched->tp_handler->runtime_event)
1690 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1695 static int perf_sched__process_fork_event(struct perf_tool *tool,
1696 union perf_event *event,
1697 struct perf_sample *sample,
1698 struct machine *machine)
1700 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1702 /* run the fork event through the perf machineruy */
1703 perf_event__process_fork(tool, event, sample, machine);
1705 /* and then run additional processing needed for this command */
1706 if (sched->tp_handler->fork_event)
1707 return sched->tp_handler->fork_event(sched, event, machine);
1712 static int process_sched_migrate_task_event(struct perf_tool *tool,
1713 struct perf_evsel *evsel,
1714 struct perf_sample *sample,
1715 struct machine *machine)
1717 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1719 if (sched->tp_handler->migrate_task_event)
1720 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1725 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1726 struct perf_evsel *evsel,
1727 struct perf_sample *sample,
1728 struct machine *machine);
1730 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1731 union perf_event *event __maybe_unused,
1732 struct perf_sample *sample,
1733 struct perf_evsel *evsel,
1734 struct machine *machine)
1738 if (evsel->handler != NULL) {
1739 tracepoint_handler f = evsel->handler;
1740 err = f(tool, evsel, sample, machine);
1746 static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1747 union perf_event *event,
1748 struct perf_sample *sample,
1749 struct machine *machine)
1751 struct thread *thread;
1752 struct thread_runtime *tr;
1755 err = perf_event__process_comm(tool, event, sample, machine);
1759 thread = machine__find_thread(machine, sample->pid, sample->tid);
1761 pr_err("Internal error: can't find thread\n");
1765 tr = thread__get_runtime(thread);
1767 thread__put(thread);
1771 tr->comm_changed = true;
1772 thread__put(thread);
1777 static int perf_sched__read_events(struct perf_sched *sched)
1779 const struct perf_evsel_str_handler handlers[] = {
1780 { "sched:sched_switch", process_sched_switch_event, },
1781 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1782 { "sched:sched_wakeup", process_sched_wakeup_event, },
1783 { "sched:sched_wakeup_new", process_sched_wakeup_event, },
1784 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1786 struct perf_session *session;
1787 struct perf_data data = {
1789 .mode = PERF_DATA_MODE_READ,
1790 .force = sched->force,
1794 session = perf_session__new(&data, false, &sched->tool);
1795 if (session == NULL) {
1796 pr_debug("No Memory for session\n");
1800 symbol__init(&session->header.env);
1802 if (perf_session__set_tracepoints_handlers(session, handlers))
1805 if (perf_session__has_traces(session, "record -R")) {
1806 int err = perf_session__process_events(session);
1808 pr_err("Failed to process events, error %d", err);
1812 sched->nr_events = session->evlist->stats.nr_events[0];
1813 sched->nr_lost_events = session->evlist->stats.total_lost;
1814 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1819 perf_session__delete(session);
1824 * scheduling times are printed as msec.usec
1826 static inline void print_sched_time(unsigned long long nsecs, int width)
1828 unsigned long msecs;
1829 unsigned long usecs;
1831 msecs = nsecs / NSEC_PER_MSEC;
1832 nsecs -= msecs * NSEC_PER_MSEC;
1833 usecs = nsecs / NSEC_PER_USEC;
1834 printf("%*lu.%03lu ", width, msecs, usecs);
1838 * returns runtime data for event, allocating memory for it the
1839 * first time it is used.
1841 static struct evsel_runtime *perf_evsel__get_runtime(struct perf_evsel *evsel)
1843 struct evsel_runtime *r = evsel->priv;
1846 r = zalloc(sizeof(struct evsel_runtime));
1854 * save last time event was seen per cpu
1856 static void perf_evsel__save_time(struct perf_evsel *evsel,
1857 u64 timestamp, u32 cpu)
1859 struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1864 if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1865 int i, n = __roundup_pow_of_two(cpu+1);
1866 void *p = r->last_time;
1868 p = realloc(r->last_time, n * sizeof(u64));
1873 for (i = r->ncpu; i < n; ++i)
1874 r->last_time[i] = (u64) 0;
1879 r->last_time[cpu] = timestamp;
1882 /* returns last time this event was seen on the given cpu */
1883 static u64 perf_evsel__get_time(struct perf_evsel *evsel, u32 cpu)
1885 struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1887 if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1890 return r->last_time[cpu];
1893 static int comm_width = 30;
1895 static char *timehist_get_commstr(struct thread *thread)
1897 static char str[32];
1898 const char *comm = thread__comm_str(thread);
1899 pid_t tid = thread->tid;
1900 pid_t pid = thread->pid_;
1904 n = scnprintf(str, sizeof(str), "%s", comm);
1906 else if (tid != pid)
1907 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1910 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1918 static void timehist_header(struct perf_sched *sched)
1920 u32 ncpus = sched->max_cpu + 1;
1923 printf("%15s %6s ", "time", "cpu");
1925 if (sched->show_cpu_visual) {
1927 for (i = 0, j = 0; i < ncpus; ++i) {
1935 printf(" %-*s %9s %9s %9s", comm_width,
1936 "task name", "wait time", "sch delay", "run time");
1938 if (sched->show_state)
1939 printf(" %s", "state");
1946 printf("%15s %-6s ", "", "");
1948 if (sched->show_cpu_visual)
1949 printf(" %*s ", ncpus, "");
1951 printf(" %-*s %9s %9s %9s", comm_width,
1952 "[tid/pid]", "(msec)", "(msec)", "(msec)");
1954 if (sched->show_state)
1962 printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1964 if (sched->show_cpu_visual)
1965 printf(" %.*s ", ncpus, graph_dotted_line);
1967 printf(" %.*s %.9s %.9s %.9s", comm_width,
1968 graph_dotted_line, graph_dotted_line, graph_dotted_line,
1971 if (sched->show_state)
1972 printf(" %.5s", graph_dotted_line);
1977 static char task_state_char(struct thread *thread, int state)
1979 static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
1980 unsigned bit = state ? ffs(state) : 0;
1983 if (thread->tid == 0)
1986 return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
1989 static void timehist_print_sample(struct perf_sched *sched,
1990 struct perf_evsel *evsel,
1991 struct perf_sample *sample,
1992 struct addr_location *al,
1993 struct thread *thread,
1996 struct thread_runtime *tr = thread__priv(thread);
1997 const char *next_comm = perf_evsel__strval(evsel, sample, "next_comm");
1998 const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1999 u32 max_cpus = sched->max_cpu + 1;
2004 timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2005 printf("%15s [%04d] ", tstr, sample->cpu);
2007 if (sched->show_cpu_visual) {
2012 for (i = 0; i < max_cpus; ++i) {
2013 /* flag idle times with 'i'; others are sched events */
2014 if (i == sample->cpu)
2015 c = (thread->tid == 0) ? 'i' : 's';
2023 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2025 wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2026 print_sched_time(wait_time, 6);
2028 print_sched_time(tr->dt_delay, 6);
2029 print_sched_time(tr->dt_run, 6);
2031 if (sched->show_state)
2032 printf(" %5c ", task_state_char(thread, state));
2034 if (sched->show_next) {
2035 snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2036 printf(" %-*s", comm_width, nstr);
2039 if (sched->show_wakeups && !sched->show_next)
2040 printf(" %-*s", comm_width, "");
2042 if (thread->tid == 0)
2045 if (sched->show_callchain)
2048 sample__fprintf_sym(sample, al, 0,
2049 EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2050 EVSEL__PRINT_CALLCHAIN_ARROW |
2051 EVSEL__PRINT_SKIP_IGNORED,
2052 &callchain_cursor, stdout);
2059 * Explanation of delta-time stats:
2061 * t = time of current schedule out event
2062 * tprev = time of previous sched out event
2063 * also time of schedule-in event for current task
2064 * last_time = time of last sched change event for current task
2065 * (i.e, time process was last scheduled out)
2066 * ready_to_run = time of wakeup for current task
2068 * -----|------------|------------|------------|------
2069 * last ready tprev t
2072 * |-------- dt_wait --------|
2073 * |- dt_delay -|-- dt_run --|
2075 * dt_run = run time of current task
2076 * dt_wait = time between last schedule out event for task and tprev
2077 * represents time spent off the cpu
2078 * dt_delay = time between wakeup and schedule-in of task
2081 static void timehist_update_runtime_stats(struct thread_runtime *r,
2091 r->dt_run = t - tprev;
2092 if (r->ready_to_run) {
2093 if (r->ready_to_run > tprev)
2094 pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2096 r->dt_delay = tprev - r->ready_to_run;
2099 if (r->last_time > tprev)
2100 pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2101 else if (r->last_time) {
2102 u64 dt_wait = tprev - r->last_time;
2104 if (r->last_state == TASK_RUNNING)
2105 r->dt_preempt = dt_wait;
2106 else if (r->last_state == TASK_UNINTERRUPTIBLE)
2107 r->dt_iowait = dt_wait;
2109 r->dt_sleep = dt_wait;
2113 update_stats(&r->run_stats, r->dt_run);
2115 r->total_run_time += r->dt_run;
2116 r->total_delay_time += r->dt_delay;
2117 r->total_sleep_time += r->dt_sleep;
2118 r->total_iowait_time += r->dt_iowait;
2119 r->total_preempt_time += r->dt_preempt;
2122 static bool is_idle_sample(struct perf_sample *sample,
2123 struct perf_evsel *evsel)
2125 /* pid 0 == swapper == idle task */
2126 if (strcmp(perf_evsel__name(evsel), "sched:sched_switch") == 0)
2127 return perf_evsel__intval(evsel, sample, "prev_pid") == 0;
2129 return sample->pid == 0;
2132 static void save_task_callchain(struct perf_sched *sched,
2133 struct perf_sample *sample,
2134 struct perf_evsel *evsel,
2135 struct machine *machine)
2137 struct callchain_cursor *cursor = &callchain_cursor;
2138 struct thread *thread;
2140 /* want main thread for process - has maps */
2141 thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2142 if (thread == NULL) {
2143 pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2147 if (!sched->show_callchain || sample->callchain == NULL)
2150 if (thread__resolve_callchain(thread, cursor, evsel, sample,
2151 NULL, NULL, sched->max_stack + 2) != 0) {
2153 pr_err("Failed to resolve callchain. Skipping\n");
2158 callchain_cursor_commit(cursor);
2161 struct callchain_cursor_node *node;
2164 node = callchain_cursor_current(cursor);
2170 if (!strcmp(sym->name, "schedule") ||
2171 !strcmp(sym->name, "__schedule") ||
2172 !strcmp(sym->name, "preempt_schedule"))
2176 callchain_cursor_advance(cursor);
2180 static int init_idle_thread(struct thread *thread)
2182 struct idle_thread_runtime *itr;
2184 thread__set_comm(thread, idle_comm, 0);
2186 itr = zalloc(sizeof(*itr));
2190 init_stats(&itr->tr.run_stats);
2191 callchain_init(&itr->callchain);
2192 callchain_cursor_reset(&itr->cursor);
2193 thread__set_priv(thread, itr);
2199 * Track idle stats per cpu by maintaining a local thread
2200 * struct for the idle task on each cpu.
2202 static int init_idle_threads(int ncpu)
2206 idle_threads = zalloc(ncpu * sizeof(struct thread *));
2210 idle_max_cpu = ncpu;
2212 /* allocate the actual thread struct if needed */
2213 for (i = 0; i < ncpu; ++i) {
2214 idle_threads[i] = thread__new(0, 0);
2215 if (idle_threads[i] == NULL)
2218 ret = init_idle_thread(idle_threads[i]);
2226 static void free_idle_threads(void)
2230 if (idle_threads == NULL)
2233 for (i = 0; i < idle_max_cpu; ++i) {
2234 if ((idle_threads[i]))
2235 thread__delete(idle_threads[i]);
2241 static struct thread *get_idle_thread(int cpu)
2244 * expand/allocate array of pointers to local thread
2247 if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2248 int i, j = __roundup_pow_of_two(cpu+1);
2251 p = realloc(idle_threads, j * sizeof(struct thread *));
2255 idle_threads = (struct thread **) p;
2256 for (i = idle_max_cpu; i < j; ++i)
2257 idle_threads[i] = NULL;
2262 /* allocate a new thread struct if needed */
2263 if (idle_threads[cpu] == NULL) {
2264 idle_threads[cpu] = thread__new(0, 0);
2266 if (idle_threads[cpu]) {
2267 if (init_idle_thread(idle_threads[cpu]) < 0)
2272 return idle_threads[cpu];
2275 static void save_idle_callchain(struct perf_sched *sched,
2276 struct idle_thread_runtime *itr,
2277 struct perf_sample *sample)
2279 if (!sched->show_callchain || sample->callchain == NULL)
2282 callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2285 static struct thread *timehist_get_thread(struct perf_sched *sched,
2286 struct perf_sample *sample,
2287 struct machine *machine,
2288 struct perf_evsel *evsel)
2290 struct thread *thread;
2292 if (is_idle_sample(sample, evsel)) {
2293 thread = get_idle_thread(sample->cpu);
2295 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2298 /* there were samples with tid 0 but non-zero pid */
2299 thread = machine__findnew_thread(machine, sample->pid,
2300 sample->tid ?: sample->pid);
2301 if (thread == NULL) {
2302 pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2306 save_task_callchain(sched, sample, evsel, machine);
2307 if (sched->idle_hist) {
2308 struct thread *idle;
2309 struct idle_thread_runtime *itr;
2311 idle = get_idle_thread(sample->cpu);
2313 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2317 itr = thread__priv(idle);
2321 itr->last_thread = thread;
2323 /* copy task callchain when entering to idle */
2324 if (perf_evsel__intval(evsel, sample, "next_pid") == 0)
2325 save_idle_callchain(sched, itr, sample);
2332 static bool timehist_skip_sample(struct perf_sched *sched,
2333 struct thread *thread,
2334 struct perf_evsel *evsel,
2335 struct perf_sample *sample)
2339 if (thread__is_filtered(thread)) {
2341 sched->skipped_samples++;
2344 if (sched->idle_hist) {
2345 if (strcmp(perf_evsel__name(evsel), "sched:sched_switch"))
2347 else if (perf_evsel__intval(evsel, sample, "prev_pid") != 0 &&
2348 perf_evsel__intval(evsel, sample, "next_pid") != 0)
2355 static void timehist_print_wakeup_event(struct perf_sched *sched,
2356 struct perf_evsel *evsel,
2357 struct perf_sample *sample,
2358 struct machine *machine,
2359 struct thread *awakened)
2361 struct thread *thread;
2364 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2368 /* show wakeup unless both awakee and awaker are filtered */
2369 if (timehist_skip_sample(sched, thread, evsel, sample) &&
2370 timehist_skip_sample(sched, awakened, evsel, sample)) {
2374 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2375 printf("%15s [%04d] ", tstr, sample->cpu);
2376 if (sched->show_cpu_visual)
2377 printf(" %*s ", sched->max_cpu + 1, "");
2379 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2382 printf(" %9s %9s %9s ", "", "", "");
2384 printf("awakened: %s", timehist_get_commstr(awakened));
2389 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2390 union perf_event *event __maybe_unused,
2391 struct perf_evsel *evsel,
2392 struct perf_sample *sample,
2393 struct machine *machine)
2395 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2396 struct thread *thread;
2397 struct thread_runtime *tr = NULL;
2398 /* want pid of awakened task not pid in sample */
2399 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2401 thread = machine__findnew_thread(machine, 0, pid);
2405 tr = thread__get_runtime(thread);
2409 if (tr->ready_to_run == 0)
2410 tr->ready_to_run = sample->time;
2412 /* show wakeups if requested */
2413 if (sched->show_wakeups &&
2414 !perf_time__skip_sample(&sched->ptime, sample->time))
2415 timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2420 static void timehist_print_migration_event(struct perf_sched *sched,
2421 struct perf_evsel *evsel,
2422 struct perf_sample *sample,
2423 struct machine *machine,
2424 struct thread *migrated)
2426 struct thread *thread;
2428 u32 max_cpus = sched->max_cpu + 1;
2431 if (sched->summary_only)
2434 max_cpus = sched->max_cpu + 1;
2435 ocpu = perf_evsel__intval(evsel, sample, "orig_cpu");
2436 dcpu = perf_evsel__intval(evsel, sample, "dest_cpu");
2438 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2442 if (timehist_skip_sample(sched, thread, evsel, sample) &&
2443 timehist_skip_sample(sched, migrated, evsel, sample)) {
2447 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2448 printf("%15s [%04d] ", tstr, sample->cpu);
2450 if (sched->show_cpu_visual) {
2455 for (i = 0; i < max_cpus; ++i) {
2456 c = (i == sample->cpu) ? 'm' : ' ';
2462 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2465 printf(" %9s %9s %9s ", "", "", "");
2467 printf("migrated: %s", timehist_get_commstr(migrated));
2468 printf(" cpu %d => %d", ocpu, dcpu);
2473 static int timehist_migrate_task_event(struct perf_tool *tool,
2474 union perf_event *event __maybe_unused,
2475 struct perf_evsel *evsel,
2476 struct perf_sample *sample,
2477 struct machine *machine)
2479 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2480 struct thread *thread;
2481 struct thread_runtime *tr = NULL;
2482 /* want pid of migrated task not pid in sample */
2483 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2485 thread = machine__findnew_thread(machine, 0, pid);
2489 tr = thread__get_runtime(thread);
2495 /* show migrations if requested */
2496 timehist_print_migration_event(sched, evsel, sample, machine, thread);
2501 static int timehist_sched_change_event(struct perf_tool *tool,
2502 union perf_event *event,
2503 struct perf_evsel *evsel,
2504 struct perf_sample *sample,
2505 struct machine *machine)
2507 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2508 struct perf_time_interval *ptime = &sched->ptime;
2509 struct addr_location al;
2510 struct thread *thread;
2511 struct thread_runtime *tr = NULL;
2512 u64 tprev, t = sample->time;
2514 int state = perf_evsel__intval(evsel, sample, "prev_state");
2517 if (machine__resolve(machine, &al, sample) < 0) {
2518 pr_err("problem processing %d event. skipping it\n",
2519 event->header.type);
2524 thread = timehist_get_thread(sched, sample, machine, evsel);
2525 if (thread == NULL) {
2530 if (timehist_skip_sample(sched, thread, evsel, sample))
2533 tr = thread__get_runtime(thread);
2539 tprev = perf_evsel__get_time(evsel, sample->cpu);
2542 * If start time given:
2543 * - sample time is under window user cares about - skip sample
2544 * - tprev is under window user cares about - reset to start of window
2546 if (ptime->start && ptime->start > t)
2549 if (tprev && ptime->start > tprev)
2550 tprev = ptime->start;
2553 * If end time given:
2554 * - previous sched event is out of window - we are done
2555 * - sample time is beyond window user cares about - reset it
2556 * to close out stats for time window interest
2559 if (tprev > ptime->end)
2566 if (!sched->idle_hist || thread->tid == 0) {
2567 timehist_update_runtime_stats(tr, t, tprev);
2569 if (sched->idle_hist) {
2570 struct idle_thread_runtime *itr = (void *)tr;
2571 struct thread_runtime *last_tr;
2573 BUG_ON(thread->tid != 0);
2575 if (itr->last_thread == NULL)
2578 /* add current idle time as last thread's runtime */
2579 last_tr = thread__get_runtime(itr->last_thread);
2580 if (last_tr == NULL)
2583 timehist_update_runtime_stats(last_tr, t, tprev);
2585 * remove delta time of last thread as it's not updated
2586 * and otherwise it will show an invalid value next
2587 * time. we only care total run time and run stat.
2589 last_tr->dt_run = 0;
2590 last_tr->dt_delay = 0;
2591 last_tr->dt_sleep = 0;
2592 last_tr->dt_iowait = 0;
2593 last_tr->dt_preempt = 0;
2596 callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2598 itr->last_thread = NULL;
2602 if (!sched->summary_only)
2603 timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2606 if (sched->hist_time.start == 0 && t >= ptime->start)
2607 sched->hist_time.start = t;
2608 if (ptime->end == 0 || t <= ptime->end)
2609 sched->hist_time.end = t;
2612 /* time of this sched_switch event becomes last time task seen */
2613 tr->last_time = sample->time;
2615 /* last state is used to determine where to account wait time */
2616 tr->last_state = state;
2618 /* sched out event for task so reset ready to run time */
2619 tr->ready_to_run = 0;
2622 perf_evsel__save_time(evsel, sample->time, sample->cpu);
2627 static int timehist_sched_switch_event(struct perf_tool *tool,
2628 union perf_event *event,
2629 struct perf_evsel *evsel,
2630 struct perf_sample *sample,
2631 struct machine *machine __maybe_unused)
2633 return timehist_sched_change_event(tool, event, evsel, sample, machine);
2636 static int process_lost(struct perf_tool *tool __maybe_unused,
2637 union perf_event *event,
2638 struct perf_sample *sample,
2639 struct machine *machine __maybe_unused)
2643 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2644 printf("%15s ", tstr);
2645 printf("lost %" PRIu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2651 static void print_thread_runtime(struct thread *t,
2652 struct thread_runtime *r)
2654 double mean = avg_stats(&r->run_stats);
2657 printf("%*s %5d %9" PRIu64 " ",
2658 comm_width, timehist_get_commstr(t), t->ppid,
2659 (u64) r->run_stats.n);
2661 print_sched_time(r->total_run_time, 8);
2662 stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2663 print_sched_time(r->run_stats.min, 6);
2665 print_sched_time((u64) mean, 6);
2667 print_sched_time(r->run_stats.max, 6);
2669 printf("%5.2f", stddev);
2670 printf(" %5" PRIu64, r->migrations);
2674 static void print_thread_waittime(struct thread *t,
2675 struct thread_runtime *r)
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 print_sched_time(r->total_sleep_time, 6);
2684 print_sched_time(r->total_iowait_time, 6);
2686 print_sched_time(r->total_preempt_time, 6);
2688 print_sched_time(r->total_delay_time, 6);
2692 struct total_run_stats {
2693 struct perf_sched *sched;
2699 static int __show_thread_runtime(struct thread *t, void *priv)
2701 struct total_run_stats *stats = priv;
2702 struct thread_runtime *r;
2704 if (thread__is_filtered(t))
2707 r = thread__priv(t);
2708 if (r && r->run_stats.n) {
2709 stats->task_count++;
2710 stats->sched_count += r->run_stats.n;
2711 stats->total_run_time += r->total_run_time;
2713 if (stats->sched->show_state)
2714 print_thread_waittime(t, r);
2716 print_thread_runtime(t, r);
2722 static int show_thread_runtime(struct thread *t, void *priv)
2727 return __show_thread_runtime(t, priv);
2730 static int show_deadthread_runtime(struct thread *t, void *priv)
2735 return __show_thread_runtime(t, priv);
2738 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2740 const char *sep = " <- ";
2741 struct callchain_list *chain;
2749 ret = callchain__fprintf_folded(fp, node->parent);
2752 list_for_each_entry(chain, &node->val, list) {
2753 if (chain->ip >= PERF_CONTEXT_MAX)
2755 if (chain->ms.sym && chain->ms.sym->ignore)
2757 ret += fprintf(fp, "%s%s", first ? "" : sep,
2758 callchain_list__sym_name(chain, bf, sizeof(bf),
2766 static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2770 struct callchain_node *chain;
2771 struct rb_node *rb_node = rb_first_cached(root);
2773 printf(" %16s %8s %s\n", "Idle time (msec)", "Count", "Callchains");
2774 printf(" %.16s %.8s %.50s\n", graph_dotted_line, graph_dotted_line,
2778 chain = rb_entry(rb_node, struct callchain_node, rb_node);
2779 rb_node = rb_next(rb_node);
2781 ret += fprintf(fp, " ");
2782 print_sched_time(chain->hit, 12);
2783 ret += 16; /* print_sched_time returns 2nd arg + 4 */
2784 ret += fprintf(fp, " %8d ", chain->count);
2785 ret += callchain__fprintf_folded(fp, chain);
2786 ret += fprintf(fp, "\n");
2792 static void timehist_print_summary(struct perf_sched *sched,
2793 struct perf_session *session)
2795 struct machine *m = &session->machines.host;
2796 struct total_run_stats totals;
2799 struct thread_runtime *r;
2801 u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2803 memset(&totals, 0, sizeof(totals));
2804 totals.sched = sched;
2806 if (sched->idle_hist) {
2807 printf("\nIdle-time summary\n");
2808 printf("%*s parent sched-out ", comm_width, "comm");
2809 printf(" idle-time min-idle avg-idle max-idle stddev migrations\n");
2810 } else if (sched->show_state) {
2811 printf("\nWait-time summary\n");
2812 printf("%*s parent sched-in ", comm_width, "comm");
2813 printf(" run-time sleep iowait preempt delay\n");
2815 printf("\nRuntime summary\n");
2816 printf("%*s parent sched-in ", comm_width, "comm");
2817 printf(" run-time min-run avg-run max-run stddev migrations\n");
2819 printf("%*s (count) ", comm_width, "");
2820 printf(" (msec) (msec) (msec) (msec) %s\n",
2821 sched->show_state ? "(msec)" : "%");
2822 printf("%.117s\n", graph_dotted_line);
2824 machine__for_each_thread(m, show_thread_runtime, &totals);
2825 task_count = totals.task_count;
2827 printf("<no still running tasks>\n");
2829 printf("\nTerminated tasks:\n");
2830 machine__for_each_thread(m, show_deadthread_runtime, &totals);
2831 if (task_count == totals.task_count)
2832 printf("<no terminated tasks>\n");
2834 /* CPU idle stats not tracked when samples were skipped */
2835 if (sched->skipped_samples && !sched->idle_hist)
2838 printf("\nIdle stats:\n");
2839 for (i = 0; i < idle_max_cpu; ++i) {
2840 t = idle_threads[i];
2844 r = thread__priv(t);
2845 if (r && r->run_stats.n) {
2846 totals.sched_count += r->run_stats.n;
2847 printf(" CPU %2d idle for ", i);
2848 print_sched_time(r->total_run_time, 6);
2849 printf(" msec (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2851 printf(" CPU %2d idle entire time window\n", i);
2854 if (sched->idle_hist && sched->show_callchain) {
2855 callchain_param.mode = CHAIN_FOLDED;
2856 callchain_param.value = CCVAL_PERIOD;
2858 callchain_register_param(&callchain_param);
2860 printf("\nIdle stats by callchain:\n");
2861 for (i = 0; i < idle_max_cpu; ++i) {
2862 struct idle_thread_runtime *itr;
2864 t = idle_threads[i];
2868 itr = thread__priv(t);
2872 callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
2873 0, &callchain_param);
2875 printf(" CPU %2d:", i);
2876 print_sched_time(itr->tr.total_run_time, 6);
2878 timehist_print_idlehist_callchain(&itr->sorted_root);
2884 " Total number of unique tasks: %" PRIu64 "\n"
2885 "Total number of context switches: %" PRIu64 "\n",
2886 totals.task_count, totals.sched_count);
2888 printf(" Total run time (msec): ");
2889 print_sched_time(totals.total_run_time, 2);
2892 printf(" Total scheduling time (msec): ");
2893 print_sched_time(hist_time, 2);
2894 printf(" (x %d)\n", sched->max_cpu);
2897 typedef int (*sched_handler)(struct perf_tool *tool,
2898 union perf_event *event,
2899 struct perf_evsel *evsel,
2900 struct perf_sample *sample,
2901 struct machine *machine);
2903 static int perf_timehist__process_sample(struct perf_tool *tool,
2904 union perf_event *event,
2905 struct perf_sample *sample,
2906 struct perf_evsel *evsel,
2907 struct machine *machine)
2909 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2911 int this_cpu = sample->cpu;
2913 if (this_cpu > sched->max_cpu)
2914 sched->max_cpu = this_cpu;
2916 if (evsel->handler != NULL) {
2917 sched_handler f = evsel->handler;
2919 err = f(tool, event, evsel, sample, machine);
2925 static int timehist_check_attr(struct perf_sched *sched,
2926 struct perf_evlist *evlist)
2928 struct perf_evsel *evsel;
2929 struct evsel_runtime *er;
2931 list_for_each_entry(evsel, &evlist->entries, node) {
2932 er = perf_evsel__get_runtime(evsel);
2934 pr_err("Failed to allocate memory for evsel runtime data\n");
2938 if (sched->show_callchain && !evsel__has_callchain(evsel)) {
2939 pr_info("Samples do not have callchains.\n");
2940 sched->show_callchain = 0;
2941 symbol_conf.use_callchain = 0;
2948 static int perf_sched__timehist(struct perf_sched *sched)
2950 const struct perf_evsel_str_handler handlers[] = {
2951 { "sched:sched_switch", timehist_sched_switch_event, },
2952 { "sched:sched_wakeup", timehist_sched_wakeup_event, },
2953 { "sched:sched_wakeup_new", timehist_sched_wakeup_event, },
2955 const struct perf_evsel_str_handler migrate_handlers[] = {
2956 { "sched:sched_migrate_task", timehist_migrate_task_event, },
2958 struct perf_data data = {
2960 .mode = PERF_DATA_MODE_READ,
2961 .force = sched->force,
2964 struct perf_session *session;
2965 struct perf_evlist *evlist;
2969 * event handlers for timehist option
2971 sched->tool.sample = perf_timehist__process_sample;
2972 sched->tool.mmap = perf_event__process_mmap;
2973 sched->tool.comm = perf_event__process_comm;
2974 sched->tool.exit = perf_event__process_exit;
2975 sched->tool.fork = perf_event__process_fork;
2976 sched->tool.lost = process_lost;
2977 sched->tool.attr = perf_event__process_attr;
2978 sched->tool.tracing_data = perf_event__process_tracing_data;
2979 sched->tool.build_id = perf_event__process_build_id;
2981 sched->tool.ordered_events = true;
2982 sched->tool.ordering_requires_timestamps = true;
2984 symbol_conf.use_callchain = sched->show_callchain;
2986 session = perf_session__new(&data, false, &sched->tool);
2987 if (session == NULL)
2990 evlist = session->evlist;
2992 symbol__init(&session->header.env);
2994 if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
2995 pr_err("Invalid time string\n");
2999 if (timehist_check_attr(sched, evlist) != 0)
3004 /* setup per-evsel handlers */
3005 if (perf_session__set_tracepoints_handlers(session, handlers))
3008 /* sched_switch event at a minimum needs to exist */
3009 if (!perf_evlist__find_tracepoint_by_name(session->evlist,
3010 "sched:sched_switch")) {
3011 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3015 if (sched->show_migrations &&
3016 perf_session__set_tracepoints_handlers(session, migrate_handlers))
3019 /* pre-allocate struct for per-CPU idle stats */
3020 sched->max_cpu = session->header.env.nr_cpus_online;
3021 if (sched->max_cpu == 0)
3023 if (init_idle_threads(sched->max_cpu))
3026 /* summary_only implies summary option, but don't overwrite summary if set */
3027 if (sched->summary_only)
3028 sched->summary = sched->summary_only;
3030 if (!sched->summary_only)
3031 timehist_header(sched);
3033 err = perf_session__process_events(session);
3035 pr_err("Failed to process events, error %d", err);
3039 sched->nr_events = evlist->stats.nr_events[0];
3040 sched->nr_lost_events = evlist->stats.total_lost;
3041 sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3044 timehist_print_summary(sched, session);
3047 free_idle_threads();
3048 perf_session__delete(session);
3054 static void print_bad_events(struct perf_sched *sched)
3056 if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3057 printf(" INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3058 (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3059 sched->nr_unordered_timestamps, sched->nr_timestamps);
3061 if (sched->nr_lost_events && sched->nr_events) {
3062 printf(" INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3063 (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3064 sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3066 if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3067 printf(" INFO: %.3f%% context switch bugs (%ld out of %ld)",
3068 (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3069 sched->nr_context_switch_bugs, sched->nr_timestamps);
3070 if (sched->nr_lost_events)
3071 printf(" (due to lost events?)");
3076 static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3078 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3079 struct work_atoms *this;
3080 const char *comm = thread__comm_str(data->thread), *this_comm;
3081 bool leftmost = true;
3086 this = container_of(*new, struct work_atoms, node);
3089 this_comm = thread__comm_str(this->thread);
3090 cmp = strcmp(comm, this_comm);
3092 new = &((*new)->rb_left);
3093 } else if (cmp < 0) {
3094 new = &((*new)->rb_right);
3098 this->total_runtime += data->total_runtime;
3099 this->nb_atoms += data->nb_atoms;
3100 this->total_lat += data->total_lat;
3101 list_splice(&data->work_list, &this->work_list);
3102 if (this->max_lat < data->max_lat) {
3103 this->max_lat = data->max_lat;
3104 this->max_lat_at = data->max_lat_at;
3112 rb_link_node(&data->node, parent, new);
3113 rb_insert_color_cached(&data->node, root, leftmost);
3116 static void perf_sched__merge_lat(struct perf_sched *sched)
3118 struct work_atoms *data;
3119 struct rb_node *node;
3121 if (sched->skip_merge)
3124 while ((node = rb_first_cached(&sched->atom_root))) {
3125 rb_erase_cached(node, &sched->atom_root);
3126 data = rb_entry(node, struct work_atoms, node);
3127 __merge_work_atoms(&sched->merged_atom_root, data);
3131 static int perf_sched__lat(struct perf_sched *sched)
3133 struct rb_node *next;
3137 if (perf_sched__read_events(sched))
3140 perf_sched__merge_lat(sched);
3141 perf_sched__sort_lat(sched);
3143 printf("\n -----------------------------------------------------------------------------------------------------------------\n");
3144 printf(" Task | Runtime ms | Switches | Average delay ms | Maximum delay ms | Maximum delay at |\n");
3145 printf(" -----------------------------------------------------------------------------------------------------------------\n");
3147 next = rb_first_cached(&sched->sorted_atom_root);
3150 struct work_atoms *work_list;
3152 work_list = rb_entry(next, struct work_atoms, node);
3153 output_lat_thread(sched, work_list);
3154 next = rb_next(next);
3155 thread__zput(work_list->thread);
3158 printf(" -----------------------------------------------------------------------------------------------------------------\n");
3159 printf(" TOTAL: |%11.3f ms |%9" PRIu64 " |\n",
3160 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3162 printf(" ---------------------------------------------------\n");
3164 print_bad_events(sched);
3170 static int setup_map_cpus(struct perf_sched *sched)
3172 struct cpu_map *map;
3174 sched->max_cpu = sysconf(_SC_NPROCESSORS_CONF);
3176 if (sched->map.comp) {
3177 sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
3178 if (!sched->map.comp_cpus)
3182 if (!sched->map.cpus_str)
3185 map = cpu_map__new(sched->map.cpus_str);
3187 pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3191 sched->map.cpus = map;
3195 static int setup_color_pids(struct perf_sched *sched)
3197 struct thread_map *map;
3199 if (!sched->map.color_pids_str)
3202 map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3204 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3208 sched->map.color_pids = map;
3212 static int setup_color_cpus(struct perf_sched *sched)
3214 struct cpu_map *map;
3216 if (!sched->map.color_cpus_str)
3219 map = cpu_map__new(sched->map.color_cpus_str);
3221 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3225 sched->map.color_cpus = map;
3229 static int perf_sched__map(struct perf_sched *sched)
3231 if (setup_map_cpus(sched))
3234 if (setup_color_pids(sched))
3237 if (setup_color_cpus(sched))
3241 if (perf_sched__read_events(sched))
3243 print_bad_events(sched);
3247 static int perf_sched__replay(struct perf_sched *sched)
3251 calibrate_run_measurement_overhead(sched);
3252 calibrate_sleep_measurement_overhead(sched);
3254 test_calibrations(sched);
3256 if (perf_sched__read_events(sched))
3259 printf("nr_run_events: %ld\n", sched->nr_run_events);
3260 printf("nr_sleep_events: %ld\n", sched->nr_sleep_events);
3261 printf("nr_wakeup_events: %ld\n", sched->nr_wakeup_events);
3263 if (sched->targetless_wakeups)
3264 printf("target-less wakeups: %ld\n", sched->targetless_wakeups);
3265 if (sched->multitarget_wakeups)
3266 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3267 if (sched->nr_run_events_optimized)
3268 printf("run atoms optimized: %ld\n",
3269 sched->nr_run_events_optimized);
3271 print_task_traces(sched);
3272 add_cross_task_wakeups(sched);
3274 create_tasks(sched);
3275 printf("------------------------------------------------------------\n");
3276 for (i = 0; i < sched->replay_repeat; i++)
3277 run_one_test(sched);
3282 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3283 const char * const usage_msg[])
3285 char *tmp, *tok, *str = strdup(sched->sort_order);
3287 for (tok = strtok_r(str, ", ", &tmp);
3288 tok; tok = strtok_r(NULL, ", ", &tmp)) {
3289 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3290 usage_with_options_msg(usage_msg, options,
3291 "Unknown --sort key: `%s'", tok);
3297 sort_dimension__add("pid", &sched->cmp_pid);
3300 static int __cmd_record(int argc, const char **argv)
3302 unsigned int rec_argc, i, j;
3303 const char **rec_argv;
3304 const char * const record_args[] = {
3310 "-e", "sched:sched_switch",
3311 "-e", "sched:sched_stat_wait",
3312 "-e", "sched:sched_stat_sleep",
3313 "-e", "sched:sched_stat_iowait",
3314 "-e", "sched:sched_stat_runtime",
3315 "-e", "sched:sched_process_fork",
3316 "-e", "sched:sched_wakeup",
3317 "-e", "sched:sched_wakeup_new",
3318 "-e", "sched:sched_migrate_task",
3321 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
3322 rec_argv = calloc(rec_argc + 1, sizeof(char *));
3324 if (rec_argv == NULL)
3327 for (i = 0; i < ARRAY_SIZE(record_args); i++)
3328 rec_argv[i] = strdup(record_args[i]);
3330 for (j = 1; j < (unsigned int)argc; j++, i++)
3331 rec_argv[i] = argv[j];
3333 BUG_ON(i != rec_argc);
3335 return cmd_record(i, rec_argv);
3338 int cmd_sched(int argc, const char **argv)
3340 static const char default_sort_order[] = "avg, max, switch, runtime";
3341 struct perf_sched sched = {
3343 .sample = perf_sched__process_tracepoint_sample,
3344 .comm = perf_sched__process_comm,
3345 .namespaces = perf_event__process_namespaces,
3346 .lost = perf_event__process_lost,
3347 .fork = perf_sched__process_fork_event,
3348 .ordered_events = true,
3350 .cmp_pid = LIST_HEAD_INIT(sched.cmp_pid),
3351 .sort_list = LIST_HEAD_INIT(sched.sort_list),
3352 .start_work_mutex = PTHREAD_MUTEX_INITIALIZER,
3353 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
3354 .sort_order = default_sort_order,
3355 .replay_repeat = 10,
3357 .next_shortname1 = 'A',
3358 .next_shortname2 = '0',
3360 .show_callchain = 1,
3363 const struct option sched_options[] = {
3364 OPT_STRING('i', "input", &input_name, "file",
3366 OPT_INCR('v', "verbose", &verbose,
3367 "be more verbose (show symbol address, etc)"),
3368 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3369 "dump raw trace in ASCII"),
3370 OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3373 const struct option latency_options[] = {
3374 OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3375 "sort by key(s): runtime, switch, avg, max"),
3376 OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3377 "CPU to profile on"),
3378 OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3379 "latency stats per pid instead of per comm"),
3380 OPT_PARENT(sched_options)
3382 const struct option replay_options[] = {
3383 OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3384 "repeat the workload replay N times (-1: infinite)"),
3385 OPT_PARENT(sched_options)
3387 const struct option map_options[] = {
3388 OPT_BOOLEAN(0, "compact", &sched.map.comp,
3389 "map output in compact mode"),
3390 OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3391 "highlight given pids in map"),
3392 OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3393 "highlight given CPUs in map"),
3394 OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3395 "display given CPUs in map"),
3396 OPT_PARENT(sched_options)
3398 const struct option timehist_options[] = {
3399 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3400 "file", "vmlinux pathname"),
3401 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3402 "file", "kallsyms pathname"),
3403 OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3404 "Display call chains if present (default on)"),
3405 OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3406 "Maximum number of functions to display backtrace."),
3407 OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3408 "Look for files with symbols relative to this directory"),
3409 OPT_BOOLEAN('s', "summary", &sched.summary_only,
3410 "Show only syscall summary with statistics"),
3411 OPT_BOOLEAN('S', "with-summary", &sched.summary,
3412 "Show all syscalls and summary with statistics"),
3413 OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3414 OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3415 OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3416 OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3417 OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3418 OPT_STRING(0, "time", &sched.time_str, "str",
3419 "Time span for analysis (start,stop)"),
3420 OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3421 OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3422 "analyze events only for given process id(s)"),
3423 OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3424 "analyze events only for given thread id(s)"),
3425 OPT_PARENT(sched_options)
3428 const char * const latency_usage[] = {
3429 "perf sched latency [<options>]",
3432 const char * const replay_usage[] = {
3433 "perf sched replay [<options>]",
3436 const char * const map_usage[] = {
3437 "perf sched map [<options>]",
3440 const char * const timehist_usage[] = {
3441 "perf sched timehist [<options>]",
3444 const char *const sched_subcommands[] = { "record", "latency", "map",
3447 const char *sched_usage[] = {
3451 struct trace_sched_handler lat_ops = {
3452 .wakeup_event = latency_wakeup_event,
3453 .switch_event = latency_switch_event,
3454 .runtime_event = latency_runtime_event,
3455 .migrate_task_event = latency_migrate_task_event,
3457 struct trace_sched_handler map_ops = {
3458 .switch_event = map_switch_event,
3460 struct trace_sched_handler replay_ops = {
3461 .wakeup_event = replay_wakeup_event,
3462 .switch_event = replay_switch_event,
3463 .fork_event = replay_fork_event,
3467 for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3468 sched.curr_pid[i] = -1;
3470 argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3471 sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3473 usage_with_options(sched_usage, sched_options);
3476 * Aliased to 'perf script' for now:
3478 if (!strcmp(argv[0], "script"))
3479 return cmd_script(argc, argv);
3481 if (!strncmp(argv[0], "rec", 3)) {
3482 return __cmd_record(argc, argv);
3483 } else if (!strncmp(argv[0], "lat", 3)) {
3484 sched.tp_handler = &lat_ops;
3486 argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3488 usage_with_options(latency_usage, latency_options);
3490 setup_sorting(&sched, latency_options, latency_usage);
3491 return perf_sched__lat(&sched);
3492 } else if (!strcmp(argv[0], "map")) {
3494 argc = parse_options(argc, argv, map_options, map_usage, 0);
3496 usage_with_options(map_usage, map_options);
3498 sched.tp_handler = &map_ops;
3499 setup_sorting(&sched, latency_options, latency_usage);
3500 return perf_sched__map(&sched);
3501 } else if (!strncmp(argv[0], "rep", 3)) {
3502 sched.tp_handler = &replay_ops;
3504 argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3506 usage_with_options(replay_usage, replay_options);
3508 return perf_sched__replay(&sched);
3509 } else if (!strcmp(argv[0], "timehist")) {
3511 argc = parse_options(argc, argv, timehist_options,
3514 usage_with_options(timehist_usage, timehist_options);
3516 if ((sched.show_wakeups || sched.show_next) &&
3517 sched.summary_only) {
3518 pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3519 parse_options_usage(timehist_usage, timehist_options, "s", true);
3520 if (sched.show_wakeups)
3521 parse_options_usage(NULL, timehist_options, "w", true);
3522 if (sched.show_next)
3523 parse_options_usage(NULL, timehist_options, "n", true);
3527 return perf_sched__timehist(&sched);
3529 usage_with_options(sched_usage, sched_options);