Loading...
1// SPDX-License-Identifier: GPL-2.0
2#include "builtin.h"
3#include "perf.h"
4#include "perf-sys.h"
5
6#include "util/cpumap.h"
7#include "util/evlist.h"
8#include "util/evsel.h"
9#include "util/evsel_fprintf.h"
10#include "util/mutex.h"
11#include "util/symbol.h"
12#include "util/thread.h"
13#include "util/header.h"
14#include "util/session.h"
15#include "util/tool.h"
16#include "util/cloexec.h"
17#include "util/thread_map.h"
18#include "util/color.h"
19#include "util/stat.h"
20#include "util/string2.h"
21#include "util/callchain.h"
22#include "util/time-utils.h"
23
24#include <subcmd/pager.h>
25#include <subcmd/parse-options.h>
26#include "util/trace-event.h"
27
28#include "util/debug.h"
29#include "util/event.h"
30
31#include <linux/kernel.h>
32#include <linux/log2.h>
33#include <linux/zalloc.h>
34#include <sys/prctl.h>
35#include <sys/resource.h>
36#include <inttypes.h>
37
38#include <errno.h>
39#include <semaphore.h>
40#include <pthread.h>
41#include <math.h>
42#include <api/fs/fs.h>
43#include <perf/cpumap.h>
44#include <linux/time64.h>
45#include <linux/err.h>
46
47#include <linux/ctype.h>
48
49#define PR_SET_NAME 15 /* Set process name */
50#define MAX_CPUS 4096
51#define COMM_LEN 20
52#define SYM_LEN 129
53#define MAX_PID 1024000
54
55static const char *cpu_list;
56static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
57
58struct sched_atom;
59
60struct task_desc {
61 unsigned long nr;
62 unsigned long pid;
63 char comm[COMM_LEN];
64
65 unsigned long nr_events;
66 unsigned long curr_event;
67 struct sched_atom **atoms;
68
69 pthread_t thread;
70 sem_t sleep_sem;
71
72 sem_t ready_for_work;
73 sem_t work_done_sem;
74
75 u64 cpu_usage;
76};
77
78enum sched_event_type {
79 SCHED_EVENT_RUN,
80 SCHED_EVENT_SLEEP,
81 SCHED_EVENT_WAKEUP,
82 SCHED_EVENT_MIGRATION,
83};
84
85struct sched_atom {
86 enum sched_event_type type;
87 int specific_wait;
88 u64 timestamp;
89 u64 duration;
90 unsigned long nr;
91 sem_t *wait_sem;
92 struct task_desc *wakee;
93};
94
95#define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
96
97/* task state bitmask, copied from include/linux/sched.h */
98#define TASK_RUNNING 0
99#define TASK_INTERRUPTIBLE 1
100#define TASK_UNINTERRUPTIBLE 2
101#define __TASK_STOPPED 4
102#define __TASK_TRACED 8
103/* in tsk->exit_state */
104#define EXIT_DEAD 16
105#define EXIT_ZOMBIE 32
106#define EXIT_TRACE (EXIT_ZOMBIE | EXIT_DEAD)
107/* in tsk->state again */
108#define TASK_DEAD 64
109#define TASK_WAKEKILL 128
110#define TASK_WAKING 256
111#define TASK_PARKED 512
112
113enum thread_state {
114 THREAD_SLEEPING = 0,
115 THREAD_WAIT_CPU,
116 THREAD_SCHED_IN,
117 THREAD_IGNORE
118};
119
120struct work_atom {
121 struct list_head list;
122 enum thread_state state;
123 u64 sched_out_time;
124 u64 wake_up_time;
125 u64 sched_in_time;
126 u64 runtime;
127};
128
129struct work_atoms {
130 struct list_head work_list;
131 struct thread *thread;
132 struct rb_node node;
133 u64 max_lat;
134 u64 max_lat_start;
135 u64 max_lat_end;
136 u64 total_lat;
137 u64 nb_atoms;
138 u64 total_runtime;
139 int num_merged;
140};
141
142typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
143
144struct perf_sched;
145
146struct trace_sched_handler {
147 int (*switch_event)(struct perf_sched *sched, struct evsel *evsel,
148 struct perf_sample *sample, struct machine *machine);
149
150 int (*runtime_event)(struct perf_sched *sched, struct evsel *evsel,
151 struct perf_sample *sample, struct machine *machine);
152
153 int (*wakeup_event)(struct perf_sched *sched, struct evsel *evsel,
154 struct perf_sample *sample, struct machine *machine);
155
156 /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
157 int (*fork_event)(struct perf_sched *sched, union perf_event *event,
158 struct machine *machine);
159
160 int (*migrate_task_event)(struct perf_sched *sched,
161 struct evsel *evsel,
162 struct perf_sample *sample,
163 struct machine *machine);
164};
165
166#define COLOR_PIDS PERF_COLOR_BLUE
167#define COLOR_CPUS PERF_COLOR_BG_RED
168
169struct perf_sched_map {
170 DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
171 struct perf_cpu *comp_cpus;
172 bool comp;
173 struct perf_thread_map *color_pids;
174 const char *color_pids_str;
175 struct perf_cpu_map *color_cpus;
176 const char *color_cpus_str;
177 struct perf_cpu_map *cpus;
178 const char *cpus_str;
179};
180
181struct perf_sched {
182 struct perf_tool tool;
183 const char *sort_order;
184 unsigned long nr_tasks;
185 struct task_desc **pid_to_task;
186 struct task_desc **tasks;
187 const struct trace_sched_handler *tp_handler;
188 struct mutex start_work_mutex;
189 struct mutex work_done_wait_mutex;
190 int profile_cpu;
191/*
192 * Track the current task - that way we can know whether there's any
193 * weird events, such as a task being switched away that is not current.
194 */
195 struct perf_cpu max_cpu;
196 u32 curr_pid[MAX_CPUS];
197 struct thread *curr_thread[MAX_CPUS];
198 char next_shortname1;
199 char next_shortname2;
200 unsigned int replay_repeat;
201 unsigned long nr_run_events;
202 unsigned long nr_sleep_events;
203 unsigned long nr_wakeup_events;
204 unsigned long nr_sleep_corrections;
205 unsigned long nr_run_events_optimized;
206 unsigned long targetless_wakeups;
207 unsigned long multitarget_wakeups;
208 unsigned long nr_runs;
209 unsigned long nr_timestamps;
210 unsigned long nr_unordered_timestamps;
211 unsigned long nr_context_switch_bugs;
212 unsigned long nr_events;
213 unsigned long nr_lost_chunks;
214 unsigned long nr_lost_events;
215 u64 run_measurement_overhead;
216 u64 sleep_measurement_overhead;
217 u64 start_time;
218 u64 cpu_usage;
219 u64 runavg_cpu_usage;
220 u64 parent_cpu_usage;
221 u64 runavg_parent_cpu_usage;
222 u64 sum_runtime;
223 u64 sum_fluct;
224 u64 run_avg;
225 u64 all_runtime;
226 u64 all_count;
227 u64 cpu_last_switched[MAX_CPUS];
228 struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
229 struct list_head sort_list, cmp_pid;
230 bool force;
231 bool skip_merge;
232 struct perf_sched_map map;
233
234 /* options for timehist command */
235 bool summary;
236 bool summary_only;
237 bool idle_hist;
238 bool show_callchain;
239 unsigned int max_stack;
240 bool show_cpu_visual;
241 bool show_wakeups;
242 bool show_next;
243 bool show_migrations;
244 bool show_state;
245 u64 skipped_samples;
246 const char *time_str;
247 struct perf_time_interval ptime;
248 struct perf_time_interval hist_time;
249 volatile bool thread_funcs_exit;
250};
251
252/* per thread run time data */
253struct thread_runtime {
254 u64 last_time; /* time of previous sched in/out event */
255 u64 dt_run; /* run time */
256 u64 dt_sleep; /* time between CPU access by sleep (off cpu) */
257 u64 dt_iowait; /* time between CPU access by iowait (off cpu) */
258 u64 dt_preempt; /* time between CPU access by preempt (off cpu) */
259 u64 dt_delay; /* time between wakeup and sched-in */
260 u64 ready_to_run; /* time of wakeup */
261
262 struct stats run_stats;
263 u64 total_run_time;
264 u64 total_sleep_time;
265 u64 total_iowait_time;
266 u64 total_preempt_time;
267 u64 total_delay_time;
268
269 int last_state;
270
271 char shortname[3];
272 bool comm_changed;
273
274 u64 migrations;
275};
276
277/* per event run time data */
278struct evsel_runtime {
279 u64 *last_time; /* time this event was last seen per cpu */
280 u32 ncpu; /* highest cpu slot allocated */
281};
282
283/* per cpu idle time data */
284struct idle_thread_runtime {
285 struct thread_runtime tr;
286 struct thread *last_thread;
287 struct rb_root_cached sorted_root;
288 struct callchain_root callchain;
289 struct callchain_cursor cursor;
290};
291
292/* track idle times per cpu */
293static struct thread **idle_threads;
294static int idle_max_cpu;
295static char idle_comm[] = "<idle>";
296
297static u64 get_nsecs(void)
298{
299 struct timespec ts;
300
301 clock_gettime(CLOCK_MONOTONIC, &ts);
302
303 return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
304}
305
306static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
307{
308 u64 T0 = get_nsecs(), T1;
309
310 do {
311 T1 = get_nsecs();
312 } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
313}
314
315static void sleep_nsecs(u64 nsecs)
316{
317 struct timespec ts;
318
319 ts.tv_nsec = nsecs % 999999999;
320 ts.tv_sec = nsecs / 999999999;
321
322 nanosleep(&ts, NULL);
323}
324
325static void calibrate_run_measurement_overhead(struct perf_sched *sched)
326{
327 u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
328 int i;
329
330 for (i = 0; i < 10; i++) {
331 T0 = get_nsecs();
332 burn_nsecs(sched, 0);
333 T1 = get_nsecs();
334 delta = T1-T0;
335 min_delta = min(min_delta, delta);
336 }
337 sched->run_measurement_overhead = min_delta;
338
339 printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
340}
341
342static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
343{
344 u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
345 int i;
346
347 for (i = 0; i < 10; i++) {
348 T0 = get_nsecs();
349 sleep_nsecs(10000);
350 T1 = get_nsecs();
351 delta = T1-T0;
352 min_delta = min(min_delta, delta);
353 }
354 min_delta -= 10000;
355 sched->sleep_measurement_overhead = min_delta;
356
357 printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
358}
359
360static struct sched_atom *
361get_new_event(struct task_desc *task, u64 timestamp)
362{
363 struct sched_atom *event = zalloc(sizeof(*event));
364 unsigned long idx = task->nr_events;
365 size_t size;
366
367 event->timestamp = timestamp;
368 event->nr = idx;
369
370 task->nr_events++;
371 size = sizeof(struct sched_atom *) * task->nr_events;
372 task->atoms = realloc(task->atoms, size);
373 BUG_ON(!task->atoms);
374
375 task->atoms[idx] = event;
376
377 return event;
378}
379
380static struct sched_atom *last_event(struct task_desc *task)
381{
382 if (!task->nr_events)
383 return NULL;
384
385 return task->atoms[task->nr_events - 1];
386}
387
388static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
389 u64 timestamp, u64 duration)
390{
391 struct sched_atom *event, *curr_event = last_event(task);
392
393 /*
394 * optimize an existing RUN event by merging this one
395 * to it:
396 */
397 if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
398 sched->nr_run_events_optimized++;
399 curr_event->duration += duration;
400 return;
401 }
402
403 event = get_new_event(task, timestamp);
404
405 event->type = SCHED_EVENT_RUN;
406 event->duration = duration;
407
408 sched->nr_run_events++;
409}
410
411static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
412 u64 timestamp, struct task_desc *wakee)
413{
414 struct sched_atom *event, *wakee_event;
415
416 event = get_new_event(task, timestamp);
417 event->type = SCHED_EVENT_WAKEUP;
418 event->wakee = wakee;
419
420 wakee_event = last_event(wakee);
421 if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
422 sched->targetless_wakeups++;
423 return;
424 }
425 if (wakee_event->wait_sem) {
426 sched->multitarget_wakeups++;
427 return;
428 }
429
430 wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
431 sem_init(wakee_event->wait_sem, 0, 0);
432 wakee_event->specific_wait = 1;
433 event->wait_sem = wakee_event->wait_sem;
434
435 sched->nr_wakeup_events++;
436}
437
438static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
439 u64 timestamp, u64 task_state __maybe_unused)
440{
441 struct sched_atom *event = get_new_event(task, timestamp);
442
443 event->type = SCHED_EVENT_SLEEP;
444
445 sched->nr_sleep_events++;
446}
447
448static struct task_desc *register_pid(struct perf_sched *sched,
449 unsigned long pid, const char *comm)
450{
451 struct task_desc *task;
452 static int pid_max;
453
454 if (sched->pid_to_task == NULL) {
455 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
456 pid_max = MAX_PID;
457 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
458 }
459 if (pid >= (unsigned long)pid_max) {
460 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
461 sizeof(struct task_desc *))) == NULL);
462 while (pid >= (unsigned long)pid_max)
463 sched->pid_to_task[pid_max++] = NULL;
464 }
465
466 task = sched->pid_to_task[pid];
467
468 if (task)
469 return task;
470
471 task = zalloc(sizeof(*task));
472 task->pid = pid;
473 task->nr = sched->nr_tasks;
474 strcpy(task->comm, comm);
475 /*
476 * every task starts in sleeping state - this gets ignored
477 * if there's no wakeup pointing to this sleep state:
478 */
479 add_sched_event_sleep(sched, task, 0, 0);
480
481 sched->pid_to_task[pid] = task;
482 sched->nr_tasks++;
483 sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
484 BUG_ON(!sched->tasks);
485 sched->tasks[task->nr] = task;
486
487 if (verbose > 0)
488 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
489
490 return task;
491}
492
493
494static void print_task_traces(struct perf_sched *sched)
495{
496 struct task_desc *task;
497 unsigned long i;
498
499 for (i = 0; i < sched->nr_tasks; i++) {
500 task = sched->tasks[i];
501 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
502 task->nr, task->comm, task->pid, task->nr_events);
503 }
504}
505
506static void add_cross_task_wakeups(struct perf_sched *sched)
507{
508 struct task_desc *task1, *task2;
509 unsigned long i, j;
510
511 for (i = 0; i < sched->nr_tasks; i++) {
512 task1 = sched->tasks[i];
513 j = i + 1;
514 if (j == sched->nr_tasks)
515 j = 0;
516 task2 = sched->tasks[j];
517 add_sched_event_wakeup(sched, task1, 0, task2);
518 }
519}
520
521static void perf_sched__process_event(struct perf_sched *sched,
522 struct sched_atom *atom)
523{
524 int ret = 0;
525
526 switch (atom->type) {
527 case SCHED_EVENT_RUN:
528 burn_nsecs(sched, atom->duration);
529 break;
530 case SCHED_EVENT_SLEEP:
531 if (atom->wait_sem)
532 ret = sem_wait(atom->wait_sem);
533 BUG_ON(ret);
534 break;
535 case SCHED_EVENT_WAKEUP:
536 if (atom->wait_sem)
537 ret = sem_post(atom->wait_sem);
538 BUG_ON(ret);
539 break;
540 case SCHED_EVENT_MIGRATION:
541 break;
542 default:
543 BUG_ON(1);
544 }
545}
546
547static u64 get_cpu_usage_nsec_parent(void)
548{
549 struct rusage ru;
550 u64 sum;
551 int err;
552
553 err = getrusage(RUSAGE_SELF, &ru);
554 BUG_ON(err);
555
556 sum = ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
557 sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
558
559 return sum;
560}
561
562static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
563{
564 struct perf_event_attr attr;
565 char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
566 int fd;
567 struct rlimit limit;
568 bool need_privilege = false;
569
570 memset(&attr, 0, sizeof(attr));
571
572 attr.type = PERF_TYPE_SOFTWARE;
573 attr.config = PERF_COUNT_SW_TASK_CLOCK;
574
575force_again:
576 fd = sys_perf_event_open(&attr, 0, -1, -1,
577 perf_event_open_cloexec_flag());
578
579 if (fd < 0) {
580 if (errno == EMFILE) {
581 if (sched->force) {
582 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
583 limit.rlim_cur += sched->nr_tasks - cur_task;
584 if (limit.rlim_cur > limit.rlim_max) {
585 limit.rlim_max = limit.rlim_cur;
586 need_privilege = true;
587 }
588 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
589 if (need_privilege && errno == EPERM)
590 strcpy(info, "Need privilege\n");
591 } else
592 goto force_again;
593 } else
594 strcpy(info, "Have a try with -f option\n");
595 }
596 pr_err("Error: sys_perf_event_open() syscall returned "
597 "with %d (%s)\n%s", fd,
598 str_error_r(errno, sbuf, sizeof(sbuf)), info);
599 exit(EXIT_FAILURE);
600 }
601 return fd;
602}
603
604static u64 get_cpu_usage_nsec_self(int fd)
605{
606 u64 runtime;
607 int ret;
608
609 ret = read(fd, &runtime, sizeof(runtime));
610 BUG_ON(ret != sizeof(runtime));
611
612 return runtime;
613}
614
615struct sched_thread_parms {
616 struct task_desc *task;
617 struct perf_sched *sched;
618 int fd;
619};
620
621static void *thread_func(void *ctx)
622{
623 struct sched_thread_parms *parms = ctx;
624 struct task_desc *this_task = parms->task;
625 struct perf_sched *sched = parms->sched;
626 u64 cpu_usage_0, cpu_usage_1;
627 unsigned long i, ret;
628 char comm2[22];
629 int fd = parms->fd;
630
631 zfree(&parms);
632
633 sprintf(comm2, ":%s", this_task->comm);
634 prctl(PR_SET_NAME, comm2);
635 if (fd < 0)
636 return NULL;
637
638 while (!sched->thread_funcs_exit) {
639 ret = sem_post(&this_task->ready_for_work);
640 BUG_ON(ret);
641 mutex_lock(&sched->start_work_mutex);
642 mutex_unlock(&sched->start_work_mutex);
643
644 cpu_usage_0 = get_cpu_usage_nsec_self(fd);
645
646 for (i = 0; i < this_task->nr_events; i++) {
647 this_task->curr_event = i;
648 perf_sched__process_event(sched, this_task->atoms[i]);
649 }
650
651 cpu_usage_1 = get_cpu_usage_nsec_self(fd);
652 this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
653 ret = sem_post(&this_task->work_done_sem);
654 BUG_ON(ret);
655
656 mutex_lock(&sched->work_done_wait_mutex);
657 mutex_unlock(&sched->work_done_wait_mutex);
658 }
659 return NULL;
660}
661
662static void create_tasks(struct perf_sched *sched)
663 EXCLUSIVE_LOCK_FUNCTION(sched->start_work_mutex)
664 EXCLUSIVE_LOCK_FUNCTION(sched->work_done_wait_mutex)
665{
666 struct task_desc *task;
667 pthread_attr_t attr;
668 unsigned long i;
669 int err;
670
671 err = pthread_attr_init(&attr);
672 BUG_ON(err);
673 err = pthread_attr_setstacksize(&attr,
674 (size_t) max(16 * 1024, (int)PTHREAD_STACK_MIN));
675 BUG_ON(err);
676 mutex_lock(&sched->start_work_mutex);
677 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);
689 BUG_ON(err);
690 }
691}
692
693static void destroy_tasks(struct perf_sched *sched)
694 UNLOCK_FUNCTION(sched->start_work_mutex)
695 UNLOCK_FUNCTION(sched->work_done_wait_mutex)
696{
697 struct task_desc *task;
698 unsigned long i;
699 int err;
700
701 mutex_unlock(&sched->start_work_mutex);
702 mutex_unlock(&sched->work_done_wait_mutex);
703 /* Get rid of threads so they won't be upset by mutex destrunction */
704 for (i = 0; i < sched->nr_tasks; i++) {
705 task = sched->tasks[i];
706 err = pthread_join(task->thread, NULL);
707 BUG_ON(err);
708 sem_destroy(&task->sleep_sem);
709 sem_destroy(&task->ready_for_work);
710 sem_destroy(&task->work_done_sem);
711 }
712}
713
714static void wait_for_tasks(struct perf_sched *sched)
715 EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
716 EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
717{
718 u64 cpu_usage_0, cpu_usage_1;
719 struct task_desc *task;
720 unsigned long i, ret;
721
722 sched->start_time = get_nsecs();
723 sched->cpu_usage = 0;
724 mutex_unlock(&sched->work_done_wait_mutex);
725
726 for (i = 0; i < sched->nr_tasks; i++) {
727 task = sched->tasks[i];
728 ret = sem_wait(&task->ready_for_work);
729 BUG_ON(ret);
730 sem_init(&task->ready_for_work, 0, 0);
731 }
732 mutex_lock(&sched->work_done_wait_mutex);
733
734 cpu_usage_0 = get_cpu_usage_nsec_parent();
735
736 mutex_unlock(&sched->start_work_mutex);
737
738 for (i = 0; i < sched->nr_tasks; i++) {
739 task = sched->tasks[i];
740 ret = sem_wait(&task->work_done_sem);
741 BUG_ON(ret);
742 sem_init(&task->work_done_sem, 0, 0);
743 sched->cpu_usage += task->cpu_usage;
744 task->cpu_usage = 0;
745 }
746
747 cpu_usage_1 = get_cpu_usage_nsec_parent();
748 if (!sched->runavg_cpu_usage)
749 sched->runavg_cpu_usage = sched->cpu_usage;
750 sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
751
752 sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
753 if (!sched->runavg_parent_cpu_usage)
754 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
755 sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
756 sched->parent_cpu_usage)/sched->replay_repeat;
757
758 mutex_lock(&sched->start_work_mutex);
759
760 for (i = 0; i < sched->nr_tasks; i++) {
761 task = sched->tasks[i];
762 sem_init(&task->sleep_sem, 0, 0);
763 task->curr_event = 0;
764 }
765}
766
767static void run_one_test(struct perf_sched *sched)
768 EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
769 EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
770{
771 u64 T0, T1, delta, avg_delta, fluct;
772
773 T0 = get_nsecs();
774 wait_for_tasks(sched);
775 T1 = get_nsecs();
776
777 delta = T1 - T0;
778 sched->sum_runtime += delta;
779 sched->nr_runs++;
780
781 avg_delta = sched->sum_runtime / sched->nr_runs;
782 if (delta < avg_delta)
783 fluct = avg_delta - delta;
784 else
785 fluct = delta - avg_delta;
786 sched->sum_fluct += fluct;
787 if (!sched->run_avg)
788 sched->run_avg = delta;
789 sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
790
791 printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
792
793 printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
794
795 printf("cpu: %0.2f / %0.2f",
796 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
797
798#if 0
799 /*
800 * rusage statistics done by the parent, these are less
801 * accurate than the sched->sum_exec_runtime based statistics:
802 */
803 printf(" [%0.2f / %0.2f]",
804 (double)sched->parent_cpu_usage / NSEC_PER_MSEC,
805 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
806#endif
807
808 printf("\n");
809
810 if (sched->nr_sleep_corrections)
811 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
812 sched->nr_sleep_corrections = 0;
813}
814
815static void test_calibrations(struct perf_sched *sched)
816{
817 u64 T0, T1;
818
819 T0 = get_nsecs();
820 burn_nsecs(sched, NSEC_PER_MSEC);
821 T1 = get_nsecs();
822
823 printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
824
825 T0 = get_nsecs();
826 sleep_nsecs(NSEC_PER_MSEC);
827 T1 = get_nsecs();
828
829 printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
830}
831
832static int
833replay_wakeup_event(struct perf_sched *sched,
834 struct evsel *evsel, struct perf_sample *sample,
835 struct machine *machine __maybe_unused)
836{
837 const char *comm = evsel__strval(evsel, sample, "comm");
838 const u32 pid = evsel__intval(evsel, sample, "pid");
839 struct task_desc *waker, *wakee;
840
841 if (verbose > 0) {
842 printf("sched_wakeup event %p\n", evsel);
843
844 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
845 }
846
847 waker = register_pid(sched, sample->tid, "<unknown>");
848 wakee = register_pid(sched, pid, comm);
849
850 add_sched_event_wakeup(sched, waker, sample->time, wakee);
851 return 0;
852}
853
854static int replay_switch_event(struct perf_sched *sched,
855 struct evsel *evsel,
856 struct perf_sample *sample,
857 struct machine *machine __maybe_unused)
858{
859 const char *prev_comm = evsel__strval(evsel, sample, "prev_comm"),
860 *next_comm = evsel__strval(evsel, sample, "next_comm");
861 const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
862 next_pid = evsel__intval(evsel, sample, "next_pid");
863 const u64 prev_state = evsel__intval(evsel, sample, "prev_state");
864 struct task_desc *prev, __maybe_unused *next;
865 u64 timestamp0, timestamp = sample->time;
866 int cpu = sample->cpu;
867 s64 delta;
868
869 if (verbose > 0)
870 printf("sched_switch event %p\n", evsel);
871
872 if (cpu >= MAX_CPUS || cpu < 0)
873 return 0;
874
875 timestamp0 = sched->cpu_last_switched[cpu];
876 if (timestamp0)
877 delta = timestamp - timestamp0;
878 else
879 delta = 0;
880
881 if (delta < 0) {
882 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
883 return -1;
884 }
885
886 pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
887 prev_comm, prev_pid, next_comm, next_pid, delta);
888
889 prev = register_pid(sched, prev_pid, prev_comm);
890 next = register_pid(sched, next_pid, next_comm);
891
892 sched->cpu_last_switched[cpu] = timestamp;
893
894 add_sched_event_run(sched, prev, timestamp, delta);
895 add_sched_event_sleep(sched, prev, timestamp, prev_state);
896
897 return 0;
898}
899
900static int replay_fork_event(struct perf_sched *sched,
901 union perf_event *event,
902 struct machine *machine)
903{
904 struct thread *child, *parent;
905
906 child = machine__findnew_thread(machine, event->fork.pid,
907 event->fork.tid);
908 parent = machine__findnew_thread(machine, event->fork.ppid,
909 event->fork.ptid);
910
911 if (child == NULL || parent == NULL) {
912 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
913 child, parent);
914 goto out_put;
915 }
916
917 if (verbose > 0) {
918 printf("fork event\n");
919 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
920 printf("... child: %s/%d\n", thread__comm_str(child), child->tid);
921 }
922
923 register_pid(sched, parent->tid, thread__comm_str(parent));
924 register_pid(sched, child->tid, thread__comm_str(child));
925out_put:
926 thread__put(child);
927 thread__put(parent);
928 return 0;
929}
930
931struct sort_dimension {
932 const char *name;
933 sort_fn_t cmp;
934 struct list_head list;
935};
936
937/*
938 * handle runtime stats saved per thread
939 */
940static struct thread_runtime *thread__init_runtime(struct thread *thread)
941{
942 struct thread_runtime *r;
943
944 r = zalloc(sizeof(struct thread_runtime));
945 if (!r)
946 return NULL;
947
948 init_stats(&r->run_stats);
949 thread__set_priv(thread, r);
950
951 return r;
952}
953
954static struct thread_runtime *thread__get_runtime(struct thread *thread)
955{
956 struct thread_runtime *tr;
957
958 tr = thread__priv(thread);
959 if (tr == NULL) {
960 tr = thread__init_runtime(thread);
961 if (tr == NULL)
962 pr_debug("Failed to malloc memory for runtime data.\n");
963 }
964
965 return tr;
966}
967
968static int
969thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
970{
971 struct sort_dimension *sort;
972 int ret = 0;
973
974 BUG_ON(list_empty(list));
975
976 list_for_each_entry(sort, list, list) {
977 ret = sort->cmp(l, r);
978 if (ret)
979 return ret;
980 }
981
982 return ret;
983}
984
985static struct work_atoms *
986thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
987 struct list_head *sort_list)
988{
989 struct rb_node *node = root->rb_root.rb_node;
990 struct work_atoms key = { .thread = thread };
991
992 while (node) {
993 struct work_atoms *atoms;
994 int cmp;
995
996 atoms = container_of(node, struct work_atoms, node);
997
998 cmp = thread_lat_cmp(sort_list, &key, atoms);
999 if (cmp > 0)
1000 node = node->rb_left;
1001 else if (cmp < 0)
1002 node = node->rb_right;
1003 else {
1004 BUG_ON(thread != atoms->thread);
1005 return atoms;
1006 }
1007 }
1008 return NULL;
1009}
1010
1011static void
1012__thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
1013 struct list_head *sort_list)
1014{
1015 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
1016 bool leftmost = true;
1017
1018 while (*new) {
1019 struct work_atoms *this;
1020 int cmp;
1021
1022 this = container_of(*new, struct work_atoms, node);
1023 parent = *new;
1024
1025 cmp = thread_lat_cmp(sort_list, data, this);
1026
1027 if (cmp > 0)
1028 new = &((*new)->rb_left);
1029 else {
1030 new = &((*new)->rb_right);
1031 leftmost = false;
1032 }
1033 }
1034
1035 rb_link_node(&data->node, parent, new);
1036 rb_insert_color_cached(&data->node, root, leftmost);
1037}
1038
1039static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1040{
1041 struct work_atoms *atoms = zalloc(sizeof(*atoms));
1042 if (!atoms) {
1043 pr_err("No memory at %s\n", __func__);
1044 return -1;
1045 }
1046
1047 atoms->thread = thread__get(thread);
1048 INIT_LIST_HEAD(&atoms->work_list);
1049 __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1050 return 0;
1051}
1052
1053static char sched_out_state(u64 prev_state)
1054{
1055 const char *str = TASK_STATE_TO_CHAR_STR;
1056
1057 return str[prev_state];
1058}
1059
1060static int
1061add_sched_out_event(struct work_atoms *atoms,
1062 char run_state,
1063 u64 timestamp)
1064{
1065 struct work_atom *atom = zalloc(sizeof(*atom));
1066 if (!atom) {
1067 pr_err("Non memory at %s", __func__);
1068 return -1;
1069 }
1070
1071 atom->sched_out_time = timestamp;
1072
1073 if (run_state == 'R') {
1074 atom->state = THREAD_WAIT_CPU;
1075 atom->wake_up_time = atom->sched_out_time;
1076 }
1077
1078 list_add_tail(&atom->list, &atoms->work_list);
1079 return 0;
1080}
1081
1082static void
1083add_runtime_event(struct work_atoms *atoms, u64 delta,
1084 u64 timestamp __maybe_unused)
1085{
1086 struct work_atom *atom;
1087
1088 BUG_ON(list_empty(&atoms->work_list));
1089
1090 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1091
1092 atom->runtime += delta;
1093 atoms->total_runtime += delta;
1094}
1095
1096static void
1097add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1098{
1099 struct work_atom *atom;
1100 u64 delta;
1101
1102 if (list_empty(&atoms->work_list))
1103 return;
1104
1105 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1106
1107 if (atom->state != THREAD_WAIT_CPU)
1108 return;
1109
1110 if (timestamp < atom->wake_up_time) {
1111 atom->state = THREAD_IGNORE;
1112 return;
1113 }
1114
1115 atom->state = THREAD_SCHED_IN;
1116 atom->sched_in_time = timestamp;
1117
1118 delta = atom->sched_in_time - atom->wake_up_time;
1119 atoms->total_lat += delta;
1120 if (delta > atoms->max_lat) {
1121 atoms->max_lat = delta;
1122 atoms->max_lat_start = atom->wake_up_time;
1123 atoms->max_lat_end = timestamp;
1124 }
1125 atoms->nb_atoms++;
1126}
1127
1128static int latency_switch_event(struct perf_sched *sched,
1129 struct evsel *evsel,
1130 struct perf_sample *sample,
1131 struct machine *machine)
1132{
1133 const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1134 next_pid = evsel__intval(evsel, sample, "next_pid");
1135 const u64 prev_state = evsel__intval(evsel, sample, "prev_state");
1136 struct work_atoms *out_events, *in_events;
1137 struct thread *sched_out, *sched_in;
1138 u64 timestamp0, timestamp = sample->time;
1139 int cpu = sample->cpu, err = -1;
1140 s64 delta;
1141
1142 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1143
1144 timestamp0 = sched->cpu_last_switched[cpu];
1145 sched->cpu_last_switched[cpu] = timestamp;
1146 if (timestamp0)
1147 delta = timestamp - timestamp0;
1148 else
1149 delta = 0;
1150
1151 if (delta < 0) {
1152 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1153 return -1;
1154 }
1155
1156 sched_out = machine__findnew_thread(machine, -1, prev_pid);
1157 sched_in = machine__findnew_thread(machine, -1, next_pid);
1158 if (sched_out == NULL || sched_in == NULL)
1159 goto out_put;
1160
1161 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1162 if (!out_events) {
1163 if (thread_atoms_insert(sched, sched_out))
1164 goto out_put;
1165 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1166 if (!out_events) {
1167 pr_err("out-event: Internal tree error");
1168 goto out_put;
1169 }
1170 }
1171 if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1172 return -1;
1173
1174 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1175 if (!in_events) {
1176 if (thread_atoms_insert(sched, sched_in))
1177 goto out_put;
1178 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1179 if (!in_events) {
1180 pr_err("in-event: Internal tree error");
1181 goto out_put;
1182 }
1183 /*
1184 * Take came in we have not heard about yet,
1185 * add in an initial atom in runnable state:
1186 */
1187 if (add_sched_out_event(in_events, 'R', timestamp))
1188 goto out_put;
1189 }
1190 add_sched_in_event(in_events, timestamp);
1191 err = 0;
1192out_put:
1193 thread__put(sched_out);
1194 thread__put(sched_in);
1195 return err;
1196}
1197
1198static int latency_runtime_event(struct perf_sched *sched,
1199 struct evsel *evsel,
1200 struct perf_sample *sample,
1201 struct machine *machine)
1202{
1203 const u32 pid = evsel__intval(evsel, sample, "pid");
1204 const u64 runtime = evsel__intval(evsel, sample, "runtime");
1205 struct thread *thread = machine__findnew_thread(machine, -1, pid);
1206 struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1207 u64 timestamp = sample->time;
1208 int cpu = sample->cpu, err = -1;
1209
1210 if (thread == NULL)
1211 return -1;
1212
1213 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1214 if (!atoms) {
1215 if (thread_atoms_insert(sched, thread))
1216 goto out_put;
1217 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1218 if (!atoms) {
1219 pr_err("in-event: Internal tree error");
1220 goto out_put;
1221 }
1222 if (add_sched_out_event(atoms, 'R', timestamp))
1223 goto out_put;
1224 }
1225
1226 add_runtime_event(atoms, runtime, timestamp);
1227 err = 0;
1228out_put:
1229 thread__put(thread);
1230 return err;
1231}
1232
1233static int latency_wakeup_event(struct perf_sched *sched,
1234 struct evsel *evsel,
1235 struct perf_sample *sample,
1236 struct machine *machine)
1237{
1238 const u32 pid = evsel__intval(evsel, sample, "pid");
1239 struct work_atoms *atoms;
1240 struct work_atom *atom;
1241 struct thread *wakee;
1242 u64 timestamp = sample->time;
1243 int err = -1;
1244
1245 wakee = machine__findnew_thread(machine, -1, pid);
1246 if (wakee == NULL)
1247 return -1;
1248 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1249 if (!atoms) {
1250 if (thread_atoms_insert(sched, wakee))
1251 goto out_put;
1252 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1253 if (!atoms) {
1254 pr_err("wakeup-event: Internal tree error");
1255 goto out_put;
1256 }
1257 if (add_sched_out_event(atoms, 'S', timestamp))
1258 goto out_put;
1259 }
1260
1261 BUG_ON(list_empty(&atoms->work_list));
1262
1263 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1264
1265 /*
1266 * As we do not guarantee the wakeup event happens when
1267 * task is out of run queue, also may happen when task is
1268 * on run queue and wakeup only change ->state to TASK_RUNNING,
1269 * then we should not set the ->wake_up_time when wake up a
1270 * task which is on run queue.
1271 *
1272 * You WILL be missing events if you've recorded only
1273 * one CPU, or are only looking at only one, so don't
1274 * skip in this case.
1275 */
1276 if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1277 goto out_ok;
1278
1279 sched->nr_timestamps++;
1280 if (atom->sched_out_time > timestamp) {
1281 sched->nr_unordered_timestamps++;
1282 goto out_ok;
1283 }
1284
1285 atom->state = THREAD_WAIT_CPU;
1286 atom->wake_up_time = timestamp;
1287out_ok:
1288 err = 0;
1289out_put:
1290 thread__put(wakee);
1291 return err;
1292}
1293
1294static int latency_migrate_task_event(struct perf_sched *sched,
1295 struct evsel *evsel,
1296 struct perf_sample *sample,
1297 struct machine *machine)
1298{
1299 const u32 pid = evsel__intval(evsel, sample, "pid");
1300 u64 timestamp = sample->time;
1301 struct work_atoms *atoms;
1302 struct work_atom *atom;
1303 struct thread *migrant;
1304 int err = -1;
1305
1306 /*
1307 * Only need to worry about migration when profiling one CPU.
1308 */
1309 if (sched->profile_cpu == -1)
1310 return 0;
1311
1312 migrant = machine__findnew_thread(machine, -1, pid);
1313 if (migrant == NULL)
1314 return -1;
1315 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1316 if (!atoms) {
1317 if (thread_atoms_insert(sched, migrant))
1318 goto out_put;
1319 register_pid(sched, migrant->tid, thread__comm_str(migrant));
1320 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1321 if (!atoms) {
1322 pr_err("migration-event: Internal tree error");
1323 goto out_put;
1324 }
1325 if (add_sched_out_event(atoms, 'R', timestamp))
1326 goto out_put;
1327 }
1328
1329 BUG_ON(list_empty(&atoms->work_list));
1330
1331 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1332 atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1333
1334 sched->nr_timestamps++;
1335
1336 if (atom->sched_out_time > timestamp)
1337 sched->nr_unordered_timestamps++;
1338 err = 0;
1339out_put:
1340 thread__put(migrant);
1341 return err;
1342}
1343
1344static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1345{
1346 int i;
1347 int ret;
1348 u64 avg;
1349 char max_lat_start[32], max_lat_end[32];
1350
1351 if (!work_list->nb_atoms)
1352 return;
1353 /*
1354 * Ignore idle threads:
1355 */
1356 if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1357 return;
1358
1359 sched->all_runtime += work_list->total_runtime;
1360 sched->all_count += work_list->nb_atoms;
1361
1362 if (work_list->num_merged > 1)
1363 ret = printf(" %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1364 else
1365 ret = printf(" %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1366
1367 for (i = 0; i < 24 - ret; i++)
1368 printf(" ");
1369
1370 avg = work_list->total_lat / work_list->nb_atoms;
1371 timestamp__scnprintf_usec(work_list->max_lat_start, max_lat_start, sizeof(max_lat_start));
1372 timestamp__scnprintf_usec(work_list->max_lat_end, max_lat_end, sizeof(max_lat_end));
1373
1374 printf("|%11.3f ms |%9" PRIu64 " | avg:%8.3f ms | max:%8.3f ms | max start: %12s s | max end: %12s s\n",
1375 (double)work_list->total_runtime / NSEC_PER_MSEC,
1376 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1377 (double)work_list->max_lat / NSEC_PER_MSEC,
1378 max_lat_start, max_lat_end);
1379}
1380
1381static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1382{
1383 if (l->thread == r->thread)
1384 return 0;
1385 if (l->thread->tid < r->thread->tid)
1386 return -1;
1387 if (l->thread->tid > r->thread->tid)
1388 return 1;
1389 return (int)(l->thread - r->thread);
1390}
1391
1392static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1393{
1394 u64 avgl, avgr;
1395
1396 if (!l->nb_atoms)
1397 return -1;
1398
1399 if (!r->nb_atoms)
1400 return 1;
1401
1402 avgl = l->total_lat / l->nb_atoms;
1403 avgr = r->total_lat / r->nb_atoms;
1404
1405 if (avgl < avgr)
1406 return -1;
1407 if (avgl > avgr)
1408 return 1;
1409
1410 return 0;
1411}
1412
1413static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1414{
1415 if (l->max_lat < r->max_lat)
1416 return -1;
1417 if (l->max_lat > r->max_lat)
1418 return 1;
1419
1420 return 0;
1421}
1422
1423static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1424{
1425 if (l->nb_atoms < r->nb_atoms)
1426 return -1;
1427 if (l->nb_atoms > r->nb_atoms)
1428 return 1;
1429
1430 return 0;
1431}
1432
1433static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1434{
1435 if (l->total_runtime < r->total_runtime)
1436 return -1;
1437 if (l->total_runtime > r->total_runtime)
1438 return 1;
1439
1440 return 0;
1441}
1442
1443static int sort_dimension__add(const char *tok, struct list_head *list)
1444{
1445 size_t i;
1446 static struct sort_dimension avg_sort_dimension = {
1447 .name = "avg",
1448 .cmp = avg_cmp,
1449 };
1450 static struct sort_dimension max_sort_dimension = {
1451 .name = "max",
1452 .cmp = max_cmp,
1453 };
1454 static struct sort_dimension pid_sort_dimension = {
1455 .name = "pid",
1456 .cmp = pid_cmp,
1457 };
1458 static struct sort_dimension runtime_sort_dimension = {
1459 .name = "runtime",
1460 .cmp = runtime_cmp,
1461 };
1462 static struct sort_dimension switch_sort_dimension = {
1463 .name = "switch",
1464 .cmp = switch_cmp,
1465 };
1466 struct sort_dimension *available_sorts[] = {
1467 &pid_sort_dimension,
1468 &avg_sort_dimension,
1469 &max_sort_dimension,
1470 &switch_sort_dimension,
1471 &runtime_sort_dimension,
1472 };
1473
1474 for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1475 if (!strcmp(available_sorts[i]->name, tok)) {
1476 list_add_tail(&available_sorts[i]->list, list);
1477
1478 return 0;
1479 }
1480 }
1481
1482 return -1;
1483}
1484
1485static void perf_sched__sort_lat(struct perf_sched *sched)
1486{
1487 struct rb_node *node;
1488 struct rb_root_cached *root = &sched->atom_root;
1489again:
1490 for (;;) {
1491 struct work_atoms *data;
1492 node = rb_first_cached(root);
1493 if (!node)
1494 break;
1495
1496 rb_erase_cached(node, root);
1497 data = rb_entry(node, struct work_atoms, node);
1498 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1499 }
1500 if (root == &sched->atom_root) {
1501 root = &sched->merged_atom_root;
1502 goto again;
1503 }
1504}
1505
1506static int process_sched_wakeup_event(struct perf_tool *tool,
1507 struct evsel *evsel,
1508 struct perf_sample *sample,
1509 struct machine *machine)
1510{
1511 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1512
1513 if (sched->tp_handler->wakeup_event)
1514 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1515
1516 return 0;
1517}
1518
1519union map_priv {
1520 void *ptr;
1521 bool color;
1522};
1523
1524static bool thread__has_color(struct thread *thread)
1525{
1526 union map_priv priv = {
1527 .ptr = thread__priv(thread),
1528 };
1529
1530 return priv.color;
1531}
1532
1533static struct thread*
1534map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1535{
1536 struct thread *thread = machine__findnew_thread(machine, pid, tid);
1537 union map_priv priv = {
1538 .color = false,
1539 };
1540
1541 if (!sched->map.color_pids || !thread || thread__priv(thread))
1542 return thread;
1543
1544 if (thread_map__has(sched->map.color_pids, tid))
1545 priv.color = true;
1546
1547 thread__set_priv(thread, priv.ptr);
1548 return thread;
1549}
1550
1551static int map_switch_event(struct perf_sched *sched, struct evsel *evsel,
1552 struct perf_sample *sample, struct machine *machine)
1553{
1554 const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
1555 struct thread *sched_in;
1556 struct thread_runtime *tr;
1557 int new_shortname;
1558 u64 timestamp0, timestamp = sample->time;
1559 s64 delta;
1560 int i;
1561 struct perf_cpu this_cpu = {
1562 .cpu = sample->cpu,
1563 };
1564 int cpus_nr;
1565 bool new_cpu = false;
1566 const char *color = PERF_COLOR_NORMAL;
1567 char stimestamp[32];
1568
1569 BUG_ON(this_cpu.cpu >= MAX_CPUS || this_cpu.cpu < 0);
1570
1571 if (this_cpu.cpu > sched->max_cpu.cpu)
1572 sched->max_cpu = this_cpu;
1573
1574 if (sched->map.comp) {
1575 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1576 if (!__test_and_set_bit(this_cpu.cpu, sched->map.comp_cpus_mask)) {
1577 sched->map.comp_cpus[cpus_nr++] = this_cpu;
1578 new_cpu = true;
1579 }
1580 } else
1581 cpus_nr = sched->max_cpu.cpu;
1582
1583 timestamp0 = sched->cpu_last_switched[this_cpu.cpu];
1584 sched->cpu_last_switched[this_cpu.cpu] = timestamp;
1585 if (timestamp0)
1586 delta = timestamp - timestamp0;
1587 else
1588 delta = 0;
1589
1590 if (delta < 0) {
1591 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1592 return -1;
1593 }
1594
1595 sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1596 if (sched_in == NULL)
1597 return -1;
1598
1599 tr = thread__get_runtime(sched_in);
1600 if (tr == NULL) {
1601 thread__put(sched_in);
1602 return -1;
1603 }
1604
1605 sched->curr_thread[this_cpu.cpu] = thread__get(sched_in);
1606
1607 printf(" ");
1608
1609 new_shortname = 0;
1610 if (!tr->shortname[0]) {
1611 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1612 /*
1613 * Don't allocate a letter-number for swapper:0
1614 * as a shortname. Instead, we use '.' for it.
1615 */
1616 tr->shortname[0] = '.';
1617 tr->shortname[1] = ' ';
1618 } else {
1619 tr->shortname[0] = sched->next_shortname1;
1620 tr->shortname[1] = sched->next_shortname2;
1621
1622 if (sched->next_shortname1 < 'Z') {
1623 sched->next_shortname1++;
1624 } else {
1625 sched->next_shortname1 = 'A';
1626 if (sched->next_shortname2 < '9')
1627 sched->next_shortname2++;
1628 else
1629 sched->next_shortname2 = '0';
1630 }
1631 }
1632 new_shortname = 1;
1633 }
1634
1635 for (i = 0; i < cpus_nr; i++) {
1636 struct perf_cpu cpu = {
1637 .cpu = sched->map.comp ? sched->map.comp_cpus[i].cpu : i,
1638 };
1639 struct thread *curr_thread = sched->curr_thread[cpu.cpu];
1640 struct thread_runtime *curr_tr;
1641 const char *pid_color = color;
1642 const char *cpu_color = color;
1643
1644 if (curr_thread && thread__has_color(curr_thread))
1645 pid_color = COLOR_PIDS;
1646
1647 if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, cpu))
1648 continue;
1649
1650 if (sched->map.color_cpus && perf_cpu_map__has(sched->map.color_cpus, cpu))
1651 cpu_color = COLOR_CPUS;
1652
1653 if (cpu.cpu != this_cpu.cpu)
1654 color_fprintf(stdout, color, " ");
1655 else
1656 color_fprintf(stdout, cpu_color, "*");
1657
1658 if (sched->curr_thread[cpu.cpu]) {
1659 curr_tr = thread__get_runtime(sched->curr_thread[cpu.cpu]);
1660 if (curr_tr == NULL) {
1661 thread__put(sched_in);
1662 return -1;
1663 }
1664 color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1665 } else
1666 color_fprintf(stdout, color, " ");
1667 }
1668
1669 if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, this_cpu))
1670 goto out;
1671
1672 timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1673 color_fprintf(stdout, color, " %12s secs ", stimestamp);
1674 if (new_shortname || tr->comm_changed || (verbose > 0 && sched_in->tid)) {
1675 const char *pid_color = color;
1676
1677 if (thread__has_color(sched_in))
1678 pid_color = COLOR_PIDS;
1679
1680 color_fprintf(stdout, pid_color, "%s => %s:%d",
1681 tr->shortname, thread__comm_str(sched_in), sched_in->tid);
1682 tr->comm_changed = false;
1683 }
1684
1685 if (sched->map.comp && new_cpu)
1686 color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1687
1688out:
1689 color_fprintf(stdout, color, "\n");
1690
1691 thread__put(sched_in);
1692
1693 return 0;
1694}
1695
1696static int process_sched_switch_event(struct perf_tool *tool,
1697 struct evsel *evsel,
1698 struct perf_sample *sample,
1699 struct machine *machine)
1700{
1701 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1702 int this_cpu = sample->cpu, err = 0;
1703 u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1704 next_pid = evsel__intval(evsel, sample, "next_pid");
1705
1706 if (sched->curr_pid[this_cpu] != (u32)-1) {
1707 /*
1708 * Are we trying to switch away a PID that is
1709 * not current?
1710 */
1711 if (sched->curr_pid[this_cpu] != prev_pid)
1712 sched->nr_context_switch_bugs++;
1713 }
1714
1715 if (sched->tp_handler->switch_event)
1716 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1717
1718 sched->curr_pid[this_cpu] = next_pid;
1719 return err;
1720}
1721
1722static int process_sched_runtime_event(struct perf_tool *tool,
1723 struct evsel *evsel,
1724 struct perf_sample *sample,
1725 struct machine *machine)
1726{
1727 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1728
1729 if (sched->tp_handler->runtime_event)
1730 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1731
1732 return 0;
1733}
1734
1735static int perf_sched__process_fork_event(struct perf_tool *tool,
1736 union perf_event *event,
1737 struct perf_sample *sample,
1738 struct machine *machine)
1739{
1740 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1741
1742 /* run the fork event through the perf machinery */
1743 perf_event__process_fork(tool, event, sample, machine);
1744
1745 /* and then run additional processing needed for this command */
1746 if (sched->tp_handler->fork_event)
1747 return sched->tp_handler->fork_event(sched, event, machine);
1748
1749 return 0;
1750}
1751
1752static int process_sched_migrate_task_event(struct perf_tool *tool,
1753 struct evsel *evsel,
1754 struct perf_sample *sample,
1755 struct machine *machine)
1756{
1757 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1758
1759 if (sched->tp_handler->migrate_task_event)
1760 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1761
1762 return 0;
1763}
1764
1765typedef int (*tracepoint_handler)(struct perf_tool *tool,
1766 struct evsel *evsel,
1767 struct perf_sample *sample,
1768 struct machine *machine);
1769
1770static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1771 union perf_event *event __maybe_unused,
1772 struct perf_sample *sample,
1773 struct evsel *evsel,
1774 struct machine *machine)
1775{
1776 int err = 0;
1777
1778 if (evsel->handler != NULL) {
1779 tracepoint_handler f = evsel->handler;
1780 err = f(tool, evsel, sample, machine);
1781 }
1782
1783 return err;
1784}
1785
1786static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1787 union perf_event *event,
1788 struct perf_sample *sample,
1789 struct machine *machine)
1790{
1791 struct thread *thread;
1792 struct thread_runtime *tr;
1793 int err;
1794
1795 err = perf_event__process_comm(tool, event, sample, machine);
1796 if (err)
1797 return err;
1798
1799 thread = machine__find_thread(machine, sample->pid, sample->tid);
1800 if (!thread) {
1801 pr_err("Internal error: can't find thread\n");
1802 return -1;
1803 }
1804
1805 tr = thread__get_runtime(thread);
1806 if (tr == NULL) {
1807 thread__put(thread);
1808 return -1;
1809 }
1810
1811 tr->comm_changed = true;
1812 thread__put(thread);
1813
1814 return 0;
1815}
1816
1817static int perf_sched__read_events(struct perf_sched *sched)
1818{
1819 const struct evsel_str_handler handlers[] = {
1820 { "sched:sched_switch", process_sched_switch_event, },
1821 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1822 { "sched:sched_wakeup", process_sched_wakeup_event, },
1823 { "sched:sched_wakeup_new", process_sched_wakeup_event, },
1824 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1825 };
1826 struct perf_session *session;
1827 struct perf_data data = {
1828 .path = input_name,
1829 .mode = PERF_DATA_MODE_READ,
1830 .force = sched->force,
1831 };
1832 int rc = -1;
1833
1834 session = perf_session__new(&data, &sched->tool);
1835 if (IS_ERR(session)) {
1836 pr_debug("Error creating perf session");
1837 return PTR_ERR(session);
1838 }
1839
1840 symbol__init(&session->header.env);
1841
1842 if (perf_session__set_tracepoints_handlers(session, handlers))
1843 goto out_delete;
1844
1845 if (perf_session__has_traces(session, "record -R")) {
1846 int err = perf_session__process_events(session);
1847 if (err) {
1848 pr_err("Failed to process events, error %d", err);
1849 goto out_delete;
1850 }
1851
1852 sched->nr_events = session->evlist->stats.nr_events[0];
1853 sched->nr_lost_events = session->evlist->stats.total_lost;
1854 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1855 }
1856
1857 rc = 0;
1858out_delete:
1859 perf_session__delete(session);
1860 return rc;
1861}
1862
1863/*
1864 * scheduling times are printed as msec.usec
1865 */
1866static inline void print_sched_time(unsigned long long nsecs, int width)
1867{
1868 unsigned long msecs;
1869 unsigned long usecs;
1870
1871 msecs = nsecs / NSEC_PER_MSEC;
1872 nsecs -= msecs * NSEC_PER_MSEC;
1873 usecs = nsecs / NSEC_PER_USEC;
1874 printf("%*lu.%03lu ", width, msecs, usecs);
1875}
1876
1877/*
1878 * returns runtime data for event, allocating memory for it the
1879 * first time it is used.
1880 */
1881static struct evsel_runtime *evsel__get_runtime(struct evsel *evsel)
1882{
1883 struct evsel_runtime *r = evsel->priv;
1884
1885 if (r == NULL) {
1886 r = zalloc(sizeof(struct evsel_runtime));
1887 evsel->priv = r;
1888 }
1889
1890 return r;
1891}
1892
1893/*
1894 * save last time event was seen per cpu
1895 */
1896static void evsel__save_time(struct evsel *evsel, u64 timestamp, u32 cpu)
1897{
1898 struct evsel_runtime *r = evsel__get_runtime(evsel);
1899
1900 if (r == NULL)
1901 return;
1902
1903 if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1904 int i, n = __roundup_pow_of_two(cpu+1);
1905 void *p = r->last_time;
1906
1907 p = realloc(r->last_time, n * sizeof(u64));
1908 if (!p)
1909 return;
1910
1911 r->last_time = p;
1912 for (i = r->ncpu; i < n; ++i)
1913 r->last_time[i] = (u64) 0;
1914
1915 r->ncpu = n;
1916 }
1917
1918 r->last_time[cpu] = timestamp;
1919}
1920
1921/* returns last time this event was seen on the given cpu */
1922static u64 evsel__get_time(struct evsel *evsel, u32 cpu)
1923{
1924 struct evsel_runtime *r = evsel__get_runtime(evsel);
1925
1926 if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1927 return 0;
1928
1929 return r->last_time[cpu];
1930}
1931
1932static int comm_width = 30;
1933
1934static char *timehist_get_commstr(struct thread *thread)
1935{
1936 static char str[32];
1937 const char *comm = thread__comm_str(thread);
1938 pid_t tid = thread->tid;
1939 pid_t pid = thread->pid_;
1940 int n;
1941
1942 if (pid == 0)
1943 n = scnprintf(str, sizeof(str), "%s", comm);
1944
1945 else if (tid != pid)
1946 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1947
1948 else
1949 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1950
1951 if (n > comm_width)
1952 comm_width = n;
1953
1954 return str;
1955}
1956
1957static void timehist_header(struct perf_sched *sched)
1958{
1959 u32 ncpus = sched->max_cpu.cpu + 1;
1960 u32 i, j;
1961
1962 printf("%15s %6s ", "time", "cpu");
1963
1964 if (sched->show_cpu_visual) {
1965 printf(" ");
1966 for (i = 0, j = 0; i < ncpus; ++i) {
1967 printf("%x", j++);
1968 if (j > 15)
1969 j = 0;
1970 }
1971 printf(" ");
1972 }
1973
1974 printf(" %-*s %9s %9s %9s", comm_width,
1975 "task name", "wait time", "sch delay", "run time");
1976
1977 if (sched->show_state)
1978 printf(" %s", "state");
1979
1980 printf("\n");
1981
1982 /*
1983 * units row
1984 */
1985 printf("%15s %-6s ", "", "");
1986
1987 if (sched->show_cpu_visual)
1988 printf(" %*s ", ncpus, "");
1989
1990 printf(" %-*s %9s %9s %9s", comm_width,
1991 "[tid/pid]", "(msec)", "(msec)", "(msec)");
1992
1993 if (sched->show_state)
1994 printf(" %5s", "");
1995
1996 printf("\n");
1997
1998 /*
1999 * separator
2000 */
2001 printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
2002
2003 if (sched->show_cpu_visual)
2004 printf(" %.*s ", ncpus, graph_dotted_line);
2005
2006 printf(" %.*s %.9s %.9s %.9s", comm_width,
2007 graph_dotted_line, graph_dotted_line, graph_dotted_line,
2008 graph_dotted_line);
2009
2010 if (sched->show_state)
2011 printf(" %.5s", graph_dotted_line);
2012
2013 printf("\n");
2014}
2015
2016static char task_state_char(struct thread *thread, int state)
2017{
2018 static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
2019 unsigned bit = state ? ffs(state) : 0;
2020
2021 /* 'I' for idle */
2022 if (thread->tid == 0)
2023 return 'I';
2024
2025 return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
2026}
2027
2028static void timehist_print_sample(struct perf_sched *sched,
2029 struct evsel *evsel,
2030 struct perf_sample *sample,
2031 struct addr_location *al,
2032 struct thread *thread,
2033 u64 t, int state)
2034{
2035 struct thread_runtime *tr = thread__priv(thread);
2036 const char *next_comm = evsel__strval(evsel, sample, "next_comm");
2037 const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
2038 u32 max_cpus = sched->max_cpu.cpu + 1;
2039 char tstr[64];
2040 char nstr[30];
2041 u64 wait_time;
2042
2043 if (cpu_list && !test_bit(sample->cpu, cpu_bitmap))
2044 return;
2045
2046 timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2047 printf("%15s [%04d] ", tstr, sample->cpu);
2048
2049 if (sched->show_cpu_visual) {
2050 u32 i;
2051 char c;
2052
2053 printf(" ");
2054 for (i = 0; i < max_cpus; ++i) {
2055 /* flag idle times with 'i'; others are sched events */
2056 if (i == sample->cpu)
2057 c = (thread->tid == 0) ? 'i' : 's';
2058 else
2059 c = ' ';
2060 printf("%c", c);
2061 }
2062 printf(" ");
2063 }
2064
2065 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2066
2067 wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2068 print_sched_time(wait_time, 6);
2069
2070 print_sched_time(tr->dt_delay, 6);
2071 print_sched_time(tr->dt_run, 6);
2072
2073 if (sched->show_state)
2074 printf(" %5c ", task_state_char(thread, state));
2075
2076 if (sched->show_next) {
2077 snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2078 printf(" %-*s", comm_width, nstr);
2079 }
2080
2081 if (sched->show_wakeups && !sched->show_next)
2082 printf(" %-*s", comm_width, "");
2083
2084 if (thread->tid == 0)
2085 goto out;
2086
2087 if (sched->show_callchain)
2088 printf(" ");
2089
2090 sample__fprintf_sym(sample, al, 0,
2091 EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2092 EVSEL__PRINT_CALLCHAIN_ARROW |
2093 EVSEL__PRINT_SKIP_IGNORED,
2094 &callchain_cursor, symbol_conf.bt_stop_list, stdout);
2095
2096out:
2097 printf("\n");
2098}
2099
2100/*
2101 * Explanation of delta-time stats:
2102 *
2103 * t = time of current schedule out event
2104 * tprev = time of previous sched out event
2105 * also time of schedule-in event for current task
2106 * last_time = time of last sched change event for current task
2107 * (i.e, time process was last scheduled out)
2108 * ready_to_run = time of wakeup for current task
2109 *
2110 * -----|------------|------------|------------|------
2111 * last ready tprev t
2112 * time to run
2113 *
2114 * |-------- dt_wait --------|
2115 * |- dt_delay -|-- dt_run --|
2116 *
2117 * dt_run = run time of current task
2118 * dt_wait = time between last schedule out event for task and tprev
2119 * represents time spent off the cpu
2120 * dt_delay = time between wakeup and schedule-in of task
2121 */
2122
2123static void timehist_update_runtime_stats(struct thread_runtime *r,
2124 u64 t, u64 tprev)
2125{
2126 r->dt_delay = 0;
2127 r->dt_sleep = 0;
2128 r->dt_iowait = 0;
2129 r->dt_preempt = 0;
2130 r->dt_run = 0;
2131
2132 if (tprev) {
2133 r->dt_run = t - tprev;
2134 if (r->ready_to_run) {
2135 if (r->ready_to_run > tprev)
2136 pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2137 else
2138 r->dt_delay = tprev - r->ready_to_run;
2139 }
2140
2141 if (r->last_time > tprev)
2142 pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2143 else if (r->last_time) {
2144 u64 dt_wait = tprev - r->last_time;
2145
2146 if (r->last_state == TASK_RUNNING)
2147 r->dt_preempt = dt_wait;
2148 else if (r->last_state == TASK_UNINTERRUPTIBLE)
2149 r->dt_iowait = dt_wait;
2150 else
2151 r->dt_sleep = dt_wait;
2152 }
2153 }
2154
2155 update_stats(&r->run_stats, r->dt_run);
2156
2157 r->total_run_time += r->dt_run;
2158 r->total_delay_time += r->dt_delay;
2159 r->total_sleep_time += r->dt_sleep;
2160 r->total_iowait_time += r->dt_iowait;
2161 r->total_preempt_time += r->dt_preempt;
2162}
2163
2164static bool is_idle_sample(struct perf_sample *sample,
2165 struct evsel *evsel)
2166{
2167 /* pid 0 == swapper == idle task */
2168 if (strcmp(evsel__name(evsel), "sched:sched_switch") == 0)
2169 return evsel__intval(evsel, sample, "prev_pid") == 0;
2170
2171 return sample->pid == 0;
2172}
2173
2174static void save_task_callchain(struct perf_sched *sched,
2175 struct perf_sample *sample,
2176 struct evsel *evsel,
2177 struct machine *machine)
2178{
2179 struct callchain_cursor *cursor = &callchain_cursor;
2180 struct thread *thread;
2181
2182 /* want main thread for process - has maps */
2183 thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2184 if (thread == NULL) {
2185 pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2186 return;
2187 }
2188
2189 if (!sched->show_callchain || sample->callchain == NULL)
2190 return;
2191
2192 if (thread__resolve_callchain(thread, cursor, evsel, sample,
2193 NULL, NULL, sched->max_stack + 2) != 0) {
2194 if (verbose > 0)
2195 pr_err("Failed to resolve callchain. Skipping\n");
2196
2197 return;
2198 }
2199
2200 callchain_cursor_commit(cursor);
2201
2202 while (true) {
2203 struct callchain_cursor_node *node;
2204 struct symbol *sym;
2205
2206 node = callchain_cursor_current(cursor);
2207 if (node == NULL)
2208 break;
2209
2210 sym = node->ms.sym;
2211 if (sym) {
2212 if (!strcmp(sym->name, "schedule") ||
2213 !strcmp(sym->name, "__schedule") ||
2214 !strcmp(sym->name, "preempt_schedule"))
2215 sym->ignore = 1;
2216 }
2217
2218 callchain_cursor_advance(cursor);
2219 }
2220}
2221
2222static int init_idle_thread(struct thread *thread)
2223{
2224 struct idle_thread_runtime *itr;
2225
2226 thread__set_comm(thread, idle_comm, 0);
2227
2228 itr = zalloc(sizeof(*itr));
2229 if (itr == NULL)
2230 return -ENOMEM;
2231
2232 init_stats(&itr->tr.run_stats);
2233 callchain_init(&itr->callchain);
2234 callchain_cursor_reset(&itr->cursor);
2235 thread__set_priv(thread, itr);
2236
2237 return 0;
2238}
2239
2240/*
2241 * Track idle stats per cpu by maintaining a local thread
2242 * struct for the idle task on each cpu.
2243 */
2244static int init_idle_threads(int ncpu)
2245{
2246 int i, ret;
2247
2248 idle_threads = zalloc(ncpu * sizeof(struct thread *));
2249 if (!idle_threads)
2250 return -ENOMEM;
2251
2252 idle_max_cpu = ncpu;
2253
2254 /* allocate the actual thread struct if needed */
2255 for (i = 0; i < ncpu; ++i) {
2256 idle_threads[i] = thread__new(0, 0);
2257 if (idle_threads[i] == NULL)
2258 return -ENOMEM;
2259
2260 ret = init_idle_thread(idle_threads[i]);
2261 if (ret < 0)
2262 return ret;
2263 }
2264
2265 return 0;
2266}
2267
2268static void free_idle_threads(void)
2269{
2270 int i;
2271
2272 if (idle_threads == NULL)
2273 return;
2274
2275 for (i = 0; i < idle_max_cpu; ++i) {
2276 if ((idle_threads[i]))
2277 thread__delete(idle_threads[i]);
2278 }
2279
2280 free(idle_threads);
2281}
2282
2283static struct thread *get_idle_thread(int cpu)
2284{
2285 /*
2286 * expand/allocate array of pointers to local thread
2287 * structs if needed
2288 */
2289 if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2290 int i, j = __roundup_pow_of_two(cpu+1);
2291 void *p;
2292
2293 p = realloc(idle_threads, j * sizeof(struct thread *));
2294 if (!p)
2295 return NULL;
2296
2297 idle_threads = (struct thread **) p;
2298 for (i = idle_max_cpu; i < j; ++i)
2299 idle_threads[i] = NULL;
2300
2301 idle_max_cpu = j;
2302 }
2303
2304 /* allocate a new thread struct if needed */
2305 if (idle_threads[cpu] == NULL) {
2306 idle_threads[cpu] = thread__new(0, 0);
2307
2308 if (idle_threads[cpu]) {
2309 if (init_idle_thread(idle_threads[cpu]) < 0)
2310 return NULL;
2311 }
2312 }
2313
2314 return idle_threads[cpu];
2315}
2316
2317static void save_idle_callchain(struct perf_sched *sched,
2318 struct idle_thread_runtime *itr,
2319 struct perf_sample *sample)
2320{
2321 if (!sched->show_callchain || sample->callchain == NULL)
2322 return;
2323
2324 callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2325}
2326
2327static struct thread *timehist_get_thread(struct perf_sched *sched,
2328 struct perf_sample *sample,
2329 struct machine *machine,
2330 struct evsel *evsel)
2331{
2332 struct thread *thread;
2333
2334 if (is_idle_sample(sample, evsel)) {
2335 thread = get_idle_thread(sample->cpu);
2336 if (thread == NULL)
2337 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2338
2339 } else {
2340 /* there were samples with tid 0 but non-zero pid */
2341 thread = machine__findnew_thread(machine, sample->pid,
2342 sample->tid ?: sample->pid);
2343 if (thread == NULL) {
2344 pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2345 sample->tid);
2346 }
2347
2348 save_task_callchain(sched, sample, evsel, machine);
2349 if (sched->idle_hist) {
2350 struct thread *idle;
2351 struct idle_thread_runtime *itr;
2352
2353 idle = get_idle_thread(sample->cpu);
2354 if (idle == NULL) {
2355 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2356 return NULL;
2357 }
2358
2359 itr = thread__priv(idle);
2360 if (itr == NULL)
2361 return NULL;
2362
2363 itr->last_thread = thread;
2364
2365 /* copy task callchain when entering to idle */
2366 if (evsel__intval(evsel, sample, "next_pid") == 0)
2367 save_idle_callchain(sched, itr, sample);
2368 }
2369 }
2370
2371 return thread;
2372}
2373
2374static bool timehist_skip_sample(struct perf_sched *sched,
2375 struct thread *thread,
2376 struct evsel *evsel,
2377 struct perf_sample *sample)
2378{
2379 bool rc = false;
2380
2381 if (thread__is_filtered(thread)) {
2382 rc = true;
2383 sched->skipped_samples++;
2384 }
2385
2386 if (sched->idle_hist) {
2387 if (strcmp(evsel__name(evsel), "sched:sched_switch"))
2388 rc = true;
2389 else if (evsel__intval(evsel, sample, "prev_pid") != 0 &&
2390 evsel__intval(evsel, sample, "next_pid") != 0)
2391 rc = true;
2392 }
2393
2394 return rc;
2395}
2396
2397static void timehist_print_wakeup_event(struct perf_sched *sched,
2398 struct evsel *evsel,
2399 struct perf_sample *sample,
2400 struct machine *machine,
2401 struct thread *awakened)
2402{
2403 struct thread *thread;
2404 char tstr[64];
2405
2406 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2407 if (thread == NULL)
2408 return;
2409
2410 /* show wakeup unless both awakee and awaker are filtered */
2411 if (timehist_skip_sample(sched, thread, evsel, sample) &&
2412 timehist_skip_sample(sched, awakened, evsel, sample)) {
2413 return;
2414 }
2415
2416 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2417 printf("%15s [%04d] ", tstr, sample->cpu);
2418 if (sched->show_cpu_visual)
2419 printf(" %*s ", sched->max_cpu.cpu + 1, "");
2420
2421 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2422
2423 /* dt spacer */
2424 printf(" %9s %9s %9s ", "", "", "");
2425
2426 printf("awakened: %s", timehist_get_commstr(awakened));
2427
2428 printf("\n");
2429}
2430
2431static int timehist_sched_wakeup_ignore(struct perf_tool *tool __maybe_unused,
2432 union perf_event *event __maybe_unused,
2433 struct evsel *evsel __maybe_unused,
2434 struct perf_sample *sample __maybe_unused,
2435 struct machine *machine __maybe_unused)
2436{
2437 return 0;
2438}
2439
2440static int timehist_sched_wakeup_event(struct perf_tool *tool,
2441 union perf_event *event __maybe_unused,
2442 struct evsel *evsel,
2443 struct perf_sample *sample,
2444 struct machine *machine)
2445{
2446 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2447 struct thread *thread;
2448 struct thread_runtime *tr = NULL;
2449 /* want pid of awakened task not pid in sample */
2450 const u32 pid = evsel__intval(evsel, sample, "pid");
2451
2452 thread = machine__findnew_thread(machine, 0, pid);
2453 if (thread == NULL)
2454 return -1;
2455
2456 tr = thread__get_runtime(thread);
2457 if (tr == NULL)
2458 return -1;
2459
2460 if (tr->ready_to_run == 0)
2461 tr->ready_to_run = sample->time;
2462
2463 /* show wakeups if requested */
2464 if (sched->show_wakeups &&
2465 !perf_time__skip_sample(&sched->ptime, sample->time))
2466 timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2467
2468 return 0;
2469}
2470
2471static void timehist_print_migration_event(struct perf_sched *sched,
2472 struct evsel *evsel,
2473 struct perf_sample *sample,
2474 struct machine *machine,
2475 struct thread *migrated)
2476{
2477 struct thread *thread;
2478 char tstr[64];
2479 u32 max_cpus;
2480 u32 ocpu, dcpu;
2481
2482 if (sched->summary_only)
2483 return;
2484
2485 max_cpus = sched->max_cpu.cpu + 1;
2486 ocpu = evsel__intval(evsel, sample, "orig_cpu");
2487 dcpu = evsel__intval(evsel, sample, "dest_cpu");
2488
2489 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2490 if (thread == NULL)
2491 return;
2492
2493 if (timehist_skip_sample(sched, thread, evsel, sample) &&
2494 timehist_skip_sample(sched, migrated, evsel, sample)) {
2495 return;
2496 }
2497
2498 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2499 printf("%15s [%04d] ", tstr, sample->cpu);
2500
2501 if (sched->show_cpu_visual) {
2502 u32 i;
2503 char c;
2504
2505 printf(" ");
2506 for (i = 0; i < max_cpus; ++i) {
2507 c = (i == sample->cpu) ? 'm' : ' ';
2508 printf("%c", c);
2509 }
2510 printf(" ");
2511 }
2512
2513 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2514
2515 /* dt spacer */
2516 printf(" %9s %9s %9s ", "", "", "");
2517
2518 printf("migrated: %s", timehist_get_commstr(migrated));
2519 printf(" cpu %d => %d", ocpu, dcpu);
2520
2521 printf("\n");
2522}
2523
2524static int timehist_migrate_task_event(struct perf_tool *tool,
2525 union perf_event *event __maybe_unused,
2526 struct evsel *evsel,
2527 struct perf_sample *sample,
2528 struct machine *machine)
2529{
2530 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2531 struct thread *thread;
2532 struct thread_runtime *tr = NULL;
2533 /* want pid of migrated task not pid in sample */
2534 const u32 pid = evsel__intval(evsel, sample, "pid");
2535
2536 thread = machine__findnew_thread(machine, 0, pid);
2537 if (thread == NULL)
2538 return -1;
2539
2540 tr = thread__get_runtime(thread);
2541 if (tr == NULL)
2542 return -1;
2543
2544 tr->migrations++;
2545
2546 /* show migrations if requested */
2547 timehist_print_migration_event(sched, evsel, sample, machine, thread);
2548
2549 return 0;
2550}
2551
2552static int timehist_sched_change_event(struct perf_tool *tool,
2553 union perf_event *event,
2554 struct evsel *evsel,
2555 struct perf_sample *sample,
2556 struct machine *machine)
2557{
2558 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2559 struct perf_time_interval *ptime = &sched->ptime;
2560 struct addr_location al;
2561 struct thread *thread;
2562 struct thread_runtime *tr = NULL;
2563 u64 tprev, t = sample->time;
2564 int rc = 0;
2565 int state = evsel__intval(evsel, sample, "prev_state");
2566
2567 if (machine__resolve(machine, &al, sample) < 0) {
2568 pr_err("problem processing %d event. skipping it\n",
2569 event->header.type);
2570 rc = -1;
2571 goto out;
2572 }
2573
2574 thread = timehist_get_thread(sched, sample, machine, evsel);
2575 if (thread == NULL) {
2576 rc = -1;
2577 goto out;
2578 }
2579
2580 if (timehist_skip_sample(sched, thread, evsel, sample))
2581 goto out;
2582
2583 tr = thread__get_runtime(thread);
2584 if (tr == NULL) {
2585 rc = -1;
2586 goto out;
2587 }
2588
2589 tprev = evsel__get_time(evsel, sample->cpu);
2590
2591 /*
2592 * If start time given:
2593 * - sample time is under window user cares about - skip sample
2594 * - tprev is under window user cares about - reset to start of window
2595 */
2596 if (ptime->start && ptime->start > t)
2597 goto out;
2598
2599 if (tprev && ptime->start > tprev)
2600 tprev = ptime->start;
2601
2602 /*
2603 * If end time given:
2604 * - previous sched event is out of window - we are done
2605 * - sample time is beyond window user cares about - reset it
2606 * to close out stats for time window interest
2607 */
2608 if (ptime->end) {
2609 if (tprev > ptime->end)
2610 goto out;
2611
2612 if (t > ptime->end)
2613 t = ptime->end;
2614 }
2615
2616 if (!sched->idle_hist || thread->tid == 0) {
2617 if (!cpu_list || test_bit(sample->cpu, cpu_bitmap))
2618 timehist_update_runtime_stats(tr, t, tprev);
2619
2620 if (sched->idle_hist) {
2621 struct idle_thread_runtime *itr = (void *)tr;
2622 struct thread_runtime *last_tr;
2623
2624 BUG_ON(thread->tid != 0);
2625
2626 if (itr->last_thread == NULL)
2627 goto out;
2628
2629 /* add current idle time as last thread's runtime */
2630 last_tr = thread__get_runtime(itr->last_thread);
2631 if (last_tr == NULL)
2632 goto out;
2633
2634 timehist_update_runtime_stats(last_tr, t, tprev);
2635 /*
2636 * remove delta time of last thread as it's not updated
2637 * and otherwise it will show an invalid value next
2638 * time. we only care total run time and run stat.
2639 */
2640 last_tr->dt_run = 0;
2641 last_tr->dt_delay = 0;
2642 last_tr->dt_sleep = 0;
2643 last_tr->dt_iowait = 0;
2644 last_tr->dt_preempt = 0;
2645
2646 if (itr->cursor.nr)
2647 callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2648
2649 itr->last_thread = NULL;
2650 }
2651 }
2652
2653 if (!sched->summary_only)
2654 timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2655
2656out:
2657 if (sched->hist_time.start == 0 && t >= ptime->start)
2658 sched->hist_time.start = t;
2659 if (ptime->end == 0 || t <= ptime->end)
2660 sched->hist_time.end = t;
2661
2662 if (tr) {
2663 /* time of this sched_switch event becomes last time task seen */
2664 tr->last_time = sample->time;
2665
2666 /* last state is used to determine where to account wait time */
2667 tr->last_state = state;
2668
2669 /* sched out event for task so reset ready to run time */
2670 tr->ready_to_run = 0;
2671 }
2672
2673 evsel__save_time(evsel, sample->time, sample->cpu);
2674
2675 return rc;
2676}
2677
2678static int timehist_sched_switch_event(struct perf_tool *tool,
2679 union perf_event *event,
2680 struct evsel *evsel,
2681 struct perf_sample *sample,
2682 struct machine *machine __maybe_unused)
2683{
2684 return timehist_sched_change_event(tool, event, evsel, sample, machine);
2685}
2686
2687static int process_lost(struct perf_tool *tool __maybe_unused,
2688 union perf_event *event,
2689 struct perf_sample *sample,
2690 struct machine *machine __maybe_unused)
2691{
2692 char tstr[64];
2693
2694 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2695 printf("%15s ", tstr);
2696 printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2697
2698 return 0;
2699}
2700
2701
2702static void print_thread_runtime(struct thread *t,
2703 struct thread_runtime *r)
2704{
2705 double mean = avg_stats(&r->run_stats);
2706 float stddev;
2707
2708 printf("%*s %5d %9" PRIu64 " ",
2709 comm_width, timehist_get_commstr(t), t->ppid,
2710 (u64) r->run_stats.n);
2711
2712 print_sched_time(r->total_run_time, 8);
2713 stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2714 print_sched_time(r->run_stats.min, 6);
2715 printf(" ");
2716 print_sched_time((u64) mean, 6);
2717 printf(" ");
2718 print_sched_time(r->run_stats.max, 6);
2719 printf(" ");
2720 printf("%5.2f", stddev);
2721 printf(" %5" PRIu64, r->migrations);
2722 printf("\n");
2723}
2724
2725static void print_thread_waittime(struct thread *t,
2726 struct thread_runtime *r)
2727{
2728 printf("%*s %5d %9" PRIu64 " ",
2729 comm_width, timehist_get_commstr(t), t->ppid,
2730 (u64) r->run_stats.n);
2731
2732 print_sched_time(r->total_run_time, 8);
2733 print_sched_time(r->total_sleep_time, 6);
2734 printf(" ");
2735 print_sched_time(r->total_iowait_time, 6);
2736 printf(" ");
2737 print_sched_time(r->total_preempt_time, 6);
2738 printf(" ");
2739 print_sched_time(r->total_delay_time, 6);
2740 printf("\n");
2741}
2742
2743struct total_run_stats {
2744 struct perf_sched *sched;
2745 u64 sched_count;
2746 u64 task_count;
2747 u64 total_run_time;
2748};
2749
2750static int __show_thread_runtime(struct thread *t, void *priv)
2751{
2752 struct total_run_stats *stats = priv;
2753 struct thread_runtime *r;
2754
2755 if (thread__is_filtered(t))
2756 return 0;
2757
2758 r = thread__priv(t);
2759 if (r && r->run_stats.n) {
2760 stats->task_count++;
2761 stats->sched_count += r->run_stats.n;
2762 stats->total_run_time += r->total_run_time;
2763
2764 if (stats->sched->show_state)
2765 print_thread_waittime(t, r);
2766 else
2767 print_thread_runtime(t, r);
2768 }
2769
2770 return 0;
2771}
2772
2773static int show_thread_runtime(struct thread *t, void *priv)
2774{
2775 if (t->dead)
2776 return 0;
2777
2778 return __show_thread_runtime(t, priv);
2779}
2780
2781static int show_deadthread_runtime(struct thread *t, void *priv)
2782{
2783 if (!t->dead)
2784 return 0;
2785
2786 return __show_thread_runtime(t, priv);
2787}
2788
2789static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2790{
2791 const char *sep = " <- ";
2792 struct callchain_list *chain;
2793 size_t ret = 0;
2794 char bf[1024];
2795 bool first;
2796
2797 if (node == NULL)
2798 return 0;
2799
2800 ret = callchain__fprintf_folded(fp, node->parent);
2801 first = (ret == 0);
2802
2803 list_for_each_entry(chain, &node->val, list) {
2804 if (chain->ip >= PERF_CONTEXT_MAX)
2805 continue;
2806 if (chain->ms.sym && chain->ms.sym->ignore)
2807 continue;
2808 ret += fprintf(fp, "%s%s", first ? "" : sep,
2809 callchain_list__sym_name(chain, bf, sizeof(bf),
2810 false));
2811 first = false;
2812 }
2813
2814 return ret;
2815}
2816
2817static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2818{
2819 size_t ret = 0;
2820 FILE *fp = stdout;
2821 struct callchain_node *chain;
2822 struct rb_node *rb_node = rb_first_cached(root);
2823
2824 printf(" %16s %8s %s\n", "Idle time (msec)", "Count", "Callchains");
2825 printf(" %.16s %.8s %.50s\n", graph_dotted_line, graph_dotted_line,
2826 graph_dotted_line);
2827
2828 while (rb_node) {
2829 chain = rb_entry(rb_node, struct callchain_node, rb_node);
2830 rb_node = rb_next(rb_node);
2831
2832 ret += fprintf(fp, " ");
2833 print_sched_time(chain->hit, 12);
2834 ret += 16; /* print_sched_time returns 2nd arg + 4 */
2835 ret += fprintf(fp, " %8d ", chain->count);
2836 ret += callchain__fprintf_folded(fp, chain);
2837 ret += fprintf(fp, "\n");
2838 }
2839
2840 return ret;
2841}
2842
2843static void timehist_print_summary(struct perf_sched *sched,
2844 struct perf_session *session)
2845{
2846 struct machine *m = &session->machines.host;
2847 struct total_run_stats totals;
2848 u64 task_count;
2849 struct thread *t;
2850 struct thread_runtime *r;
2851 int i;
2852 u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2853
2854 memset(&totals, 0, sizeof(totals));
2855 totals.sched = sched;
2856
2857 if (sched->idle_hist) {
2858 printf("\nIdle-time summary\n");
2859 printf("%*s parent sched-out ", comm_width, "comm");
2860 printf(" idle-time min-idle avg-idle max-idle stddev migrations\n");
2861 } else if (sched->show_state) {
2862 printf("\nWait-time summary\n");
2863 printf("%*s parent sched-in ", comm_width, "comm");
2864 printf(" run-time sleep iowait preempt delay\n");
2865 } else {
2866 printf("\nRuntime summary\n");
2867 printf("%*s parent sched-in ", comm_width, "comm");
2868 printf(" run-time min-run avg-run max-run stddev migrations\n");
2869 }
2870 printf("%*s (count) ", comm_width, "");
2871 printf(" (msec) (msec) (msec) (msec) %s\n",
2872 sched->show_state ? "(msec)" : "%");
2873 printf("%.117s\n", graph_dotted_line);
2874
2875 machine__for_each_thread(m, show_thread_runtime, &totals);
2876 task_count = totals.task_count;
2877 if (!task_count)
2878 printf("<no still running tasks>\n");
2879
2880 printf("\nTerminated tasks:\n");
2881 machine__for_each_thread(m, show_deadthread_runtime, &totals);
2882 if (task_count == totals.task_count)
2883 printf("<no terminated tasks>\n");
2884
2885 /* CPU idle stats not tracked when samples were skipped */
2886 if (sched->skipped_samples && !sched->idle_hist)
2887 return;
2888
2889 printf("\nIdle stats:\n");
2890 for (i = 0; i < idle_max_cpu; ++i) {
2891 if (cpu_list && !test_bit(i, cpu_bitmap))
2892 continue;
2893
2894 t = idle_threads[i];
2895 if (!t)
2896 continue;
2897
2898 r = thread__priv(t);
2899 if (r && r->run_stats.n) {
2900 totals.sched_count += r->run_stats.n;
2901 printf(" CPU %2d idle for ", i);
2902 print_sched_time(r->total_run_time, 6);
2903 printf(" msec (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2904 } else
2905 printf(" CPU %2d idle entire time window\n", i);
2906 }
2907
2908 if (sched->idle_hist && sched->show_callchain) {
2909 callchain_param.mode = CHAIN_FOLDED;
2910 callchain_param.value = CCVAL_PERIOD;
2911
2912 callchain_register_param(&callchain_param);
2913
2914 printf("\nIdle stats by callchain:\n");
2915 for (i = 0; i < idle_max_cpu; ++i) {
2916 struct idle_thread_runtime *itr;
2917
2918 t = idle_threads[i];
2919 if (!t)
2920 continue;
2921
2922 itr = thread__priv(t);
2923 if (itr == NULL)
2924 continue;
2925
2926 callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
2927 0, &callchain_param);
2928
2929 printf(" CPU %2d:", i);
2930 print_sched_time(itr->tr.total_run_time, 6);
2931 printf(" msec\n");
2932 timehist_print_idlehist_callchain(&itr->sorted_root);
2933 printf("\n");
2934 }
2935 }
2936
2937 printf("\n"
2938 " Total number of unique tasks: %" PRIu64 "\n"
2939 "Total number of context switches: %" PRIu64 "\n",
2940 totals.task_count, totals.sched_count);
2941
2942 printf(" Total run time (msec): ");
2943 print_sched_time(totals.total_run_time, 2);
2944 printf("\n");
2945
2946 printf(" Total scheduling time (msec): ");
2947 print_sched_time(hist_time, 2);
2948 printf(" (x %d)\n", sched->max_cpu.cpu);
2949}
2950
2951typedef int (*sched_handler)(struct perf_tool *tool,
2952 union perf_event *event,
2953 struct evsel *evsel,
2954 struct perf_sample *sample,
2955 struct machine *machine);
2956
2957static int perf_timehist__process_sample(struct perf_tool *tool,
2958 union perf_event *event,
2959 struct perf_sample *sample,
2960 struct evsel *evsel,
2961 struct machine *machine)
2962{
2963 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2964 int err = 0;
2965 struct perf_cpu this_cpu = {
2966 .cpu = sample->cpu,
2967 };
2968
2969 if (this_cpu.cpu > sched->max_cpu.cpu)
2970 sched->max_cpu = this_cpu;
2971
2972 if (evsel->handler != NULL) {
2973 sched_handler f = evsel->handler;
2974
2975 err = f(tool, event, evsel, sample, machine);
2976 }
2977
2978 return err;
2979}
2980
2981static int timehist_check_attr(struct perf_sched *sched,
2982 struct evlist *evlist)
2983{
2984 struct evsel *evsel;
2985 struct evsel_runtime *er;
2986
2987 list_for_each_entry(evsel, &evlist->core.entries, core.node) {
2988 er = evsel__get_runtime(evsel);
2989 if (er == NULL) {
2990 pr_err("Failed to allocate memory for evsel runtime data\n");
2991 return -1;
2992 }
2993
2994 if (sched->show_callchain && !evsel__has_callchain(evsel)) {
2995 pr_info("Samples do not have callchains.\n");
2996 sched->show_callchain = 0;
2997 symbol_conf.use_callchain = 0;
2998 }
2999 }
3000
3001 return 0;
3002}
3003
3004static int perf_sched__timehist(struct perf_sched *sched)
3005{
3006 struct evsel_str_handler handlers[] = {
3007 { "sched:sched_switch", timehist_sched_switch_event, },
3008 { "sched:sched_wakeup", timehist_sched_wakeup_event, },
3009 { "sched:sched_waking", timehist_sched_wakeup_event, },
3010 { "sched:sched_wakeup_new", timehist_sched_wakeup_event, },
3011 };
3012 const struct evsel_str_handler migrate_handlers[] = {
3013 { "sched:sched_migrate_task", timehist_migrate_task_event, },
3014 };
3015 struct perf_data data = {
3016 .path = input_name,
3017 .mode = PERF_DATA_MODE_READ,
3018 .force = sched->force,
3019 };
3020
3021 struct perf_session *session;
3022 struct evlist *evlist;
3023 int err = -1;
3024
3025 /*
3026 * event handlers for timehist option
3027 */
3028 sched->tool.sample = perf_timehist__process_sample;
3029 sched->tool.mmap = perf_event__process_mmap;
3030 sched->tool.comm = perf_event__process_comm;
3031 sched->tool.exit = perf_event__process_exit;
3032 sched->tool.fork = perf_event__process_fork;
3033 sched->tool.lost = process_lost;
3034 sched->tool.attr = perf_event__process_attr;
3035 sched->tool.tracing_data = perf_event__process_tracing_data;
3036 sched->tool.build_id = perf_event__process_build_id;
3037
3038 sched->tool.ordered_events = true;
3039 sched->tool.ordering_requires_timestamps = true;
3040
3041 symbol_conf.use_callchain = sched->show_callchain;
3042
3043 session = perf_session__new(&data, &sched->tool);
3044 if (IS_ERR(session))
3045 return PTR_ERR(session);
3046
3047 if (cpu_list) {
3048 err = perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap);
3049 if (err < 0)
3050 goto out;
3051 }
3052
3053 evlist = session->evlist;
3054
3055 symbol__init(&session->header.env);
3056
3057 if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
3058 pr_err("Invalid time string\n");
3059 return -EINVAL;
3060 }
3061
3062 if (timehist_check_attr(sched, evlist) != 0)
3063 goto out;
3064
3065 setup_pager();
3066
3067 /* prefer sched_waking if it is captured */
3068 if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
3069 handlers[1].handler = timehist_sched_wakeup_ignore;
3070
3071 /* setup per-evsel handlers */
3072 if (perf_session__set_tracepoints_handlers(session, handlers))
3073 goto out;
3074
3075 /* sched_switch event at a minimum needs to exist */
3076 if (!evlist__find_tracepoint_by_name(session->evlist, "sched:sched_switch")) {
3077 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3078 goto out;
3079 }
3080
3081 if (sched->show_migrations &&
3082 perf_session__set_tracepoints_handlers(session, migrate_handlers))
3083 goto out;
3084
3085 /* pre-allocate struct for per-CPU idle stats */
3086 sched->max_cpu.cpu = session->header.env.nr_cpus_online;
3087 if (sched->max_cpu.cpu == 0)
3088 sched->max_cpu.cpu = 4;
3089 if (init_idle_threads(sched->max_cpu.cpu))
3090 goto out;
3091
3092 /* summary_only implies summary option, but don't overwrite summary if set */
3093 if (sched->summary_only)
3094 sched->summary = sched->summary_only;
3095
3096 if (!sched->summary_only)
3097 timehist_header(sched);
3098
3099 err = perf_session__process_events(session);
3100 if (err) {
3101 pr_err("Failed to process events, error %d", err);
3102 goto out;
3103 }
3104
3105 sched->nr_events = evlist->stats.nr_events[0];
3106 sched->nr_lost_events = evlist->stats.total_lost;
3107 sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3108
3109 if (sched->summary)
3110 timehist_print_summary(sched, session);
3111
3112out:
3113 free_idle_threads();
3114 perf_session__delete(session);
3115
3116 return err;
3117}
3118
3119
3120static void print_bad_events(struct perf_sched *sched)
3121{
3122 if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3123 printf(" INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3124 (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3125 sched->nr_unordered_timestamps, sched->nr_timestamps);
3126 }
3127 if (sched->nr_lost_events && sched->nr_events) {
3128 printf(" INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3129 (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3130 sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3131 }
3132 if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3133 printf(" INFO: %.3f%% context switch bugs (%ld out of %ld)",
3134 (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3135 sched->nr_context_switch_bugs, sched->nr_timestamps);
3136 if (sched->nr_lost_events)
3137 printf(" (due to lost events?)");
3138 printf("\n");
3139 }
3140}
3141
3142static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3143{
3144 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3145 struct work_atoms *this;
3146 const char *comm = thread__comm_str(data->thread), *this_comm;
3147 bool leftmost = true;
3148
3149 while (*new) {
3150 int cmp;
3151
3152 this = container_of(*new, struct work_atoms, node);
3153 parent = *new;
3154
3155 this_comm = thread__comm_str(this->thread);
3156 cmp = strcmp(comm, this_comm);
3157 if (cmp > 0) {
3158 new = &((*new)->rb_left);
3159 } else if (cmp < 0) {
3160 new = &((*new)->rb_right);
3161 leftmost = false;
3162 } else {
3163 this->num_merged++;
3164 this->total_runtime += data->total_runtime;
3165 this->nb_atoms += data->nb_atoms;
3166 this->total_lat += data->total_lat;
3167 list_splice(&data->work_list, &this->work_list);
3168 if (this->max_lat < data->max_lat) {
3169 this->max_lat = data->max_lat;
3170 this->max_lat_start = data->max_lat_start;
3171 this->max_lat_end = data->max_lat_end;
3172 }
3173 zfree(&data);
3174 return;
3175 }
3176 }
3177
3178 data->num_merged++;
3179 rb_link_node(&data->node, parent, new);
3180 rb_insert_color_cached(&data->node, root, leftmost);
3181}
3182
3183static void perf_sched__merge_lat(struct perf_sched *sched)
3184{
3185 struct work_atoms *data;
3186 struct rb_node *node;
3187
3188 if (sched->skip_merge)
3189 return;
3190
3191 while ((node = rb_first_cached(&sched->atom_root))) {
3192 rb_erase_cached(node, &sched->atom_root);
3193 data = rb_entry(node, struct work_atoms, node);
3194 __merge_work_atoms(&sched->merged_atom_root, data);
3195 }
3196}
3197
3198static int perf_sched__lat(struct perf_sched *sched)
3199{
3200 struct rb_node *next;
3201
3202 setup_pager();
3203
3204 if (perf_sched__read_events(sched))
3205 return -1;
3206
3207 perf_sched__merge_lat(sched);
3208 perf_sched__sort_lat(sched);
3209
3210 printf("\n -------------------------------------------------------------------------------------------------------------------------------------------\n");
3211 printf(" Task | Runtime ms | Switches | Avg delay ms | Max delay ms | Max delay start | Max delay end |\n");
3212 printf(" -------------------------------------------------------------------------------------------------------------------------------------------\n");
3213
3214 next = rb_first_cached(&sched->sorted_atom_root);
3215
3216 while (next) {
3217 struct work_atoms *work_list;
3218
3219 work_list = rb_entry(next, struct work_atoms, node);
3220 output_lat_thread(sched, work_list);
3221 next = rb_next(next);
3222 thread__zput(work_list->thread);
3223 }
3224
3225 printf(" -----------------------------------------------------------------------------------------------------------------\n");
3226 printf(" TOTAL: |%11.3f ms |%9" PRIu64 " |\n",
3227 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3228
3229 printf(" ---------------------------------------------------\n");
3230
3231 print_bad_events(sched);
3232 printf("\n");
3233
3234 return 0;
3235}
3236
3237static int setup_map_cpus(struct perf_sched *sched)
3238{
3239 struct perf_cpu_map *map;
3240
3241 sched->max_cpu.cpu = sysconf(_SC_NPROCESSORS_CONF);
3242
3243 if (sched->map.comp) {
3244 sched->map.comp_cpus = zalloc(sched->max_cpu.cpu * sizeof(int));
3245 if (!sched->map.comp_cpus)
3246 return -1;
3247 }
3248
3249 if (!sched->map.cpus_str)
3250 return 0;
3251
3252 map = perf_cpu_map__new(sched->map.cpus_str);
3253 if (!map) {
3254 pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3255 return -1;
3256 }
3257
3258 sched->map.cpus = map;
3259 return 0;
3260}
3261
3262static int setup_color_pids(struct perf_sched *sched)
3263{
3264 struct perf_thread_map *map;
3265
3266 if (!sched->map.color_pids_str)
3267 return 0;
3268
3269 map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3270 if (!map) {
3271 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3272 return -1;
3273 }
3274
3275 sched->map.color_pids = map;
3276 return 0;
3277}
3278
3279static int setup_color_cpus(struct perf_sched *sched)
3280{
3281 struct perf_cpu_map *map;
3282
3283 if (!sched->map.color_cpus_str)
3284 return 0;
3285
3286 map = perf_cpu_map__new(sched->map.color_cpus_str);
3287 if (!map) {
3288 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3289 return -1;
3290 }
3291
3292 sched->map.color_cpus = map;
3293 return 0;
3294}
3295
3296static int perf_sched__map(struct perf_sched *sched)
3297{
3298 if (setup_map_cpus(sched))
3299 return -1;
3300
3301 if (setup_color_pids(sched))
3302 return -1;
3303
3304 if (setup_color_cpus(sched))
3305 return -1;
3306
3307 setup_pager();
3308 if (perf_sched__read_events(sched))
3309 return -1;
3310 print_bad_events(sched);
3311 return 0;
3312}
3313
3314static int perf_sched__replay(struct perf_sched *sched)
3315{
3316 unsigned long i;
3317
3318 calibrate_run_measurement_overhead(sched);
3319 calibrate_sleep_measurement_overhead(sched);
3320
3321 test_calibrations(sched);
3322
3323 if (perf_sched__read_events(sched))
3324 return -1;
3325
3326 printf("nr_run_events: %ld\n", sched->nr_run_events);
3327 printf("nr_sleep_events: %ld\n", sched->nr_sleep_events);
3328 printf("nr_wakeup_events: %ld\n", sched->nr_wakeup_events);
3329
3330 if (sched->targetless_wakeups)
3331 printf("target-less wakeups: %ld\n", sched->targetless_wakeups);
3332 if (sched->multitarget_wakeups)
3333 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3334 if (sched->nr_run_events_optimized)
3335 printf("run atoms optimized: %ld\n",
3336 sched->nr_run_events_optimized);
3337
3338 print_task_traces(sched);
3339 add_cross_task_wakeups(sched);
3340
3341 sched->thread_funcs_exit = false;
3342 create_tasks(sched);
3343 printf("------------------------------------------------------------\n");
3344 for (i = 0; i < sched->replay_repeat; i++)
3345 run_one_test(sched);
3346
3347 sched->thread_funcs_exit = true;
3348 destroy_tasks(sched);
3349 return 0;
3350}
3351
3352static void setup_sorting(struct perf_sched *sched, const struct option *options,
3353 const char * const usage_msg[])
3354{
3355 char *tmp, *tok, *str = strdup(sched->sort_order);
3356
3357 for (tok = strtok_r(str, ", ", &tmp);
3358 tok; tok = strtok_r(NULL, ", ", &tmp)) {
3359 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3360 usage_with_options_msg(usage_msg, options,
3361 "Unknown --sort key: `%s'", tok);
3362 }
3363 }
3364
3365 free(str);
3366
3367 sort_dimension__add("pid", &sched->cmp_pid);
3368}
3369
3370static bool schedstat_events_exposed(void)
3371{
3372 /*
3373 * Select "sched:sched_stat_wait" event to check
3374 * whether schedstat tracepoints are exposed.
3375 */
3376 return IS_ERR(trace_event__tp_format("sched", "sched_stat_wait")) ?
3377 false : true;
3378}
3379
3380static int __cmd_record(int argc, const char **argv)
3381{
3382 unsigned int rec_argc, i, j;
3383 char **rec_argv;
3384 const char **rec_argv_copy;
3385 const char * const record_args[] = {
3386 "record",
3387 "-a",
3388 "-R",
3389 "-m", "1024",
3390 "-c", "1",
3391 "-e", "sched:sched_switch",
3392 "-e", "sched:sched_stat_runtime",
3393 "-e", "sched:sched_process_fork",
3394 "-e", "sched:sched_wakeup_new",
3395 "-e", "sched:sched_migrate_task",
3396 };
3397
3398 /*
3399 * The tracepoints trace_sched_stat_{wait, sleep, iowait}
3400 * are not exposed to user if CONFIG_SCHEDSTATS is not set,
3401 * to prevent "perf sched record" execution failure, determine
3402 * whether to record schedstat events according to actual situation.
3403 */
3404 const char * const schedstat_args[] = {
3405 "-e", "sched:sched_stat_wait",
3406 "-e", "sched:sched_stat_sleep",
3407 "-e", "sched:sched_stat_iowait",
3408 };
3409 unsigned int schedstat_argc = schedstat_events_exposed() ?
3410 ARRAY_SIZE(schedstat_args) : 0;
3411
3412 struct tep_event *waking_event;
3413 int ret;
3414
3415 /*
3416 * +2 for either "-e", "sched:sched_wakeup" or
3417 * "-e", "sched:sched_waking"
3418 */
3419 rec_argc = ARRAY_SIZE(record_args) + 2 + schedstat_argc + argc - 1;
3420 rec_argv = calloc(rec_argc + 1, sizeof(char *));
3421 if (rec_argv == NULL)
3422 return -ENOMEM;
3423 rec_argv_copy = calloc(rec_argc + 1, sizeof(char *));
3424 if (rec_argv_copy == NULL) {
3425 free(rec_argv);
3426 return -ENOMEM;
3427 }
3428
3429 for (i = 0; i < ARRAY_SIZE(record_args); i++)
3430 rec_argv[i] = strdup(record_args[i]);
3431
3432 rec_argv[i++] = strdup("-e");
3433 waking_event = trace_event__tp_format("sched", "sched_waking");
3434 if (!IS_ERR(waking_event))
3435 rec_argv[i++] = strdup("sched:sched_waking");
3436 else
3437 rec_argv[i++] = strdup("sched:sched_wakeup");
3438
3439 for (j = 0; j < schedstat_argc; j++)
3440 rec_argv[i++] = strdup(schedstat_args[j]);
3441
3442 for (j = 1; j < (unsigned int)argc; j++, i++)
3443 rec_argv[i] = strdup(argv[j]);
3444
3445 BUG_ON(i != rec_argc);
3446
3447 memcpy(rec_argv_copy, rec_argv, sizeof(char *) * rec_argc);
3448 ret = cmd_record(rec_argc, rec_argv_copy);
3449
3450 for (i = 0; i < rec_argc; i++)
3451 free(rec_argv[i]);
3452 free(rec_argv);
3453 free(rec_argv_copy);
3454
3455 return ret;
3456}
3457
3458int cmd_sched(int argc, const char **argv)
3459{
3460 static const char default_sort_order[] = "avg, max, switch, runtime";
3461 struct perf_sched sched = {
3462 .tool = {
3463 .sample = perf_sched__process_tracepoint_sample,
3464 .comm = perf_sched__process_comm,
3465 .namespaces = perf_event__process_namespaces,
3466 .lost = perf_event__process_lost,
3467 .fork = perf_sched__process_fork_event,
3468 .ordered_events = true,
3469 },
3470 .cmp_pid = LIST_HEAD_INIT(sched.cmp_pid),
3471 .sort_list = LIST_HEAD_INIT(sched.sort_list),
3472 .sort_order = default_sort_order,
3473 .replay_repeat = 10,
3474 .profile_cpu = -1,
3475 .next_shortname1 = 'A',
3476 .next_shortname2 = '0',
3477 .skip_merge = 0,
3478 .show_callchain = 1,
3479 .max_stack = 5,
3480 };
3481 const struct option sched_options[] = {
3482 OPT_STRING('i', "input", &input_name, "file",
3483 "input file name"),
3484 OPT_INCR('v', "verbose", &verbose,
3485 "be more verbose (show symbol address, etc)"),
3486 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3487 "dump raw trace in ASCII"),
3488 OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3489 OPT_END()
3490 };
3491 const struct option latency_options[] = {
3492 OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3493 "sort by key(s): runtime, switch, avg, max"),
3494 OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3495 "CPU to profile on"),
3496 OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3497 "latency stats per pid instead of per comm"),
3498 OPT_PARENT(sched_options)
3499 };
3500 const struct option replay_options[] = {
3501 OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3502 "repeat the workload replay N times (-1: infinite)"),
3503 OPT_PARENT(sched_options)
3504 };
3505 const struct option map_options[] = {
3506 OPT_BOOLEAN(0, "compact", &sched.map.comp,
3507 "map output in compact mode"),
3508 OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3509 "highlight given pids in map"),
3510 OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3511 "highlight given CPUs in map"),
3512 OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3513 "display given CPUs in map"),
3514 OPT_PARENT(sched_options)
3515 };
3516 const struct option timehist_options[] = {
3517 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3518 "file", "vmlinux pathname"),
3519 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3520 "file", "kallsyms pathname"),
3521 OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3522 "Display call chains if present (default on)"),
3523 OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3524 "Maximum number of functions to display backtrace."),
3525 OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3526 "Look for files with symbols relative to this directory"),
3527 OPT_BOOLEAN('s', "summary", &sched.summary_only,
3528 "Show only syscall summary with statistics"),
3529 OPT_BOOLEAN('S', "with-summary", &sched.summary,
3530 "Show all syscalls and summary with statistics"),
3531 OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3532 OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3533 OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3534 OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3535 OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3536 OPT_STRING(0, "time", &sched.time_str, "str",
3537 "Time span for analysis (start,stop)"),
3538 OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3539 OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3540 "analyze events only for given process id(s)"),
3541 OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3542 "analyze events only for given thread id(s)"),
3543 OPT_STRING('C', "cpu", &cpu_list, "cpu", "list of cpus to profile"),
3544 OPT_PARENT(sched_options)
3545 };
3546
3547 const char * const latency_usage[] = {
3548 "perf sched latency [<options>]",
3549 NULL
3550 };
3551 const char * const replay_usage[] = {
3552 "perf sched replay [<options>]",
3553 NULL
3554 };
3555 const char * const map_usage[] = {
3556 "perf sched map [<options>]",
3557 NULL
3558 };
3559 const char * const timehist_usage[] = {
3560 "perf sched timehist [<options>]",
3561 NULL
3562 };
3563 const char *const sched_subcommands[] = { "record", "latency", "map",
3564 "replay", "script",
3565 "timehist", NULL };
3566 const char *sched_usage[] = {
3567 NULL,
3568 NULL
3569 };
3570 struct trace_sched_handler lat_ops = {
3571 .wakeup_event = latency_wakeup_event,
3572 .switch_event = latency_switch_event,
3573 .runtime_event = latency_runtime_event,
3574 .migrate_task_event = latency_migrate_task_event,
3575 };
3576 struct trace_sched_handler map_ops = {
3577 .switch_event = map_switch_event,
3578 };
3579 struct trace_sched_handler replay_ops = {
3580 .wakeup_event = replay_wakeup_event,
3581 .switch_event = replay_switch_event,
3582 .fork_event = replay_fork_event,
3583 };
3584 unsigned int i;
3585 int ret = 0;
3586
3587 mutex_init(&sched.start_work_mutex);
3588 mutex_init(&sched.work_done_wait_mutex);
3589 for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3590 sched.curr_pid[i] = -1;
3591
3592 argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3593 sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3594 if (!argc)
3595 usage_with_options(sched_usage, sched_options);
3596
3597 /*
3598 * Aliased to 'perf script' for now:
3599 */
3600 if (!strcmp(argv[0], "script")) {
3601 ret = cmd_script(argc, argv);
3602 } else if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
3603 ret = __cmd_record(argc, argv);
3604 } else if (strlen(argv[0]) > 2 && strstarts("latency", argv[0])) {
3605 sched.tp_handler = &lat_ops;
3606 if (argc > 1) {
3607 argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3608 if (argc)
3609 usage_with_options(latency_usage, latency_options);
3610 }
3611 setup_sorting(&sched, latency_options, latency_usage);
3612 ret = perf_sched__lat(&sched);
3613 } else if (!strcmp(argv[0], "map")) {
3614 if (argc) {
3615 argc = parse_options(argc, argv, map_options, map_usage, 0);
3616 if (argc)
3617 usage_with_options(map_usage, map_options);
3618 }
3619 sched.tp_handler = &map_ops;
3620 setup_sorting(&sched, latency_options, latency_usage);
3621 ret = perf_sched__map(&sched);
3622 } else if (strlen(argv[0]) > 2 && strstarts("replay", argv[0])) {
3623 sched.tp_handler = &replay_ops;
3624 if (argc) {
3625 argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3626 if (argc)
3627 usage_with_options(replay_usage, replay_options);
3628 }
3629 ret = perf_sched__replay(&sched);
3630 } else if (!strcmp(argv[0], "timehist")) {
3631 if (argc) {
3632 argc = parse_options(argc, argv, timehist_options,
3633 timehist_usage, 0);
3634 if (argc)
3635 usage_with_options(timehist_usage, timehist_options);
3636 }
3637 if ((sched.show_wakeups || sched.show_next) &&
3638 sched.summary_only) {
3639 pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3640 parse_options_usage(timehist_usage, timehist_options, "s", true);
3641 if (sched.show_wakeups)
3642 parse_options_usage(NULL, timehist_options, "w", true);
3643 if (sched.show_next)
3644 parse_options_usage(NULL, timehist_options, "n", true);
3645 ret = -EINVAL;
3646 goto out;
3647 }
3648 ret = symbol__validate_sym_arguments();
3649 if (ret)
3650 goto out;
3651
3652 ret = perf_sched__timehist(&sched);
3653 } else {
3654 usage_with_options(sched_usage, sched_options);
3655 }
3656
3657out:
3658 mutex_destroy(&sched.start_work_mutex);
3659 mutex_destroy(&sched.work_done_wait_mutex);
3660
3661 return ret;
3662}
1// SPDX-License-Identifier: GPL-2.0
2#include "builtin.h"
3#include "perf-sys.h"
4
5#include "util/cpumap.h"
6#include "util/evlist.h"
7#include "util/evsel.h"
8#include "util/evsel_fprintf.h"
9#include "util/mutex.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"
22
23#include <subcmd/pager.h>
24#include <subcmd/parse-options.h>
25#include "util/trace-event.h"
26
27#include "util/debug.h"
28#include "util/event.h"
29#include "util/util.h"
30
31#include <linux/kernel.h>
32#include <linux/log2.h>
33#include <linux/zalloc.h>
34#include <sys/prctl.h>
35#include <sys/resource.h>
36#include <inttypes.h>
37
38#include <errno.h>
39#include <semaphore.h>
40#include <pthread.h>
41#include <math.h>
42#include <api/fs/fs.h>
43#include <perf/cpumap.h>
44#include <linux/time64.h>
45#include <linux/err.h>
46
47#include <linux/ctype.h>
48
49#define PR_SET_NAME 15 /* Set process name */
50#define MAX_CPUS 4096
51#define COMM_LEN 20
52#define SYM_LEN 129
53#define MAX_PID 1024000
54#define MAX_PRIO 140
55
56static const char *cpu_list;
57static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
58
59struct sched_atom;
60
61struct task_desc {
62 unsigned long nr;
63 unsigned long pid;
64 char comm[COMM_LEN];
65
66 unsigned long nr_events;
67 unsigned long curr_event;
68 struct sched_atom **atoms;
69
70 pthread_t thread;
71
72 sem_t ready_for_work;
73 sem_t work_done_sem;
74
75 u64 cpu_usage;
76};
77
78enum sched_event_type {
79 SCHED_EVENT_RUN,
80 SCHED_EVENT_SLEEP,
81 SCHED_EVENT_WAKEUP,
82};
83
84struct sched_atom {
85 enum sched_event_type type;
86 u64 timestamp;
87 u64 duration;
88 unsigned long nr;
89 sem_t *wait_sem;
90 struct task_desc *wakee;
91};
92
93enum thread_state {
94 THREAD_SLEEPING = 0,
95 THREAD_WAIT_CPU,
96 THREAD_SCHED_IN,
97 THREAD_IGNORE
98};
99
100struct work_atom {
101 struct list_head list;
102 enum thread_state state;
103 u64 sched_out_time;
104 u64 wake_up_time;
105 u64 sched_in_time;
106 u64 runtime;
107};
108
109struct work_atoms {
110 struct list_head work_list;
111 struct thread *thread;
112 struct rb_node node;
113 u64 max_lat;
114 u64 max_lat_start;
115 u64 max_lat_end;
116 u64 total_lat;
117 u64 nb_atoms;
118 u64 total_runtime;
119 int num_merged;
120};
121
122typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
123
124struct perf_sched;
125
126struct trace_sched_handler {
127 int (*switch_event)(struct perf_sched *sched, struct evsel *evsel,
128 struct perf_sample *sample, struct machine *machine);
129
130 int (*runtime_event)(struct perf_sched *sched, struct evsel *evsel,
131 struct perf_sample *sample, struct machine *machine);
132
133 int (*wakeup_event)(struct perf_sched *sched, struct evsel *evsel,
134 struct perf_sample *sample, struct machine *machine);
135
136 /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
137 int (*fork_event)(struct perf_sched *sched, union perf_event *event,
138 struct machine *machine);
139
140 int (*migrate_task_event)(struct perf_sched *sched,
141 struct evsel *evsel,
142 struct perf_sample *sample,
143 struct machine *machine);
144};
145
146#define COLOR_PIDS PERF_COLOR_BLUE
147#define COLOR_CPUS PERF_COLOR_BG_RED
148
149struct perf_sched_map {
150 DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
151 struct perf_cpu *comp_cpus;
152 bool comp;
153 struct perf_thread_map *color_pids;
154 const char *color_pids_str;
155 struct perf_cpu_map *color_cpus;
156 const char *color_cpus_str;
157 const char *task_name;
158 struct strlist *task_names;
159 bool fuzzy;
160 struct perf_cpu_map *cpus;
161 const char *cpus_str;
162};
163
164struct perf_sched {
165 struct perf_tool tool;
166 const char *sort_order;
167 unsigned long nr_tasks;
168 struct task_desc **pid_to_task;
169 struct task_desc **tasks;
170 const struct trace_sched_handler *tp_handler;
171 struct mutex start_work_mutex;
172 struct mutex work_done_wait_mutex;
173 int profile_cpu;
174/*
175 * Track the current task - that way we can know whether there's any
176 * weird events, such as a task being switched away that is not current.
177 */
178 struct perf_cpu max_cpu;
179 u32 *curr_pid;
180 struct thread **curr_thread;
181 struct thread **curr_out_thread;
182 char next_shortname1;
183 char next_shortname2;
184 unsigned int replay_repeat;
185 unsigned long nr_run_events;
186 unsigned long nr_sleep_events;
187 unsigned long nr_wakeup_events;
188 unsigned long nr_sleep_corrections;
189 unsigned long nr_run_events_optimized;
190 unsigned long targetless_wakeups;
191 unsigned long multitarget_wakeups;
192 unsigned long nr_runs;
193 unsigned long nr_timestamps;
194 unsigned long nr_unordered_timestamps;
195 unsigned long nr_context_switch_bugs;
196 unsigned long nr_events;
197 unsigned long nr_lost_chunks;
198 unsigned long nr_lost_events;
199 u64 run_measurement_overhead;
200 u64 sleep_measurement_overhead;
201 u64 start_time;
202 u64 cpu_usage;
203 u64 runavg_cpu_usage;
204 u64 parent_cpu_usage;
205 u64 runavg_parent_cpu_usage;
206 u64 sum_runtime;
207 u64 sum_fluct;
208 u64 run_avg;
209 u64 all_runtime;
210 u64 all_count;
211 u64 *cpu_last_switched;
212 struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
213 struct list_head sort_list, cmp_pid;
214 bool force;
215 bool skip_merge;
216 struct perf_sched_map map;
217
218 /* options for timehist command */
219 bool summary;
220 bool summary_only;
221 bool idle_hist;
222 bool show_callchain;
223 unsigned int max_stack;
224 bool show_cpu_visual;
225 bool show_wakeups;
226 bool show_next;
227 bool show_migrations;
228 bool pre_migrations;
229 bool show_state;
230 bool show_prio;
231 u64 skipped_samples;
232 const char *time_str;
233 struct perf_time_interval ptime;
234 struct perf_time_interval hist_time;
235 volatile bool thread_funcs_exit;
236 const char *prio_str;
237 DECLARE_BITMAP(prio_bitmap, MAX_PRIO);
238};
239
240/* per thread run time data */
241struct thread_runtime {
242 u64 last_time; /* time of previous sched in/out event */
243 u64 dt_run; /* run time */
244 u64 dt_sleep; /* time between CPU access by sleep (off cpu) */
245 u64 dt_iowait; /* time between CPU access by iowait (off cpu) */
246 u64 dt_preempt; /* time between CPU access by preempt (off cpu) */
247 u64 dt_delay; /* time between wakeup and sched-in */
248 u64 dt_pre_mig; /* time between migration and wakeup */
249 u64 ready_to_run; /* time of wakeup */
250 u64 migrated; /* time when a thread is migrated */
251
252 struct stats run_stats;
253 u64 total_run_time;
254 u64 total_sleep_time;
255 u64 total_iowait_time;
256 u64 total_preempt_time;
257 u64 total_delay_time;
258 u64 total_pre_mig_time;
259
260 char last_state;
261
262 char shortname[3];
263 bool comm_changed;
264
265 u64 migrations;
266
267 int prio;
268};
269
270/* per event run time data */
271struct evsel_runtime {
272 u64 *last_time; /* time this event was last seen per cpu */
273 u32 ncpu; /* highest cpu slot allocated */
274};
275
276/* per cpu idle time data */
277struct idle_thread_runtime {
278 struct thread_runtime tr;
279 struct thread *last_thread;
280 struct rb_root_cached sorted_root;
281 struct callchain_root callchain;
282 struct callchain_cursor cursor;
283};
284
285/* track idle times per cpu */
286static struct thread **idle_threads;
287static int idle_max_cpu;
288static char idle_comm[] = "<idle>";
289
290static u64 get_nsecs(void)
291{
292 struct timespec ts;
293
294 clock_gettime(CLOCK_MONOTONIC, &ts);
295
296 return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
297}
298
299static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
300{
301 u64 T0 = get_nsecs(), T1;
302
303 do {
304 T1 = get_nsecs();
305 } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
306}
307
308static void sleep_nsecs(u64 nsecs)
309{
310 struct timespec ts;
311
312 ts.tv_nsec = nsecs % 999999999;
313 ts.tv_sec = nsecs / 999999999;
314
315 nanosleep(&ts, NULL);
316}
317
318static void calibrate_run_measurement_overhead(struct perf_sched *sched)
319{
320 u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
321 int i;
322
323 for (i = 0; i < 10; i++) {
324 T0 = get_nsecs();
325 burn_nsecs(sched, 0);
326 T1 = get_nsecs();
327 delta = T1-T0;
328 min_delta = min(min_delta, delta);
329 }
330 sched->run_measurement_overhead = min_delta;
331
332 printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
333}
334
335static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
336{
337 u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
338 int i;
339
340 for (i = 0; i < 10; i++) {
341 T0 = get_nsecs();
342 sleep_nsecs(10000);
343 T1 = get_nsecs();
344 delta = T1-T0;
345 min_delta = min(min_delta, delta);
346 }
347 min_delta -= 10000;
348 sched->sleep_measurement_overhead = min_delta;
349
350 printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
351}
352
353static struct sched_atom *
354get_new_event(struct task_desc *task, u64 timestamp)
355{
356 struct sched_atom *event = zalloc(sizeof(*event));
357 unsigned long idx = task->nr_events;
358 size_t size;
359
360 event->timestamp = timestamp;
361 event->nr = idx;
362
363 task->nr_events++;
364 size = sizeof(struct sched_atom *) * task->nr_events;
365 task->atoms = realloc(task->atoms, size);
366 BUG_ON(!task->atoms);
367
368 task->atoms[idx] = event;
369
370 return event;
371}
372
373static struct sched_atom *last_event(struct task_desc *task)
374{
375 if (!task->nr_events)
376 return NULL;
377
378 return task->atoms[task->nr_events - 1];
379}
380
381static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
382 u64 timestamp, u64 duration)
383{
384 struct sched_atom *event, *curr_event = last_event(task);
385
386 /*
387 * optimize an existing RUN event by merging this one
388 * to it:
389 */
390 if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
391 sched->nr_run_events_optimized++;
392 curr_event->duration += duration;
393 return;
394 }
395
396 event = get_new_event(task, timestamp);
397
398 event->type = SCHED_EVENT_RUN;
399 event->duration = duration;
400
401 sched->nr_run_events++;
402}
403
404static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
405 u64 timestamp, struct task_desc *wakee)
406{
407 struct sched_atom *event, *wakee_event;
408
409 event = get_new_event(task, timestamp);
410 event->type = SCHED_EVENT_WAKEUP;
411 event->wakee = wakee;
412
413 wakee_event = last_event(wakee);
414 if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
415 sched->targetless_wakeups++;
416 return;
417 }
418 if (wakee_event->wait_sem) {
419 sched->multitarget_wakeups++;
420 return;
421 }
422
423 wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
424 sem_init(wakee_event->wait_sem, 0, 0);
425 event->wait_sem = wakee_event->wait_sem;
426
427 sched->nr_wakeup_events++;
428}
429
430static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
431 u64 timestamp)
432{
433 struct sched_atom *event = get_new_event(task, timestamp);
434
435 event->type = SCHED_EVENT_SLEEP;
436
437 sched->nr_sleep_events++;
438}
439
440static struct task_desc *register_pid(struct perf_sched *sched,
441 unsigned long pid, const char *comm)
442{
443 struct task_desc *task;
444 static int pid_max;
445
446 if (sched->pid_to_task == NULL) {
447 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
448 pid_max = MAX_PID;
449 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
450 }
451 if (pid >= (unsigned long)pid_max) {
452 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
453 sizeof(struct task_desc *))) == NULL);
454 while (pid >= (unsigned long)pid_max)
455 sched->pid_to_task[pid_max++] = NULL;
456 }
457
458 task = sched->pid_to_task[pid];
459
460 if (task)
461 return task;
462
463 task = zalloc(sizeof(*task));
464 task->pid = pid;
465 task->nr = sched->nr_tasks;
466 strcpy(task->comm, comm);
467 /*
468 * every task starts in sleeping state - this gets ignored
469 * if there's no wakeup pointing to this sleep state:
470 */
471 add_sched_event_sleep(sched, task, 0);
472
473 sched->pid_to_task[pid] = task;
474 sched->nr_tasks++;
475 sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
476 BUG_ON(!sched->tasks);
477 sched->tasks[task->nr] = task;
478
479 if (verbose > 0)
480 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
481
482 return task;
483}
484
485
486static void print_task_traces(struct perf_sched *sched)
487{
488 struct task_desc *task;
489 unsigned long i;
490
491 for (i = 0; i < sched->nr_tasks; i++) {
492 task = sched->tasks[i];
493 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
494 task->nr, task->comm, task->pid, task->nr_events);
495 }
496}
497
498static void add_cross_task_wakeups(struct perf_sched *sched)
499{
500 struct task_desc *task1, *task2;
501 unsigned long i, j;
502
503 for (i = 0; i < sched->nr_tasks; i++) {
504 task1 = sched->tasks[i];
505 j = i + 1;
506 if (j == sched->nr_tasks)
507 j = 0;
508 task2 = sched->tasks[j];
509 add_sched_event_wakeup(sched, task1, 0, task2);
510 }
511}
512
513static void perf_sched__process_event(struct perf_sched *sched,
514 struct sched_atom *atom)
515{
516 int ret = 0;
517
518 switch (atom->type) {
519 case SCHED_EVENT_RUN:
520 burn_nsecs(sched, atom->duration);
521 break;
522 case SCHED_EVENT_SLEEP:
523 if (atom->wait_sem)
524 ret = sem_wait(atom->wait_sem);
525 BUG_ON(ret);
526 break;
527 case SCHED_EVENT_WAKEUP:
528 if (atom->wait_sem)
529 ret = sem_post(atom->wait_sem);
530 BUG_ON(ret);
531 break;
532 default:
533 BUG_ON(1);
534 }
535}
536
537static u64 get_cpu_usage_nsec_parent(void)
538{
539 struct rusage ru;
540 u64 sum;
541 int err;
542
543 err = getrusage(RUSAGE_SELF, &ru);
544 BUG_ON(err);
545
546 sum = ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
547 sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
548
549 return sum;
550}
551
552static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
553{
554 struct perf_event_attr attr;
555 char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
556 int fd;
557 struct rlimit limit;
558 bool need_privilege = false;
559
560 memset(&attr, 0, sizeof(attr));
561
562 attr.type = PERF_TYPE_SOFTWARE;
563 attr.config = PERF_COUNT_SW_TASK_CLOCK;
564
565force_again:
566 fd = sys_perf_event_open(&attr, 0, -1, -1,
567 perf_event_open_cloexec_flag());
568
569 if (fd < 0) {
570 if (errno == EMFILE) {
571 if (sched->force) {
572 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
573 limit.rlim_cur += sched->nr_tasks - cur_task;
574 if (limit.rlim_cur > limit.rlim_max) {
575 limit.rlim_max = limit.rlim_cur;
576 need_privilege = true;
577 }
578 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
579 if (need_privilege && errno == EPERM)
580 strcpy(info, "Need privilege\n");
581 } else
582 goto force_again;
583 } else
584 strcpy(info, "Have a try with -f option\n");
585 }
586 pr_err("Error: sys_perf_event_open() syscall returned "
587 "with %d (%s)\n%s", fd,
588 str_error_r(errno, sbuf, sizeof(sbuf)), info);
589 exit(EXIT_FAILURE);
590 }
591 return fd;
592}
593
594static u64 get_cpu_usage_nsec_self(int fd)
595{
596 u64 runtime;
597 int ret;
598
599 ret = read(fd, &runtime, sizeof(runtime));
600 BUG_ON(ret != sizeof(runtime));
601
602 return runtime;
603}
604
605struct sched_thread_parms {
606 struct task_desc *task;
607 struct perf_sched *sched;
608 int fd;
609};
610
611static void *thread_func(void *ctx)
612{
613 struct sched_thread_parms *parms = ctx;
614 struct task_desc *this_task = parms->task;
615 struct perf_sched *sched = parms->sched;
616 u64 cpu_usage_0, cpu_usage_1;
617 unsigned long i, ret;
618 char comm2[22];
619 int fd = parms->fd;
620
621 zfree(&parms);
622
623 sprintf(comm2, ":%s", this_task->comm);
624 prctl(PR_SET_NAME, comm2);
625 if (fd < 0)
626 return NULL;
627
628 while (!sched->thread_funcs_exit) {
629 ret = sem_post(&this_task->ready_for_work);
630 BUG_ON(ret);
631 mutex_lock(&sched->start_work_mutex);
632 mutex_unlock(&sched->start_work_mutex);
633
634 cpu_usage_0 = get_cpu_usage_nsec_self(fd);
635
636 for (i = 0; i < this_task->nr_events; i++) {
637 this_task->curr_event = i;
638 perf_sched__process_event(sched, this_task->atoms[i]);
639 }
640
641 cpu_usage_1 = get_cpu_usage_nsec_self(fd);
642 this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
643 ret = sem_post(&this_task->work_done_sem);
644 BUG_ON(ret);
645
646 mutex_lock(&sched->work_done_wait_mutex);
647 mutex_unlock(&sched->work_done_wait_mutex);
648 }
649 return NULL;
650}
651
652static void create_tasks(struct perf_sched *sched)
653 EXCLUSIVE_LOCK_FUNCTION(sched->start_work_mutex)
654 EXCLUSIVE_LOCK_FUNCTION(sched->work_done_wait_mutex)
655{
656 struct task_desc *task;
657 pthread_attr_t attr;
658 unsigned long i;
659 int err;
660
661 err = pthread_attr_init(&attr);
662 BUG_ON(err);
663 err = pthread_attr_setstacksize(&attr,
664 (size_t) max(16 * 1024, (int)PTHREAD_STACK_MIN));
665 BUG_ON(err);
666 mutex_lock(&sched->start_work_mutex);
667 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->ready_for_work, 0, 0);
675 sem_init(&task->work_done_sem, 0, 0);
676 task->curr_event = 0;
677 err = pthread_create(&task->thread, &attr, thread_func, parms);
678 BUG_ON(err);
679 }
680}
681
682static void destroy_tasks(struct perf_sched *sched)
683 UNLOCK_FUNCTION(sched->start_work_mutex)
684 UNLOCK_FUNCTION(sched->work_done_wait_mutex)
685{
686 struct task_desc *task;
687 unsigned long i;
688 int err;
689
690 mutex_unlock(&sched->start_work_mutex);
691 mutex_unlock(&sched->work_done_wait_mutex);
692 /* Get rid of threads so they won't be upset by mutex destrunction */
693 for (i = 0; i < sched->nr_tasks; i++) {
694 task = sched->tasks[i];
695 err = pthread_join(task->thread, NULL);
696 BUG_ON(err);
697 sem_destroy(&task->ready_for_work);
698 sem_destroy(&task->work_done_sem);
699 }
700}
701
702static void wait_for_tasks(struct perf_sched *sched)
703 EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
704 EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
705{
706 u64 cpu_usage_0, cpu_usage_1;
707 struct task_desc *task;
708 unsigned long i, ret;
709
710 sched->start_time = get_nsecs();
711 sched->cpu_usage = 0;
712 mutex_unlock(&sched->work_done_wait_mutex);
713
714 for (i = 0; i < sched->nr_tasks; i++) {
715 task = sched->tasks[i];
716 ret = sem_wait(&task->ready_for_work);
717 BUG_ON(ret);
718 sem_init(&task->ready_for_work, 0, 0);
719 }
720 mutex_lock(&sched->work_done_wait_mutex);
721
722 cpu_usage_0 = get_cpu_usage_nsec_parent();
723
724 mutex_unlock(&sched->start_work_mutex);
725
726 for (i = 0; i < sched->nr_tasks; i++) {
727 task = sched->tasks[i];
728 ret = sem_wait(&task->work_done_sem);
729 BUG_ON(ret);
730 sem_init(&task->work_done_sem, 0, 0);
731 sched->cpu_usage += task->cpu_usage;
732 task->cpu_usage = 0;
733 }
734
735 cpu_usage_1 = get_cpu_usage_nsec_parent();
736 if (!sched->runavg_cpu_usage)
737 sched->runavg_cpu_usage = sched->cpu_usage;
738 sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
739
740 sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
741 if (!sched->runavg_parent_cpu_usage)
742 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
743 sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
744 sched->parent_cpu_usage)/sched->replay_repeat;
745
746 mutex_lock(&sched->start_work_mutex);
747
748 for (i = 0; i < sched->nr_tasks; i++) {
749 task = sched->tasks[i];
750 task->curr_event = 0;
751 }
752}
753
754static void run_one_test(struct perf_sched *sched)
755 EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
756 EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
757{
758 u64 T0, T1, delta, avg_delta, fluct;
759
760 T0 = get_nsecs();
761 wait_for_tasks(sched);
762 T1 = get_nsecs();
763
764 delta = T1 - T0;
765 sched->sum_runtime += delta;
766 sched->nr_runs++;
767
768 avg_delta = sched->sum_runtime / sched->nr_runs;
769 if (delta < avg_delta)
770 fluct = avg_delta - delta;
771 else
772 fluct = delta - avg_delta;
773 sched->sum_fluct += fluct;
774 if (!sched->run_avg)
775 sched->run_avg = delta;
776 sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
777
778 printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
779
780 printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
781
782 printf("cpu: %0.2f / %0.2f",
783 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
784
785#if 0
786 /*
787 * rusage statistics done by the parent, these are less
788 * accurate than the sched->sum_exec_runtime based statistics:
789 */
790 printf(" [%0.2f / %0.2f]",
791 (double)sched->parent_cpu_usage / NSEC_PER_MSEC,
792 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
793#endif
794
795 printf("\n");
796
797 if (sched->nr_sleep_corrections)
798 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
799 sched->nr_sleep_corrections = 0;
800}
801
802static void test_calibrations(struct perf_sched *sched)
803{
804 u64 T0, T1;
805
806 T0 = get_nsecs();
807 burn_nsecs(sched, NSEC_PER_MSEC);
808 T1 = get_nsecs();
809
810 printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
811
812 T0 = get_nsecs();
813 sleep_nsecs(NSEC_PER_MSEC);
814 T1 = get_nsecs();
815
816 printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
817}
818
819static int
820replay_wakeup_event(struct perf_sched *sched,
821 struct evsel *evsel, struct perf_sample *sample,
822 struct machine *machine __maybe_unused)
823{
824 const char *comm = evsel__strval(evsel, sample, "comm");
825 const u32 pid = evsel__intval(evsel, sample, "pid");
826 struct task_desc *waker, *wakee;
827
828 if (verbose > 0) {
829 printf("sched_wakeup event %p\n", evsel);
830
831 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
832 }
833
834 waker = register_pid(sched, sample->tid, "<unknown>");
835 wakee = register_pid(sched, pid, comm);
836
837 add_sched_event_wakeup(sched, waker, sample->time, wakee);
838 return 0;
839}
840
841static int replay_switch_event(struct perf_sched *sched,
842 struct evsel *evsel,
843 struct perf_sample *sample,
844 struct machine *machine __maybe_unused)
845{
846 const char *prev_comm = evsel__strval(evsel, sample, "prev_comm"),
847 *next_comm = evsel__strval(evsel, sample, "next_comm");
848 const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
849 next_pid = evsel__intval(evsel, sample, "next_pid");
850 struct task_desc *prev, __maybe_unused *next;
851 u64 timestamp0, timestamp = sample->time;
852 int cpu = sample->cpu;
853 s64 delta;
854
855 if (verbose > 0)
856 printf("sched_switch event %p\n", evsel);
857
858 if (cpu >= MAX_CPUS || cpu < 0)
859 return 0;
860
861 timestamp0 = sched->cpu_last_switched[cpu];
862 if (timestamp0)
863 delta = timestamp - timestamp0;
864 else
865 delta = 0;
866
867 if (delta < 0) {
868 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
869 return -1;
870 }
871
872 pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
873 prev_comm, prev_pid, next_comm, next_pid, delta);
874
875 prev = register_pid(sched, prev_pid, prev_comm);
876 next = register_pid(sched, next_pid, next_comm);
877
878 sched->cpu_last_switched[cpu] = timestamp;
879
880 add_sched_event_run(sched, prev, timestamp, delta);
881 add_sched_event_sleep(sched, prev, timestamp);
882
883 return 0;
884}
885
886static int replay_fork_event(struct perf_sched *sched,
887 union perf_event *event,
888 struct machine *machine)
889{
890 struct thread *child, *parent;
891
892 child = machine__findnew_thread(machine, event->fork.pid,
893 event->fork.tid);
894 parent = machine__findnew_thread(machine, event->fork.ppid,
895 event->fork.ptid);
896
897 if (child == NULL || parent == NULL) {
898 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
899 child, parent);
900 goto out_put;
901 }
902
903 if (verbose > 0) {
904 printf("fork event\n");
905 printf("... parent: %s/%d\n", thread__comm_str(parent), thread__tid(parent));
906 printf("... child: %s/%d\n", thread__comm_str(child), thread__tid(child));
907 }
908
909 register_pid(sched, thread__tid(parent), thread__comm_str(parent));
910 register_pid(sched, thread__tid(child), thread__comm_str(child));
911out_put:
912 thread__put(child);
913 thread__put(parent);
914 return 0;
915}
916
917struct sort_dimension {
918 const char *name;
919 sort_fn_t cmp;
920 struct list_head list;
921};
922
923static inline void init_prio(struct thread_runtime *r)
924{
925 r->prio = -1;
926}
927
928/*
929 * handle runtime stats saved per thread
930 */
931static struct thread_runtime *thread__init_runtime(struct thread *thread)
932{
933 struct thread_runtime *r;
934
935 r = zalloc(sizeof(struct thread_runtime));
936 if (!r)
937 return NULL;
938
939 init_stats(&r->run_stats);
940 init_prio(r);
941 thread__set_priv(thread, r);
942
943 return r;
944}
945
946static struct thread_runtime *thread__get_runtime(struct thread *thread)
947{
948 struct thread_runtime *tr;
949
950 tr = thread__priv(thread);
951 if (tr == NULL) {
952 tr = thread__init_runtime(thread);
953 if (tr == NULL)
954 pr_debug("Failed to malloc memory for runtime data.\n");
955 }
956
957 return tr;
958}
959
960static int
961thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
962{
963 struct sort_dimension *sort;
964 int ret = 0;
965
966 BUG_ON(list_empty(list));
967
968 list_for_each_entry(sort, list, list) {
969 ret = sort->cmp(l, r);
970 if (ret)
971 return ret;
972 }
973
974 return ret;
975}
976
977static struct work_atoms *
978thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
979 struct list_head *sort_list)
980{
981 struct rb_node *node = root->rb_root.rb_node;
982 struct work_atoms key = { .thread = thread };
983
984 while (node) {
985 struct work_atoms *atoms;
986 int cmp;
987
988 atoms = container_of(node, struct work_atoms, node);
989
990 cmp = thread_lat_cmp(sort_list, &key, atoms);
991 if (cmp > 0)
992 node = node->rb_left;
993 else if (cmp < 0)
994 node = node->rb_right;
995 else {
996 BUG_ON(thread != atoms->thread);
997 return atoms;
998 }
999 }
1000 return NULL;
1001}
1002
1003static void
1004__thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
1005 struct list_head *sort_list)
1006{
1007 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
1008 bool leftmost = true;
1009
1010 while (*new) {
1011 struct work_atoms *this;
1012 int cmp;
1013
1014 this = container_of(*new, struct work_atoms, node);
1015 parent = *new;
1016
1017 cmp = thread_lat_cmp(sort_list, data, this);
1018
1019 if (cmp > 0)
1020 new = &((*new)->rb_left);
1021 else {
1022 new = &((*new)->rb_right);
1023 leftmost = false;
1024 }
1025 }
1026
1027 rb_link_node(&data->node, parent, new);
1028 rb_insert_color_cached(&data->node, root, leftmost);
1029}
1030
1031static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1032{
1033 struct work_atoms *atoms = zalloc(sizeof(*atoms));
1034 if (!atoms) {
1035 pr_err("No memory at %s\n", __func__);
1036 return -1;
1037 }
1038
1039 atoms->thread = thread__get(thread);
1040 INIT_LIST_HEAD(&atoms->work_list);
1041 __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1042 return 0;
1043}
1044
1045static int
1046add_sched_out_event(struct work_atoms *atoms,
1047 char run_state,
1048 u64 timestamp)
1049{
1050 struct work_atom *atom = zalloc(sizeof(*atom));
1051 if (!atom) {
1052 pr_err("Non memory at %s", __func__);
1053 return -1;
1054 }
1055
1056 atom->sched_out_time = timestamp;
1057
1058 if (run_state == 'R') {
1059 atom->state = THREAD_WAIT_CPU;
1060 atom->wake_up_time = atom->sched_out_time;
1061 }
1062
1063 list_add_tail(&atom->list, &atoms->work_list);
1064 return 0;
1065}
1066
1067static void
1068add_runtime_event(struct work_atoms *atoms, u64 delta,
1069 u64 timestamp __maybe_unused)
1070{
1071 struct work_atom *atom;
1072
1073 BUG_ON(list_empty(&atoms->work_list));
1074
1075 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1076
1077 atom->runtime += delta;
1078 atoms->total_runtime += delta;
1079}
1080
1081static void
1082add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1083{
1084 struct work_atom *atom;
1085 u64 delta;
1086
1087 if (list_empty(&atoms->work_list))
1088 return;
1089
1090 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1091
1092 if (atom->state != THREAD_WAIT_CPU)
1093 return;
1094
1095 if (timestamp < atom->wake_up_time) {
1096 atom->state = THREAD_IGNORE;
1097 return;
1098 }
1099
1100 atom->state = THREAD_SCHED_IN;
1101 atom->sched_in_time = timestamp;
1102
1103 delta = atom->sched_in_time - atom->wake_up_time;
1104 atoms->total_lat += delta;
1105 if (delta > atoms->max_lat) {
1106 atoms->max_lat = delta;
1107 atoms->max_lat_start = atom->wake_up_time;
1108 atoms->max_lat_end = timestamp;
1109 }
1110 atoms->nb_atoms++;
1111}
1112
1113static int latency_switch_event(struct perf_sched *sched,
1114 struct evsel *evsel,
1115 struct perf_sample *sample,
1116 struct machine *machine)
1117{
1118 const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1119 next_pid = evsel__intval(evsel, sample, "next_pid");
1120 const char prev_state = evsel__taskstate(evsel, sample, "prev_state");
1121 struct work_atoms *out_events, *in_events;
1122 struct thread *sched_out, *sched_in;
1123 u64 timestamp0, timestamp = sample->time;
1124 int cpu = sample->cpu, err = -1;
1125 s64 delta;
1126
1127 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1128
1129 timestamp0 = sched->cpu_last_switched[cpu];
1130 sched->cpu_last_switched[cpu] = timestamp;
1131 if (timestamp0)
1132 delta = timestamp - timestamp0;
1133 else
1134 delta = 0;
1135
1136 if (delta < 0) {
1137 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1138 return -1;
1139 }
1140
1141 sched_out = machine__findnew_thread(machine, -1, prev_pid);
1142 sched_in = machine__findnew_thread(machine, -1, next_pid);
1143 if (sched_out == NULL || sched_in == NULL)
1144 goto out_put;
1145
1146 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1147 if (!out_events) {
1148 if (thread_atoms_insert(sched, sched_out))
1149 goto out_put;
1150 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1151 if (!out_events) {
1152 pr_err("out-event: Internal tree error");
1153 goto out_put;
1154 }
1155 }
1156 if (add_sched_out_event(out_events, prev_state, timestamp))
1157 return -1;
1158
1159 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1160 if (!in_events) {
1161 if (thread_atoms_insert(sched, sched_in))
1162 goto out_put;
1163 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1164 if (!in_events) {
1165 pr_err("in-event: Internal tree error");
1166 goto out_put;
1167 }
1168 /*
1169 * Take came in we have not heard about yet,
1170 * add in an initial atom in runnable state:
1171 */
1172 if (add_sched_out_event(in_events, 'R', timestamp))
1173 goto out_put;
1174 }
1175 add_sched_in_event(in_events, timestamp);
1176 err = 0;
1177out_put:
1178 thread__put(sched_out);
1179 thread__put(sched_in);
1180 return err;
1181}
1182
1183static int latency_runtime_event(struct perf_sched *sched,
1184 struct evsel *evsel,
1185 struct perf_sample *sample,
1186 struct machine *machine)
1187{
1188 const u32 pid = evsel__intval(evsel, sample, "pid");
1189 const u64 runtime = evsel__intval(evsel, sample, "runtime");
1190 struct thread *thread = machine__findnew_thread(machine, -1, pid);
1191 struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1192 u64 timestamp = sample->time;
1193 int cpu = sample->cpu, err = -1;
1194
1195 if (thread == NULL)
1196 return -1;
1197
1198 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1199 if (!atoms) {
1200 if (thread_atoms_insert(sched, thread))
1201 goto out_put;
1202 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1203 if (!atoms) {
1204 pr_err("in-event: Internal tree error");
1205 goto out_put;
1206 }
1207 if (add_sched_out_event(atoms, 'R', timestamp))
1208 goto out_put;
1209 }
1210
1211 add_runtime_event(atoms, runtime, timestamp);
1212 err = 0;
1213out_put:
1214 thread__put(thread);
1215 return err;
1216}
1217
1218static int latency_wakeup_event(struct perf_sched *sched,
1219 struct evsel *evsel,
1220 struct perf_sample *sample,
1221 struct machine *machine)
1222{
1223 const u32 pid = evsel__intval(evsel, sample, "pid");
1224 struct work_atoms *atoms;
1225 struct work_atom *atom;
1226 struct thread *wakee;
1227 u64 timestamp = sample->time;
1228 int err = -1;
1229
1230 wakee = machine__findnew_thread(machine, -1, pid);
1231 if (wakee == NULL)
1232 return -1;
1233 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1234 if (!atoms) {
1235 if (thread_atoms_insert(sched, wakee))
1236 goto out_put;
1237 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1238 if (!atoms) {
1239 pr_err("wakeup-event: Internal tree error");
1240 goto out_put;
1241 }
1242 if (add_sched_out_event(atoms, 'S', timestamp))
1243 goto out_put;
1244 }
1245
1246 BUG_ON(list_empty(&atoms->work_list));
1247
1248 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1249
1250 /*
1251 * As we do not guarantee the wakeup event happens when
1252 * task is out of run queue, also may happen when task is
1253 * on run queue and wakeup only change ->state to TASK_RUNNING,
1254 * then we should not set the ->wake_up_time when wake up a
1255 * task which is on run queue.
1256 *
1257 * You WILL be missing events if you've recorded only
1258 * one CPU, or are only looking at only one, so don't
1259 * skip in this case.
1260 */
1261 if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1262 goto out_ok;
1263
1264 sched->nr_timestamps++;
1265 if (atom->sched_out_time > timestamp) {
1266 sched->nr_unordered_timestamps++;
1267 goto out_ok;
1268 }
1269
1270 atom->state = THREAD_WAIT_CPU;
1271 atom->wake_up_time = timestamp;
1272out_ok:
1273 err = 0;
1274out_put:
1275 thread__put(wakee);
1276 return err;
1277}
1278
1279static int latency_migrate_task_event(struct perf_sched *sched,
1280 struct evsel *evsel,
1281 struct perf_sample *sample,
1282 struct machine *machine)
1283{
1284 const u32 pid = evsel__intval(evsel, sample, "pid");
1285 u64 timestamp = sample->time;
1286 struct work_atoms *atoms;
1287 struct work_atom *atom;
1288 struct thread *migrant;
1289 int err = -1;
1290
1291 /*
1292 * Only need to worry about migration when profiling one CPU.
1293 */
1294 if (sched->profile_cpu == -1)
1295 return 0;
1296
1297 migrant = machine__findnew_thread(machine, -1, pid);
1298 if (migrant == NULL)
1299 return -1;
1300 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1301 if (!atoms) {
1302 if (thread_atoms_insert(sched, migrant))
1303 goto out_put;
1304 register_pid(sched, thread__tid(migrant), thread__comm_str(migrant));
1305 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1306 if (!atoms) {
1307 pr_err("migration-event: Internal tree error");
1308 goto out_put;
1309 }
1310 if (add_sched_out_event(atoms, 'R', timestamp))
1311 goto out_put;
1312 }
1313
1314 BUG_ON(list_empty(&atoms->work_list));
1315
1316 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1317 atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1318
1319 sched->nr_timestamps++;
1320
1321 if (atom->sched_out_time > timestamp)
1322 sched->nr_unordered_timestamps++;
1323 err = 0;
1324out_put:
1325 thread__put(migrant);
1326 return err;
1327}
1328
1329static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1330{
1331 int i;
1332 int ret;
1333 u64 avg;
1334 char max_lat_start[32], max_lat_end[32];
1335
1336 if (!work_list->nb_atoms)
1337 return;
1338 /*
1339 * Ignore idle threads:
1340 */
1341 if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1342 return;
1343
1344 sched->all_runtime += work_list->total_runtime;
1345 sched->all_count += work_list->nb_atoms;
1346
1347 if (work_list->num_merged > 1) {
1348 ret = printf(" %s:(%d) ", thread__comm_str(work_list->thread),
1349 work_list->num_merged);
1350 } else {
1351 ret = printf(" %s:%d ", thread__comm_str(work_list->thread),
1352 thread__tid(work_list->thread));
1353 }
1354
1355 for (i = 0; i < 24 - ret; i++)
1356 printf(" ");
1357
1358 avg = work_list->total_lat / work_list->nb_atoms;
1359 timestamp__scnprintf_usec(work_list->max_lat_start, max_lat_start, sizeof(max_lat_start));
1360 timestamp__scnprintf_usec(work_list->max_lat_end, max_lat_end, sizeof(max_lat_end));
1361
1362 printf("|%11.3f ms |%9" PRIu64 " | avg:%8.3f ms | max:%8.3f ms | max start: %12s s | max end: %12s s\n",
1363 (double)work_list->total_runtime / NSEC_PER_MSEC,
1364 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1365 (double)work_list->max_lat / NSEC_PER_MSEC,
1366 max_lat_start, max_lat_end);
1367}
1368
1369static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1370{
1371 pid_t l_tid, r_tid;
1372
1373 if (RC_CHK_EQUAL(l->thread, r->thread))
1374 return 0;
1375 l_tid = thread__tid(l->thread);
1376 r_tid = thread__tid(r->thread);
1377 if (l_tid < r_tid)
1378 return -1;
1379 if (l_tid > r_tid)
1380 return 1;
1381 return (int)(RC_CHK_ACCESS(l->thread) - RC_CHK_ACCESS(r->thread));
1382}
1383
1384static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1385{
1386 u64 avgl, avgr;
1387
1388 if (!l->nb_atoms)
1389 return -1;
1390
1391 if (!r->nb_atoms)
1392 return 1;
1393
1394 avgl = l->total_lat / l->nb_atoms;
1395 avgr = r->total_lat / r->nb_atoms;
1396
1397 if (avgl < avgr)
1398 return -1;
1399 if (avgl > avgr)
1400 return 1;
1401
1402 return 0;
1403}
1404
1405static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1406{
1407 if (l->max_lat < r->max_lat)
1408 return -1;
1409 if (l->max_lat > r->max_lat)
1410 return 1;
1411
1412 return 0;
1413}
1414
1415static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1416{
1417 if (l->nb_atoms < r->nb_atoms)
1418 return -1;
1419 if (l->nb_atoms > r->nb_atoms)
1420 return 1;
1421
1422 return 0;
1423}
1424
1425static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1426{
1427 if (l->total_runtime < r->total_runtime)
1428 return -1;
1429 if (l->total_runtime > r->total_runtime)
1430 return 1;
1431
1432 return 0;
1433}
1434
1435static int sort_dimension__add(const char *tok, struct list_head *list)
1436{
1437 size_t i;
1438 static struct sort_dimension avg_sort_dimension = {
1439 .name = "avg",
1440 .cmp = avg_cmp,
1441 };
1442 static struct sort_dimension max_sort_dimension = {
1443 .name = "max",
1444 .cmp = max_cmp,
1445 };
1446 static struct sort_dimension pid_sort_dimension = {
1447 .name = "pid",
1448 .cmp = pid_cmp,
1449 };
1450 static struct sort_dimension runtime_sort_dimension = {
1451 .name = "runtime",
1452 .cmp = runtime_cmp,
1453 };
1454 static struct sort_dimension switch_sort_dimension = {
1455 .name = "switch",
1456 .cmp = switch_cmp,
1457 };
1458 struct sort_dimension *available_sorts[] = {
1459 &pid_sort_dimension,
1460 &avg_sort_dimension,
1461 &max_sort_dimension,
1462 &switch_sort_dimension,
1463 &runtime_sort_dimension,
1464 };
1465
1466 for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1467 if (!strcmp(available_sorts[i]->name, tok)) {
1468 list_add_tail(&available_sorts[i]->list, list);
1469
1470 return 0;
1471 }
1472 }
1473
1474 return -1;
1475}
1476
1477static void perf_sched__sort_lat(struct perf_sched *sched)
1478{
1479 struct rb_node *node;
1480 struct rb_root_cached *root = &sched->atom_root;
1481again:
1482 for (;;) {
1483 struct work_atoms *data;
1484 node = rb_first_cached(root);
1485 if (!node)
1486 break;
1487
1488 rb_erase_cached(node, root);
1489 data = rb_entry(node, struct work_atoms, node);
1490 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1491 }
1492 if (root == &sched->atom_root) {
1493 root = &sched->merged_atom_root;
1494 goto again;
1495 }
1496}
1497
1498static int process_sched_wakeup_event(const struct perf_tool *tool,
1499 struct evsel *evsel,
1500 struct perf_sample *sample,
1501 struct machine *machine)
1502{
1503 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1504
1505 if (sched->tp_handler->wakeup_event)
1506 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1507
1508 return 0;
1509}
1510
1511static int process_sched_wakeup_ignore(const struct perf_tool *tool __maybe_unused,
1512 struct evsel *evsel __maybe_unused,
1513 struct perf_sample *sample __maybe_unused,
1514 struct machine *machine __maybe_unused)
1515{
1516 return 0;
1517}
1518
1519union map_priv {
1520 void *ptr;
1521 bool color;
1522};
1523
1524static bool thread__has_color(struct thread *thread)
1525{
1526 union map_priv priv = {
1527 .ptr = thread__priv(thread),
1528 };
1529
1530 return priv.color;
1531}
1532
1533static struct thread*
1534map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1535{
1536 struct thread *thread = machine__findnew_thread(machine, pid, tid);
1537 union map_priv priv = {
1538 .color = false,
1539 };
1540
1541 if (!sched->map.color_pids || !thread || thread__priv(thread))
1542 return thread;
1543
1544 if (thread_map__has(sched->map.color_pids, tid))
1545 priv.color = true;
1546
1547 thread__set_priv(thread, priv.ptr);
1548 return thread;
1549}
1550
1551static bool sched_match_task(struct perf_sched *sched, const char *comm_str)
1552{
1553 bool fuzzy_match = sched->map.fuzzy;
1554 struct strlist *task_names = sched->map.task_names;
1555 struct str_node *node;
1556
1557 strlist__for_each_entry(node, task_names) {
1558 bool match_found = fuzzy_match ? !!strstr(comm_str, node->s) :
1559 !strcmp(comm_str, node->s);
1560 if (match_found)
1561 return true;
1562 }
1563
1564 return false;
1565}
1566
1567static void print_sched_map(struct perf_sched *sched, struct perf_cpu this_cpu, int cpus_nr,
1568 const char *color, bool sched_out)
1569{
1570 for (int i = 0; i < cpus_nr; i++) {
1571 struct perf_cpu cpu = {
1572 .cpu = sched->map.comp ? sched->map.comp_cpus[i].cpu : i,
1573 };
1574 struct thread *curr_thread = sched->curr_thread[cpu.cpu];
1575 struct thread *curr_out_thread = sched->curr_out_thread[cpu.cpu];
1576 struct thread_runtime *curr_tr;
1577 const char *pid_color = color;
1578 const char *cpu_color = color;
1579 char symbol = ' ';
1580 struct thread *thread_to_check = sched_out ? curr_out_thread : curr_thread;
1581
1582 if (thread_to_check && thread__has_color(thread_to_check))
1583 pid_color = COLOR_PIDS;
1584
1585 if (sched->map.color_cpus && perf_cpu_map__has(sched->map.color_cpus, cpu))
1586 cpu_color = COLOR_CPUS;
1587
1588 if (cpu.cpu == this_cpu.cpu)
1589 symbol = '*';
1590
1591 color_fprintf(stdout, cpu.cpu != this_cpu.cpu ? color : cpu_color, "%c", symbol);
1592
1593 thread_to_check = sched_out ? sched->curr_out_thread[cpu.cpu] :
1594 sched->curr_thread[cpu.cpu];
1595
1596 if (thread_to_check) {
1597 curr_tr = thread__get_runtime(thread_to_check);
1598 if (curr_tr == NULL)
1599 return;
1600
1601 if (sched_out) {
1602 if (cpu.cpu == this_cpu.cpu)
1603 color_fprintf(stdout, color, "- ");
1604 else {
1605 curr_tr = thread__get_runtime(sched->curr_thread[cpu.cpu]);
1606 if (curr_tr != NULL)
1607 color_fprintf(stdout, pid_color, "%2s ",
1608 curr_tr->shortname);
1609 }
1610 } else
1611 color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1612 } else
1613 color_fprintf(stdout, color, " ");
1614 }
1615}
1616
1617static int map_switch_event(struct perf_sched *sched, struct evsel *evsel,
1618 struct perf_sample *sample, struct machine *machine)
1619{
1620 const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
1621 const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid");
1622 struct thread *sched_in, *sched_out;
1623 struct thread_runtime *tr;
1624 int new_shortname;
1625 u64 timestamp0, timestamp = sample->time;
1626 s64 delta;
1627 struct perf_cpu this_cpu = {
1628 .cpu = sample->cpu,
1629 };
1630 int cpus_nr;
1631 int proceed;
1632 bool new_cpu = false;
1633 const char *color = PERF_COLOR_NORMAL;
1634 char stimestamp[32];
1635 const char *str;
1636
1637 BUG_ON(this_cpu.cpu >= MAX_CPUS || this_cpu.cpu < 0);
1638
1639 if (this_cpu.cpu > sched->max_cpu.cpu)
1640 sched->max_cpu = this_cpu;
1641
1642 if (sched->map.comp) {
1643 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1644 if (!__test_and_set_bit(this_cpu.cpu, sched->map.comp_cpus_mask)) {
1645 sched->map.comp_cpus[cpus_nr++] = this_cpu;
1646 new_cpu = true;
1647 }
1648 } else
1649 cpus_nr = sched->max_cpu.cpu;
1650
1651 timestamp0 = sched->cpu_last_switched[this_cpu.cpu];
1652 sched->cpu_last_switched[this_cpu.cpu] = timestamp;
1653 if (timestamp0)
1654 delta = timestamp - timestamp0;
1655 else
1656 delta = 0;
1657
1658 if (delta < 0) {
1659 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1660 return -1;
1661 }
1662
1663 sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1664 sched_out = map__findnew_thread(sched, machine, -1, prev_pid);
1665 if (sched_in == NULL || sched_out == NULL)
1666 return -1;
1667
1668 tr = thread__get_runtime(sched_in);
1669 if (tr == NULL) {
1670 thread__put(sched_in);
1671 return -1;
1672 }
1673
1674 sched->curr_thread[this_cpu.cpu] = thread__get(sched_in);
1675 sched->curr_out_thread[this_cpu.cpu] = thread__get(sched_out);
1676
1677 str = thread__comm_str(sched_in);
1678 new_shortname = 0;
1679 if (!tr->shortname[0]) {
1680 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1681 /*
1682 * Don't allocate a letter-number for swapper:0
1683 * as a shortname. Instead, we use '.' for it.
1684 */
1685 tr->shortname[0] = '.';
1686 tr->shortname[1] = ' ';
1687 } else if (!sched->map.task_name || sched_match_task(sched, str)) {
1688 tr->shortname[0] = sched->next_shortname1;
1689 tr->shortname[1] = sched->next_shortname2;
1690
1691 if (sched->next_shortname1 < 'Z') {
1692 sched->next_shortname1++;
1693 } else {
1694 sched->next_shortname1 = 'A';
1695 if (sched->next_shortname2 < '9')
1696 sched->next_shortname2++;
1697 else
1698 sched->next_shortname2 = '0';
1699 }
1700 } else {
1701 tr->shortname[0] = '-';
1702 tr->shortname[1] = ' ';
1703 }
1704 new_shortname = 1;
1705 }
1706
1707 if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, this_cpu))
1708 goto out;
1709
1710 proceed = 0;
1711 str = thread__comm_str(sched_in);
1712 /*
1713 * Check which of sched_in and sched_out matches the passed --task-name
1714 * arguments and call the corresponding print_sched_map.
1715 */
1716 if (sched->map.task_name && !sched_match_task(sched, str)) {
1717 if (!sched_match_task(sched, thread__comm_str(sched_out)))
1718 goto out;
1719 else
1720 goto sched_out;
1721
1722 } else {
1723 str = thread__comm_str(sched_out);
1724 if (!(sched->map.task_name && !sched_match_task(sched, str)))
1725 proceed = 1;
1726 }
1727
1728 printf(" ");
1729
1730 print_sched_map(sched, this_cpu, cpus_nr, color, false);
1731
1732 timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1733 color_fprintf(stdout, color, " %12s secs ", stimestamp);
1734 if (new_shortname || tr->comm_changed || (verbose > 0 && thread__tid(sched_in))) {
1735 const char *pid_color = color;
1736
1737 if (thread__has_color(sched_in))
1738 pid_color = COLOR_PIDS;
1739
1740 color_fprintf(stdout, pid_color, "%s => %s:%d",
1741 tr->shortname, thread__comm_str(sched_in), thread__tid(sched_in));
1742 tr->comm_changed = false;
1743 }
1744
1745 if (sched->map.comp && new_cpu)
1746 color_fprintf(stdout, color, " (CPU %d)", this_cpu.cpu);
1747
1748 if (proceed != 1) {
1749 color_fprintf(stdout, color, "\n");
1750 goto out;
1751 }
1752
1753sched_out:
1754 if (sched->map.task_name) {
1755 tr = thread__get_runtime(sched->curr_out_thread[this_cpu.cpu]);
1756 if (strcmp(tr->shortname, "") == 0)
1757 goto out;
1758
1759 if (proceed == 1)
1760 color_fprintf(stdout, color, "\n");
1761
1762 printf(" ");
1763 print_sched_map(sched, this_cpu, cpus_nr, color, true);
1764 timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1765 color_fprintf(stdout, color, " %12s secs ", stimestamp);
1766 }
1767
1768 color_fprintf(stdout, color, "\n");
1769
1770out:
1771 if (sched->map.task_name)
1772 thread__put(sched_out);
1773
1774 thread__put(sched_in);
1775
1776 return 0;
1777}
1778
1779static int process_sched_switch_event(const struct perf_tool *tool,
1780 struct evsel *evsel,
1781 struct perf_sample *sample,
1782 struct machine *machine)
1783{
1784 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1785 int this_cpu = sample->cpu, err = 0;
1786 u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1787 next_pid = evsel__intval(evsel, sample, "next_pid");
1788
1789 if (sched->curr_pid[this_cpu] != (u32)-1) {
1790 /*
1791 * Are we trying to switch away a PID that is
1792 * not current?
1793 */
1794 if (sched->curr_pid[this_cpu] != prev_pid)
1795 sched->nr_context_switch_bugs++;
1796 }
1797
1798 if (sched->tp_handler->switch_event)
1799 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1800
1801 sched->curr_pid[this_cpu] = next_pid;
1802 return err;
1803}
1804
1805static int process_sched_runtime_event(const struct perf_tool *tool,
1806 struct evsel *evsel,
1807 struct perf_sample *sample,
1808 struct machine *machine)
1809{
1810 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1811
1812 if (sched->tp_handler->runtime_event)
1813 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1814
1815 return 0;
1816}
1817
1818static int perf_sched__process_fork_event(const struct perf_tool *tool,
1819 union perf_event *event,
1820 struct perf_sample *sample,
1821 struct machine *machine)
1822{
1823 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1824
1825 /* run the fork event through the perf machinery */
1826 perf_event__process_fork(tool, event, sample, machine);
1827
1828 /* and then run additional processing needed for this command */
1829 if (sched->tp_handler->fork_event)
1830 return sched->tp_handler->fork_event(sched, event, machine);
1831
1832 return 0;
1833}
1834
1835static int process_sched_migrate_task_event(const struct perf_tool *tool,
1836 struct evsel *evsel,
1837 struct perf_sample *sample,
1838 struct machine *machine)
1839{
1840 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1841
1842 if (sched->tp_handler->migrate_task_event)
1843 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1844
1845 return 0;
1846}
1847
1848typedef int (*tracepoint_handler)(const struct perf_tool *tool,
1849 struct evsel *evsel,
1850 struct perf_sample *sample,
1851 struct machine *machine);
1852
1853static int perf_sched__process_tracepoint_sample(const struct perf_tool *tool __maybe_unused,
1854 union perf_event *event __maybe_unused,
1855 struct perf_sample *sample,
1856 struct evsel *evsel,
1857 struct machine *machine)
1858{
1859 int err = 0;
1860
1861 if (evsel->handler != NULL) {
1862 tracepoint_handler f = evsel->handler;
1863 err = f(tool, evsel, sample, machine);
1864 }
1865
1866 return err;
1867}
1868
1869static int perf_sched__process_comm(const struct perf_tool *tool __maybe_unused,
1870 union perf_event *event,
1871 struct perf_sample *sample,
1872 struct machine *machine)
1873{
1874 struct thread *thread;
1875 struct thread_runtime *tr;
1876 int err;
1877
1878 err = perf_event__process_comm(tool, event, sample, machine);
1879 if (err)
1880 return err;
1881
1882 thread = machine__find_thread(machine, sample->pid, sample->tid);
1883 if (!thread) {
1884 pr_err("Internal error: can't find thread\n");
1885 return -1;
1886 }
1887
1888 tr = thread__get_runtime(thread);
1889 if (tr == NULL) {
1890 thread__put(thread);
1891 return -1;
1892 }
1893
1894 tr->comm_changed = true;
1895 thread__put(thread);
1896
1897 return 0;
1898}
1899
1900static int perf_sched__read_events(struct perf_sched *sched)
1901{
1902 struct evsel_str_handler handlers[] = {
1903 { "sched:sched_switch", process_sched_switch_event, },
1904 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1905 { "sched:sched_wakeup", process_sched_wakeup_event, },
1906 { "sched:sched_waking", process_sched_wakeup_event, },
1907 { "sched:sched_wakeup_new", process_sched_wakeup_event, },
1908 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1909 };
1910 struct perf_session *session;
1911 struct perf_data data = {
1912 .path = input_name,
1913 .mode = PERF_DATA_MODE_READ,
1914 .force = sched->force,
1915 };
1916 int rc = -1;
1917
1918 session = perf_session__new(&data, &sched->tool);
1919 if (IS_ERR(session)) {
1920 pr_debug("Error creating perf session");
1921 return PTR_ERR(session);
1922 }
1923
1924 symbol__init(&session->header.env);
1925
1926 /* prefer sched_waking if it is captured */
1927 if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
1928 handlers[2].handler = process_sched_wakeup_ignore;
1929
1930 if (perf_session__set_tracepoints_handlers(session, handlers))
1931 goto out_delete;
1932
1933 if (perf_session__has_traces(session, "record -R")) {
1934 int err = perf_session__process_events(session);
1935 if (err) {
1936 pr_err("Failed to process events, error %d", err);
1937 goto out_delete;
1938 }
1939
1940 sched->nr_events = session->evlist->stats.nr_events[0];
1941 sched->nr_lost_events = session->evlist->stats.total_lost;
1942 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1943 }
1944
1945 rc = 0;
1946out_delete:
1947 perf_session__delete(session);
1948 return rc;
1949}
1950
1951/*
1952 * scheduling times are printed as msec.usec
1953 */
1954static inline void print_sched_time(unsigned long long nsecs, int width)
1955{
1956 unsigned long msecs;
1957 unsigned long usecs;
1958
1959 msecs = nsecs / NSEC_PER_MSEC;
1960 nsecs -= msecs * NSEC_PER_MSEC;
1961 usecs = nsecs / NSEC_PER_USEC;
1962 printf("%*lu.%03lu ", width, msecs, usecs);
1963}
1964
1965/*
1966 * returns runtime data for event, allocating memory for it the
1967 * first time it is used.
1968 */
1969static struct evsel_runtime *evsel__get_runtime(struct evsel *evsel)
1970{
1971 struct evsel_runtime *r = evsel->priv;
1972
1973 if (r == NULL) {
1974 r = zalloc(sizeof(struct evsel_runtime));
1975 evsel->priv = r;
1976 }
1977
1978 return r;
1979}
1980
1981/*
1982 * save last time event was seen per cpu
1983 */
1984static void evsel__save_time(struct evsel *evsel, u64 timestamp, u32 cpu)
1985{
1986 struct evsel_runtime *r = evsel__get_runtime(evsel);
1987
1988 if (r == NULL)
1989 return;
1990
1991 if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1992 int i, n = __roundup_pow_of_two(cpu+1);
1993 void *p = r->last_time;
1994
1995 p = realloc(r->last_time, n * sizeof(u64));
1996 if (!p)
1997 return;
1998
1999 r->last_time = p;
2000 for (i = r->ncpu; i < n; ++i)
2001 r->last_time[i] = (u64) 0;
2002
2003 r->ncpu = n;
2004 }
2005
2006 r->last_time[cpu] = timestamp;
2007}
2008
2009/* returns last time this event was seen on the given cpu */
2010static u64 evsel__get_time(struct evsel *evsel, u32 cpu)
2011{
2012 struct evsel_runtime *r = evsel__get_runtime(evsel);
2013
2014 if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
2015 return 0;
2016
2017 return r->last_time[cpu];
2018}
2019
2020static int comm_width = 30;
2021
2022static char *timehist_get_commstr(struct thread *thread)
2023{
2024 static char str[32];
2025 const char *comm = thread__comm_str(thread);
2026 pid_t tid = thread__tid(thread);
2027 pid_t pid = thread__pid(thread);
2028 int n;
2029
2030 if (pid == 0)
2031 n = scnprintf(str, sizeof(str), "%s", comm);
2032
2033 else if (tid != pid)
2034 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
2035
2036 else
2037 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
2038
2039 if (n > comm_width)
2040 comm_width = n;
2041
2042 return str;
2043}
2044
2045/* prio field format: xxx or xxx->yyy */
2046#define MAX_PRIO_STR_LEN 8
2047static char *timehist_get_priostr(struct evsel *evsel,
2048 struct thread *thread,
2049 struct perf_sample *sample)
2050{
2051 static char prio_str[16];
2052 int prev_prio = (int)evsel__intval(evsel, sample, "prev_prio");
2053 struct thread_runtime *tr = thread__priv(thread);
2054
2055 if (tr->prio != prev_prio && tr->prio != -1)
2056 scnprintf(prio_str, sizeof(prio_str), "%d->%d", tr->prio, prev_prio);
2057 else
2058 scnprintf(prio_str, sizeof(prio_str), "%d", prev_prio);
2059
2060 return prio_str;
2061}
2062
2063static void timehist_header(struct perf_sched *sched)
2064{
2065 u32 ncpus = sched->max_cpu.cpu + 1;
2066 u32 i, j;
2067
2068 printf("%15s %6s ", "time", "cpu");
2069
2070 if (sched->show_cpu_visual) {
2071 printf(" ");
2072 for (i = 0, j = 0; i < ncpus; ++i) {
2073 printf("%x", j++);
2074 if (j > 15)
2075 j = 0;
2076 }
2077 printf(" ");
2078 }
2079
2080 printf(" %-*s", comm_width, "task name");
2081
2082 if (sched->show_prio)
2083 printf(" %-*s", MAX_PRIO_STR_LEN, "prio");
2084
2085 printf(" %9s %9s %9s", "wait time", "sch delay", "run time");
2086
2087 if (sched->pre_migrations)
2088 printf(" %9s", "pre-mig time");
2089
2090 if (sched->show_state)
2091 printf(" %s", "state");
2092
2093 printf("\n");
2094
2095 /*
2096 * units row
2097 */
2098 printf("%15s %-6s ", "", "");
2099
2100 if (sched->show_cpu_visual)
2101 printf(" %*s ", ncpus, "");
2102
2103 printf(" %-*s", comm_width, "[tid/pid]");
2104
2105 if (sched->show_prio)
2106 printf(" %-*s", MAX_PRIO_STR_LEN, "");
2107
2108 printf(" %9s %9s %9s", "(msec)", "(msec)", "(msec)");
2109
2110 if (sched->pre_migrations)
2111 printf(" %9s", "(msec)");
2112
2113 printf("\n");
2114
2115 /*
2116 * separator
2117 */
2118 printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
2119
2120 if (sched->show_cpu_visual)
2121 printf(" %.*s ", ncpus, graph_dotted_line);
2122
2123 printf(" %.*s", comm_width, graph_dotted_line);
2124
2125 if (sched->show_prio)
2126 printf(" %.*s", MAX_PRIO_STR_LEN, graph_dotted_line);
2127
2128 printf(" %.9s %.9s %.9s", graph_dotted_line, graph_dotted_line, graph_dotted_line);
2129
2130 if (sched->pre_migrations)
2131 printf(" %.9s", graph_dotted_line);
2132
2133 if (sched->show_state)
2134 printf(" %.5s", graph_dotted_line);
2135
2136 printf("\n");
2137}
2138
2139static void timehist_print_sample(struct perf_sched *sched,
2140 struct evsel *evsel,
2141 struct perf_sample *sample,
2142 struct addr_location *al,
2143 struct thread *thread,
2144 u64 t, const char state)
2145{
2146 struct thread_runtime *tr = thread__priv(thread);
2147 const char *next_comm = evsel__strval(evsel, sample, "next_comm");
2148 const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
2149 u32 max_cpus = sched->max_cpu.cpu + 1;
2150 char tstr[64];
2151 char nstr[30];
2152 u64 wait_time;
2153
2154 if (cpu_list && !test_bit(sample->cpu, cpu_bitmap))
2155 return;
2156
2157 timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2158 printf("%15s [%04d] ", tstr, sample->cpu);
2159
2160 if (sched->show_cpu_visual) {
2161 u32 i;
2162 char c;
2163
2164 printf(" ");
2165 for (i = 0; i < max_cpus; ++i) {
2166 /* flag idle times with 'i'; others are sched events */
2167 if (i == sample->cpu)
2168 c = (thread__tid(thread) == 0) ? 'i' : 's';
2169 else
2170 c = ' ';
2171 printf("%c", c);
2172 }
2173 printf(" ");
2174 }
2175
2176 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2177
2178 if (sched->show_prio)
2179 printf(" %-*s ", MAX_PRIO_STR_LEN, timehist_get_priostr(evsel, thread, sample));
2180
2181 wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2182 print_sched_time(wait_time, 6);
2183
2184 print_sched_time(tr->dt_delay, 6);
2185 print_sched_time(tr->dt_run, 6);
2186 if (sched->pre_migrations)
2187 print_sched_time(tr->dt_pre_mig, 6);
2188
2189 if (sched->show_state)
2190 printf(" %5c ", thread__tid(thread) == 0 ? 'I' : state);
2191
2192 if (sched->show_next) {
2193 snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2194 printf(" %-*s", comm_width, nstr);
2195 }
2196
2197 if (sched->show_wakeups && !sched->show_next)
2198 printf(" %-*s", comm_width, "");
2199
2200 if (thread__tid(thread) == 0)
2201 goto out;
2202
2203 if (sched->show_callchain)
2204 printf(" ");
2205
2206 sample__fprintf_sym(sample, al, 0,
2207 EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2208 EVSEL__PRINT_CALLCHAIN_ARROW |
2209 EVSEL__PRINT_SKIP_IGNORED,
2210 get_tls_callchain_cursor(), symbol_conf.bt_stop_list, stdout);
2211
2212out:
2213 printf("\n");
2214}
2215
2216/*
2217 * Explanation of delta-time stats:
2218 *
2219 * t = time of current schedule out event
2220 * tprev = time of previous sched out event
2221 * also time of schedule-in event for current task
2222 * last_time = time of last sched change event for current task
2223 * (i.e, time process was last scheduled out)
2224 * ready_to_run = time of wakeup for current task
2225 * migrated = time of task migration to another CPU
2226 *
2227 * -----|-------------|-------------|-------------|-------------|-----
2228 * last ready migrated tprev t
2229 * time to run
2230 *
2231 * |---------------- dt_wait ----------------|
2232 * |--------- dt_delay ---------|-- dt_run --|
2233 * |- dt_pre_mig -|
2234 *
2235 * dt_run = run time of current task
2236 * dt_wait = time between last schedule out event for task and tprev
2237 * represents time spent off the cpu
2238 * dt_delay = time between wakeup and schedule-in of task
2239 * dt_pre_mig = time between wakeup and migration to another CPU
2240 */
2241
2242static void timehist_update_runtime_stats(struct thread_runtime *r,
2243 u64 t, u64 tprev)
2244{
2245 r->dt_delay = 0;
2246 r->dt_sleep = 0;
2247 r->dt_iowait = 0;
2248 r->dt_preempt = 0;
2249 r->dt_run = 0;
2250 r->dt_pre_mig = 0;
2251
2252 if (tprev) {
2253 r->dt_run = t - tprev;
2254 if (r->ready_to_run) {
2255 if (r->ready_to_run > tprev)
2256 pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2257 else
2258 r->dt_delay = tprev - r->ready_to_run;
2259
2260 if ((r->migrated > r->ready_to_run) && (r->migrated < tprev))
2261 r->dt_pre_mig = r->migrated - r->ready_to_run;
2262 }
2263
2264 if (r->last_time > tprev)
2265 pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2266 else if (r->last_time) {
2267 u64 dt_wait = tprev - r->last_time;
2268
2269 if (r->last_state == 'R')
2270 r->dt_preempt = dt_wait;
2271 else if (r->last_state == 'D')
2272 r->dt_iowait = dt_wait;
2273 else
2274 r->dt_sleep = dt_wait;
2275 }
2276 }
2277
2278 update_stats(&r->run_stats, r->dt_run);
2279
2280 r->total_run_time += r->dt_run;
2281 r->total_delay_time += r->dt_delay;
2282 r->total_sleep_time += r->dt_sleep;
2283 r->total_iowait_time += r->dt_iowait;
2284 r->total_preempt_time += r->dt_preempt;
2285 r->total_pre_mig_time += r->dt_pre_mig;
2286}
2287
2288static bool is_idle_sample(struct perf_sample *sample,
2289 struct evsel *evsel)
2290{
2291 /* pid 0 == swapper == idle task */
2292 if (evsel__name_is(evsel, "sched:sched_switch"))
2293 return evsel__intval(evsel, sample, "prev_pid") == 0;
2294
2295 return sample->pid == 0;
2296}
2297
2298static void save_task_callchain(struct perf_sched *sched,
2299 struct perf_sample *sample,
2300 struct evsel *evsel,
2301 struct machine *machine)
2302{
2303 struct callchain_cursor *cursor;
2304 struct thread *thread;
2305
2306 /* want main thread for process - has maps */
2307 thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2308 if (thread == NULL) {
2309 pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2310 return;
2311 }
2312
2313 if (!sched->show_callchain || sample->callchain == NULL)
2314 return;
2315
2316 cursor = get_tls_callchain_cursor();
2317
2318 if (thread__resolve_callchain(thread, cursor, evsel, sample,
2319 NULL, NULL, sched->max_stack + 2) != 0) {
2320 if (verbose > 0)
2321 pr_err("Failed to resolve callchain. Skipping\n");
2322
2323 return;
2324 }
2325
2326 callchain_cursor_commit(cursor);
2327
2328 while (true) {
2329 struct callchain_cursor_node *node;
2330 struct symbol *sym;
2331
2332 node = callchain_cursor_current(cursor);
2333 if (node == NULL)
2334 break;
2335
2336 sym = node->ms.sym;
2337 if (sym) {
2338 if (!strcmp(sym->name, "schedule") ||
2339 !strcmp(sym->name, "__schedule") ||
2340 !strcmp(sym->name, "preempt_schedule"))
2341 sym->ignore = 1;
2342 }
2343
2344 callchain_cursor_advance(cursor);
2345 }
2346}
2347
2348static int init_idle_thread(struct thread *thread)
2349{
2350 struct idle_thread_runtime *itr;
2351
2352 thread__set_comm(thread, idle_comm, 0);
2353
2354 itr = zalloc(sizeof(*itr));
2355 if (itr == NULL)
2356 return -ENOMEM;
2357
2358 init_prio(&itr->tr);
2359 init_stats(&itr->tr.run_stats);
2360 callchain_init(&itr->callchain);
2361 callchain_cursor_reset(&itr->cursor);
2362 thread__set_priv(thread, itr);
2363
2364 return 0;
2365}
2366
2367/*
2368 * Track idle stats per cpu by maintaining a local thread
2369 * struct for the idle task on each cpu.
2370 */
2371static int init_idle_threads(int ncpu)
2372{
2373 int i, ret;
2374
2375 idle_threads = zalloc(ncpu * sizeof(struct thread *));
2376 if (!idle_threads)
2377 return -ENOMEM;
2378
2379 idle_max_cpu = ncpu;
2380
2381 /* allocate the actual thread struct if needed */
2382 for (i = 0; i < ncpu; ++i) {
2383 idle_threads[i] = thread__new(0, 0);
2384 if (idle_threads[i] == NULL)
2385 return -ENOMEM;
2386
2387 ret = init_idle_thread(idle_threads[i]);
2388 if (ret < 0)
2389 return ret;
2390 }
2391
2392 return 0;
2393}
2394
2395static void free_idle_threads(void)
2396{
2397 int i;
2398
2399 if (idle_threads == NULL)
2400 return;
2401
2402 for (i = 0; i < idle_max_cpu; ++i) {
2403 if ((idle_threads[i]))
2404 thread__delete(idle_threads[i]);
2405 }
2406
2407 free(idle_threads);
2408}
2409
2410static struct thread *get_idle_thread(int cpu)
2411{
2412 /*
2413 * expand/allocate array of pointers to local thread
2414 * structs if needed
2415 */
2416 if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2417 int i, j = __roundup_pow_of_two(cpu+1);
2418 void *p;
2419
2420 p = realloc(idle_threads, j * sizeof(struct thread *));
2421 if (!p)
2422 return NULL;
2423
2424 idle_threads = (struct thread **) p;
2425 for (i = idle_max_cpu; i < j; ++i)
2426 idle_threads[i] = NULL;
2427
2428 idle_max_cpu = j;
2429 }
2430
2431 /* allocate a new thread struct if needed */
2432 if (idle_threads[cpu] == NULL) {
2433 idle_threads[cpu] = thread__new(0, 0);
2434
2435 if (idle_threads[cpu]) {
2436 if (init_idle_thread(idle_threads[cpu]) < 0)
2437 return NULL;
2438 }
2439 }
2440
2441 return idle_threads[cpu];
2442}
2443
2444static void save_idle_callchain(struct perf_sched *sched,
2445 struct idle_thread_runtime *itr,
2446 struct perf_sample *sample)
2447{
2448 struct callchain_cursor *cursor;
2449
2450 if (!sched->show_callchain || sample->callchain == NULL)
2451 return;
2452
2453 cursor = get_tls_callchain_cursor();
2454 if (cursor == NULL)
2455 return;
2456
2457 callchain_cursor__copy(&itr->cursor, cursor);
2458}
2459
2460static struct thread *timehist_get_thread(struct perf_sched *sched,
2461 struct perf_sample *sample,
2462 struct machine *machine,
2463 struct evsel *evsel)
2464{
2465 struct thread *thread;
2466
2467 if (is_idle_sample(sample, evsel)) {
2468 thread = get_idle_thread(sample->cpu);
2469 if (thread == NULL)
2470 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2471
2472 } else {
2473 /* there were samples with tid 0 but non-zero pid */
2474 thread = machine__findnew_thread(machine, sample->pid,
2475 sample->tid ?: sample->pid);
2476 if (thread == NULL) {
2477 pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2478 sample->tid);
2479 }
2480
2481 save_task_callchain(sched, sample, evsel, machine);
2482 if (sched->idle_hist) {
2483 struct thread *idle;
2484 struct idle_thread_runtime *itr;
2485
2486 idle = get_idle_thread(sample->cpu);
2487 if (idle == NULL) {
2488 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2489 return NULL;
2490 }
2491
2492 itr = thread__priv(idle);
2493 if (itr == NULL)
2494 return NULL;
2495
2496 itr->last_thread = thread;
2497
2498 /* copy task callchain when entering to idle */
2499 if (evsel__intval(evsel, sample, "next_pid") == 0)
2500 save_idle_callchain(sched, itr, sample);
2501 }
2502 }
2503
2504 return thread;
2505}
2506
2507static bool timehist_skip_sample(struct perf_sched *sched,
2508 struct thread *thread,
2509 struct evsel *evsel,
2510 struct perf_sample *sample)
2511{
2512 bool rc = false;
2513 int prio = -1;
2514 struct thread_runtime *tr = NULL;
2515
2516 if (thread__is_filtered(thread)) {
2517 rc = true;
2518 sched->skipped_samples++;
2519 }
2520
2521 if (sched->prio_str) {
2522 /*
2523 * Because priority may be changed during task execution,
2524 * first read priority from prev sched_in event for current task.
2525 * If prev sched_in event is not saved, then read priority from
2526 * current task sched_out event.
2527 */
2528 tr = thread__get_runtime(thread);
2529 if (tr && tr->prio != -1)
2530 prio = tr->prio;
2531 else if (evsel__name_is(evsel, "sched:sched_switch"))
2532 prio = evsel__intval(evsel, sample, "prev_prio");
2533
2534 if (prio != -1 && !test_bit(prio, sched->prio_bitmap)) {
2535 rc = true;
2536 sched->skipped_samples++;
2537 }
2538 }
2539
2540 if (sched->idle_hist) {
2541 if (!evsel__name_is(evsel, "sched:sched_switch"))
2542 rc = true;
2543 else if (evsel__intval(evsel, sample, "prev_pid") != 0 &&
2544 evsel__intval(evsel, sample, "next_pid") != 0)
2545 rc = true;
2546 }
2547
2548 return rc;
2549}
2550
2551static void timehist_print_wakeup_event(struct perf_sched *sched,
2552 struct evsel *evsel,
2553 struct perf_sample *sample,
2554 struct machine *machine,
2555 struct thread *awakened)
2556{
2557 struct thread *thread;
2558 char tstr[64];
2559
2560 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2561 if (thread == NULL)
2562 return;
2563
2564 /* show wakeup unless both awakee and awaker are filtered */
2565 if (timehist_skip_sample(sched, thread, evsel, sample) &&
2566 timehist_skip_sample(sched, awakened, evsel, sample)) {
2567 return;
2568 }
2569
2570 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2571 printf("%15s [%04d] ", tstr, sample->cpu);
2572 if (sched->show_cpu_visual)
2573 printf(" %*s ", sched->max_cpu.cpu + 1, "");
2574
2575 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2576
2577 /* dt spacer */
2578 printf(" %9s %9s %9s ", "", "", "");
2579
2580 printf("awakened: %s", timehist_get_commstr(awakened));
2581
2582 printf("\n");
2583}
2584
2585static int timehist_sched_wakeup_ignore(const struct perf_tool *tool __maybe_unused,
2586 union perf_event *event __maybe_unused,
2587 struct evsel *evsel __maybe_unused,
2588 struct perf_sample *sample __maybe_unused,
2589 struct machine *machine __maybe_unused)
2590{
2591 return 0;
2592}
2593
2594static int timehist_sched_wakeup_event(const struct perf_tool *tool,
2595 union perf_event *event __maybe_unused,
2596 struct evsel *evsel,
2597 struct perf_sample *sample,
2598 struct machine *machine)
2599{
2600 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2601 struct thread *thread;
2602 struct thread_runtime *tr = NULL;
2603 /* want pid of awakened task not pid in sample */
2604 const u32 pid = evsel__intval(evsel, sample, "pid");
2605
2606 thread = machine__findnew_thread(machine, 0, pid);
2607 if (thread == NULL)
2608 return -1;
2609
2610 tr = thread__get_runtime(thread);
2611 if (tr == NULL)
2612 return -1;
2613
2614 if (tr->ready_to_run == 0)
2615 tr->ready_to_run = sample->time;
2616
2617 /* show wakeups if requested */
2618 if (sched->show_wakeups &&
2619 !perf_time__skip_sample(&sched->ptime, sample->time))
2620 timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2621
2622 return 0;
2623}
2624
2625static void timehist_print_migration_event(struct perf_sched *sched,
2626 struct evsel *evsel,
2627 struct perf_sample *sample,
2628 struct machine *machine,
2629 struct thread *migrated)
2630{
2631 struct thread *thread;
2632 char tstr[64];
2633 u32 max_cpus;
2634 u32 ocpu, dcpu;
2635
2636 if (sched->summary_only)
2637 return;
2638
2639 max_cpus = sched->max_cpu.cpu + 1;
2640 ocpu = evsel__intval(evsel, sample, "orig_cpu");
2641 dcpu = evsel__intval(evsel, sample, "dest_cpu");
2642
2643 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2644 if (thread == NULL)
2645 return;
2646
2647 if (timehist_skip_sample(sched, thread, evsel, sample) &&
2648 timehist_skip_sample(sched, migrated, evsel, sample)) {
2649 return;
2650 }
2651
2652 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2653 printf("%15s [%04d] ", tstr, sample->cpu);
2654
2655 if (sched->show_cpu_visual) {
2656 u32 i;
2657 char c;
2658
2659 printf(" ");
2660 for (i = 0; i < max_cpus; ++i) {
2661 c = (i == sample->cpu) ? 'm' : ' ';
2662 printf("%c", c);
2663 }
2664 printf(" ");
2665 }
2666
2667 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2668
2669 /* dt spacer */
2670 printf(" %9s %9s %9s ", "", "", "");
2671
2672 printf("migrated: %s", timehist_get_commstr(migrated));
2673 printf(" cpu %d => %d", ocpu, dcpu);
2674
2675 printf("\n");
2676}
2677
2678static int timehist_migrate_task_event(const struct perf_tool *tool,
2679 union perf_event *event __maybe_unused,
2680 struct evsel *evsel,
2681 struct perf_sample *sample,
2682 struct machine *machine)
2683{
2684 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2685 struct thread *thread;
2686 struct thread_runtime *tr = NULL;
2687 /* want pid of migrated task not pid in sample */
2688 const u32 pid = evsel__intval(evsel, sample, "pid");
2689
2690 thread = machine__findnew_thread(machine, 0, pid);
2691 if (thread == NULL)
2692 return -1;
2693
2694 tr = thread__get_runtime(thread);
2695 if (tr == NULL)
2696 return -1;
2697
2698 tr->migrations++;
2699 tr->migrated = sample->time;
2700
2701 /* show migrations if requested */
2702 if (sched->show_migrations) {
2703 timehist_print_migration_event(sched, evsel, sample,
2704 machine, thread);
2705 }
2706
2707 return 0;
2708}
2709
2710static void timehist_update_task_prio(struct evsel *evsel,
2711 struct perf_sample *sample,
2712 struct machine *machine)
2713{
2714 struct thread *thread;
2715 struct thread_runtime *tr = NULL;
2716 const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
2717 const u32 next_prio = evsel__intval(evsel, sample, "next_prio");
2718
2719 if (next_pid == 0)
2720 thread = get_idle_thread(sample->cpu);
2721 else
2722 thread = machine__findnew_thread(machine, -1, next_pid);
2723
2724 if (thread == NULL)
2725 return;
2726
2727 tr = thread__get_runtime(thread);
2728 if (tr == NULL)
2729 return;
2730
2731 tr->prio = next_prio;
2732}
2733
2734static int timehist_sched_change_event(const struct perf_tool *tool,
2735 union perf_event *event,
2736 struct evsel *evsel,
2737 struct perf_sample *sample,
2738 struct machine *machine)
2739{
2740 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2741 struct perf_time_interval *ptime = &sched->ptime;
2742 struct addr_location al;
2743 struct thread *thread;
2744 struct thread_runtime *tr = NULL;
2745 u64 tprev, t = sample->time;
2746 int rc = 0;
2747 const char state = evsel__taskstate(evsel, sample, "prev_state");
2748
2749 addr_location__init(&al);
2750 if (machine__resolve(machine, &al, sample) < 0) {
2751 pr_err("problem processing %d event. skipping it\n",
2752 event->header.type);
2753 rc = -1;
2754 goto out;
2755 }
2756
2757 if (sched->show_prio || sched->prio_str)
2758 timehist_update_task_prio(evsel, sample, machine);
2759
2760 thread = timehist_get_thread(sched, sample, machine, evsel);
2761 if (thread == NULL) {
2762 rc = -1;
2763 goto out;
2764 }
2765
2766 if (timehist_skip_sample(sched, thread, evsel, sample))
2767 goto out;
2768
2769 tr = thread__get_runtime(thread);
2770 if (tr == NULL) {
2771 rc = -1;
2772 goto out;
2773 }
2774
2775 tprev = evsel__get_time(evsel, sample->cpu);
2776
2777 /*
2778 * If start time given:
2779 * - sample time is under window user cares about - skip sample
2780 * - tprev is under window user cares about - reset to start of window
2781 */
2782 if (ptime->start && ptime->start > t)
2783 goto out;
2784
2785 if (tprev && ptime->start > tprev)
2786 tprev = ptime->start;
2787
2788 /*
2789 * If end time given:
2790 * - previous sched event is out of window - we are done
2791 * - sample time is beyond window user cares about - reset it
2792 * to close out stats for time window interest
2793 * - If tprev is 0, that is, sched_in event for current task is
2794 * not recorded, cannot determine whether sched_in event is
2795 * within time window interest - ignore it
2796 */
2797 if (ptime->end) {
2798 if (!tprev || tprev > ptime->end)
2799 goto out;
2800
2801 if (t > ptime->end)
2802 t = ptime->end;
2803 }
2804
2805 if (!sched->idle_hist || thread__tid(thread) == 0) {
2806 if (!cpu_list || test_bit(sample->cpu, cpu_bitmap))
2807 timehist_update_runtime_stats(tr, t, tprev);
2808
2809 if (sched->idle_hist) {
2810 struct idle_thread_runtime *itr = (void *)tr;
2811 struct thread_runtime *last_tr;
2812
2813 if (itr->last_thread == NULL)
2814 goto out;
2815
2816 /* add current idle time as last thread's runtime */
2817 last_tr = thread__get_runtime(itr->last_thread);
2818 if (last_tr == NULL)
2819 goto out;
2820
2821 timehist_update_runtime_stats(last_tr, t, tprev);
2822 /*
2823 * remove delta time of last thread as it's not updated
2824 * and otherwise it will show an invalid value next
2825 * time. we only care total run time and run stat.
2826 */
2827 last_tr->dt_run = 0;
2828 last_tr->dt_delay = 0;
2829 last_tr->dt_sleep = 0;
2830 last_tr->dt_iowait = 0;
2831 last_tr->dt_preempt = 0;
2832
2833 if (itr->cursor.nr)
2834 callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2835
2836 itr->last_thread = NULL;
2837 }
2838
2839 if (!sched->summary_only)
2840 timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2841 }
2842
2843out:
2844 if (sched->hist_time.start == 0 && t >= ptime->start)
2845 sched->hist_time.start = t;
2846 if (ptime->end == 0 || t <= ptime->end)
2847 sched->hist_time.end = t;
2848
2849 if (tr) {
2850 /* time of this sched_switch event becomes last time task seen */
2851 tr->last_time = sample->time;
2852
2853 /* last state is used to determine where to account wait time */
2854 tr->last_state = state;
2855
2856 /* sched out event for task so reset ready to run time and migrated time */
2857 if (state == 'R')
2858 tr->ready_to_run = t;
2859 else
2860 tr->ready_to_run = 0;
2861
2862 tr->migrated = 0;
2863 }
2864
2865 evsel__save_time(evsel, sample->time, sample->cpu);
2866
2867 addr_location__exit(&al);
2868 return rc;
2869}
2870
2871static int timehist_sched_switch_event(const struct perf_tool *tool,
2872 union perf_event *event,
2873 struct evsel *evsel,
2874 struct perf_sample *sample,
2875 struct machine *machine __maybe_unused)
2876{
2877 return timehist_sched_change_event(tool, event, evsel, sample, machine);
2878}
2879
2880static int process_lost(const struct perf_tool *tool __maybe_unused,
2881 union perf_event *event,
2882 struct perf_sample *sample,
2883 struct machine *machine __maybe_unused)
2884{
2885 char tstr[64];
2886
2887 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2888 printf("%15s ", tstr);
2889 printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2890
2891 return 0;
2892}
2893
2894
2895static void print_thread_runtime(struct thread *t,
2896 struct thread_runtime *r)
2897{
2898 double mean = avg_stats(&r->run_stats);
2899 float stddev;
2900
2901 printf("%*s %5d %9" PRIu64 " ",
2902 comm_width, timehist_get_commstr(t), thread__ppid(t),
2903 (u64) r->run_stats.n);
2904
2905 print_sched_time(r->total_run_time, 8);
2906 stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2907 print_sched_time(r->run_stats.min, 6);
2908 printf(" ");
2909 print_sched_time((u64) mean, 6);
2910 printf(" ");
2911 print_sched_time(r->run_stats.max, 6);
2912 printf(" ");
2913 printf("%5.2f", stddev);
2914 printf(" %5" PRIu64, r->migrations);
2915 printf("\n");
2916}
2917
2918static void print_thread_waittime(struct thread *t,
2919 struct thread_runtime *r)
2920{
2921 printf("%*s %5d %9" PRIu64 " ",
2922 comm_width, timehist_get_commstr(t), thread__ppid(t),
2923 (u64) r->run_stats.n);
2924
2925 print_sched_time(r->total_run_time, 8);
2926 print_sched_time(r->total_sleep_time, 6);
2927 printf(" ");
2928 print_sched_time(r->total_iowait_time, 6);
2929 printf(" ");
2930 print_sched_time(r->total_preempt_time, 6);
2931 printf(" ");
2932 print_sched_time(r->total_delay_time, 6);
2933 printf("\n");
2934}
2935
2936struct total_run_stats {
2937 struct perf_sched *sched;
2938 u64 sched_count;
2939 u64 task_count;
2940 u64 total_run_time;
2941};
2942
2943static int show_thread_runtime(struct thread *t, void *priv)
2944{
2945 struct total_run_stats *stats = priv;
2946 struct thread_runtime *r;
2947
2948 if (thread__is_filtered(t))
2949 return 0;
2950
2951 r = thread__priv(t);
2952 if (r && r->run_stats.n) {
2953 stats->task_count++;
2954 stats->sched_count += r->run_stats.n;
2955 stats->total_run_time += r->total_run_time;
2956
2957 if (stats->sched->show_state)
2958 print_thread_waittime(t, r);
2959 else
2960 print_thread_runtime(t, r);
2961 }
2962
2963 return 0;
2964}
2965
2966static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2967{
2968 const char *sep = " <- ";
2969 struct callchain_list *chain;
2970 size_t ret = 0;
2971 char bf[1024];
2972 bool first;
2973
2974 if (node == NULL)
2975 return 0;
2976
2977 ret = callchain__fprintf_folded(fp, node->parent);
2978 first = (ret == 0);
2979
2980 list_for_each_entry(chain, &node->val, list) {
2981 if (chain->ip >= PERF_CONTEXT_MAX)
2982 continue;
2983 if (chain->ms.sym && chain->ms.sym->ignore)
2984 continue;
2985 ret += fprintf(fp, "%s%s", first ? "" : sep,
2986 callchain_list__sym_name(chain, bf, sizeof(bf),
2987 false));
2988 first = false;
2989 }
2990
2991 return ret;
2992}
2993
2994static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2995{
2996 size_t ret = 0;
2997 FILE *fp = stdout;
2998 struct callchain_node *chain;
2999 struct rb_node *rb_node = rb_first_cached(root);
3000
3001 printf(" %16s %8s %s\n", "Idle time (msec)", "Count", "Callchains");
3002 printf(" %.16s %.8s %.50s\n", graph_dotted_line, graph_dotted_line,
3003 graph_dotted_line);
3004
3005 while (rb_node) {
3006 chain = rb_entry(rb_node, struct callchain_node, rb_node);
3007 rb_node = rb_next(rb_node);
3008
3009 ret += fprintf(fp, " ");
3010 print_sched_time(chain->hit, 12);
3011 ret += 16; /* print_sched_time returns 2nd arg + 4 */
3012 ret += fprintf(fp, " %8d ", chain->count);
3013 ret += callchain__fprintf_folded(fp, chain);
3014 ret += fprintf(fp, "\n");
3015 }
3016
3017 return ret;
3018}
3019
3020static void timehist_print_summary(struct perf_sched *sched,
3021 struct perf_session *session)
3022{
3023 struct machine *m = &session->machines.host;
3024 struct total_run_stats totals;
3025 u64 task_count;
3026 struct thread *t;
3027 struct thread_runtime *r;
3028 int i;
3029 u64 hist_time = sched->hist_time.end - sched->hist_time.start;
3030
3031 memset(&totals, 0, sizeof(totals));
3032 totals.sched = sched;
3033
3034 if (sched->idle_hist) {
3035 printf("\nIdle-time summary\n");
3036 printf("%*s parent sched-out ", comm_width, "comm");
3037 printf(" idle-time min-idle avg-idle max-idle stddev migrations\n");
3038 } else if (sched->show_state) {
3039 printf("\nWait-time summary\n");
3040 printf("%*s parent sched-in ", comm_width, "comm");
3041 printf(" run-time sleep iowait preempt delay\n");
3042 } else {
3043 printf("\nRuntime summary\n");
3044 printf("%*s parent sched-in ", comm_width, "comm");
3045 printf(" run-time min-run avg-run max-run stddev migrations\n");
3046 }
3047 printf("%*s (count) ", comm_width, "");
3048 printf(" (msec) (msec) (msec) (msec) %s\n",
3049 sched->show_state ? "(msec)" : "%");
3050 printf("%.117s\n", graph_dotted_line);
3051
3052 machine__for_each_thread(m, show_thread_runtime, &totals);
3053 task_count = totals.task_count;
3054 if (!task_count)
3055 printf("<no still running tasks>\n");
3056
3057 /* CPU idle stats not tracked when samples were skipped */
3058 if (sched->skipped_samples && !sched->idle_hist)
3059 return;
3060
3061 printf("\nIdle stats:\n");
3062 for (i = 0; i < idle_max_cpu; ++i) {
3063 if (cpu_list && !test_bit(i, cpu_bitmap))
3064 continue;
3065
3066 t = idle_threads[i];
3067 if (!t)
3068 continue;
3069
3070 r = thread__priv(t);
3071 if (r && r->run_stats.n) {
3072 totals.sched_count += r->run_stats.n;
3073 printf(" CPU %2d idle for ", i);
3074 print_sched_time(r->total_run_time, 6);
3075 printf(" msec (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
3076 } else
3077 printf(" CPU %2d idle entire time window\n", i);
3078 }
3079
3080 if (sched->idle_hist && sched->show_callchain) {
3081 callchain_param.mode = CHAIN_FOLDED;
3082 callchain_param.value = CCVAL_PERIOD;
3083
3084 callchain_register_param(&callchain_param);
3085
3086 printf("\nIdle stats by callchain:\n");
3087 for (i = 0; i < idle_max_cpu; ++i) {
3088 struct idle_thread_runtime *itr;
3089
3090 t = idle_threads[i];
3091 if (!t)
3092 continue;
3093
3094 itr = thread__priv(t);
3095 if (itr == NULL)
3096 continue;
3097
3098 callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
3099 0, &callchain_param);
3100
3101 printf(" CPU %2d:", i);
3102 print_sched_time(itr->tr.total_run_time, 6);
3103 printf(" msec\n");
3104 timehist_print_idlehist_callchain(&itr->sorted_root);
3105 printf("\n");
3106 }
3107 }
3108
3109 printf("\n"
3110 " Total number of unique tasks: %" PRIu64 "\n"
3111 "Total number of context switches: %" PRIu64 "\n",
3112 totals.task_count, totals.sched_count);
3113
3114 printf(" Total run time (msec): ");
3115 print_sched_time(totals.total_run_time, 2);
3116 printf("\n");
3117
3118 printf(" Total scheduling time (msec): ");
3119 print_sched_time(hist_time, 2);
3120 printf(" (x %d)\n", sched->max_cpu.cpu);
3121}
3122
3123typedef int (*sched_handler)(const struct perf_tool *tool,
3124 union perf_event *event,
3125 struct evsel *evsel,
3126 struct perf_sample *sample,
3127 struct machine *machine);
3128
3129static int perf_timehist__process_sample(const struct perf_tool *tool,
3130 union perf_event *event,
3131 struct perf_sample *sample,
3132 struct evsel *evsel,
3133 struct machine *machine)
3134{
3135 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
3136 int err = 0;
3137 struct perf_cpu this_cpu = {
3138 .cpu = sample->cpu,
3139 };
3140
3141 if (this_cpu.cpu > sched->max_cpu.cpu)
3142 sched->max_cpu = this_cpu;
3143
3144 if (evsel->handler != NULL) {
3145 sched_handler f = evsel->handler;
3146
3147 err = f(tool, event, evsel, sample, machine);
3148 }
3149
3150 return err;
3151}
3152
3153static int timehist_check_attr(struct perf_sched *sched,
3154 struct evlist *evlist)
3155{
3156 struct evsel *evsel;
3157 struct evsel_runtime *er;
3158
3159 list_for_each_entry(evsel, &evlist->core.entries, core.node) {
3160 er = evsel__get_runtime(evsel);
3161 if (er == NULL) {
3162 pr_err("Failed to allocate memory for evsel runtime data\n");
3163 return -1;
3164 }
3165
3166 /* only need to save callchain related to sched_switch event */
3167 if (sched->show_callchain &&
3168 evsel__name_is(evsel, "sched:sched_switch") &&
3169 !evsel__has_callchain(evsel)) {
3170 pr_info("Samples of sched_switch event do not have callchains.\n");
3171 sched->show_callchain = 0;
3172 symbol_conf.use_callchain = 0;
3173 }
3174 }
3175
3176 return 0;
3177}
3178
3179static int timehist_parse_prio_str(struct perf_sched *sched)
3180{
3181 char *p;
3182 unsigned long start_prio, end_prio;
3183 const char *str = sched->prio_str;
3184
3185 if (!str)
3186 return 0;
3187
3188 while (isdigit(*str)) {
3189 p = NULL;
3190 start_prio = strtoul(str, &p, 0);
3191 if (start_prio >= MAX_PRIO || (*p != '\0' && *p != ',' && *p != '-'))
3192 return -1;
3193
3194 if (*p == '-') {
3195 str = ++p;
3196 p = NULL;
3197 end_prio = strtoul(str, &p, 0);
3198
3199 if (end_prio >= MAX_PRIO || (*p != '\0' && *p != ','))
3200 return -1;
3201
3202 if (end_prio < start_prio)
3203 return -1;
3204 } else {
3205 end_prio = start_prio;
3206 }
3207
3208 for (; start_prio <= end_prio; start_prio++)
3209 __set_bit(start_prio, sched->prio_bitmap);
3210
3211 if (*p)
3212 ++p;
3213
3214 str = p;
3215 }
3216
3217 return 0;
3218}
3219
3220static int perf_sched__timehist(struct perf_sched *sched)
3221{
3222 struct evsel_str_handler handlers[] = {
3223 { "sched:sched_switch", timehist_sched_switch_event, },
3224 { "sched:sched_wakeup", timehist_sched_wakeup_event, },
3225 { "sched:sched_waking", timehist_sched_wakeup_event, },
3226 { "sched:sched_wakeup_new", timehist_sched_wakeup_event, },
3227 };
3228 const struct evsel_str_handler migrate_handlers[] = {
3229 { "sched:sched_migrate_task", timehist_migrate_task_event, },
3230 };
3231 struct perf_data data = {
3232 .path = input_name,
3233 .mode = PERF_DATA_MODE_READ,
3234 .force = sched->force,
3235 };
3236
3237 struct perf_session *session;
3238 struct evlist *evlist;
3239 int err = -1;
3240
3241 /*
3242 * event handlers for timehist option
3243 */
3244 sched->tool.sample = perf_timehist__process_sample;
3245 sched->tool.mmap = perf_event__process_mmap;
3246 sched->tool.comm = perf_event__process_comm;
3247 sched->tool.exit = perf_event__process_exit;
3248 sched->tool.fork = perf_event__process_fork;
3249 sched->tool.lost = process_lost;
3250 sched->tool.attr = perf_event__process_attr;
3251 sched->tool.tracing_data = perf_event__process_tracing_data;
3252 sched->tool.build_id = perf_event__process_build_id;
3253
3254 sched->tool.ordering_requires_timestamps = true;
3255
3256 symbol_conf.use_callchain = sched->show_callchain;
3257
3258 session = perf_session__new(&data, &sched->tool);
3259 if (IS_ERR(session))
3260 return PTR_ERR(session);
3261
3262 if (cpu_list) {
3263 err = perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap);
3264 if (err < 0)
3265 goto out;
3266 }
3267
3268 evlist = session->evlist;
3269
3270 symbol__init(&session->header.env);
3271
3272 if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
3273 pr_err("Invalid time string\n");
3274 err = -EINVAL;
3275 goto out;
3276 }
3277
3278 if (timehist_check_attr(sched, evlist) != 0)
3279 goto out;
3280
3281 if (timehist_parse_prio_str(sched) != 0) {
3282 pr_err("Invalid prio string\n");
3283 goto out;
3284 }
3285
3286 setup_pager();
3287
3288 /* prefer sched_waking if it is captured */
3289 if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
3290 handlers[1].handler = timehist_sched_wakeup_ignore;
3291
3292 /* setup per-evsel handlers */
3293 if (perf_session__set_tracepoints_handlers(session, handlers))
3294 goto out;
3295
3296 /* sched_switch event at a minimum needs to exist */
3297 if (!evlist__find_tracepoint_by_name(session->evlist, "sched:sched_switch")) {
3298 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3299 goto out;
3300 }
3301
3302 if ((sched->show_migrations || sched->pre_migrations) &&
3303 perf_session__set_tracepoints_handlers(session, migrate_handlers))
3304 goto out;
3305
3306 /* pre-allocate struct for per-CPU idle stats */
3307 sched->max_cpu.cpu = session->header.env.nr_cpus_online;
3308 if (sched->max_cpu.cpu == 0)
3309 sched->max_cpu.cpu = 4;
3310 if (init_idle_threads(sched->max_cpu.cpu))
3311 goto out;
3312
3313 /* summary_only implies summary option, but don't overwrite summary if set */
3314 if (sched->summary_only)
3315 sched->summary = sched->summary_only;
3316
3317 if (!sched->summary_only)
3318 timehist_header(sched);
3319
3320 err = perf_session__process_events(session);
3321 if (err) {
3322 pr_err("Failed to process events, error %d", err);
3323 goto out;
3324 }
3325
3326 sched->nr_events = evlist->stats.nr_events[0];
3327 sched->nr_lost_events = evlist->stats.total_lost;
3328 sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3329
3330 if (sched->summary)
3331 timehist_print_summary(sched, session);
3332
3333out:
3334 free_idle_threads();
3335 perf_session__delete(session);
3336
3337 return err;
3338}
3339
3340
3341static void print_bad_events(struct perf_sched *sched)
3342{
3343 if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3344 printf(" INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3345 (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3346 sched->nr_unordered_timestamps, sched->nr_timestamps);
3347 }
3348 if (sched->nr_lost_events && sched->nr_events) {
3349 printf(" INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3350 (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3351 sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3352 }
3353 if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3354 printf(" INFO: %.3f%% context switch bugs (%ld out of %ld)",
3355 (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3356 sched->nr_context_switch_bugs, sched->nr_timestamps);
3357 if (sched->nr_lost_events)
3358 printf(" (due to lost events?)");
3359 printf("\n");
3360 }
3361}
3362
3363static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3364{
3365 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3366 struct work_atoms *this;
3367 const char *comm = thread__comm_str(data->thread), *this_comm;
3368 bool leftmost = true;
3369
3370 while (*new) {
3371 int cmp;
3372
3373 this = container_of(*new, struct work_atoms, node);
3374 parent = *new;
3375
3376 this_comm = thread__comm_str(this->thread);
3377 cmp = strcmp(comm, this_comm);
3378 if (cmp > 0) {
3379 new = &((*new)->rb_left);
3380 } else if (cmp < 0) {
3381 new = &((*new)->rb_right);
3382 leftmost = false;
3383 } else {
3384 this->num_merged++;
3385 this->total_runtime += data->total_runtime;
3386 this->nb_atoms += data->nb_atoms;
3387 this->total_lat += data->total_lat;
3388 list_splice(&data->work_list, &this->work_list);
3389 if (this->max_lat < data->max_lat) {
3390 this->max_lat = data->max_lat;
3391 this->max_lat_start = data->max_lat_start;
3392 this->max_lat_end = data->max_lat_end;
3393 }
3394 zfree(&data);
3395 return;
3396 }
3397 }
3398
3399 data->num_merged++;
3400 rb_link_node(&data->node, parent, new);
3401 rb_insert_color_cached(&data->node, root, leftmost);
3402}
3403
3404static void perf_sched__merge_lat(struct perf_sched *sched)
3405{
3406 struct work_atoms *data;
3407 struct rb_node *node;
3408
3409 if (sched->skip_merge)
3410 return;
3411
3412 while ((node = rb_first_cached(&sched->atom_root))) {
3413 rb_erase_cached(node, &sched->atom_root);
3414 data = rb_entry(node, struct work_atoms, node);
3415 __merge_work_atoms(&sched->merged_atom_root, data);
3416 }
3417}
3418
3419static int setup_cpus_switch_event(struct perf_sched *sched)
3420{
3421 unsigned int i;
3422
3423 sched->cpu_last_switched = calloc(MAX_CPUS, sizeof(*(sched->cpu_last_switched)));
3424 if (!sched->cpu_last_switched)
3425 return -1;
3426
3427 sched->curr_pid = malloc(MAX_CPUS * sizeof(*(sched->curr_pid)));
3428 if (!sched->curr_pid) {
3429 zfree(&sched->cpu_last_switched);
3430 return -1;
3431 }
3432
3433 for (i = 0; i < MAX_CPUS; i++)
3434 sched->curr_pid[i] = -1;
3435
3436 return 0;
3437}
3438
3439static void free_cpus_switch_event(struct perf_sched *sched)
3440{
3441 zfree(&sched->curr_pid);
3442 zfree(&sched->cpu_last_switched);
3443}
3444
3445static int perf_sched__lat(struct perf_sched *sched)
3446{
3447 int rc = -1;
3448 struct rb_node *next;
3449
3450 setup_pager();
3451
3452 if (setup_cpus_switch_event(sched))
3453 return rc;
3454
3455 if (perf_sched__read_events(sched))
3456 goto out_free_cpus_switch_event;
3457
3458 perf_sched__merge_lat(sched);
3459 perf_sched__sort_lat(sched);
3460
3461 printf("\n -------------------------------------------------------------------------------------------------------------------------------------------\n");
3462 printf(" Task | Runtime ms | Count | Avg delay ms | Max delay ms | Max delay start | Max delay end |\n");
3463 printf(" -------------------------------------------------------------------------------------------------------------------------------------------\n");
3464
3465 next = rb_first_cached(&sched->sorted_atom_root);
3466
3467 while (next) {
3468 struct work_atoms *work_list;
3469
3470 work_list = rb_entry(next, struct work_atoms, node);
3471 output_lat_thread(sched, work_list);
3472 next = rb_next(next);
3473 thread__zput(work_list->thread);
3474 }
3475
3476 printf(" -----------------------------------------------------------------------------------------------------------------\n");
3477 printf(" TOTAL: |%11.3f ms |%9" PRIu64 " |\n",
3478 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3479
3480 printf(" ---------------------------------------------------\n");
3481
3482 print_bad_events(sched);
3483 printf("\n");
3484
3485 rc = 0;
3486
3487out_free_cpus_switch_event:
3488 free_cpus_switch_event(sched);
3489 return rc;
3490}
3491
3492static int setup_map_cpus(struct perf_sched *sched)
3493{
3494 sched->max_cpu.cpu = sysconf(_SC_NPROCESSORS_CONF);
3495
3496 if (sched->map.comp) {
3497 sched->map.comp_cpus = zalloc(sched->max_cpu.cpu * sizeof(int));
3498 if (!sched->map.comp_cpus)
3499 return -1;
3500 }
3501
3502 if (sched->map.cpus_str) {
3503 sched->map.cpus = perf_cpu_map__new(sched->map.cpus_str);
3504 if (!sched->map.cpus) {
3505 pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3506 zfree(&sched->map.comp_cpus);
3507 return -1;
3508 }
3509 }
3510
3511 return 0;
3512}
3513
3514static int setup_color_pids(struct perf_sched *sched)
3515{
3516 struct perf_thread_map *map;
3517
3518 if (!sched->map.color_pids_str)
3519 return 0;
3520
3521 map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3522 if (!map) {
3523 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3524 return -1;
3525 }
3526
3527 sched->map.color_pids = map;
3528 return 0;
3529}
3530
3531static int setup_color_cpus(struct perf_sched *sched)
3532{
3533 struct perf_cpu_map *map;
3534
3535 if (!sched->map.color_cpus_str)
3536 return 0;
3537
3538 map = perf_cpu_map__new(sched->map.color_cpus_str);
3539 if (!map) {
3540 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3541 return -1;
3542 }
3543
3544 sched->map.color_cpus = map;
3545 return 0;
3546}
3547
3548static int perf_sched__map(struct perf_sched *sched)
3549{
3550 int rc = -1;
3551
3552 sched->curr_thread = calloc(MAX_CPUS, sizeof(*(sched->curr_thread)));
3553 if (!sched->curr_thread)
3554 return rc;
3555
3556 sched->curr_out_thread = calloc(MAX_CPUS, sizeof(*(sched->curr_out_thread)));
3557 if (!sched->curr_out_thread)
3558 return rc;
3559
3560 if (setup_cpus_switch_event(sched))
3561 goto out_free_curr_thread;
3562
3563 if (setup_map_cpus(sched))
3564 goto out_free_cpus_switch_event;
3565
3566 if (setup_color_pids(sched))
3567 goto out_put_map_cpus;
3568
3569 if (setup_color_cpus(sched))
3570 goto out_put_color_pids;
3571
3572 setup_pager();
3573 if (perf_sched__read_events(sched))
3574 goto out_put_color_cpus;
3575
3576 rc = 0;
3577 print_bad_events(sched);
3578
3579out_put_color_cpus:
3580 perf_cpu_map__put(sched->map.color_cpus);
3581
3582out_put_color_pids:
3583 perf_thread_map__put(sched->map.color_pids);
3584
3585out_put_map_cpus:
3586 zfree(&sched->map.comp_cpus);
3587 perf_cpu_map__put(sched->map.cpus);
3588
3589out_free_cpus_switch_event:
3590 free_cpus_switch_event(sched);
3591
3592out_free_curr_thread:
3593 zfree(&sched->curr_thread);
3594 return rc;
3595}
3596
3597static int perf_sched__replay(struct perf_sched *sched)
3598{
3599 int ret;
3600 unsigned long i;
3601
3602 mutex_init(&sched->start_work_mutex);
3603 mutex_init(&sched->work_done_wait_mutex);
3604
3605 ret = setup_cpus_switch_event(sched);
3606 if (ret)
3607 goto out_mutex_destroy;
3608
3609 calibrate_run_measurement_overhead(sched);
3610 calibrate_sleep_measurement_overhead(sched);
3611
3612 test_calibrations(sched);
3613
3614 ret = perf_sched__read_events(sched);
3615 if (ret)
3616 goto out_free_cpus_switch_event;
3617
3618 printf("nr_run_events: %ld\n", sched->nr_run_events);
3619 printf("nr_sleep_events: %ld\n", sched->nr_sleep_events);
3620 printf("nr_wakeup_events: %ld\n", sched->nr_wakeup_events);
3621
3622 if (sched->targetless_wakeups)
3623 printf("target-less wakeups: %ld\n", sched->targetless_wakeups);
3624 if (sched->multitarget_wakeups)
3625 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3626 if (sched->nr_run_events_optimized)
3627 printf("run atoms optimized: %ld\n",
3628 sched->nr_run_events_optimized);
3629
3630 print_task_traces(sched);
3631 add_cross_task_wakeups(sched);
3632
3633 sched->thread_funcs_exit = false;
3634 create_tasks(sched);
3635 printf("------------------------------------------------------------\n");
3636 if (sched->replay_repeat == 0)
3637 sched->replay_repeat = UINT_MAX;
3638
3639 for (i = 0; i < sched->replay_repeat; i++)
3640 run_one_test(sched);
3641
3642 sched->thread_funcs_exit = true;
3643 destroy_tasks(sched);
3644
3645out_free_cpus_switch_event:
3646 free_cpus_switch_event(sched);
3647
3648out_mutex_destroy:
3649 mutex_destroy(&sched->start_work_mutex);
3650 mutex_destroy(&sched->work_done_wait_mutex);
3651 return ret;
3652}
3653
3654static void setup_sorting(struct perf_sched *sched, const struct option *options,
3655 const char * const usage_msg[])
3656{
3657 char *tmp, *tok, *str = strdup(sched->sort_order);
3658
3659 for (tok = strtok_r(str, ", ", &tmp);
3660 tok; tok = strtok_r(NULL, ", ", &tmp)) {
3661 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3662 usage_with_options_msg(usage_msg, options,
3663 "Unknown --sort key: `%s'", tok);
3664 }
3665 }
3666
3667 free(str);
3668
3669 sort_dimension__add("pid", &sched->cmp_pid);
3670}
3671
3672static bool schedstat_events_exposed(void)
3673{
3674 /*
3675 * Select "sched:sched_stat_wait" event to check
3676 * whether schedstat tracepoints are exposed.
3677 */
3678 return IS_ERR(trace_event__tp_format("sched", "sched_stat_wait")) ?
3679 false : true;
3680}
3681
3682static int __cmd_record(int argc, const char **argv)
3683{
3684 unsigned int rec_argc, i, j;
3685 char **rec_argv;
3686 const char **rec_argv_copy;
3687 const char * const record_args[] = {
3688 "record",
3689 "-a",
3690 "-R",
3691 "-m", "1024",
3692 "-c", "1",
3693 "-e", "sched:sched_switch",
3694 "-e", "sched:sched_stat_runtime",
3695 "-e", "sched:sched_process_fork",
3696 "-e", "sched:sched_wakeup_new",
3697 "-e", "sched:sched_migrate_task",
3698 };
3699
3700 /*
3701 * The tracepoints trace_sched_stat_{wait, sleep, iowait}
3702 * are not exposed to user if CONFIG_SCHEDSTATS is not set,
3703 * to prevent "perf sched record" execution failure, determine
3704 * whether to record schedstat events according to actual situation.
3705 */
3706 const char * const schedstat_args[] = {
3707 "-e", "sched:sched_stat_wait",
3708 "-e", "sched:sched_stat_sleep",
3709 "-e", "sched:sched_stat_iowait",
3710 };
3711 unsigned int schedstat_argc = schedstat_events_exposed() ?
3712 ARRAY_SIZE(schedstat_args) : 0;
3713
3714 struct tep_event *waking_event;
3715 int ret;
3716
3717 /*
3718 * +2 for either "-e", "sched:sched_wakeup" or
3719 * "-e", "sched:sched_waking"
3720 */
3721 rec_argc = ARRAY_SIZE(record_args) + 2 + schedstat_argc + argc - 1;
3722 rec_argv = calloc(rec_argc + 1, sizeof(char *));
3723 if (rec_argv == NULL)
3724 return -ENOMEM;
3725 rec_argv_copy = calloc(rec_argc + 1, sizeof(char *));
3726 if (rec_argv_copy == NULL) {
3727 free(rec_argv);
3728 return -ENOMEM;
3729 }
3730
3731 for (i = 0; i < ARRAY_SIZE(record_args); i++)
3732 rec_argv[i] = strdup(record_args[i]);
3733
3734 rec_argv[i++] = strdup("-e");
3735 waking_event = trace_event__tp_format("sched", "sched_waking");
3736 if (!IS_ERR(waking_event))
3737 rec_argv[i++] = strdup("sched:sched_waking");
3738 else
3739 rec_argv[i++] = strdup("sched:sched_wakeup");
3740
3741 for (j = 0; j < schedstat_argc; j++)
3742 rec_argv[i++] = strdup(schedstat_args[j]);
3743
3744 for (j = 1; j < (unsigned int)argc; j++, i++)
3745 rec_argv[i] = strdup(argv[j]);
3746
3747 BUG_ON(i != rec_argc);
3748
3749 memcpy(rec_argv_copy, rec_argv, sizeof(char *) * rec_argc);
3750 ret = cmd_record(rec_argc, rec_argv_copy);
3751
3752 for (i = 0; i < rec_argc; i++)
3753 free(rec_argv[i]);
3754 free(rec_argv);
3755 free(rec_argv_copy);
3756
3757 return ret;
3758}
3759
3760int cmd_sched(int argc, const char **argv)
3761{
3762 static const char default_sort_order[] = "avg, max, switch, runtime";
3763 struct perf_sched sched = {
3764 .cmp_pid = LIST_HEAD_INIT(sched.cmp_pid),
3765 .sort_list = LIST_HEAD_INIT(sched.sort_list),
3766 .sort_order = default_sort_order,
3767 .replay_repeat = 10,
3768 .profile_cpu = -1,
3769 .next_shortname1 = 'A',
3770 .next_shortname2 = '0',
3771 .skip_merge = 0,
3772 .show_callchain = 1,
3773 .max_stack = 5,
3774 };
3775 const struct option sched_options[] = {
3776 OPT_STRING('i', "input", &input_name, "file",
3777 "input file name"),
3778 OPT_INCR('v', "verbose", &verbose,
3779 "be more verbose (show symbol address, etc)"),
3780 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3781 "dump raw trace in ASCII"),
3782 OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3783 OPT_END()
3784 };
3785 const struct option latency_options[] = {
3786 OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3787 "sort by key(s): runtime, switch, avg, max"),
3788 OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3789 "CPU to profile on"),
3790 OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3791 "latency stats per pid instead of per comm"),
3792 OPT_PARENT(sched_options)
3793 };
3794 const struct option replay_options[] = {
3795 OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3796 "repeat the workload replay N times (0: infinite)"),
3797 OPT_PARENT(sched_options)
3798 };
3799 const struct option map_options[] = {
3800 OPT_BOOLEAN(0, "compact", &sched.map.comp,
3801 "map output in compact mode"),
3802 OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3803 "highlight given pids in map"),
3804 OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3805 "highlight given CPUs in map"),
3806 OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3807 "display given CPUs in map"),
3808 OPT_STRING(0, "task-name", &sched.map.task_name, "task",
3809 "map output only for the given task name(s)."),
3810 OPT_BOOLEAN(0, "fuzzy-name", &sched.map.fuzzy,
3811 "given command name can be partially matched (fuzzy matching)"),
3812 OPT_PARENT(sched_options)
3813 };
3814 const struct option timehist_options[] = {
3815 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3816 "file", "vmlinux pathname"),
3817 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3818 "file", "kallsyms pathname"),
3819 OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3820 "Display call chains if present (default on)"),
3821 OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3822 "Maximum number of functions to display backtrace."),
3823 OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3824 "Look for files with symbols relative to this directory"),
3825 OPT_BOOLEAN('s', "summary", &sched.summary_only,
3826 "Show only syscall summary with statistics"),
3827 OPT_BOOLEAN('S', "with-summary", &sched.summary,
3828 "Show all syscalls and summary with statistics"),
3829 OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3830 OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3831 OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3832 OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3833 OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3834 OPT_STRING(0, "time", &sched.time_str, "str",
3835 "Time span for analysis (start,stop)"),
3836 OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3837 OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3838 "analyze events only for given process id(s)"),
3839 OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3840 "analyze events only for given thread id(s)"),
3841 OPT_STRING('C', "cpu", &cpu_list, "cpu", "list of cpus to profile"),
3842 OPT_BOOLEAN(0, "show-prio", &sched.show_prio, "Show task priority"),
3843 OPT_STRING(0, "prio", &sched.prio_str, "prio",
3844 "analyze events only for given task priority(ies)"),
3845 OPT_BOOLEAN('P', "pre-migrations", &sched.pre_migrations, "Show pre-migration wait time"),
3846 OPT_PARENT(sched_options)
3847 };
3848
3849 const char * const latency_usage[] = {
3850 "perf sched latency [<options>]",
3851 NULL
3852 };
3853 const char * const replay_usage[] = {
3854 "perf sched replay [<options>]",
3855 NULL
3856 };
3857 const char * const map_usage[] = {
3858 "perf sched map [<options>]",
3859 NULL
3860 };
3861 const char * const timehist_usage[] = {
3862 "perf sched timehist [<options>]",
3863 NULL
3864 };
3865 const char *const sched_subcommands[] = { "record", "latency", "map",
3866 "replay", "script",
3867 "timehist", NULL };
3868 const char *sched_usage[] = {
3869 NULL,
3870 NULL
3871 };
3872 struct trace_sched_handler lat_ops = {
3873 .wakeup_event = latency_wakeup_event,
3874 .switch_event = latency_switch_event,
3875 .runtime_event = latency_runtime_event,
3876 .migrate_task_event = latency_migrate_task_event,
3877 };
3878 struct trace_sched_handler map_ops = {
3879 .switch_event = map_switch_event,
3880 };
3881 struct trace_sched_handler replay_ops = {
3882 .wakeup_event = replay_wakeup_event,
3883 .switch_event = replay_switch_event,
3884 .fork_event = replay_fork_event,
3885 };
3886 int ret;
3887
3888 perf_tool__init(&sched.tool, /*ordered_events=*/true);
3889 sched.tool.sample = perf_sched__process_tracepoint_sample;
3890 sched.tool.comm = perf_sched__process_comm;
3891 sched.tool.namespaces = perf_event__process_namespaces;
3892 sched.tool.lost = perf_event__process_lost;
3893 sched.tool.fork = perf_sched__process_fork_event;
3894
3895 argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3896 sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3897 if (!argc)
3898 usage_with_options(sched_usage, sched_options);
3899
3900 /*
3901 * Aliased to 'perf script' for now:
3902 */
3903 if (!strcmp(argv[0], "script")) {
3904 return cmd_script(argc, argv);
3905 } else if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
3906 return __cmd_record(argc, argv);
3907 } else if (strlen(argv[0]) > 2 && strstarts("latency", argv[0])) {
3908 sched.tp_handler = &lat_ops;
3909 if (argc > 1) {
3910 argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3911 if (argc)
3912 usage_with_options(latency_usage, latency_options);
3913 }
3914 setup_sorting(&sched, latency_options, latency_usage);
3915 return perf_sched__lat(&sched);
3916 } else if (!strcmp(argv[0], "map")) {
3917 if (argc) {
3918 argc = parse_options(argc, argv, map_options, map_usage, 0);
3919 if (argc)
3920 usage_with_options(map_usage, map_options);
3921
3922 if (sched.map.task_name) {
3923 sched.map.task_names = strlist__new(sched.map.task_name, NULL);
3924 if (sched.map.task_names == NULL) {
3925 fprintf(stderr, "Failed to parse task names\n");
3926 return -1;
3927 }
3928 }
3929 }
3930 sched.tp_handler = &map_ops;
3931 setup_sorting(&sched, latency_options, latency_usage);
3932 return perf_sched__map(&sched);
3933 } else if (strlen(argv[0]) > 2 && strstarts("replay", argv[0])) {
3934 sched.tp_handler = &replay_ops;
3935 if (argc) {
3936 argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3937 if (argc)
3938 usage_with_options(replay_usage, replay_options);
3939 }
3940 return perf_sched__replay(&sched);
3941 } else if (!strcmp(argv[0], "timehist")) {
3942 if (argc) {
3943 argc = parse_options(argc, argv, timehist_options,
3944 timehist_usage, 0);
3945 if (argc)
3946 usage_with_options(timehist_usage, timehist_options);
3947 }
3948 if ((sched.show_wakeups || sched.show_next) &&
3949 sched.summary_only) {
3950 pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3951 parse_options_usage(timehist_usage, timehist_options, "s", true);
3952 if (sched.show_wakeups)
3953 parse_options_usage(NULL, timehist_options, "w", true);
3954 if (sched.show_next)
3955 parse_options_usage(NULL, timehist_options, "n", true);
3956 return -EINVAL;
3957 }
3958 ret = symbol__validate_sym_arguments();
3959 if (ret)
3960 return ret;
3961
3962 return perf_sched__timehist(&sched);
3963 } else {
3964 usage_with_options(sched_usage, sched_options);
3965 }
3966
3967 /* free usage string allocated by parse_options_subcommand */
3968 free((void *)sched_usage[0]);
3969
3970 return 0;
3971}