Loading...
1// SPDX-License-Identifier: GPL-2.0
2#include "builtin.h"
3#include "perf.h"
4
5#include "util/util.h"
6#include "util/evlist.h"
7#include "util/cache.h"
8#include "util/evsel.h"
9#include "util/symbol.h"
10#include "util/thread.h"
11#include "util/header.h"
12#include "util/session.h"
13#include "util/tool.h"
14#include "util/cloexec.h"
15#include "util/thread_map.h"
16#include "util/color.h"
17#include "util/stat.h"
18#include "util/callchain.h"
19#include "util/time-utils.h"
20
21#include <subcmd/parse-options.h>
22#include "util/trace-event.h"
23
24#include "util/debug.h"
25
26#include <linux/kernel.h>
27#include <linux/log2.h>
28#include <sys/prctl.h>
29#include <sys/resource.h>
30#include <inttypes.h>
31
32#include <errno.h>
33#include <semaphore.h>
34#include <pthread.h>
35#include <math.h>
36#include <api/fs/fs.h>
37#include <linux/time64.h>
38
39#include "sane_ctype.h"
40
41#define PR_SET_NAME 15 /* Set process name */
42#define MAX_CPUS 4096
43#define COMM_LEN 20
44#define SYM_LEN 129
45#define MAX_PID 1024000
46
47struct sched_atom;
48
49struct task_desc {
50 unsigned long nr;
51 unsigned long pid;
52 char comm[COMM_LEN];
53
54 unsigned long nr_events;
55 unsigned long curr_event;
56 struct sched_atom **atoms;
57
58 pthread_t thread;
59 sem_t sleep_sem;
60
61 sem_t ready_for_work;
62 sem_t work_done_sem;
63
64 u64 cpu_usage;
65};
66
67enum sched_event_type {
68 SCHED_EVENT_RUN,
69 SCHED_EVENT_SLEEP,
70 SCHED_EVENT_WAKEUP,
71 SCHED_EVENT_MIGRATION,
72};
73
74struct sched_atom {
75 enum sched_event_type type;
76 int specific_wait;
77 u64 timestamp;
78 u64 duration;
79 unsigned long nr;
80 sem_t *wait_sem;
81 struct task_desc *wakee;
82};
83
84#define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
85
86/* task state bitmask, copied from include/linux/sched.h */
87#define TASK_RUNNING 0
88#define TASK_INTERRUPTIBLE 1
89#define TASK_UNINTERRUPTIBLE 2
90#define __TASK_STOPPED 4
91#define __TASK_TRACED 8
92/* in tsk->exit_state */
93#define EXIT_DEAD 16
94#define EXIT_ZOMBIE 32
95#define EXIT_TRACE (EXIT_ZOMBIE | EXIT_DEAD)
96/* in tsk->state again */
97#define TASK_DEAD 64
98#define TASK_WAKEKILL 128
99#define TASK_WAKING 256
100#define TASK_PARKED 512
101
102enum thread_state {
103 THREAD_SLEEPING = 0,
104 THREAD_WAIT_CPU,
105 THREAD_SCHED_IN,
106 THREAD_IGNORE
107};
108
109struct work_atom {
110 struct list_head list;
111 enum thread_state state;
112 u64 sched_out_time;
113 u64 wake_up_time;
114 u64 sched_in_time;
115 u64 runtime;
116};
117
118struct work_atoms {
119 struct list_head work_list;
120 struct thread *thread;
121 struct rb_node node;
122 u64 max_lat;
123 u64 max_lat_at;
124 u64 total_lat;
125 u64 nb_atoms;
126 u64 total_runtime;
127 int num_merged;
128};
129
130typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
131
132struct perf_sched;
133
134struct trace_sched_handler {
135 int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
136 struct perf_sample *sample, struct machine *machine);
137
138 int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
139 struct perf_sample *sample, struct machine *machine);
140
141 int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
142 struct perf_sample *sample, struct machine *machine);
143
144 /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
145 int (*fork_event)(struct perf_sched *sched, union perf_event *event,
146 struct machine *machine);
147
148 int (*migrate_task_event)(struct perf_sched *sched,
149 struct perf_evsel *evsel,
150 struct perf_sample *sample,
151 struct machine *machine);
152};
153
154#define COLOR_PIDS PERF_COLOR_BLUE
155#define COLOR_CPUS PERF_COLOR_BG_RED
156
157struct perf_sched_map {
158 DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
159 int *comp_cpus;
160 bool comp;
161 struct thread_map *color_pids;
162 const char *color_pids_str;
163 struct cpu_map *color_cpus;
164 const char *color_cpus_str;
165 struct cpu_map *cpus;
166 const char *cpus_str;
167};
168
169struct perf_sched {
170 struct perf_tool tool;
171 const char *sort_order;
172 unsigned long nr_tasks;
173 struct task_desc **pid_to_task;
174 struct task_desc **tasks;
175 const struct trace_sched_handler *tp_handler;
176 pthread_mutex_t start_work_mutex;
177 pthread_mutex_t work_done_wait_mutex;
178 int profile_cpu;
179/*
180 * Track the current task - that way we can know whether there's any
181 * weird events, such as a task being switched away that is not current.
182 */
183 int max_cpu;
184 u32 curr_pid[MAX_CPUS];
185 struct thread *curr_thread[MAX_CPUS];
186 char next_shortname1;
187 char next_shortname2;
188 unsigned int replay_repeat;
189 unsigned long nr_run_events;
190 unsigned long nr_sleep_events;
191 unsigned long nr_wakeup_events;
192 unsigned long nr_sleep_corrections;
193 unsigned long nr_run_events_optimized;
194 unsigned long targetless_wakeups;
195 unsigned long multitarget_wakeups;
196 unsigned long nr_runs;
197 unsigned long nr_timestamps;
198 unsigned long nr_unordered_timestamps;
199 unsigned long nr_context_switch_bugs;
200 unsigned long nr_events;
201 unsigned long nr_lost_chunks;
202 unsigned long nr_lost_events;
203 u64 run_measurement_overhead;
204 u64 sleep_measurement_overhead;
205 u64 start_time;
206 u64 cpu_usage;
207 u64 runavg_cpu_usage;
208 u64 parent_cpu_usage;
209 u64 runavg_parent_cpu_usage;
210 u64 sum_runtime;
211 u64 sum_fluct;
212 u64 run_avg;
213 u64 all_runtime;
214 u64 all_count;
215 u64 cpu_last_switched[MAX_CPUS];
216 struct rb_root atom_root, sorted_atom_root, merged_atom_root;
217 struct list_head sort_list, cmp_pid;
218 bool force;
219 bool skip_merge;
220 struct perf_sched_map map;
221
222 /* options for timehist command */
223 bool summary;
224 bool summary_only;
225 bool idle_hist;
226 bool show_callchain;
227 unsigned int max_stack;
228 bool show_cpu_visual;
229 bool show_wakeups;
230 bool show_next;
231 bool show_migrations;
232 bool show_state;
233 u64 skipped_samples;
234 const char *time_str;
235 struct perf_time_interval ptime;
236 struct perf_time_interval hist_time;
237};
238
239/* per thread run time data */
240struct thread_runtime {
241 u64 last_time; /* time of previous sched in/out event */
242 u64 dt_run; /* run time */
243 u64 dt_sleep; /* time between CPU access by sleep (off cpu) */
244 u64 dt_iowait; /* time between CPU access by iowait (off cpu) */
245 u64 dt_preempt; /* time between CPU access by preempt (off cpu) */
246 u64 dt_delay; /* time between wakeup and sched-in */
247 u64 ready_to_run; /* time of wakeup */
248
249 struct stats run_stats;
250 u64 total_run_time;
251 u64 total_sleep_time;
252 u64 total_iowait_time;
253 u64 total_preempt_time;
254 u64 total_delay_time;
255
256 int last_state;
257
258 char shortname[3];
259 bool comm_changed;
260
261 u64 migrations;
262};
263
264/* per event run time data */
265struct evsel_runtime {
266 u64 *last_time; /* time this event was last seen per cpu */
267 u32 ncpu; /* highest cpu slot allocated */
268};
269
270/* per cpu idle time data */
271struct idle_thread_runtime {
272 struct thread_runtime tr;
273 struct thread *last_thread;
274 struct rb_root sorted_root;
275 struct callchain_root callchain;
276 struct callchain_cursor cursor;
277};
278
279/* track idle times per cpu */
280static struct thread **idle_threads;
281static int idle_max_cpu;
282static char idle_comm[] = "<idle>";
283
284static u64 get_nsecs(void)
285{
286 struct timespec ts;
287
288 clock_gettime(CLOCK_MONOTONIC, &ts);
289
290 return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
291}
292
293static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
294{
295 u64 T0 = get_nsecs(), T1;
296
297 do {
298 T1 = get_nsecs();
299 } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
300}
301
302static void sleep_nsecs(u64 nsecs)
303{
304 struct timespec ts;
305
306 ts.tv_nsec = nsecs % 999999999;
307 ts.tv_sec = nsecs / 999999999;
308
309 nanosleep(&ts, NULL);
310}
311
312static void calibrate_run_measurement_overhead(struct perf_sched *sched)
313{
314 u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
315 int i;
316
317 for (i = 0; i < 10; i++) {
318 T0 = get_nsecs();
319 burn_nsecs(sched, 0);
320 T1 = get_nsecs();
321 delta = T1-T0;
322 min_delta = min(min_delta, delta);
323 }
324 sched->run_measurement_overhead = min_delta;
325
326 printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
327}
328
329static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
330{
331 u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
332 int i;
333
334 for (i = 0; i < 10; i++) {
335 T0 = get_nsecs();
336 sleep_nsecs(10000);
337 T1 = get_nsecs();
338 delta = T1-T0;
339 min_delta = min(min_delta, delta);
340 }
341 min_delta -= 10000;
342 sched->sleep_measurement_overhead = min_delta;
343
344 printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
345}
346
347static struct sched_atom *
348get_new_event(struct task_desc *task, u64 timestamp)
349{
350 struct sched_atom *event = zalloc(sizeof(*event));
351 unsigned long idx = task->nr_events;
352 size_t size;
353
354 event->timestamp = timestamp;
355 event->nr = idx;
356
357 task->nr_events++;
358 size = sizeof(struct sched_atom *) * task->nr_events;
359 task->atoms = realloc(task->atoms, size);
360 BUG_ON(!task->atoms);
361
362 task->atoms[idx] = event;
363
364 return event;
365}
366
367static struct sched_atom *last_event(struct task_desc *task)
368{
369 if (!task->nr_events)
370 return NULL;
371
372 return task->atoms[task->nr_events - 1];
373}
374
375static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
376 u64 timestamp, u64 duration)
377{
378 struct sched_atom *event, *curr_event = last_event(task);
379
380 /*
381 * optimize an existing RUN event by merging this one
382 * to it:
383 */
384 if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
385 sched->nr_run_events_optimized++;
386 curr_event->duration += duration;
387 return;
388 }
389
390 event = get_new_event(task, timestamp);
391
392 event->type = SCHED_EVENT_RUN;
393 event->duration = duration;
394
395 sched->nr_run_events++;
396}
397
398static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
399 u64 timestamp, struct task_desc *wakee)
400{
401 struct sched_atom *event, *wakee_event;
402
403 event = get_new_event(task, timestamp);
404 event->type = SCHED_EVENT_WAKEUP;
405 event->wakee = wakee;
406
407 wakee_event = last_event(wakee);
408 if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
409 sched->targetless_wakeups++;
410 return;
411 }
412 if (wakee_event->wait_sem) {
413 sched->multitarget_wakeups++;
414 return;
415 }
416
417 wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
418 sem_init(wakee_event->wait_sem, 0, 0);
419 wakee_event->specific_wait = 1;
420 event->wait_sem = wakee_event->wait_sem;
421
422 sched->nr_wakeup_events++;
423}
424
425static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
426 u64 timestamp, u64 task_state __maybe_unused)
427{
428 struct sched_atom *event = get_new_event(task, timestamp);
429
430 event->type = SCHED_EVENT_SLEEP;
431
432 sched->nr_sleep_events++;
433}
434
435static struct task_desc *register_pid(struct perf_sched *sched,
436 unsigned long pid, const char *comm)
437{
438 struct task_desc *task;
439 static int pid_max;
440
441 if (sched->pid_to_task == NULL) {
442 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
443 pid_max = MAX_PID;
444 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
445 }
446 if (pid >= (unsigned long)pid_max) {
447 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
448 sizeof(struct task_desc *))) == NULL);
449 while (pid >= (unsigned long)pid_max)
450 sched->pid_to_task[pid_max++] = NULL;
451 }
452
453 task = sched->pid_to_task[pid];
454
455 if (task)
456 return task;
457
458 task = zalloc(sizeof(*task));
459 task->pid = pid;
460 task->nr = sched->nr_tasks;
461 strcpy(task->comm, comm);
462 /*
463 * every task starts in sleeping state - this gets ignored
464 * if there's no wakeup pointing to this sleep state:
465 */
466 add_sched_event_sleep(sched, task, 0, 0);
467
468 sched->pid_to_task[pid] = task;
469 sched->nr_tasks++;
470 sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
471 BUG_ON(!sched->tasks);
472 sched->tasks[task->nr] = task;
473
474 if (verbose > 0)
475 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
476
477 return task;
478}
479
480
481static void print_task_traces(struct perf_sched *sched)
482{
483 struct task_desc *task;
484 unsigned long i;
485
486 for (i = 0; i < sched->nr_tasks; i++) {
487 task = sched->tasks[i];
488 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
489 task->nr, task->comm, task->pid, task->nr_events);
490 }
491}
492
493static void add_cross_task_wakeups(struct perf_sched *sched)
494{
495 struct task_desc *task1, *task2;
496 unsigned long i, j;
497
498 for (i = 0; i < sched->nr_tasks; i++) {
499 task1 = sched->tasks[i];
500 j = i + 1;
501 if (j == sched->nr_tasks)
502 j = 0;
503 task2 = sched->tasks[j];
504 add_sched_event_wakeup(sched, task1, 0, task2);
505 }
506}
507
508static void perf_sched__process_event(struct perf_sched *sched,
509 struct sched_atom *atom)
510{
511 int ret = 0;
512
513 switch (atom->type) {
514 case SCHED_EVENT_RUN:
515 burn_nsecs(sched, atom->duration);
516 break;
517 case SCHED_EVENT_SLEEP:
518 if (atom->wait_sem)
519 ret = sem_wait(atom->wait_sem);
520 BUG_ON(ret);
521 break;
522 case SCHED_EVENT_WAKEUP:
523 if (atom->wait_sem)
524 ret = sem_post(atom->wait_sem);
525 BUG_ON(ret);
526 break;
527 case SCHED_EVENT_MIGRATION:
528 break;
529 default:
530 BUG_ON(1);
531 }
532}
533
534static u64 get_cpu_usage_nsec_parent(void)
535{
536 struct rusage ru;
537 u64 sum;
538 int err;
539
540 err = getrusage(RUSAGE_SELF, &ru);
541 BUG_ON(err);
542
543 sum = ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
544 sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
545
546 return sum;
547}
548
549static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
550{
551 struct perf_event_attr attr;
552 char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
553 int fd;
554 struct rlimit limit;
555 bool need_privilege = false;
556
557 memset(&attr, 0, sizeof(attr));
558
559 attr.type = PERF_TYPE_SOFTWARE;
560 attr.config = PERF_COUNT_SW_TASK_CLOCK;
561
562force_again:
563 fd = sys_perf_event_open(&attr, 0, -1, -1,
564 perf_event_open_cloexec_flag());
565
566 if (fd < 0) {
567 if (errno == EMFILE) {
568 if (sched->force) {
569 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
570 limit.rlim_cur += sched->nr_tasks - cur_task;
571 if (limit.rlim_cur > limit.rlim_max) {
572 limit.rlim_max = limit.rlim_cur;
573 need_privilege = true;
574 }
575 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
576 if (need_privilege && errno == EPERM)
577 strcpy(info, "Need privilege\n");
578 } else
579 goto force_again;
580 } else
581 strcpy(info, "Have a try with -f option\n");
582 }
583 pr_err("Error: sys_perf_event_open() syscall returned "
584 "with %d (%s)\n%s", fd,
585 str_error_r(errno, sbuf, sizeof(sbuf)), info);
586 exit(EXIT_FAILURE);
587 }
588 return fd;
589}
590
591static u64 get_cpu_usage_nsec_self(int fd)
592{
593 u64 runtime;
594 int ret;
595
596 ret = read(fd, &runtime, sizeof(runtime));
597 BUG_ON(ret != sizeof(runtime));
598
599 return runtime;
600}
601
602struct sched_thread_parms {
603 struct task_desc *task;
604 struct perf_sched *sched;
605 int fd;
606};
607
608static void *thread_func(void *ctx)
609{
610 struct sched_thread_parms *parms = ctx;
611 struct task_desc *this_task = parms->task;
612 struct perf_sched *sched = parms->sched;
613 u64 cpu_usage_0, cpu_usage_1;
614 unsigned long i, ret;
615 char comm2[22];
616 int fd = parms->fd;
617
618 zfree(&parms);
619
620 sprintf(comm2, ":%s", this_task->comm);
621 prctl(PR_SET_NAME, comm2);
622 if (fd < 0)
623 return NULL;
624again:
625 ret = sem_post(&this_task->ready_for_work);
626 BUG_ON(ret);
627 ret = pthread_mutex_lock(&sched->start_work_mutex);
628 BUG_ON(ret);
629 ret = pthread_mutex_unlock(&sched->start_work_mutex);
630 BUG_ON(ret);
631
632 cpu_usage_0 = get_cpu_usage_nsec_self(fd);
633
634 for (i = 0; i < this_task->nr_events; i++) {
635 this_task->curr_event = i;
636 perf_sched__process_event(sched, this_task->atoms[i]);
637 }
638
639 cpu_usage_1 = get_cpu_usage_nsec_self(fd);
640 this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
641 ret = sem_post(&this_task->work_done_sem);
642 BUG_ON(ret);
643
644 ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
645 BUG_ON(ret);
646 ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
647 BUG_ON(ret);
648
649 goto again;
650}
651
652static void create_tasks(struct perf_sched *sched)
653{
654 struct task_desc *task;
655 pthread_attr_t attr;
656 unsigned long i;
657 int err;
658
659 err = pthread_attr_init(&attr);
660 BUG_ON(err);
661 err = pthread_attr_setstacksize(&attr,
662 (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
663 BUG_ON(err);
664 err = pthread_mutex_lock(&sched->start_work_mutex);
665 BUG_ON(err);
666 err = pthread_mutex_lock(&sched->work_done_wait_mutex);
667 BUG_ON(err);
668 for (i = 0; i < sched->nr_tasks; i++) {
669 struct sched_thread_parms *parms = malloc(sizeof(*parms));
670 BUG_ON(parms == NULL);
671 parms->task = task = sched->tasks[i];
672 parms->sched = sched;
673 parms->fd = self_open_counters(sched, i);
674 sem_init(&task->sleep_sem, 0, 0);
675 sem_init(&task->ready_for_work, 0, 0);
676 sem_init(&task->work_done_sem, 0, 0);
677 task->curr_event = 0;
678 err = pthread_create(&task->thread, &attr, thread_func, parms);
679 BUG_ON(err);
680 }
681}
682
683static void wait_for_tasks(struct perf_sched *sched)
684{
685 u64 cpu_usage_0, cpu_usage_1;
686 struct task_desc *task;
687 unsigned long i, ret;
688
689 sched->start_time = get_nsecs();
690 sched->cpu_usage = 0;
691 pthread_mutex_unlock(&sched->work_done_wait_mutex);
692
693 for (i = 0; i < sched->nr_tasks; i++) {
694 task = sched->tasks[i];
695 ret = sem_wait(&task->ready_for_work);
696 BUG_ON(ret);
697 sem_init(&task->ready_for_work, 0, 0);
698 }
699 ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
700 BUG_ON(ret);
701
702 cpu_usage_0 = get_cpu_usage_nsec_parent();
703
704 pthread_mutex_unlock(&sched->start_work_mutex);
705
706 for (i = 0; i < sched->nr_tasks; i++) {
707 task = sched->tasks[i];
708 ret = sem_wait(&task->work_done_sem);
709 BUG_ON(ret);
710 sem_init(&task->work_done_sem, 0, 0);
711 sched->cpu_usage += task->cpu_usage;
712 task->cpu_usage = 0;
713 }
714
715 cpu_usage_1 = get_cpu_usage_nsec_parent();
716 if (!sched->runavg_cpu_usage)
717 sched->runavg_cpu_usage = sched->cpu_usage;
718 sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
719
720 sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
721 if (!sched->runavg_parent_cpu_usage)
722 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
723 sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
724 sched->parent_cpu_usage)/sched->replay_repeat;
725
726 ret = pthread_mutex_lock(&sched->start_work_mutex);
727 BUG_ON(ret);
728
729 for (i = 0; i < sched->nr_tasks; i++) {
730 task = sched->tasks[i];
731 sem_init(&task->sleep_sem, 0, 0);
732 task->curr_event = 0;
733 }
734}
735
736static void run_one_test(struct perf_sched *sched)
737{
738 u64 T0, T1, delta, avg_delta, fluct;
739
740 T0 = get_nsecs();
741 wait_for_tasks(sched);
742 T1 = get_nsecs();
743
744 delta = T1 - T0;
745 sched->sum_runtime += delta;
746 sched->nr_runs++;
747
748 avg_delta = sched->sum_runtime / sched->nr_runs;
749 if (delta < avg_delta)
750 fluct = avg_delta - delta;
751 else
752 fluct = delta - avg_delta;
753 sched->sum_fluct += fluct;
754 if (!sched->run_avg)
755 sched->run_avg = delta;
756 sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
757
758 printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
759
760 printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
761
762 printf("cpu: %0.2f / %0.2f",
763 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
764
765#if 0
766 /*
767 * rusage statistics done by the parent, these are less
768 * accurate than the sched->sum_exec_runtime based statistics:
769 */
770 printf(" [%0.2f / %0.2f]",
771 (double)sched->parent_cpu_usage / NSEC_PER_MSEC,
772 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
773#endif
774
775 printf("\n");
776
777 if (sched->nr_sleep_corrections)
778 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
779 sched->nr_sleep_corrections = 0;
780}
781
782static void test_calibrations(struct perf_sched *sched)
783{
784 u64 T0, T1;
785
786 T0 = get_nsecs();
787 burn_nsecs(sched, NSEC_PER_MSEC);
788 T1 = get_nsecs();
789
790 printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
791
792 T0 = get_nsecs();
793 sleep_nsecs(NSEC_PER_MSEC);
794 T1 = get_nsecs();
795
796 printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
797}
798
799static int
800replay_wakeup_event(struct perf_sched *sched,
801 struct perf_evsel *evsel, struct perf_sample *sample,
802 struct machine *machine __maybe_unused)
803{
804 const char *comm = perf_evsel__strval(evsel, sample, "comm");
805 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
806 struct task_desc *waker, *wakee;
807
808 if (verbose > 0) {
809 printf("sched_wakeup event %p\n", evsel);
810
811 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
812 }
813
814 waker = register_pid(sched, sample->tid, "<unknown>");
815 wakee = register_pid(sched, pid, comm);
816
817 add_sched_event_wakeup(sched, waker, sample->time, wakee);
818 return 0;
819}
820
821static int replay_switch_event(struct perf_sched *sched,
822 struct perf_evsel *evsel,
823 struct perf_sample *sample,
824 struct machine *machine __maybe_unused)
825{
826 const char *prev_comm = perf_evsel__strval(evsel, sample, "prev_comm"),
827 *next_comm = perf_evsel__strval(evsel, sample, "next_comm");
828 const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
829 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
830 const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
831 struct task_desc *prev, __maybe_unused *next;
832 u64 timestamp0, timestamp = sample->time;
833 int cpu = sample->cpu;
834 s64 delta;
835
836 if (verbose > 0)
837 printf("sched_switch event %p\n", evsel);
838
839 if (cpu >= MAX_CPUS || cpu < 0)
840 return 0;
841
842 timestamp0 = sched->cpu_last_switched[cpu];
843 if (timestamp0)
844 delta = timestamp - timestamp0;
845 else
846 delta = 0;
847
848 if (delta < 0) {
849 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
850 return -1;
851 }
852
853 pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
854 prev_comm, prev_pid, next_comm, next_pid, delta);
855
856 prev = register_pid(sched, prev_pid, prev_comm);
857 next = register_pid(sched, next_pid, next_comm);
858
859 sched->cpu_last_switched[cpu] = timestamp;
860
861 add_sched_event_run(sched, prev, timestamp, delta);
862 add_sched_event_sleep(sched, prev, timestamp, prev_state);
863
864 return 0;
865}
866
867static int replay_fork_event(struct perf_sched *sched,
868 union perf_event *event,
869 struct machine *machine)
870{
871 struct thread *child, *parent;
872
873 child = machine__findnew_thread(machine, event->fork.pid,
874 event->fork.tid);
875 parent = machine__findnew_thread(machine, event->fork.ppid,
876 event->fork.ptid);
877
878 if (child == NULL || parent == NULL) {
879 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
880 child, parent);
881 goto out_put;
882 }
883
884 if (verbose > 0) {
885 printf("fork event\n");
886 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
887 printf("... child: %s/%d\n", thread__comm_str(child), child->tid);
888 }
889
890 register_pid(sched, parent->tid, thread__comm_str(parent));
891 register_pid(sched, child->tid, thread__comm_str(child));
892out_put:
893 thread__put(child);
894 thread__put(parent);
895 return 0;
896}
897
898struct sort_dimension {
899 const char *name;
900 sort_fn_t cmp;
901 struct list_head list;
902};
903
904/*
905 * handle runtime stats saved per thread
906 */
907static struct thread_runtime *thread__init_runtime(struct thread *thread)
908{
909 struct thread_runtime *r;
910
911 r = zalloc(sizeof(struct thread_runtime));
912 if (!r)
913 return NULL;
914
915 init_stats(&r->run_stats);
916 thread__set_priv(thread, r);
917
918 return r;
919}
920
921static struct thread_runtime *thread__get_runtime(struct thread *thread)
922{
923 struct thread_runtime *tr;
924
925 tr = thread__priv(thread);
926 if (tr == NULL) {
927 tr = thread__init_runtime(thread);
928 if (tr == NULL)
929 pr_debug("Failed to malloc memory for runtime data.\n");
930 }
931
932 return tr;
933}
934
935static int
936thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
937{
938 struct sort_dimension *sort;
939 int ret = 0;
940
941 BUG_ON(list_empty(list));
942
943 list_for_each_entry(sort, list, list) {
944 ret = sort->cmp(l, r);
945 if (ret)
946 return ret;
947 }
948
949 return ret;
950}
951
952static struct work_atoms *
953thread_atoms_search(struct rb_root *root, struct thread *thread,
954 struct list_head *sort_list)
955{
956 struct rb_node *node = root->rb_node;
957 struct work_atoms key = { .thread = thread };
958
959 while (node) {
960 struct work_atoms *atoms;
961 int cmp;
962
963 atoms = container_of(node, struct work_atoms, node);
964
965 cmp = thread_lat_cmp(sort_list, &key, atoms);
966 if (cmp > 0)
967 node = node->rb_left;
968 else if (cmp < 0)
969 node = node->rb_right;
970 else {
971 BUG_ON(thread != atoms->thread);
972 return atoms;
973 }
974 }
975 return NULL;
976}
977
978static void
979__thread_latency_insert(struct rb_root *root, struct work_atoms *data,
980 struct list_head *sort_list)
981{
982 struct rb_node **new = &(root->rb_node), *parent = NULL;
983
984 while (*new) {
985 struct work_atoms *this;
986 int cmp;
987
988 this = container_of(*new, struct work_atoms, node);
989 parent = *new;
990
991 cmp = thread_lat_cmp(sort_list, data, this);
992
993 if (cmp > 0)
994 new = &((*new)->rb_left);
995 else
996 new = &((*new)->rb_right);
997 }
998
999 rb_link_node(&data->node, parent, new);
1000 rb_insert_color(&data->node, root);
1001}
1002
1003static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1004{
1005 struct work_atoms *atoms = zalloc(sizeof(*atoms));
1006 if (!atoms) {
1007 pr_err("No memory at %s\n", __func__);
1008 return -1;
1009 }
1010
1011 atoms->thread = thread__get(thread);
1012 INIT_LIST_HEAD(&atoms->work_list);
1013 __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1014 return 0;
1015}
1016
1017static char sched_out_state(u64 prev_state)
1018{
1019 const char *str = TASK_STATE_TO_CHAR_STR;
1020
1021 return str[prev_state];
1022}
1023
1024static int
1025add_sched_out_event(struct work_atoms *atoms,
1026 char run_state,
1027 u64 timestamp)
1028{
1029 struct work_atom *atom = zalloc(sizeof(*atom));
1030 if (!atom) {
1031 pr_err("Non memory at %s", __func__);
1032 return -1;
1033 }
1034
1035 atom->sched_out_time = timestamp;
1036
1037 if (run_state == 'R') {
1038 atom->state = THREAD_WAIT_CPU;
1039 atom->wake_up_time = atom->sched_out_time;
1040 }
1041
1042 list_add_tail(&atom->list, &atoms->work_list);
1043 return 0;
1044}
1045
1046static void
1047add_runtime_event(struct work_atoms *atoms, u64 delta,
1048 u64 timestamp __maybe_unused)
1049{
1050 struct work_atom *atom;
1051
1052 BUG_ON(list_empty(&atoms->work_list));
1053
1054 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1055
1056 atom->runtime += delta;
1057 atoms->total_runtime += delta;
1058}
1059
1060static void
1061add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1062{
1063 struct work_atom *atom;
1064 u64 delta;
1065
1066 if (list_empty(&atoms->work_list))
1067 return;
1068
1069 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1070
1071 if (atom->state != THREAD_WAIT_CPU)
1072 return;
1073
1074 if (timestamp < atom->wake_up_time) {
1075 atom->state = THREAD_IGNORE;
1076 return;
1077 }
1078
1079 atom->state = THREAD_SCHED_IN;
1080 atom->sched_in_time = timestamp;
1081
1082 delta = atom->sched_in_time - atom->wake_up_time;
1083 atoms->total_lat += delta;
1084 if (delta > atoms->max_lat) {
1085 atoms->max_lat = delta;
1086 atoms->max_lat_at = timestamp;
1087 }
1088 atoms->nb_atoms++;
1089}
1090
1091static int latency_switch_event(struct perf_sched *sched,
1092 struct perf_evsel *evsel,
1093 struct perf_sample *sample,
1094 struct machine *machine)
1095{
1096 const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1097 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1098 const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
1099 struct work_atoms *out_events, *in_events;
1100 struct thread *sched_out, *sched_in;
1101 u64 timestamp0, timestamp = sample->time;
1102 int cpu = sample->cpu, err = -1;
1103 s64 delta;
1104
1105 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1106
1107 timestamp0 = sched->cpu_last_switched[cpu];
1108 sched->cpu_last_switched[cpu] = timestamp;
1109 if (timestamp0)
1110 delta = timestamp - timestamp0;
1111 else
1112 delta = 0;
1113
1114 if (delta < 0) {
1115 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1116 return -1;
1117 }
1118
1119 sched_out = machine__findnew_thread(machine, -1, prev_pid);
1120 sched_in = machine__findnew_thread(machine, -1, next_pid);
1121 if (sched_out == NULL || sched_in == NULL)
1122 goto out_put;
1123
1124 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1125 if (!out_events) {
1126 if (thread_atoms_insert(sched, sched_out))
1127 goto out_put;
1128 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1129 if (!out_events) {
1130 pr_err("out-event: Internal tree error");
1131 goto out_put;
1132 }
1133 }
1134 if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1135 return -1;
1136
1137 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1138 if (!in_events) {
1139 if (thread_atoms_insert(sched, sched_in))
1140 goto out_put;
1141 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1142 if (!in_events) {
1143 pr_err("in-event: Internal tree error");
1144 goto out_put;
1145 }
1146 /*
1147 * Take came in we have not heard about yet,
1148 * add in an initial atom in runnable state:
1149 */
1150 if (add_sched_out_event(in_events, 'R', timestamp))
1151 goto out_put;
1152 }
1153 add_sched_in_event(in_events, timestamp);
1154 err = 0;
1155out_put:
1156 thread__put(sched_out);
1157 thread__put(sched_in);
1158 return err;
1159}
1160
1161static int latency_runtime_event(struct perf_sched *sched,
1162 struct perf_evsel *evsel,
1163 struct perf_sample *sample,
1164 struct machine *machine)
1165{
1166 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1167 const u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
1168 struct thread *thread = machine__findnew_thread(machine, -1, pid);
1169 struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1170 u64 timestamp = sample->time;
1171 int cpu = sample->cpu, err = -1;
1172
1173 if (thread == NULL)
1174 return -1;
1175
1176 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1177 if (!atoms) {
1178 if (thread_atoms_insert(sched, thread))
1179 goto out_put;
1180 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1181 if (!atoms) {
1182 pr_err("in-event: Internal tree error");
1183 goto out_put;
1184 }
1185 if (add_sched_out_event(atoms, 'R', timestamp))
1186 goto out_put;
1187 }
1188
1189 add_runtime_event(atoms, runtime, timestamp);
1190 err = 0;
1191out_put:
1192 thread__put(thread);
1193 return err;
1194}
1195
1196static int latency_wakeup_event(struct perf_sched *sched,
1197 struct perf_evsel *evsel,
1198 struct perf_sample *sample,
1199 struct machine *machine)
1200{
1201 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1202 struct work_atoms *atoms;
1203 struct work_atom *atom;
1204 struct thread *wakee;
1205 u64 timestamp = sample->time;
1206 int err = -1;
1207
1208 wakee = machine__findnew_thread(machine, -1, pid);
1209 if (wakee == NULL)
1210 return -1;
1211 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1212 if (!atoms) {
1213 if (thread_atoms_insert(sched, wakee))
1214 goto out_put;
1215 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1216 if (!atoms) {
1217 pr_err("wakeup-event: Internal tree error");
1218 goto out_put;
1219 }
1220 if (add_sched_out_event(atoms, 'S', timestamp))
1221 goto out_put;
1222 }
1223
1224 BUG_ON(list_empty(&atoms->work_list));
1225
1226 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1227
1228 /*
1229 * As we do not guarantee the wakeup event happens when
1230 * task is out of run queue, also may happen when task is
1231 * on run queue and wakeup only change ->state to TASK_RUNNING,
1232 * then we should not set the ->wake_up_time when wake up a
1233 * task which is on run queue.
1234 *
1235 * You WILL be missing events if you've recorded only
1236 * one CPU, or are only looking at only one, so don't
1237 * skip in this case.
1238 */
1239 if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1240 goto out_ok;
1241
1242 sched->nr_timestamps++;
1243 if (atom->sched_out_time > timestamp) {
1244 sched->nr_unordered_timestamps++;
1245 goto out_ok;
1246 }
1247
1248 atom->state = THREAD_WAIT_CPU;
1249 atom->wake_up_time = timestamp;
1250out_ok:
1251 err = 0;
1252out_put:
1253 thread__put(wakee);
1254 return err;
1255}
1256
1257static int latency_migrate_task_event(struct perf_sched *sched,
1258 struct perf_evsel *evsel,
1259 struct perf_sample *sample,
1260 struct machine *machine)
1261{
1262 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1263 u64 timestamp = sample->time;
1264 struct work_atoms *atoms;
1265 struct work_atom *atom;
1266 struct thread *migrant;
1267 int err = -1;
1268
1269 /*
1270 * Only need to worry about migration when profiling one CPU.
1271 */
1272 if (sched->profile_cpu == -1)
1273 return 0;
1274
1275 migrant = machine__findnew_thread(machine, -1, pid);
1276 if (migrant == NULL)
1277 return -1;
1278 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1279 if (!atoms) {
1280 if (thread_atoms_insert(sched, migrant))
1281 goto out_put;
1282 register_pid(sched, migrant->tid, thread__comm_str(migrant));
1283 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1284 if (!atoms) {
1285 pr_err("migration-event: Internal tree error");
1286 goto out_put;
1287 }
1288 if (add_sched_out_event(atoms, 'R', timestamp))
1289 goto out_put;
1290 }
1291
1292 BUG_ON(list_empty(&atoms->work_list));
1293
1294 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1295 atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1296
1297 sched->nr_timestamps++;
1298
1299 if (atom->sched_out_time > timestamp)
1300 sched->nr_unordered_timestamps++;
1301 err = 0;
1302out_put:
1303 thread__put(migrant);
1304 return err;
1305}
1306
1307static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1308{
1309 int i;
1310 int ret;
1311 u64 avg;
1312 char max_lat_at[32];
1313
1314 if (!work_list->nb_atoms)
1315 return;
1316 /*
1317 * Ignore idle threads:
1318 */
1319 if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1320 return;
1321
1322 sched->all_runtime += work_list->total_runtime;
1323 sched->all_count += work_list->nb_atoms;
1324
1325 if (work_list->num_merged > 1)
1326 ret = printf(" %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1327 else
1328 ret = printf(" %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1329
1330 for (i = 0; i < 24 - ret; i++)
1331 printf(" ");
1332
1333 avg = work_list->total_lat / work_list->nb_atoms;
1334 timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at));
1335
1336 printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n",
1337 (double)work_list->total_runtime / NSEC_PER_MSEC,
1338 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1339 (double)work_list->max_lat / NSEC_PER_MSEC,
1340 max_lat_at);
1341}
1342
1343static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1344{
1345 if (l->thread == r->thread)
1346 return 0;
1347 if (l->thread->tid < r->thread->tid)
1348 return -1;
1349 if (l->thread->tid > r->thread->tid)
1350 return 1;
1351 return (int)(l->thread - r->thread);
1352}
1353
1354static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1355{
1356 u64 avgl, avgr;
1357
1358 if (!l->nb_atoms)
1359 return -1;
1360
1361 if (!r->nb_atoms)
1362 return 1;
1363
1364 avgl = l->total_lat / l->nb_atoms;
1365 avgr = r->total_lat / r->nb_atoms;
1366
1367 if (avgl < avgr)
1368 return -1;
1369 if (avgl > avgr)
1370 return 1;
1371
1372 return 0;
1373}
1374
1375static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1376{
1377 if (l->max_lat < r->max_lat)
1378 return -1;
1379 if (l->max_lat > r->max_lat)
1380 return 1;
1381
1382 return 0;
1383}
1384
1385static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1386{
1387 if (l->nb_atoms < r->nb_atoms)
1388 return -1;
1389 if (l->nb_atoms > r->nb_atoms)
1390 return 1;
1391
1392 return 0;
1393}
1394
1395static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1396{
1397 if (l->total_runtime < r->total_runtime)
1398 return -1;
1399 if (l->total_runtime > r->total_runtime)
1400 return 1;
1401
1402 return 0;
1403}
1404
1405static int sort_dimension__add(const char *tok, struct list_head *list)
1406{
1407 size_t i;
1408 static struct sort_dimension avg_sort_dimension = {
1409 .name = "avg",
1410 .cmp = avg_cmp,
1411 };
1412 static struct sort_dimension max_sort_dimension = {
1413 .name = "max",
1414 .cmp = max_cmp,
1415 };
1416 static struct sort_dimension pid_sort_dimension = {
1417 .name = "pid",
1418 .cmp = pid_cmp,
1419 };
1420 static struct sort_dimension runtime_sort_dimension = {
1421 .name = "runtime",
1422 .cmp = runtime_cmp,
1423 };
1424 static struct sort_dimension switch_sort_dimension = {
1425 .name = "switch",
1426 .cmp = switch_cmp,
1427 };
1428 struct sort_dimension *available_sorts[] = {
1429 &pid_sort_dimension,
1430 &avg_sort_dimension,
1431 &max_sort_dimension,
1432 &switch_sort_dimension,
1433 &runtime_sort_dimension,
1434 };
1435
1436 for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1437 if (!strcmp(available_sorts[i]->name, tok)) {
1438 list_add_tail(&available_sorts[i]->list, list);
1439
1440 return 0;
1441 }
1442 }
1443
1444 return -1;
1445}
1446
1447static void perf_sched__sort_lat(struct perf_sched *sched)
1448{
1449 struct rb_node *node;
1450 struct rb_root *root = &sched->atom_root;
1451again:
1452 for (;;) {
1453 struct work_atoms *data;
1454 node = rb_first(root);
1455 if (!node)
1456 break;
1457
1458 rb_erase(node, root);
1459 data = rb_entry(node, struct work_atoms, node);
1460 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1461 }
1462 if (root == &sched->atom_root) {
1463 root = &sched->merged_atom_root;
1464 goto again;
1465 }
1466}
1467
1468static int process_sched_wakeup_event(struct perf_tool *tool,
1469 struct perf_evsel *evsel,
1470 struct perf_sample *sample,
1471 struct machine *machine)
1472{
1473 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1474
1475 if (sched->tp_handler->wakeup_event)
1476 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1477
1478 return 0;
1479}
1480
1481union map_priv {
1482 void *ptr;
1483 bool color;
1484};
1485
1486static bool thread__has_color(struct thread *thread)
1487{
1488 union map_priv priv = {
1489 .ptr = thread__priv(thread),
1490 };
1491
1492 return priv.color;
1493}
1494
1495static struct thread*
1496map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1497{
1498 struct thread *thread = machine__findnew_thread(machine, pid, tid);
1499 union map_priv priv = {
1500 .color = false,
1501 };
1502
1503 if (!sched->map.color_pids || !thread || thread__priv(thread))
1504 return thread;
1505
1506 if (thread_map__has(sched->map.color_pids, tid))
1507 priv.color = true;
1508
1509 thread__set_priv(thread, priv.ptr);
1510 return thread;
1511}
1512
1513static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1514 struct perf_sample *sample, struct machine *machine)
1515{
1516 const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1517 struct thread *sched_in;
1518 struct thread_runtime *tr;
1519 int new_shortname;
1520 u64 timestamp0, timestamp = sample->time;
1521 s64 delta;
1522 int i, this_cpu = sample->cpu;
1523 int cpus_nr;
1524 bool new_cpu = false;
1525 const char *color = PERF_COLOR_NORMAL;
1526 char stimestamp[32];
1527
1528 BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1529
1530 if (this_cpu > sched->max_cpu)
1531 sched->max_cpu = this_cpu;
1532
1533 if (sched->map.comp) {
1534 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1535 if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1536 sched->map.comp_cpus[cpus_nr++] = this_cpu;
1537 new_cpu = true;
1538 }
1539 } else
1540 cpus_nr = sched->max_cpu;
1541
1542 timestamp0 = sched->cpu_last_switched[this_cpu];
1543 sched->cpu_last_switched[this_cpu] = timestamp;
1544 if (timestamp0)
1545 delta = timestamp - timestamp0;
1546 else
1547 delta = 0;
1548
1549 if (delta < 0) {
1550 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1551 return -1;
1552 }
1553
1554 sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1555 if (sched_in == NULL)
1556 return -1;
1557
1558 tr = thread__get_runtime(sched_in);
1559 if (tr == NULL) {
1560 thread__put(sched_in);
1561 return -1;
1562 }
1563
1564 sched->curr_thread[this_cpu] = thread__get(sched_in);
1565
1566 printf(" ");
1567
1568 new_shortname = 0;
1569 if (!tr->shortname[0]) {
1570 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1571 /*
1572 * Don't allocate a letter-number for swapper:0
1573 * as a shortname. Instead, we use '.' for it.
1574 */
1575 tr->shortname[0] = '.';
1576 tr->shortname[1] = ' ';
1577 } else {
1578 tr->shortname[0] = sched->next_shortname1;
1579 tr->shortname[1] = sched->next_shortname2;
1580
1581 if (sched->next_shortname1 < 'Z') {
1582 sched->next_shortname1++;
1583 } else {
1584 sched->next_shortname1 = 'A';
1585 if (sched->next_shortname2 < '9')
1586 sched->next_shortname2++;
1587 else
1588 sched->next_shortname2 = '0';
1589 }
1590 }
1591 new_shortname = 1;
1592 }
1593
1594 for (i = 0; i < cpus_nr; i++) {
1595 int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
1596 struct thread *curr_thread = sched->curr_thread[cpu];
1597 struct thread_runtime *curr_tr;
1598 const char *pid_color = color;
1599 const char *cpu_color = color;
1600
1601 if (curr_thread && thread__has_color(curr_thread))
1602 pid_color = COLOR_PIDS;
1603
1604 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1605 continue;
1606
1607 if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1608 cpu_color = COLOR_CPUS;
1609
1610 if (cpu != this_cpu)
1611 color_fprintf(stdout, color, " ");
1612 else
1613 color_fprintf(stdout, cpu_color, "*");
1614
1615 if (sched->curr_thread[cpu]) {
1616 curr_tr = thread__get_runtime(sched->curr_thread[cpu]);
1617 if (curr_tr == NULL) {
1618 thread__put(sched_in);
1619 return -1;
1620 }
1621 color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1622 } else
1623 color_fprintf(stdout, color, " ");
1624 }
1625
1626 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1627 goto out;
1628
1629 timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1630 color_fprintf(stdout, color, " %12s secs ", stimestamp);
1631 if (new_shortname || tr->comm_changed || (verbose > 0 && sched_in->tid)) {
1632 const char *pid_color = color;
1633
1634 if (thread__has_color(sched_in))
1635 pid_color = COLOR_PIDS;
1636
1637 color_fprintf(stdout, pid_color, "%s => %s:%d",
1638 tr->shortname, thread__comm_str(sched_in), sched_in->tid);
1639 tr->comm_changed = false;
1640 }
1641
1642 if (sched->map.comp && new_cpu)
1643 color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1644
1645out:
1646 color_fprintf(stdout, color, "\n");
1647
1648 thread__put(sched_in);
1649
1650 return 0;
1651}
1652
1653static int process_sched_switch_event(struct perf_tool *tool,
1654 struct perf_evsel *evsel,
1655 struct perf_sample *sample,
1656 struct machine *machine)
1657{
1658 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1659 int this_cpu = sample->cpu, err = 0;
1660 u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1661 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1662
1663 if (sched->curr_pid[this_cpu] != (u32)-1) {
1664 /*
1665 * Are we trying to switch away a PID that is
1666 * not current?
1667 */
1668 if (sched->curr_pid[this_cpu] != prev_pid)
1669 sched->nr_context_switch_bugs++;
1670 }
1671
1672 if (sched->tp_handler->switch_event)
1673 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1674
1675 sched->curr_pid[this_cpu] = next_pid;
1676 return err;
1677}
1678
1679static int process_sched_runtime_event(struct perf_tool *tool,
1680 struct perf_evsel *evsel,
1681 struct perf_sample *sample,
1682 struct machine *machine)
1683{
1684 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1685
1686 if (sched->tp_handler->runtime_event)
1687 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1688
1689 return 0;
1690}
1691
1692static int perf_sched__process_fork_event(struct perf_tool *tool,
1693 union perf_event *event,
1694 struct perf_sample *sample,
1695 struct machine *machine)
1696{
1697 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1698
1699 /* run the fork event through the perf machineruy */
1700 perf_event__process_fork(tool, event, sample, machine);
1701
1702 /* and then run additional processing needed for this command */
1703 if (sched->tp_handler->fork_event)
1704 return sched->tp_handler->fork_event(sched, event, machine);
1705
1706 return 0;
1707}
1708
1709static int process_sched_migrate_task_event(struct perf_tool *tool,
1710 struct perf_evsel *evsel,
1711 struct perf_sample *sample,
1712 struct machine *machine)
1713{
1714 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1715
1716 if (sched->tp_handler->migrate_task_event)
1717 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1718
1719 return 0;
1720}
1721
1722typedef int (*tracepoint_handler)(struct perf_tool *tool,
1723 struct perf_evsel *evsel,
1724 struct perf_sample *sample,
1725 struct machine *machine);
1726
1727static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1728 union perf_event *event __maybe_unused,
1729 struct perf_sample *sample,
1730 struct perf_evsel *evsel,
1731 struct machine *machine)
1732{
1733 int err = 0;
1734
1735 if (evsel->handler != NULL) {
1736 tracepoint_handler f = evsel->handler;
1737 err = f(tool, evsel, sample, machine);
1738 }
1739
1740 return err;
1741}
1742
1743static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1744 union perf_event *event,
1745 struct perf_sample *sample,
1746 struct machine *machine)
1747{
1748 struct thread *thread;
1749 struct thread_runtime *tr;
1750 int err;
1751
1752 err = perf_event__process_comm(tool, event, sample, machine);
1753 if (err)
1754 return err;
1755
1756 thread = machine__find_thread(machine, sample->pid, sample->tid);
1757 if (!thread) {
1758 pr_err("Internal error: can't find thread\n");
1759 return -1;
1760 }
1761
1762 tr = thread__get_runtime(thread);
1763 if (tr == NULL) {
1764 thread__put(thread);
1765 return -1;
1766 }
1767
1768 tr->comm_changed = true;
1769 thread__put(thread);
1770
1771 return 0;
1772}
1773
1774static int perf_sched__read_events(struct perf_sched *sched)
1775{
1776 const struct perf_evsel_str_handler handlers[] = {
1777 { "sched:sched_switch", process_sched_switch_event, },
1778 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1779 { "sched:sched_wakeup", process_sched_wakeup_event, },
1780 { "sched:sched_wakeup_new", process_sched_wakeup_event, },
1781 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1782 };
1783 struct perf_session *session;
1784 struct perf_data data = {
1785 .file = {
1786 .path = input_name,
1787 },
1788 .mode = PERF_DATA_MODE_READ,
1789 .force = sched->force,
1790 };
1791 int rc = -1;
1792
1793 session = perf_session__new(&data, false, &sched->tool);
1794 if (session == NULL) {
1795 pr_debug("No Memory for session\n");
1796 return -1;
1797 }
1798
1799 symbol__init(&session->header.env);
1800
1801 if (perf_session__set_tracepoints_handlers(session, handlers))
1802 goto out_delete;
1803
1804 if (perf_session__has_traces(session, "record -R")) {
1805 int err = perf_session__process_events(session);
1806 if (err) {
1807 pr_err("Failed to process events, error %d", err);
1808 goto out_delete;
1809 }
1810
1811 sched->nr_events = session->evlist->stats.nr_events[0];
1812 sched->nr_lost_events = session->evlist->stats.total_lost;
1813 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1814 }
1815
1816 rc = 0;
1817out_delete:
1818 perf_session__delete(session);
1819 return rc;
1820}
1821
1822/*
1823 * scheduling times are printed as msec.usec
1824 */
1825static inline void print_sched_time(unsigned long long nsecs, int width)
1826{
1827 unsigned long msecs;
1828 unsigned long usecs;
1829
1830 msecs = nsecs / NSEC_PER_MSEC;
1831 nsecs -= msecs * NSEC_PER_MSEC;
1832 usecs = nsecs / NSEC_PER_USEC;
1833 printf("%*lu.%03lu ", width, msecs, usecs);
1834}
1835
1836/*
1837 * returns runtime data for event, allocating memory for it the
1838 * first time it is used.
1839 */
1840static struct evsel_runtime *perf_evsel__get_runtime(struct perf_evsel *evsel)
1841{
1842 struct evsel_runtime *r = evsel->priv;
1843
1844 if (r == NULL) {
1845 r = zalloc(sizeof(struct evsel_runtime));
1846 evsel->priv = r;
1847 }
1848
1849 return r;
1850}
1851
1852/*
1853 * save last time event was seen per cpu
1854 */
1855static void perf_evsel__save_time(struct perf_evsel *evsel,
1856 u64 timestamp, u32 cpu)
1857{
1858 struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1859
1860 if (r == NULL)
1861 return;
1862
1863 if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1864 int i, n = __roundup_pow_of_two(cpu+1);
1865 void *p = r->last_time;
1866
1867 p = realloc(r->last_time, n * sizeof(u64));
1868 if (!p)
1869 return;
1870
1871 r->last_time = p;
1872 for (i = r->ncpu; i < n; ++i)
1873 r->last_time[i] = (u64) 0;
1874
1875 r->ncpu = n;
1876 }
1877
1878 r->last_time[cpu] = timestamp;
1879}
1880
1881/* returns last time this event was seen on the given cpu */
1882static u64 perf_evsel__get_time(struct perf_evsel *evsel, u32 cpu)
1883{
1884 struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1885
1886 if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1887 return 0;
1888
1889 return r->last_time[cpu];
1890}
1891
1892static int comm_width = 30;
1893
1894static char *timehist_get_commstr(struct thread *thread)
1895{
1896 static char str[32];
1897 const char *comm = thread__comm_str(thread);
1898 pid_t tid = thread->tid;
1899 pid_t pid = thread->pid_;
1900 int n;
1901
1902 if (pid == 0)
1903 n = scnprintf(str, sizeof(str), "%s", comm);
1904
1905 else if (tid != pid)
1906 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1907
1908 else
1909 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1910
1911 if (n > comm_width)
1912 comm_width = n;
1913
1914 return str;
1915}
1916
1917static void timehist_header(struct perf_sched *sched)
1918{
1919 u32 ncpus = sched->max_cpu + 1;
1920 u32 i, j;
1921
1922 printf("%15s %6s ", "time", "cpu");
1923
1924 if (sched->show_cpu_visual) {
1925 printf(" ");
1926 for (i = 0, j = 0; i < ncpus; ++i) {
1927 printf("%x", j++);
1928 if (j > 15)
1929 j = 0;
1930 }
1931 printf(" ");
1932 }
1933
1934 printf(" %-*s %9s %9s %9s", comm_width,
1935 "task name", "wait time", "sch delay", "run time");
1936
1937 if (sched->show_state)
1938 printf(" %s", "state");
1939
1940 printf("\n");
1941
1942 /*
1943 * units row
1944 */
1945 printf("%15s %-6s ", "", "");
1946
1947 if (sched->show_cpu_visual)
1948 printf(" %*s ", ncpus, "");
1949
1950 printf(" %-*s %9s %9s %9s", comm_width,
1951 "[tid/pid]", "(msec)", "(msec)", "(msec)");
1952
1953 if (sched->show_state)
1954 printf(" %5s", "");
1955
1956 printf("\n");
1957
1958 /*
1959 * separator
1960 */
1961 printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1962
1963 if (sched->show_cpu_visual)
1964 printf(" %.*s ", ncpus, graph_dotted_line);
1965
1966 printf(" %.*s %.9s %.9s %.9s", comm_width,
1967 graph_dotted_line, graph_dotted_line, graph_dotted_line,
1968 graph_dotted_line);
1969
1970 if (sched->show_state)
1971 printf(" %.5s", graph_dotted_line);
1972
1973 printf("\n");
1974}
1975
1976static char task_state_char(struct thread *thread, int state)
1977{
1978 static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
1979 unsigned bit = state ? ffs(state) : 0;
1980
1981 /* 'I' for idle */
1982 if (thread->tid == 0)
1983 return 'I';
1984
1985 return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
1986}
1987
1988static void timehist_print_sample(struct perf_sched *sched,
1989 struct perf_evsel *evsel,
1990 struct perf_sample *sample,
1991 struct addr_location *al,
1992 struct thread *thread,
1993 u64 t, int state)
1994{
1995 struct thread_runtime *tr = thread__priv(thread);
1996 const char *next_comm = perf_evsel__strval(evsel, sample, "next_comm");
1997 const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1998 u32 max_cpus = sched->max_cpu + 1;
1999 char tstr[64];
2000 char nstr[30];
2001 u64 wait_time;
2002
2003 timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2004 printf("%15s [%04d] ", tstr, sample->cpu);
2005
2006 if (sched->show_cpu_visual) {
2007 u32 i;
2008 char c;
2009
2010 printf(" ");
2011 for (i = 0; i < max_cpus; ++i) {
2012 /* flag idle times with 'i'; others are sched events */
2013 if (i == sample->cpu)
2014 c = (thread->tid == 0) ? 'i' : 's';
2015 else
2016 c = ' ';
2017 printf("%c", c);
2018 }
2019 printf(" ");
2020 }
2021
2022 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2023
2024 wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2025 print_sched_time(wait_time, 6);
2026
2027 print_sched_time(tr->dt_delay, 6);
2028 print_sched_time(tr->dt_run, 6);
2029
2030 if (sched->show_state)
2031 printf(" %5c ", task_state_char(thread, state));
2032
2033 if (sched->show_next) {
2034 snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2035 printf(" %-*s", comm_width, nstr);
2036 }
2037
2038 if (sched->show_wakeups && !sched->show_next)
2039 printf(" %-*s", comm_width, "");
2040
2041 if (thread->tid == 0)
2042 goto out;
2043
2044 if (sched->show_callchain)
2045 printf(" ");
2046
2047 sample__fprintf_sym(sample, al, 0,
2048 EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2049 EVSEL__PRINT_CALLCHAIN_ARROW |
2050 EVSEL__PRINT_SKIP_IGNORED,
2051 &callchain_cursor, stdout);
2052
2053out:
2054 printf("\n");
2055}
2056
2057/*
2058 * Explanation of delta-time stats:
2059 *
2060 * t = time of current schedule out event
2061 * tprev = time of previous sched out event
2062 * also time of schedule-in event for current task
2063 * last_time = time of last sched change event for current task
2064 * (i.e, time process was last scheduled out)
2065 * ready_to_run = time of wakeup for current task
2066 *
2067 * -----|------------|------------|------------|------
2068 * last ready tprev t
2069 * time to run
2070 *
2071 * |-------- dt_wait --------|
2072 * |- dt_delay -|-- dt_run --|
2073 *
2074 * dt_run = run time of current task
2075 * dt_wait = time between last schedule out event for task and tprev
2076 * represents time spent off the cpu
2077 * dt_delay = time between wakeup and schedule-in of task
2078 */
2079
2080static void timehist_update_runtime_stats(struct thread_runtime *r,
2081 u64 t, u64 tprev)
2082{
2083 r->dt_delay = 0;
2084 r->dt_sleep = 0;
2085 r->dt_iowait = 0;
2086 r->dt_preempt = 0;
2087 r->dt_run = 0;
2088
2089 if (tprev) {
2090 r->dt_run = t - tprev;
2091 if (r->ready_to_run) {
2092 if (r->ready_to_run > tprev)
2093 pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2094 else
2095 r->dt_delay = tprev - r->ready_to_run;
2096 }
2097
2098 if (r->last_time > tprev)
2099 pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2100 else if (r->last_time) {
2101 u64 dt_wait = tprev - r->last_time;
2102
2103 if (r->last_state == TASK_RUNNING)
2104 r->dt_preempt = dt_wait;
2105 else if (r->last_state == TASK_UNINTERRUPTIBLE)
2106 r->dt_iowait = dt_wait;
2107 else
2108 r->dt_sleep = dt_wait;
2109 }
2110 }
2111
2112 update_stats(&r->run_stats, r->dt_run);
2113
2114 r->total_run_time += r->dt_run;
2115 r->total_delay_time += r->dt_delay;
2116 r->total_sleep_time += r->dt_sleep;
2117 r->total_iowait_time += r->dt_iowait;
2118 r->total_preempt_time += r->dt_preempt;
2119}
2120
2121static bool is_idle_sample(struct perf_sample *sample,
2122 struct perf_evsel *evsel)
2123{
2124 /* pid 0 == swapper == idle task */
2125 if (strcmp(perf_evsel__name(evsel), "sched:sched_switch") == 0)
2126 return perf_evsel__intval(evsel, sample, "prev_pid") == 0;
2127
2128 return sample->pid == 0;
2129}
2130
2131static void save_task_callchain(struct perf_sched *sched,
2132 struct perf_sample *sample,
2133 struct perf_evsel *evsel,
2134 struct machine *machine)
2135{
2136 struct callchain_cursor *cursor = &callchain_cursor;
2137 struct thread *thread;
2138
2139 /* want main thread for process - has maps */
2140 thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2141 if (thread == NULL) {
2142 pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2143 return;
2144 }
2145
2146 if (!symbol_conf.use_callchain || sample->callchain == NULL)
2147 return;
2148
2149 if (thread__resolve_callchain(thread, cursor, evsel, sample,
2150 NULL, NULL, sched->max_stack + 2) != 0) {
2151 if (verbose > 0)
2152 pr_err("Failed to resolve callchain. Skipping\n");
2153
2154 return;
2155 }
2156
2157 callchain_cursor_commit(cursor);
2158
2159 while (true) {
2160 struct callchain_cursor_node *node;
2161 struct symbol *sym;
2162
2163 node = callchain_cursor_current(cursor);
2164 if (node == NULL)
2165 break;
2166
2167 sym = node->sym;
2168 if (sym) {
2169 if (!strcmp(sym->name, "schedule") ||
2170 !strcmp(sym->name, "__schedule") ||
2171 !strcmp(sym->name, "preempt_schedule"))
2172 sym->ignore = 1;
2173 }
2174
2175 callchain_cursor_advance(cursor);
2176 }
2177}
2178
2179static int init_idle_thread(struct thread *thread)
2180{
2181 struct idle_thread_runtime *itr;
2182
2183 thread__set_comm(thread, idle_comm, 0);
2184
2185 itr = zalloc(sizeof(*itr));
2186 if (itr == NULL)
2187 return -ENOMEM;
2188
2189 init_stats(&itr->tr.run_stats);
2190 callchain_init(&itr->callchain);
2191 callchain_cursor_reset(&itr->cursor);
2192 thread__set_priv(thread, itr);
2193
2194 return 0;
2195}
2196
2197/*
2198 * Track idle stats per cpu by maintaining a local thread
2199 * struct for the idle task on each cpu.
2200 */
2201static int init_idle_threads(int ncpu)
2202{
2203 int i, ret;
2204
2205 idle_threads = zalloc(ncpu * sizeof(struct thread *));
2206 if (!idle_threads)
2207 return -ENOMEM;
2208
2209 idle_max_cpu = ncpu;
2210
2211 /* allocate the actual thread struct if needed */
2212 for (i = 0; i < ncpu; ++i) {
2213 idle_threads[i] = thread__new(0, 0);
2214 if (idle_threads[i] == NULL)
2215 return -ENOMEM;
2216
2217 ret = init_idle_thread(idle_threads[i]);
2218 if (ret < 0)
2219 return ret;
2220 }
2221
2222 return 0;
2223}
2224
2225static void free_idle_threads(void)
2226{
2227 int i;
2228
2229 if (idle_threads == NULL)
2230 return;
2231
2232 for (i = 0; i < idle_max_cpu; ++i) {
2233 if ((idle_threads[i]))
2234 thread__delete(idle_threads[i]);
2235 }
2236
2237 free(idle_threads);
2238}
2239
2240static struct thread *get_idle_thread(int cpu)
2241{
2242 /*
2243 * expand/allocate array of pointers to local thread
2244 * structs if needed
2245 */
2246 if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2247 int i, j = __roundup_pow_of_two(cpu+1);
2248 void *p;
2249
2250 p = realloc(idle_threads, j * sizeof(struct thread *));
2251 if (!p)
2252 return NULL;
2253
2254 idle_threads = (struct thread **) p;
2255 for (i = idle_max_cpu; i < j; ++i)
2256 idle_threads[i] = NULL;
2257
2258 idle_max_cpu = j;
2259 }
2260
2261 /* allocate a new thread struct if needed */
2262 if (idle_threads[cpu] == NULL) {
2263 idle_threads[cpu] = thread__new(0, 0);
2264
2265 if (idle_threads[cpu]) {
2266 if (init_idle_thread(idle_threads[cpu]) < 0)
2267 return NULL;
2268 }
2269 }
2270
2271 return idle_threads[cpu];
2272}
2273
2274static void save_idle_callchain(struct idle_thread_runtime *itr,
2275 struct perf_sample *sample)
2276{
2277 if (!symbol_conf.use_callchain || sample->callchain == NULL)
2278 return;
2279
2280 callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2281}
2282
2283static struct thread *timehist_get_thread(struct perf_sched *sched,
2284 struct perf_sample *sample,
2285 struct machine *machine,
2286 struct perf_evsel *evsel)
2287{
2288 struct thread *thread;
2289
2290 if (is_idle_sample(sample, evsel)) {
2291 thread = get_idle_thread(sample->cpu);
2292 if (thread == NULL)
2293 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2294
2295 } else {
2296 /* there were samples with tid 0 but non-zero pid */
2297 thread = machine__findnew_thread(machine, sample->pid,
2298 sample->tid ?: sample->pid);
2299 if (thread == NULL) {
2300 pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2301 sample->tid);
2302 }
2303
2304 save_task_callchain(sched, sample, evsel, machine);
2305 if (sched->idle_hist) {
2306 struct thread *idle;
2307 struct idle_thread_runtime *itr;
2308
2309 idle = get_idle_thread(sample->cpu);
2310 if (idle == NULL) {
2311 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2312 return NULL;
2313 }
2314
2315 itr = thread__priv(idle);
2316 if (itr == NULL)
2317 return NULL;
2318
2319 itr->last_thread = thread;
2320
2321 /* copy task callchain when entering to idle */
2322 if (perf_evsel__intval(evsel, sample, "next_pid") == 0)
2323 save_idle_callchain(itr, sample);
2324 }
2325 }
2326
2327 return thread;
2328}
2329
2330static bool timehist_skip_sample(struct perf_sched *sched,
2331 struct thread *thread,
2332 struct perf_evsel *evsel,
2333 struct perf_sample *sample)
2334{
2335 bool rc = false;
2336
2337 if (thread__is_filtered(thread)) {
2338 rc = true;
2339 sched->skipped_samples++;
2340 }
2341
2342 if (sched->idle_hist) {
2343 if (strcmp(perf_evsel__name(evsel), "sched:sched_switch"))
2344 rc = true;
2345 else if (perf_evsel__intval(evsel, sample, "prev_pid") != 0 &&
2346 perf_evsel__intval(evsel, sample, "next_pid") != 0)
2347 rc = true;
2348 }
2349
2350 return rc;
2351}
2352
2353static void timehist_print_wakeup_event(struct perf_sched *sched,
2354 struct perf_evsel *evsel,
2355 struct perf_sample *sample,
2356 struct machine *machine,
2357 struct thread *awakened)
2358{
2359 struct thread *thread;
2360 char tstr[64];
2361
2362 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2363 if (thread == NULL)
2364 return;
2365
2366 /* show wakeup unless both awakee and awaker are filtered */
2367 if (timehist_skip_sample(sched, thread, evsel, sample) &&
2368 timehist_skip_sample(sched, awakened, evsel, sample)) {
2369 return;
2370 }
2371
2372 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2373 printf("%15s [%04d] ", tstr, sample->cpu);
2374 if (sched->show_cpu_visual)
2375 printf(" %*s ", sched->max_cpu + 1, "");
2376
2377 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2378
2379 /* dt spacer */
2380 printf(" %9s %9s %9s ", "", "", "");
2381
2382 printf("awakened: %s", timehist_get_commstr(awakened));
2383
2384 printf("\n");
2385}
2386
2387static int timehist_sched_wakeup_event(struct perf_tool *tool,
2388 union perf_event *event __maybe_unused,
2389 struct perf_evsel *evsel,
2390 struct perf_sample *sample,
2391 struct machine *machine)
2392{
2393 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2394 struct thread *thread;
2395 struct thread_runtime *tr = NULL;
2396 /* want pid of awakened task not pid in sample */
2397 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2398
2399 thread = machine__findnew_thread(machine, 0, pid);
2400 if (thread == NULL)
2401 return -1;
2402
2403 tr = thread__get_runtime(thread);
2404 if (tr == NULL)
2405 return -1;
2406
2407 if (tr->ready_to_run == 0)
2408 tr->ready_to_run = sample->time;
2409
2410 /* show wakeups if requested */
2411 if (sched->show_wakeups &&
2412 !perf_time__skip_sample(&sched->ptime, sample->time))
2413 timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2414
2415 return 0;
2416}
2417
2418static void timehist_print_migration_event(struct perf_sched *sched,
2419 struct perf_evsel *evsel,
2420 struct perf_sample *sample,
2421 struct machine *machine,
2422 struct thread *migrated)
2423{
2424 struct thread *thread;
2425 char tstr[64];
2426 u32 max_cpus = sched->max_cpu + 1;
2427 u32 ocpu, dcpu;
2428
2429 if (sched->summary_only)
2430 return;
2431
2432 max_cpus = sched->max_cpu + 1;
2433 ocpu = perf_evsel__intval(evsel, sample, "orig_cpu");
2434 dcpu = perf_evsel__intval(evsel, sample, "dest_cpu");
2435
2436 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2437 if (thread == NULL)
2438 return;
2439
2440 if (timehist_skip_sample(sched, thread, evsel, sample) &&
2441 timehist_skip_sample(sched, migrated, evsel, sample)) {
2442 return;
2443 }
2444
2445 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2446 printf("%15s [%04d] ", tstr, sample->cpu);
2447
2448 if (sched->show_cpu_visual) {
2449 u32 i;
2450 char c;
2451
2452 printf(" ");
2453 for (i = 0; i < max_cpus; ++i) {
2454 c = (i == sample->cpu) ? 'm' : ' ';
2455 printf("%c", c);
2456 }
2457 printf(" ");
2458 }
2459
2460 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2461
2462 /* dt spacer */
2463 printf(" %9s %9s %9s ", "", "", "");
2464
2465 printf("migrated: %s", timehist_get_commstr(migrated));
2466 printf(" cpu %d => %d", ocpu, dcpu);
2467
2468 printf("\n");
2469}
2470
2471static int timehist_migrate_task_event(struct perf_tool *tool,
2472 union perf_event *event __maybe_unused,
2473 struct perf_evsel *evsel,
2474 struct perf_sample *sample,
2475 struct machine *machine)
2476{
2477 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2478 struct thread *thread;
2479 struct thread_runtime *tr = NULL;
2480 /* want pid of migrated task not pid in sample */
2481 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2482
2483 thread = machine__findnew_thread(machine, 0, pid);
2484 if (thread == NULL)
2485 return -1;
2486
2487 tr = thread__get_runtime(thread);
2488 if (tr == NULL)
2489 return -1;
2490
2491 tr->migrations++;
2492
2493 /* show migrations if requested */
2494 timehist_print_migration_event(sched, evsel, sample, machine, thread);
2495
2496 return 0;
2497}
2498
2499static int timehist_sched_change_event(struct perf_tool *tool,
2500 union perf_event *event,
2501 struct perf_evsel *evsel,
2502 struct perf_sample *sample,
2503 struct machine *machine)
2504{
2505 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2506 struct perf_time_interval *ptime = &sched->ptime;
2507 struct addr_location al;
2508 struct thread *thread;
2509 struct thread_runtime *tr = NULL;
2510 u64 tprev, t = sample->time;
2511 int rc = 0;
2512 int state = perf_evsel__intval(evsel, sample, "prev_state");
2513
2514
2515 if (machine__resolve(machine, &al, sample) < 0) {
2516 pr_err("problem processing %d event. skipping it\n",
2517 event->header.type);
2518 rc = -1;
2519 goto out;
2520 }
2521
2522 thread = timehist_get_thread(sched, sample, machine, evsel);
2523 if (thread == NULL) {
2524 rc = -1;
2525 goto out;
2526 }
2527
2528 if (timehist_skip_sample(sched, thread, evsel, sample))
2529 goto out;
2530
2531 tr = thread__get_runtime(thread);
2532 if (tr == NULL) {
2533 rc = -1;
2534 goto out;
2535 }
2536
2537 tprev = perf_evsel__get_time(evsel, sample->cpu);
2538
2539 /*
2540 * If start time given:
2541 * - sample time is under window user cares about - skip sample
2542 * - tprev is under window user cares about - reset to start of window
2543 */
2544 if (ptime->start && ptime->start > t)
2545 goto out;
2546
2547 if (tprev && ptime->start > tprev)
2548 tprev = ptime->start;
2549
2550 /*
2551 * If end time given:
2552 * - previous sched event is out of window - we are done
2553 * - sample time is beyond window user cares about - reset it
2554 * to close out stats for time window interest
2555 */
2556 if (ptime->end) {
2557 if (tprev > ptime->end)
2558 goto out;
2559
2560 if (t > ptime->end)
2561 t = ptime->end;
2562 }
2563
2564 if (!sched->idle_hist || thread->tid == 0) {
2565 timehist_update_runtime_stats(tr, t, tprev);
2566
2567 if (sched->idle_hist) {
2568 struct idle_thread_runtime *itr = (void *)tr;
2569 struct thread_runtime *last_tr;
2570
2571 BUG_ON(thread->tid != 0);
2572
2573 if (itr->last_thread == NULL)
2574 goto out;
2575
2576 /* add current idle time as last thread's runtime */
2577 last_tr = thread__get_runtime(itr->last_thread);
2578 if (last_tr == NULL)
2579 goto out;
2580
2581 timehist_update_runtime_stats(last_tr, t, tprev);
2582 /*
2583 * remove delta time of last thread as it's not updated
2584 * and otherwise it will show an invalid value next
2585 * time. we only care total run time and run stat.
2586 */
2587 last_tr->dt_run = 0;
2588 last_tr->dt_delay = 0;
2589 last_tr->dt_sleep = 0;
2590 last_tr->dt_iowait = 0;
2591 last_tr->dt_preempt = 0;
2592
2593 if (itr->cursor.nr)
2594 callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2595
2596 itr->last_thread = NULL;
2597 }
2598 }
2599
2600 if (!sched->summary_only)
2601 timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2602
2603out:
2604 if (sched->hist_time.start == 0 && t >= ptime->start)
2605 sched->hist_time.start = t;
2606 if (ptime->end == 0 || t <= ptime->end)
2607 sched->hist_time.end = t;
2608
2609 if (tr) {
2610 /* time of this sched_switch event becomes last time task seen */
2611 tr->last_time = sample->time;
2612
2613 /* last state is used to determine where to account wait time */
2614 tr->last_state = state;
2615
2616 /* sched out event for task so reset ready to run time */
2617 tr->ready_to_run = 0;
2618 }
2619
2620 perf_evsel__save_time(evsel, sample->time, sample->cpu);
2621
2622 return rc;
2623}
2624
2625static int timehist_sched_switch_event(struct perf_tool *tool,
2626 union perf_event *event,
2627 struct perf_evsel *evsel,
2628 struct perf_sample *sample,
2629 struct machine *machine __maybe_unused)
2630{
2631 return timehist_sched_change_event(tool, event, evsel, sample, machine);
2632}
2633
2634static int process_lost(struct perf_tool *tool __maybe_unused,
2635 union perf_event *event,
2636 struct perf_sample *sample,
2637 struct machine *machine __maybe_unused)
2638{
2639 char tstr[64];
2640
2641 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2642 printf("%15s ", tstr);
2643 printf("lost %" PRIu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2644
2645 return 0;
2646}
2647
2648
2649static void print_thread_runtime(struct thread *t,
2650 struct thread_runtime *r)
2651{
2652 double mean = avg_stats(&r->run_stats);
2653 float stddev;
2654
2655 printf("%*s %5d %9" PRIu64 " ",
2656 comm_width, timehist_get_commstr(t), t->ppid,
2657 (u64) r->run_stats.n);
2658
2659 print_sched_time(r->total_run_time, 8);
2660 stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2661 print_sched_time(r->run_stats.min, 6);
2662 printf(" ");
2663 print_sched_time((u64) mean, 6);
2664 printf(" ");
2665 print_sched_time(r->run_stats.max, 6);
2666 printf(" ");
2667 printf("%5.2f", stddev);
2668 printf(" %5" PRIu64, r->migrations);
2669 printf("\n");
2670}
2671
2672static void print_thread_waittime(struct thread *t,
2673 struct thread_runtime *r)
2674{
2675 printf("%*s %5d %9" PRIu64 " ",
2676 comm_width, timehist_get_commstr(t), t->ppid,
2677 (u64) r->run_stats.n);
2678
2679 print_sched_time(r->total_run_time, 8);
2680 print_sched_time(r->total_sleep_time, 6);
2681 printf(" ");
2682 print_sched_time(r->total_iowait_time, 6);
2683 printf(" ");
2684 print_sched_time(r->total_preempt_time, 6);
2685 printf(" ");
2686 print_sched_time(r->total_delay_time, 6);
2687 printf("\n");
2688}
2689
2690struct total_run_stats {
2691 struct perf_sched *sched;
2692 u64 sched_count;
2693 u64 task_count;
2694 u64 total_run_time;
2695};
2696
2697static int __show_thread_runtime(struct thread *t, void *priv)
2698{
2699 struct total_run_stats *stats = priv;
2700 struct thread_runtime *r;
2701
2702 if (thread__is_filtered(t))
2703 return 0;
2704
2705 r = thread__priv(t);
2706 if (r && r->run_stats.n) {
2707 stats->task_count++;
2708 stats->sched_count += r->run_stats.n;
2709 stats->total_run_time += r->total_run_time;
2710
2711 if (stats->sched->show_state)
2712 print_thread_waittime(t, r);
2713 else
2714 print_thread_runtime(t, r);
2715 }
2716
2717 return 0;
2718}
2719
2720static int show_thread_runtime(struct thread *t, void *priv)
2721{
2722 if (t->dead)
2723 return 0;
2724
2725 return __show_thread_runtime(t, priv);
2726}
2727
2728static int show_deadthread_runtime(struct thread *t, void *priv)
2729{
2730 if (!t->dead)
2731 return 0;
2732
2733 return __show_thread_runtime(t, priv);
2734}
2735
2736static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2737{
2738 const char *sep = " <- ";
2739 struct callchain_list *chain;
2740 size_t ret = 0;
2741 char bf[1024];
2742 bool first;
2743
2744 if (node == NULL)
2745 return 0;
2746
2747 ret = callchain__fprintf_folded(fp, node->parent);
2748 first = (ret == 0);
2749
2750 list_for_each_entry(chain, &node->val, list) {
2751 if (chain->ip >= PERF_CONTEXT_MAX)
2752 continue;
2753 if (chain->ms.sym && chain->ms.sym->ignore)
2754 continue;
2755 ret += fprintf(fp, "%s%s", first ? "" : sep,
2756 callchain_list__sym_name(chain, bf, sizeof(bf),
2757 false));
2758 first = false;
2759 }
2760
2761 return ret;
2762}
2763
2764static size_t timehist_print_idlehist_callchain(struct rb_root *root)
2765{
2766 size_t ret = 0;
2767 FILE *fp = stdout;
2768 struct callchain_node *chain;
2769 struct rb_node *rb_node = rb_first(root);
2770
2771 printf(" %16s %8s %s\n", "Idle time (msec)", "Count", "Callchains");
2772 printf(" %.16s %.8s %.50s\n", graph_dotted_line, graph_dotted_line,
2773 graph_dotted_line);
2774
2775 while (rb_node) {
2776 chain = rb_entry(rb_node, struct callchain_node, rb_node);
2777 rb_node = rb_next(rb_node);
2778
2779 ret += fprintf(fp, " ");
2780 print_sched_time(chain->hit, 12);
2781 ret += 16; /* print_sched_time returns 2nd arg + 4 */
2782 ret += fprintf(fp, " %8d ", chain->count);
2783 ret += callchain__fprintf_folded(fp, chain);
2784 ret += fprintf(fp, "\n");
2785 }
2786
2787 return ret;
2788}
2789
2790static void timehist_print_summary(struct perf_sched *sched,
2791 struct perf_session *session)
2792{
2793 struct machine *m = &session->machines.host;
2794 struct total_run_stats totals;
2795 u64 task_count;
2796 struct thread *t;
2797 struct thread_runtime *r;
2798 int i;
2799 u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2800
2801 memset(&totals, 0, sizeof(totals));
2802 totals.sched = sched;
2803
2804 if (sched->idle_hist) {
2805 printf("\nIdle-time summary\n");
2806 printf("%*s parent sched-out ", comm_width, "comm");
2807 printf(" idle-time min-idle avg-idle max-idle stddev migrations\n");
2808 } else if (sched->show_state) {
2809 printf("\nWait-time summary\n");
2810 printf("%*s parent sched-in ", comm_width, "comm");
2811 printf(" run-time sleep iowait preempt delay\n");
2812 } else {
2813 printf("\nRuntime summary\n");
2814 printf("%*s parent sched-in ", comm_width, "comm");
2815 printf(" run-time min-run avg-run max-run stddev migrations\n");
2816 }
2817 printf("%*s (count) ", comm_width, "");
2818 printf(" (msec) (msec) (msec) (msec) %s\n",
2819 sched->show_state ? "(msec)" : "%");
2820 printf("%.117s\n", graph_dotted_line);
2821
2822 machine__for_each_thread(m, show_thread_runtime, &totals);
2823 task_count = totals.task_count;
2824 if (!task_count)
2825 printf("<no still running tasks>\n");
2826
2827 printf("\nTerminated tasks:\n");
2828 machine__for_each_thread(m, show_deadthread_runtime, &totals);
2829 if (task_count == totals.task_count)
2830 printf("<no terminated tasks>\n");
2831
2832 /* CPU idle stats not tracked when samples were skipped */
2833 if (sched->skipped_samples && !sched->idle_hist)
2834 return;
2835
2836 printf("\nIdle stats:\n");
2837 for (i = 0; i < idle_max_cpu; ++i) {
2838 t = idle_threads[i];
2839 if (!t)
2840 continue;
2841
2842 r = thread__priv(t);
2843 if (r && r->run_stats.n) {
2844 totals.sched_count += r->run_stats.n;
2845 printf(" CPU %2d idle for ", i);
2846 print_sched_time(r->total_run_time, 6);
2847 printf(" msec (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2848 } else
2849 printf(" CPU %2d idle entire time window\n", i);
2850 }
2851
2852 if (sched->idle_hist && symbol_conf.use_callchain) {
2853 callchain_param.mode = CHAIN_FOLDED;
2854 callchain_param.value = CCVAL_PERIOD;
2855
2856 callchain_register_param(&callchain_param);
2857
2858 printf("\nIdle stats by callchain:\n");
2859 for (i = 0; i < idle_max_cpu; ++i) {
2860 struct idle_thread_runtime *itr;
2861
2862 t = idle_threads[i];
2863 if (!t)
2864 continue;
2865
2866 itr = thread__priv(t);
2867 if (itr == NULL)
2868 continue;
2869
2870 callchain_param.sort(&itr->sorted_root, &itr->callchain,
2871 0, &callchain_param);
2872
2873 printf(" CPU %2d:", i);
2874 print_sched_time(itr->tr.total_run_time, 6);
2875 printf(" msec\n");
2876 timehist_print_idlehist_callchain(&itr->sorted_root);
2877 printf("\n");
2878 }
2879 }
2880
2881 printf("\n"
2882 " Total number of unique tasks: %" PRIu64 "\n"
2883 "Total number of context switches: %" PRIu64 "\n",
2884 totals.task_count, totals.sched_count);
2885
2886 printf(" Total run time (msec): ");
2887 print_sched_time(totals.total_run_time, 2);
2888 printf("\n");
2889
2890 printf(" Total scheduling time (msec): ");
2891 print_sched_time(hist_time, 2);
2892 printf(" (x %d)\n", sched->max_cpu);
2893}
2894
2895typedef int (*sched_handler)(struct perf_tool *tool,
2896 union perf_event *event,
2897 struct perf_evsel *evsel,
2898 struct perf_sample *sample,
2899 struct machine *machine);
2900
2901static int perf_timehist__process_sample(struct perf_tool *tool,
2902 union perf_event *event,
2903 struct perf_sample *sample,
2904 struct perf_evsel *evsel,
2905 struct machine *machine)
2906{
2907 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2908 int err = 0;
2909 int this_cpu = sample->cpu;
2910
2911 if (this_cpu > sched->max_cpu)
2912 sched->max_cpu = this_cpu;
2913
2914 if (evsel->handler != NULL) {
2915 sched_handler f = evsel->handler;
2916
2917 err = f(tool, event, evsel, sample, machine);
2918 }
2919
2920 return err;
2921}
2922
2923static int timehist_check_attr(struct perf_sched *sched,
2924 struct perf_evlist *evlist)
2925{
2926 struct perf_evsel *evsel;
2927 struct evsel_runtime *er;
2928
2929 list_for_each_entry(evsel, &evlist->entries, node) {
2930 er = perf_evsel__get_runtime(evsel);
2931 if (er == NULL) {
2932 pr_err("Failed to allocate memory for evsel runtime data\n");
2933 return -1;
2934 }
2935
2936 if (sched->show_callchain &&
2937 !(evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN)) {
2938 pr_info("Samples do not have callchains.\n");
2939 sched->show_callchain = 0;
2940 symbol_conf.use_callchain = 0;
2941 }
2942 }
2943
2944 return 0;
2945}
2946
2947static int perf_sched__timehist(struct perf_sched *sched)
2948{
2949 const struct perf_evsel_str_handler handlers[] = {
2950 { "sched:sched_switch", timehist_sched_switch_event, },
2951 { "sched:sched_wakeup", timehist_sched_wakeup_event, },
2952 { "sched:sched_wakeup_new", timehist_sched_wakeup_event, },
2953 };
2954 const struct perf_evsel_str_handler migrate_handlers[] = {
2955 { "sched:sched_migrate_task", timehist_migrate_task_event, },
2956 };
2957 struct perf_data data = {
2958 .file = {
2959 .path = input_name,
2960 },
2961 .mode = PERF_DATA_MODE_READ,
2962 .force = sched->force,
2963 };
2964
2965 struct perf_session *session;
2966 struct perf_evlist *evlist;
2967 int err = -1;
2968
2969 /*
2970 * event handlers for timehist option
2971 */
2972 sched->tool.sample = perf_timehist__process_sample;
2973 sched->tool.mmap = perf_event__process_mmap;
2974 sched->tool.comm = perf_event__process_comm;
2975 sched->tool.exit = perf_event__process_exit;
2976 sched->tool.fork = perf_event__process_fork;
2977 sched->tool.lost = process_lost;
2978 sched->tool.attr = perf_event__process_attr;
2979 sched->tool.tracing_data = perf_event__process_tracing_data;
2980 sched->tool.build_id = perf_event__process_build_id;
2981
2982 sched->tool.ordered_events = true;
2983 sched->tool.ordering_requires_timestamps = true;
2984
2985 symbol_conf.use_callchain = sched->show_callchain;
2986
2987 session = perf_session__new(&data, false, &sched->tool);
2988 if (session == NULL)
2989 return -ENOMEM;
2990
2991 evlist = session->evlist;
2992
2993 symbol__init(&session->header.env);
2994
2995 if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
2996 pr_err("Invalid time string\n");
2997 return -EINVAL;
2998 }
2999
3000 if (timehist_check_attr(sched, evlist) != 0)
3001 goto out;
3002
3003 setup_pager();
3004
3005 /* setup per-evsel handlers */
3006 if (perf_session__set_tracepoints_handlers(session, handlers))
3007 goto out;
3008
3009 /* sched_switch event at a minimum needs to exist */
3010 if (!perf_evlist__find_tracepoint_by_name(session->evlist,
3011 "sched:sched_switch")) {
3012 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3013 goto out;
3014 }
3015
3016 if (sched->show_migrations &&
3017 perf_session__set_tracepoints_handlers(session, migrate_handlers))
3018 goto out;
3019
3020 /* pre-allocate struct for per-CPU idle stats */
3021 sched->max_cpu = session->header.env.nr_cpus_online;
3022 if (sched->max_cpu == 0)
3023 sched->max_cpu = 4;
3024 if (init_idle_threads(sched->max_cpu))
3025 goto out;
3026
3027 /* summary_only implies summary option, but don't overwrite summary if set */
3028 if (sched->summary_only)
3029 sched->summary = sched->summary_only;
3030
3031 if (!sched->summary_only)
3032 timehist_header(sched);
3033
3034 err = perf_session__process_events(session);
3035 if (err) {
3036 pr_err("Failed to process events, error %d", err);
3037 goto out;
3038 }
3039
3040 sched->nr_events = evlist->stats.nr_events[0];
3041 sched->nr_lost_events = evlist->stats.total_lost;
3042 sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3043
3044 if (sched->summary)
3045 timehist_print_summary(sched, session);
3046
3047out:
3048 free_idle_threads();
3049 perf_session__delete(session);
3050
3051 return err;
3052}
3053
3054
3055static void print_bad_events(struct perf_sched *sched)
3056{
3057 if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3058 printf(" INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3059 (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3060 sched->nr_unordered_timestamps, sched->nr_timestamps);
3061 }
3062 if (sched->nr_lost_events && sched->nr_events) {
3063 printf(" INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3064 (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3065 sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3066 }
3067 if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3068 printf(" INFO: %.3f%% context switch bugs (%ld out of %ld)",
3069 (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3070 sched->nr_context_switch_bugs, sched->nr_timestamps);
3071 if (sched->nr_lost_events)
3072 printf(" (due to lost events?)");
3073 printf("\n");
3074 }
3075}
3076
3077static void __merge_work_atoms(struct rb_root *root, struct work_atoms *data)
3078{
3079 struct rb_node **new = &(root->rb_node), *parent = NULL;
3080 struct work_atoms *this;
3081 const char *comm = thread__comm_str(data->thread), *this_comm;
3082
3083 while (*new) {
3084 int cmp;
3085
3086 this = container_of(*new, struct work_atoms, node);
3087 parent = *new;
3088
3089 this_comm = thread__comm_str(this->thread);
3090 cmp = strcmp(comm, this_comm);
3091 if (cmp > 0) {
3092 new = &((*new)->rb_left);
3093 } else if (cmp < 0) {
3094 new = &((*new)->rb_right);
3095 } else {
3096 this->num_merged++;
3097 this->total_runtime += data->total_runtime;
3098 this->nb_atoms += data->nb_atoms;
3099 this->total_lat += data->total_lat;
3100 list_splice(&data->work_list, &this->work_list);
3101 if (this->max_lat < data->max_lat) {
3102 this->max_lat = data->max_lat;
3103 this->max_lat_at = data->max_lat_at;
3104 }
3105 zfree(&data);
3106 return;
3107 }
3108 }
3109
3110 data->num_merged++;
3111 rb_link_node(&data->node, parent, new);
3112 rb_insert_color(&data->node, root);
3113}
3114
3115static void perf_sched__merge_lat(struct perf_sched *sched)
3116{
3117 struct work_atoms *data;
3118 struct rb_node *node;
3119
3120 if (sched->skip_merge)
3121 return;
3122
3123 while ((node = rb_first(&sched->atom_root))) {
3124 rb_erase(node, &sched->atom_root);
3125 data = rb_entry(node, struct work_atoms, node);
3126 __merge_work_atoms(&sched->merged_atom_root, data);
3127 }
3128}
3129
3130static int perf_sched__lat(struct perf_sched *sched)
3131{
3132 struct rb_node *next;
3133
3134 setup_pager();
3135
3136 if (perf_sched__read_events(sched))
3137 return -1;
3138
3139 perf_sched__merge_lat(sched);
3140 perf_sched__sort_lat(sched);
3141
3142 printf("\n -----------------------------------------------------------------------------------------------------------------\n");
3143 printf(" Task | Runtime ms | Switches | Average delay ms | Maximum delay ms | Maximum delay at |\n");
3144 printf(" -----------------------------------------------------------------------------------------------------------------\n");
3145
3146 next = rb_first(&sched->sorted_atom_root);
3147
3148 while (next) {
3149 struct work_atoms *work_list;
3150
3151 work_list = rb_entry(next, struct work_atoms, node);
3152 output_lat_thread(sched, work_list);
3153 next = rb_next(next);
3154 thread__zput(work_list->thread);
3155 }
3156
3157 printf(" -----------------------------------------------------------------------------------------------------------------\n");
3158 printf(" TOTAL: |%11.3f ms |%9" PRIu64 " |\n",
3159 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3160
3161 printf(" ---------------------------------------------------\n");
3162
3163 print_bad_events(sched);
3164 printf("\n");
3165
3166 return 0;
3167}
3168
3169static int setup_map_cpus(struct perf_sched *sched)
3170{
3171 struct cpu_map *map;
3172
3173 sched->max_cpu = sysconf(_SC_NPROCESSORS_CONF);
3174
3175 if (sched->map.comp) {
3176 sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
3177 if (!sched->map.comp_cpus)
3178 return -1;
3179 }
3180
3181 if (!sched->map.cpus_str)
3182 return 0;
3183
3184 map = cpu_map__new(sched->map.cpus_str);
3185 if (!map) {
3186 pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3187 return -1;
3188 }
3189
3190 sched->map.cpus = map;
3191 return 0;
3192}
3193
3194static int setup_color_pids(struct perf_sched *sched)
3195{
3196 struct thread_map *map;
3197
3198 if (!sched->map.color_pids_str)
3199 return 0;
3200
3201 map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3202 if (!map) {
3203 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3204 return -1;
3205 }
3206
3207 sched->map.color_pids = map;
3208 return 0;
3209}
3210
3211static int setup_color_cpus(struct perf_sched *sched)
3212{
3213 struct cpu_map *map;
3214
3215 if (!sched->map.color_cpus_str)
3216 return 0;
3217
3218 map = cpu_map__new(sched->map.color_cpus_str);
3219 if (!map) {
3220 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3221 return -1;
3222 }
3223
3224 sched->map.color_cpus = map;
3225 return 0;
3226}
3227
3228static int perf_sched__map(struct perf_sched *sched)
3229{
3230 if (setup_map_cpus(sched))
3231 return -1;
3232
3233 if (setup_color_pids(sched))
3234 return -1;
3235
3236 if (setup_color_cpus(sched))
3237 return -1;
3238
3239 setup_pager();
3240 if (perf_sched__read_events(sched))
3241 return -1;
3242 print_bad_events(sched);
3243 return 0;
3244}
3245
3246static int perf_sched__replay(struct perf_sched *sched)
3247{
3248 unsigned long i;
3249
3250 calibrate_run_measurement_overhead(sched);
3251 calibrate_sleep_measurement_overhead(sched);
3252
3253 test_calibrations(sched);
3254
3255 if (perf_sched__read_events(sched))
3256 return -1;
3257
3258 printf("nr_run_events: %ld\n", sched->nr_run_events);
3259 printf("nr_sleep_events: %ld\n", sched->nr_sleep_events);
3260 printf("nr_wakeup_events: %ld\n", sched->nr_wakeup_events);
3261
3262 if (sched->targetless_wakeups)
3263 printf("target-less wakeups: %ld\n", sched->targetless_wakeups);
3264 if (sched->multitarget_wakeups)
3265 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3266 if (sched->nr_run_events_optimized)
3267 printf("run atoms optimized: %ld\n",
3268 sched->nr_run_events_optimized);
3269
3270 print_task_traces(sched);
3271 add_cross_task_wakeups(sched);
3272
3273 create_tasks(sched);
3274 printf("------------------------------------------------------------\n");
3275 for (i = 0; i < sched->replay_repeat; i++)
3276 run_one_test(sched);
3277
3278 return 0;
3279}
3280
3281static void setup_sorting(struct perf_sched *sched, const struct option *options,
3282 const char * const usage_msg[])
3283{
3284 char *tmp, *tok, *str = strdup(sched->sort_order);
3285
3286 for (tok = strtok_r(str, ", ", &tmp);
3287 tok; tok = strtok_r(NULL, ", ", &tmp)) {
3288 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3289 usage_with_options_msg(usage_msg, options,
3290 "Unknown --sort key: `%s'", tok);
3291 }
3292 }
3293
3294 free(str);
3295
3296 sort_dimension__add("pid", &sched->cmp_pid);
3297}
3298
3299static int __cmd_record(int argc, const char **argv)
3300{
3301 unsigned int rec_argc, i, j;
3302 const char **rec_argv;
3303 const char * const record_args[] = {
3304 "record",
3305 "-a",
3306 "-R",
3307 "-m", "1024",
3308 "-c", "1",
3309 "-e", "sched:sched_switch",
3310 "-e", "sched:sched_stat_wait",
3311 "-e", "sched:sched_stat_sleep",
3312 "-e", "sched:sched_stat_iowait",
3313 "-e", "sched:sched_stat_runtime",
3314 "-e", "sched:sched_process_fork",
3315 "-e", "sched:sched_wakeup",
3316 "-e", "sched:sched_wakeup_new",
3317 "-e", "sched:sched_migrate_task",
3318 };
3319
3320 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
3321 rec_argv = calloc(rec_argc + 1, sizeof(char *));
3322
3323 if (rec_argv == NULL)
3324 return -ENOMEM;
3325
3326 for (i = 0; i < ARRAY_SIZE(record_args); i++)
3327 rec_argv[i] = strdup(record_args[i]);
3328
3329 for (j = 1; j < (unsigned int)argc; j++, i++)
3330 rec_argv[i] = argv[j];
3331
3332 BUG_ON(i != rec_argc);
3333
3334 return cmd_record(i, rec_argv);
3335}
3336
3337int cmd_sched(int argc, const char **argv)
3338{
3339 const char default_sort_order[] = "avg, max, switch, runtime";
3340 struct perf_sched sched = {
3341 .tool = {
3342 .sample = perf_sched__process_tracepoint_sample,
3343 .comm = perf_sched__process_comm,
3344 .namespaces = perf_event__process_namespaces,
3345 .lost = perf_event__process_lost,
3346 .fork = perf_sched__process_fork_event,
3347 .ordered_events = true,
3348 },
3349 .cmp_pid = LIST_HEAD_INIT(sched.cmp_pid),
3350 .sort_list = LIST_HEAD_INIT(sched.sort_list),
3351 .start_work_mutex = PTHREAD_MUTEX_INITIALIZER,
3352 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
3353 .sort_order = default_sort_order,
3354 .replay_repeat = 10,
3355 .profile_cpu = -1,
3356 .next_shortname1 = 'A',
3357 .next_shortname2 = '0',
3358 .skip_merge = 0,
3359 .show_callchain = 1,
3360 .max_stack = 5,
3361 };
3362 const struct option sched_options[] = {
3363 OPT_STRING('i', "input", &input_name, "file",
3364 "input file name"),
3365 OPT_INCR('v', "verbose", &verbose,
3366 "be more verbose (show symbol address, etc)"),
3367 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3368 "dump raw trace in ASCII"),
3369 OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3370 OPT_END()
3371 };
3372 const struct option latency_options[] = {
3373 OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3374 "sort by key(s): runtime, switch, avg, max"),
3375 OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3376 "CPU to profile on"),
3377 OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3378 "latency stats per pid instead of per comm"),
3379 OPT_PARENT(sched_options)
3380 };
3381 const struct option replay_options[] = {
3382 OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3383 "repeat the workload replay N times (-1: infinite)"),
3384 OPT_PARENT(sched_options)
3385 };
3386 const struct option map_options[] = {
3387 OPT_BOOLEAN(0, "compact", &sched.map.comp,
3388 "map output in compact mode"),
3389 OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3390 "highlight given pids in map"),
3391 OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3392 "highlight given CPUs in map"),
3393 OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3394 "display given CPUs in map"),
3395 OPT_PARENT(sched_options)
3396 };
3397 const struct option timehist_options[] = {
3398 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3399 "file", "vmlinux pathname"),
3400 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3401 "file", "kallsyms pathname"),
3402 OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3403 "Display call chains if present (default on)"),
3404 OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3405 "Maximum number of functions to display backtrace."),
3406 OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3407 "Look for files with symbols relative to this directory"),
3408 OPT_BOOLEAN('s', "summary", &sched.summary_only,
3409 "Show only syscall summary with statistics"),
3410 OPT_BOOLEAN('S', "with-summary", &sched.summary,
3411 "Show all syscalls and summary with statistics"),
3412 OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3413 OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3414 OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3415 OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3416 OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3417 OPT_STRING(0, "time", &sched.time_str, "str",
3418 "Time span for analysis (start,stop)"),
3419 OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3420 OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3421 "analyze events only for given process id(s)"),
3422 OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3423 "analyze events only for given thread id(s)"),
3424 OPT_PARENT(sched_options)
3425 };
3426
3427 const char * const latency_usage[] = {
3428 "perf sched latency [<options>]",
3429 NULL
3430 };
3431 const char * const replay_usage[] = {
3432 "perf sched replay [<options>]",
3433 NULL
3434 };
3435 const char * const map_usage[] = {
3436 "perf sched map [<options>]",
3437 NULL
3438 };
3439 const char * const timehist_usage[] = {
3440 "perf sched timehist [<options>]",
3441 NULL
3442 };
3443 const char *const sched_subcommands[] = { "record", "latency", "map",
3444 "replay", "script",
3445 "timehist", NULL };
3446 const char *sched_usage[] = {
3447 NULL,
3448 NULL
3449 };
3450 struct trace_sched_handler lat_ops = {
3451 .wakeup_event = latency_wakeup_event,
3452 .switch_event = latency_switch_event,
3453 .runtime_event = latency_runtime_event,
3454 .migrate_task_event = latency_migrate_task_event,
3455 };
3456 struct trace_sched_handler map_ops = {
3457 .switch_event = map_switch_event,
3458 };
3459 struct trace_sched_handler replay_ops = {
3460 .wakeup_event = replay_wakeup_event,
3461 .switch_event = replay_switch_event,
3462 .fork_event = replay_fork_event,
3463 };
3464 unsigned int i;
3465
3466 for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3467 sched.curr_pid[i] = -1;
3468
3469 argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3470 sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3471 if (!argc)
3472 usage_with_options(sched_usage, sched_options);
3473
3474 /*
3475 * Aliased to 'perf script' for now:
3476 */
3477 if (!strcmp(argv[0], "script"))
3478 return cmd_script(argc, argv);
3479
3480 if (!strncmp(argv[0], "rec", 3)) {
3481 return __cmd_record(argc, argv);
3482 } else if (!strncmp(argv[0], "lat", 3)) {
3483 sched.tp_handler = &lat_ops;
3484 if (argc > 1) {
3485 argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3486 if (argc)
3487 usage_with_options(latency_usage, latency_options);
3488 }
3489 setup_sorting(&sched, latency_options, latency_usage);
3490 return perf_sched__lat(&sched);
3491 } else if (!strcmp(argv[0], "map")) {
3492 if (argc) {
3493 argc = parse_options(argc, argv, map_options, map_usage, 0);
3494 if (argc)
3495 usage_with_options(map_usage, map_options);
3496 }
3497 sched.tp_handler = &map_ops;
3498 setup_sorting(&sched, latency_options, latency_usage);
3499 return perf_sched__map(&sched);
3500 } else if (!strncmp(argv[0], "rep", 3)) {
3501 sched.tp_handler = &replay_ops;
3502 if (argc) {
3503 argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3504 if (argc)
3505 usage_with_options(replay_usage, replay_options);
3506 }
3507 return perf_sched__replay(&sched);
3508 } else if (!strcmp(argv[0], "timehist")) {
3509 if (argc) {
3510 argc = parse_options(argc, argv, timehist_options,
3511 timehist_usage, 0);
3512 if (argc)
3513 usage_with_options(timehist_usage, timehist_options);
3514 }
3515 if ((sched.show_wakeups || sched.show_next) &&
3516 sched.summary_only) {
3517 pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3518 parse_options_usage(timehist_usage, timehist_options, "s", true);
3519 if (sched.show_wakeups)
3520 parse_options_usage(NULL, timehist_options, "w", true);
3521 if (sched.show_next)
3522 parse_options_usage(NULL, timehist_options, "n", true);
3523 return -EINVAL;
3524 }
3525
3526 return perf_sched__timehist(&sched);
3527 } else {
3528 usage_with_options(sched_usage, sched_options);
3529 }
3530
3531 return 0;
3532}
1#include "builtin.h"
2#include "perf.h"
3
4#include "util/util.h"
5#include "util/evlist.h"
6#include "util/cache.h"
7#include "util/evsel.h"
8#include "util/symbol.h"
9#include "util/thread.h"
10#include "util/header.h"
11#include "util/session.h"
12#include "util/tool.h"
13#include "util/cloexec.h"
14
15#include <subcmd/parse-options.h>
16#include "util/trace-event.h"
17
18#include "util/debug.h"
19
20#include <sys/prctl.h>
21#include <sys/resource.h>
22
23#include <semaphore.h>
24#include <pthread.h>
25#include <math.h>
26#include <api/fs/fs.h>
27
28#define PR_SET_NAME 15 /* Set process name */
29#define MAX_CPUS 4096
30#define COMM_LEN 20
31#define SYM_LEN 129
32#define MAX_PID 1024000
33
34struct sched_atom;
35
36struct task_desc {
37 unsigned long nr;
38 unsigned long pid;
39 char comm[COMM_LEN];
40
41 unsigned long nr_events;
42 unsigned long curr_event;
43 struct sched_atom **atoms;
44
45 pthread_t thread;
46 sem_t sleep_sem;
47
48 sem_t ready_for_work;
49 sem_t work_done_sem;
50
51 u64 cpu_usage;
52};
53
54enum sched_event_type {
55 SCHED_EVENT_RUN,
56 SCHED_EVENT_SLEEP,
57 SCHED_EVENT_WAKEUP,
58 SCHED_EVENT_MIGRATION,
59};
60
61struct sched_atom {
62 enum sched_event_type type;
63 int specific_wait;
64 u64 timestamp;
65 u64 duration;
66 unsigned long nr;
67 sem_t *wait_sem;
68 struct task_desc *wakee;
69};
70
71#define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
72
73enum thread_state {
74 THREAD_SLEEPING = 0,
75 THREAD_WAIT_CPU,
76 THREAD_SCHED_IN,
77 THREAD_IGNORE
78};
79
80struct work_atom {
81 struct list_head list;
82 enum thread_state state;
83 u64 sched_out_time;
84 u64 wake_up_time;
85 u64 sched_in_time;
86 u64 runtime;
87};
88
89struct work_atoms {
90 struct list_head work_list;
91 struct thread *thread;
92 struct rb_node node;
93 u64 max_lat;
94 u64 max_lat_at;
95 u64 total_lat;
96 u64 nb_atoms;
97 u64 total_runtime;
98 int num_merged;
99};
100
101typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
102
103struct perf_sched;
104
105struct trace_sched_handler {
106 int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
107 struct perf_sample *sample, struct machine *machine);
108
109 int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
110 struct perf_sample *sample, struct machine *machine);
111
112 int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
113 struct perf_sample *sample, struct machine *machine);
114
115 /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
116 int (*fork_event)(struct perf_sched *sched, union perf_event *event,
117 struct machine *machine);
118
119 int (*migrate_task_event)(struct perf_sched *sched,
120 struct perf_evsel *evsel,
121 struct perf_sample *sample,
122 struct machine *machine);
123};
124
125struct perf_sched {
126 struct perf_tool tool;
127 const char *sort_order;
128 unsigned long nr_tasks;
129 struct task_desc **pid_to_task;
130 struct task_desc **tasks;
131 const struct trace_sched_handler *tp_handler;
132 pthread_mutex_t start_work_mutex;
133 pthread_mutex_t work_done_wait_mutex;
134 int profile_cpu;
135/*
136 * Track the current task - that way we can know whether there's any
137 * weird events, such as a task being switched away that is not current.
138 */
139 int max_cpu;
140 u32 curr_pid[MAX_CPUS];
141 struct thread *curr_thread[MAX_CPUS];
142 char next_shortname1;
143 char next_shortname2;
144 unsigned int replay_repeat;
145 unsigned long nr_run_events;
146 unsigned long nr_sleep_events;
147 unsigned long nr_wakeup_events;
148 unsigned long nr_sleep_corrections;
149 unsigned long nr_run_events_optimized;
150 unsigned long targetless_wakeups;
151 unsigned long multitarget_wakeups;
152 unsigned long nr_runs;
153 unsigned long nr_timestamps;
154 unsigned long nr_unordered_timestamps;
155 unsigned long nr_context_switch_bugs;
156 unsigned long nr_events;
157 unsigned long nr_lost_chunks;
158 unsigned long nr_lost_events;
159 u64 run_measurement_overhead;
160 u64 sleep_measurement_overhead;
161 u64 start_time;
162 u64 cpu_usage;
163 u64 runavg_cpu_usage;
164 u64 parent_cpu_usage;
165 u64 runavg_parent_cpu_usage;
166 u64 sum_runtime;
167 u64 sum_fluct;
168 u64 run_avg;
169 u64 all_runtime;
170 u64 all_count;
171 u64 cpu_last_switched[MAX_CPUS];
172 struct rb_root atom_root, sorted_atom_root, merged_atom_root;
173 struct list_head sort_list, cmp_pid;
174 bool force;
175 bool skip_merge;
176};
177
178static u64 get_nsecs(void)
179{
180 struct timespec ts;
181
182 clock_gettime(CLOCK_MONOTONIC, &ts);
183
184 return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
185}
186
187static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
188{
189 u64 T0 = get_nsecs(), T1;
190
191 do {
192 T1 = get_nsecs();
193 } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
194}
195
196static void sleep_nsecs(u64 nsecs)
197{
198 struct timespec ts;
199
200 ts.tv_nsec = nsecs % 999999999;
201 ts.tv_sec = nsecs / 999999999;
202
203 nanosleep(&ts, NULL);
204}
205
206static void calibrate_run_measurement_overhead(struct perf_sched *sched)
207{
208 u64 T0, T1, delta, min_delta = 1000000000ULL;
209 int i;
210
211 for (i = 0; i < 10; i++) {
212 T0 = get_nsecs();
213 burn_nsecs(sched, 0);
214 T1 = get_nsecs();
215 delta = T1-T0;
216 min_delta = min(min_delta, delta);
217 }
218 sched->run_measurement_overhead = min_delta;
219
220 printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
221}
222
223static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
224{
225 u64 T0, T1, delta, min_delta = 1000000000ULL;
226 int i;
227
228 for (i = 0; i < 10; i++) {
229 T0 = get_nsecs();
230 sleep_nsecs(10000);
231 T1 = get_nsecs();
232 delta = T1-T0;
233 min_delta = min(min_delta, delta);
234 }
235 min_delta -= 10000;
236 sched->sleep_measurement_overhead = min_delta;
237
238 printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
239}
240
241static struct sched_atom *
242get_new_event(struct task_desc *task, u64 timestamp)
243{
244 struct sched_atom *event = zalloc(sizeof(*event));
245 unsigned long idx = task->nr_events;
246 size_t size;
247
248 event->timestamp = timestamp;
249 event->nr = idx;
250
251 task->nr_events++;
252 size = sizeof(struct sched_atom *) * task->nr_events;
253 task->atoms = realloc(task->atoms, size);
254 BUG_ON(!task->atoms);
255
256 task->atoms[idx] = event;
257
258 return event;
259}
260
261static struct sched_atom *last_event(struct task_desc *task)
262{
263 if (!task->nr_events)
264 return NULL;
265
266 return task->atoms[task->nr_events - 1];
267}
268
269static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
270 u64 timestamp, u64 duration)
271{
272 struct sched_atom *event, *curr_event = last_event(task);
273
274 /*
275 * optimize an existing RUN event by merging this one
276 * to it:
277 */
278 if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
279 sched->nr_run_events_optimized++;
280 curr_event->duration += duration;
281 return;
282 }
283
284 event = get_new_event(task, timestamp);
285
286 event->type = SCHED_EVENT_RUN;
287 event->duration = duration;
288
289 sched->nr_run_events++;
290}
291
292static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
293 u64 timestamp, struct task_desc *wakee)
294{
295 struct sched_atom *event, *wakee_event;
296
297 event = get_new_event(task, timestamp);
298 event->type = SCHED_EVENT_WAKEUP;
299 event->wakee = wakee;
300
301 wakee_event = last_event(wakee);
302 if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
303 sched->targetless_wakeups++;
304 return;
305 }
306 if (wakee_event->wait_sem) {
307 sched->multitarget_wakeups++;
308 return;
309 }
310
311 wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
312 sem_init(wakee_event->wait_sem, 0, 0);
313 wakee_event->specific_wait = 1;
314 event->wait_sem = wakee_event->wait_sem;
315
316 sched->nr_wakeup_events++;
317}
318
319static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
320 u64 timestamp, u64 task_state __maybe_unused)
321{
322 struct sched_atom *event = get_new_event(task, timestamp);
323
324 event->type = SCHED_EVENT_SLEEP;
325
326 sched->nr_sleep_events++;
327}
328
329static struct task_desc *register_pid(struct perf_sched *sched,
330 unsigned long pid, const char *comm)
331{
332 struct task_desc *task;
333 static int pid_max;
334
335 if (sched->pid_to_task == NULL) {
336 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
337 pid_max = MAX_PID;
338 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
339 }
340 if (pid >= (unsigned long)pid_max) {
341 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
342 sizeof(struct task_desc *))) == NULL);
343 while (pid >= (unsigned long)pid_max)
344 sched->pid_to_task[pid_max++] = NULL;
345 }
346
347 task = sched->pid_to_task[pid];
348
349 if (task)
350 return task;
351
352 task = zalloc(sizeof(*task));
353 task->pid = pid;
354 task->nr = sched->nr_tasks;
355 strcpy(task->comm, comm);
356 /*
357 * every task starts in sleeping state - this gets ignored
358 * if there's no wakeup pointing to this sleep state:
359 */
360 add_sched_event_sleep(sched, task, 0, 0);
361
362 sched->pid_to_task[pid] = task;
363 sched->nr_tasks++;
364 sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
365 BUG_ON(!sched->tasks);
366 sched->tasks[task->nr] = task;
367
368 if (verbose)
369 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
370
371 return task;
372}
373
374
375static void print_task_traces(struct perf_sched *sched)
376{
377 struct task_desc *task;
378 unsigned long i;
379
380 for (i = 0; i < sched->nr_tasks; i++) {
381 task = sched->tasks[i];
382 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
383 task->nr, task->comm, task->pid, task->nr_events);
384 }
385}
386
387static void add_cross_task_wakeups(struct perf_sched *sched)
388{
389 struct task_desc *task1, *task2;
390 unsigned long i, j;
391
392 for (i = 0; i < sched->nr_tasks; i++) {
393 task1 = sched->tasks[i];
394 j = i + 1;
395 if (j == sched->nr_tasks)
396 j = 0;
397 task2 = sched->tasks[j];
398 add_sched_event_wakeup(sched, task1, 0, task2);
399 }
400}
401
402static void perf_sched__process_event(struct perf_sched *sched,
403 struct sched_atom *atom)
404{
405 int ret = 0;
406
407 switch (atom->type) {
408 case SCHED_EVENT_RUN:
409 burn_nsecs(sched, atom->duration);
410 break;
411 case SCHED_EVENT_SLEEP:
412 if (atom->wait_sem)
413 ret = sem_wait(atom->wait_sem);
414 BUG_ON(ret);
415 break;
416 case SCHED_EVENT_WAKEUP:
417 if (atom->wait_sem)
418 ret = sem_post(atom->wait_sem);
419 BUG_ON(ret);
420 break;
421 case SCHED_EVENT_MIGRATION:
422 break;
423 default:
424 BUG_ON(1);
425 }
426}
427
428static u64 get_cpu_usage_nsec_parent(void)
429{
430 struct rusage ru;
431 u64 sum;
432 int err;
433
434 err = getrusage(RUSAGE_SELF, &ru);
435 BUG_ON(err);
436
437 sum = ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
438 sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
439
440 return sum;
441}
442
443static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
444{
445 struct perf_event_attr attr;
446 char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
447 int fd;
448 struct rlimit limit;
449 bool need_privilege = false;
450
451 memset(&attr, 0, sizeof(attr));
452
453 attr.type = PERF_TYPE_SOFTWARE;
454 attr.config = PERF_COUNT_SW_TASK_CLOCK;
455
456force_again:
457 fd = sys_perf_event_open(&attr, 0, -1, -1,
458 perf_event_open_cloexec_flag());
459
460 if (fd < 0) {
461 if (errno == EMFILE) {
462 if (sched->force) {
463 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
464 limit.rlim_cur += sched->nr_tasks - cur_task;
465 if (limit.rlim_cur > limit.rlim_max) {
466 limit.rlim_max = limit.rlim_cur;
467 need_privilege = true;
468 }
469 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
470 if (need_privilege && errno == EPERM)
471 strcpy(info, "Need privilege\n");
472 } else
473 goto force_again;
474 } else
475 strcpy(info, "Have a try with -f option\n");
476 }
477 pr_err("Error: sys_perf_event_open() syscall returned "
478 "with %d (%s)\n%s", fd,
479 strerror_r(errno, sbuf, sizeof(sbuf)), info);
480 exit(EXIT_FAILURE);
481 }
482 return fd;
483}
484
485static u64 get_cpu_usage_nsec_self(int fd)
486{
487 u64 runtime;
488 int ret;
489
490 ret = read(fd, &runtime, sizeof(runtime));
491 BUG_ON(ret != sizeof(runtime));
492
493 return runtime;
494}
495
496struct sched_thread_parms {
497 struct task_desc *task;
498 struct perf_sched *sched;
499 int fd;
500};
501
502static void *thread_func(void *ctx)
503{
504 struct sched_thread_parms *parms = ctx;
505 struct task_desc *this_task = parms->task;
506 struct perf_sched *sched = parms->sched;
507 u64 cpu_usage_0, cpu_usage_1;
508 unsigned long i, ret;
509 char comm2[22];
510 int fd = parms->fd;
511
512 zfree(&parms);
513
514 sprintf(comm2, ":%s", this_task->comm);
515 prctl(PR_SET_NAME, comm2);
516 if (fd < 0)
517 return NULL;
518again:
519 ret = sem_post(&this_task->ready_for_work);
520 BUG_ON(ret);
521 ret = pthread_mutex_lock(&sched->start_work_mutex);
522 BUG_ON(ret);
523 ret = pthread_mutex_unlock(&sched->start_work_mutex);
524 BUG_ON(ret);
525
526 cpu_usage_0 = get_cpu_usage_nsec_self(fd);
527
528 for (i = 0; i < this_task->nr_events; i++) {
529 this_task->curr_event = i;
530 perf_sched__process_event(sched, this_task->atoms[i]);
531 }
532
533 cpu_usage_1 = get_cpu_usage_nsec_self(fd);
534 this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
535 ret = sem_post(&this_task->work_done_sem);
536 BUG_ON(ret);
537
538 ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
539 BUG_ON(ret);
540 ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
541 BUG_ON(ret);
542
543 goto again;
544}
545
546static void create_tasks(struct perf_sched *sched)
547{
548 struct task_desc *task;
549 pthread_attr_t attr;
550 unsigned long i;
551 int err;
552
553 err = pthread_attr_init(&attr);
554 BUG_ON(err);
555 err = pthread_attr_setstacksize(&attr,
556 (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
557 BUG_ON(err);
558 err = pthread_mutex_lock(&sched->start_work_mutex);
559 BUG_ON(err);
560 err = pthread_mutex_lock(&sched->work_done_wait_mutex);
561 BUG_ON(err);
562 for (i = 0; i < sched->nr_tasks; i++) {
563 struct sched_thread_parms *parms = malloc(sizeof(*parms));
564 BUG_ON(parms == NULL);
565 parms->task = task = sched->tasks[i];
566 parms->sched = sched;
567 parms->fd = self_open_counters(sched, i);
568 sem_init(&task->sleep_sem, 0, 0);
569 sem_init(&task->ready_for_work, 0, 0);
570 sem_init(&task->work_done_sem, 0, 0);
571 task->curr_event = 0;
572 err = pthread_create(&task->thread, &attr, thread_func, parms);
573 BUG_ON(err);
574 }
575}
576
577static void wait_for_tasks(struct perf_sched *sched)
578{
579 u64 cpu_usage_0, cpu_usage_1;
580 struct task_desc *task;
581 unsigned long i, ret;
582
583 sched->start_time = get_nsecs();
584 sched->cpu_usage = 0;
585 pthread_mutex_unlock(&sched->work_done_wait_mutex);
586
587 for (i = 0; i < sched->nr_tasks; i++) {
588 task = sched->tasks[i];
589 ret = sem_wait(&task->ready_for_work);
590 BUG_ON(ret);
591 sem_init(&task->ready_for_work, 0, 0);
592 }
593 ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
594 BUG_ON(ret);
595
596 cpu_usage_0 = get_cpu_usage_nsec_parent();
597
598 pthread_mutex_unlock(&sched->start_work_mutex);
599
600 for (i = 0; i < sched->nr_tasks; i++) {
601 task = sched->tasks[i];
602 ret = sem_wait(&task->work_done_sem);
603 BUG_ON(ret);
604 sem_init(&task->work_done_sem, 0, 0);
605 sched->cpu_usage += task->cpu_usage;
606 task->cpu_usage = 0;
607 }
608
609 cpu_usage_1 = get_cpu_usage_nsec_parent();
610 if (!sched->runavg_cpu_usage)
611 sched->runavg_cpu_usage = sched->cpu_usage;
612 sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
613
614 sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
615 if (!sched->runavg_parent_cpu_usage)
616 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
617 sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
618 sched->parent_cpu_usage)/sched->replay_repeat;
619
620 ret = pthread_mutex_lock(&sched->start_work_mutex);
621 BUG_ON(ret);
622
623 for (i = 0; i < sched->nr_tasks; i++) {
624 task = sched->tasks[i];
625 sem_init(&task->sleep_sem, 0, 0);
626 task->curr_event = 0;
627 }
628}
629
630static void run_one_test(struct perf_sched *sched)
631{
632 u64 T0, T1, delta, avg_delta, fluct;
633
634 T0 = get_nsecs();
635 wait_for_tasks(sched);
636 T1 = get_nsecs();
637
638 delta = T1 - T0;
639 sched->sum_runtime += delta;
640 sched->nr_runs++;
641
642 avg_delta = sched->sum_runtime / sched->nr_runs;
643 if (delta < avg_delta)
644 fluct = avg_delta - delta;
645 else
646 fluct = delta - avg_delta;
647 sched->sum_fluct += fluct;
648 if (!sched->run_avg)
649 sched->run_avg = delta;
650 sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
651
652 printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / 1000000.0);
653
654 printf("ravg: %0.2f, ", (double)sched->run_avg / 1e6);
655
656 printf("cpu: %0.2f / %0.2f",
657 (double)sched->cpu_usage / 1e6, (double)sched->runavg_cpu_usage / 1e6);
658
659#if 0
660 /*
661 * rusage statistics done by the parent, these are less
662 * accurate than the sched->sum_exec_runtime based statistics:
663 */
664 printf(" [%0.2f / %0.2f]",
665 (double)sched->parent_cpu_usage/1e6,
666 (double)sched->runavg_parent_cpu_usage/1e6);
667#endif
668
669 printf("\n");
670
671 if (sched->nr_sleep_corrections)
672 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
673 sched->nr_sleep_corrections = 0;
674}
675
676static void test_calibrations(struct perf_sched *sched)
677{
678 u64 T0, T1;
679
680 T0 = get_nsecs();
681 burn_nsecs(sched, 1e6);
682 T1 = get_nsecs();
683
684 printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
685
686 T0 = get_nsecs();
687 sleep_nsecs(1e6);
688 T1 = get_nsecs();
689
690 printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
691}
692
693static int
694replay_wakeup_event(struct perf_sched *sched,
695 struct perf_evsel *evsel, struct perf_sample *sample,
696 struct machine *machine __maybe_unused)
697{
698 const char *comm = perf_evsel__strval(evsel, sample, "comm");
699 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
700 struct task_desc *waker, *wakee;
701
702 if (verbose) {
703 printf("sched_wakeup event %p\n", evsel);
704
705 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
706 }
707
708 waker = register_pid(sched, sample->tid, "<unknown>");
709 wakee = register_pid(sched, pid, comm);
710
711 add_sched_event_wakeup(sched, waker, sample->time, wakee);
712 return 0;
713}
714
715static int replay_switch_event(struct perf_sched *sched,
716 struct perf_evsel *evsel,
717 struct perf_sample *sample,
718 struct machine *machine __maybe_unused)
719{
720 const char *prev_comm = perf_evsel__strval(evsel, sample, "prev_comm"),
721 *next_comm = perf_evsel__strval(evsel, sample, "next_comm");
722 const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
723 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
724 const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
725 struct task_desc *prev, __maybe_unused *next;
726 u64 timestamp0, timestamp = sample->time;
727 int cpu = sample->cpu;
728 s64 delta;
729
730 if (verbose)
731 printf("sched_switch event %p\n", evsel);
732
733 if (cpu >= MAX_CPUS || cpu < 0)
734 return 0;
735
736 timestamp0 = sched->cpu_last_switched[cpu];
737 if (timestamp0)
738 delta = timestamp - timestamp0;
739 else
740 delta = 0;
741
742 if (delta < 0) {
743 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
744 return -1;
745 }
746
747 pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
748 prev_comm, prev_pid, next_comm, next_pid, delta);
749
750 prev = register_pid(sched, prev_pid, prev_comm);
751 next = register_pid(sched, next_pid, next_comm);
752
753 sched->cpu_last_switched[cpu] = timestamp;
754
755 add_sched_event_run(sched, prev, timestamp, delta);
756 add_sched_event_sleep(sched, prev, timestamp, prev_state);
757
758 return 0;
759}
760
761static int replay_fork_event(struct perf_sched *sched,
762 union perf_event *event,
763 struct machine *machine)
764{
765 struct thread *child, *parent;
766
767 child = machine__findnew_thread(machine, event->fork.pid,
768 event->fork.tid);
769 parent = machine__findnew_thread(machine, event->fork.ppid,
770 event->fork.ptid);
771
772 if (child == NULL || parent == NULL) {
773 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
774 child, parent);
775 goto out_put;
776 }
777
778 if (verbose) {
779 printf("fork event\n");
780 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
781 printf("... child: %s/%d\n", thread__comm_str(child), child->tid);
782 }
783
784 register_pid(sched, parent->tid, thread__comm_str(parent));
785 register_pid(sched, child->tid, thread__comm_str(child));
786out_put:
787 thread__put(child);
788 thread__put(parent);
789 return 0;
790}
791
792struct sort_dimension {
793 const char *name;
794 sort_fn_t cmp;
795 struct list_head list;
796};
797
798static int
799thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
800{
801 struct sort_dimension *sort;
802 int ret = 0;
803
804 BUG_ON(list_empty(list));
805
806 list_for_each_entry(sort, list, list) {
807 ret = sort->cmp(l, r);
808 if (ret)
809 return ret;
810 }
811
812 return ret;
813}
814
815static struct work_atoms *
816thread_atoms_search(struct rb_root *root, struct thread *thread,
817 struct list_head *sort_list)
818{
819 struct rb_node *node = root->rb_node;
820 struct work_atoms key = { .thread = thread };
821
822 while (node) {
823 struct work_atoms *atoms;
824 int cmp;
825
826 atoms = container_of(node, struct work_atoms, node);
827
828 cmp = thread_lat_cmp(sort_list, &key, atoms);
829 if (cmp > 0)
830 node = node->rb_left;
831 else if (cmp < 0)
832 node = node->rb_right;
833 else {
834 BUG_ON(thread != atoms->thread);
835 return atoms;
836 }
837 }
838 return NULL;
839}
840
841static void
842__thread_latency_insert(struct rb_root *root, struct work_atoms *data,
843 struct list_head *sort_list)
844{
845 struct rb_node **new = &(root->rb_node), *parent = NULL;
846
847 while (*new) {
848 struct work_atoms *this;
849 int cmp;
850
851 this = container_of(*new, struct work_atoms, node);
852 parent = *new;
853
854 cmp = thread_lat_cmp(sort_list, data, this);
855
856 if (cmp > 0)
857 new = &((*new)->rb_left);
858 else
859 new = &((*new)->rb_right);
860 }
861
862 rb_link_node(&data->node, parent, new);
863 rb_insert_color(&data->node, root);
864}
865
866static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
867{
868 struct work_atoms *atoms = zalloc(sizeof(*atoms));
869 if (!atoms) {
870 pr_err("No memory at %s\n", __func__);
871 return -1;
872 }
873
874 atoms->thread = thread__get(thread);
875 INIT_LIST_HEAD(&atoms->work_list);
876 __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
877 return 0;
878}
879
880static char sched_out_state(u64 prev_state)
881{
882 const char *str = TASK_STATE_TO_CHAR_STR;
883
884 return str[prev_state];
885}
886
887static int
888add_sched_out_event(struct work_atoms *atoms,
889 char run_state,
890 u64 timestamp)
891{
892 struct work_atom *atom = zalloc(sizeof(*atom));
893 if (!atom) {
894 pr_err("Non memory at %s", __func__);
895 return -1;
896 }
897
898 atom->sched_out_time = timestamp;
899
900 if (run_state == 'R') {
901 atom->state = THREAD_WAIT_CPU;
902 atom->wake_up_time = atom->sched_out_time;
903 }
904
905 list_add_tail(&atom->list, &atoms->work_list);
906 return 0;
907}
908
909static void
910add_runtime_event(struct work_atoms *atoms, u64 delta,
911 u64 timestamp __maybe_unused)
912{
913 struct work_atom *atom;
914
915 BUG_ON(list_empty(&atoms->work_list));
916
917 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
918
919 atom->runtime += delta;
920 atoms->total_runtime += delta;
921}
922
923static void
924add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
925{
926 struct work_atom *atom;
927 u64 delta;
928
929 if (list_empty(&atoms->work_list))
930 return;
931
932 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
933
934 if (atom->state != THREAD_WAIT_CPU)
935 return;
936
937 if (timestamp < atom->wake_up_time) {
938 atom->state = THREAD_IGNORE;
939 return;
940 }
941
942 atom->state = THREAD_SCHED_IN;
943 atom->sched_in_time = timestamp;
944
945 delta = atom->sched_in_time - atom->wake_up_time;
946 atoms->total_lat += delta;
947 if (delta > atoms->max_lat) {
948 atoms->max_lat = delta;
949 atoms->max_lat_at = timestamp;
950 }
951 atoms->nb_atoms++;
952}
953
954static int latency_switch_event(struct perf_sched *sched,
955 struct perf_evsel *evsel,
956 struct perf_sample *sample,
957 struct machine *machine)
958{
959 const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
960 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
961 const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
962 struct work_atoms *out_events, *in_events;
963 struct thread *sched_out, *sched_in;
964 u64 timestamp0, timestamp = sample->time;
965 int cpu = sample->cpu, err = -1;
966 s64 delta;
967
968 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
969
970 timestamp0 = sched->cpu_last_switched[cpu];
971 sched->cpu_last_switched[cpu] = timestamp;
972 if (timestamp0)
973 delta = timestamp - timestamp0;
974 else
975 delta = 0;
976
977 if (delta < 0) {
978 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
979 return -1;
980 }
981
982 sched_out = machine__findnew_thread(machine, -1, prev_pid);
983 sched_in = machine__findnew_thread(machine, -1, next_pid);
984 if (sched_out == NULL || sched_in == NULL)
985 goto out_put;
986
987 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
988 if (!out_events) {
989 if (thread_atoms_insert(sched, sched_out))
990 goto out_put;
991 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
992 if (!out_events) {
993 pr_err("out-event: Internal tree error");
994 goto out_put;
995 }
996 }
997 if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
998 return -1;
999
1000 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1001 if (!in_events) {
1002 if (thread_atoms_insert(sched, sched_in))
1003 goto out_put;
1004 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1005 if (!in_events) {
1006 pr_err("in-event: Internal tree error");
1007 goto out_put;
1008 }
1009 /*
1010 * Take came in we have not heard about yet,
1011 * add in an initial atom in runnable state:
1012 */
1013 if (add_sched_out_event(in_events, 'R', timestamp))
1014 goto out_put;
1015 }
1016 add_sched_in_event(in_events, timestamp);
1017 err = 0;
1018out_put:
1019 thread__put(sched_out);
1020 thread__put(sched_in);
1021 return err;
1022}
1023
1024static int latency_runtime_event(struct perf_sched *sched,
1025 struct perf_evsel *evsel,
1026 struct perf_sample *sample,
1027 struct machine *machine)
1028{
1029 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1030 const u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
1031 struct thread *thread = machine__findnew_thread(machine, -1, pid);
1032 struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1033 u64 timestamp = sample->time;
1034 int cpu = sample->cpu, err = -1;
1035
1036 if (thread == NULL)
1037 return -1;
1038
1039 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1040 if (!atoms) {
1041 if (thread_atoms_insert(sched, thread))
1042 goto out_put;
1043 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1044 if (!atoms) {
1045 pr_err("in-event: Internal tree error");
1046 goto out_put;
1047 }
1048 if (add_sched_out_event(atoms, 'R', timestamp))
1049 goto out_put;
1050 }
1051
1052 add_runtime_event(atoms, runtime, timestamp);
1053 err = 0;
1054out_put:
1055 thread__put(thread);
1056 return err;
1057}
1058
1059static int latency_wakeup_event(struct perf_sched *sched,
1060 struct perf_evsel *evsel,
1061 struct perf_sample *sample,
1062 struct machine *machine)
1063{
1064 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1065 struct work_atoms *atoms;
1066 struct work_atom *atom;
1067 struct thread *wakee;
1068 u64 timestamp = sample->time;
1069 int err = -1;
1070
1071 wakee = machine__findnew_thread(machine, -1, pid);
1072 if (wakee == NULL)
1073 return -1;
1074 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1075 if (!atoms) {
1076 if (thread_atoms_insert(sched, wakee))
1077 goto out_put;
1078 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1079 if (!atoms) {
1080 pr_err("wakeup-event: Internal tree error");
1081 goto out_put;
1082 }
1083 if (add_sched_out_event(atoms, 'S', timestamp))
1084 goto out_put;
1085 }
1086
1087 BUG_ON(list_empty(&atoms->work_list));
1088
1089 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1090
1091 /*
1092 * As we do not guarantee the wakeup event happens when
1093 * task is out of run queue, also may happen when task is
1094 * on run queue and wakeup only change ->state to TASK_RUNNING,
1095 * then we should not set the ->wake_up_time when wake up a
1096 * task which is on run queue.
1097 *
1098 * You WILL be missing events if you've recorded only
1099 * one CPU, or are only looking at only one, so don't
1100 * skip in this case.
1101 */
1102 if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1103 goto out_ok;
1104
1105 sched->nr_timestamps++;
1106 if (atom->sched_out_time > timestamp) {
1107 sched->nr_unordered_timestamps++;
1108 goto out_ok;
1109 }
1110
1111 atom->state = THREAD_WAIT_CPU;
1112 atom->wake_up_time = timestamp;
1113out_ok:
1114 err = 0;
1115out_put:
1116 thread__put(wakee);
1117 return err;
1118}
1119
1120static int latency_migrate_task_event(struct perf_sched *sched,
1121 struct perf_evsel *evsel,
1122 struct perf_sample *sample,
1123 struct machine *machine)
1124{
1125 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1126 u64 timestamp = sample->time;
1127 struct work_atoms *atoms;
1128 struct work_atom *atom;
1129 struct thread *migrant;
1130 int err = -1;
1131
1132 /*
1133 * Only need to worry about migration when profiling one CPU.
1134 */
1135 if (sched->profile_cpu == -1)
1136 return 0;
1137
1138 migrant = machine__findnew_thread(machine, -1, pid);
1139 if (migrant == NULL)
1140 return -1;
1141 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1142 if (!atoms) {
1143 if (thread_atoms_insert(sched, migrant))
1144 goto out_put;
1145 register_pid(sched, migrant->tid, thread__comm_str(migrant));
1146 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1147 if (!atoms) {
1148 pr_err("migration-event: Internal tree error");
1149 goto out_put;
1150 }
1151 if (add_sched_out_event(atoms, 'R', timestamp))
1152 goto out_put;
1153 }
1154
1155 BUG_ON(list_empty(&atoms->work_list));
1156
1157 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1158 atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1159
1160 sched->nr_timestamps++;
1161
1162 if (atom->sched_out_time > timestamp)
1163 sched->nr_unordered_timestamps++;
1164 err = 0;
1165out_put:
1166 thread__put(migrant);
1167 return err;
1168}
1169
1170static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1171{
1172 int i;
1173 int ret;
1174 u64 avg;
1175
1176 if (!work_list->nb_atoms)
1177 return;
1178 /*
1179 * Ignore idle threads:
1180 */
1181 if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1182 return;
1183
1184 sched->all_runtime += work_list->total_runtime;
1185 sched->all_count += work_list->nb_atoms;
1186
1187 if (work_list->num_merged > 1)
1188 ret = printf(" %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1189 else
1190 ret = printf(" %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1191
1192 for (i = 0; i < 24 - ret; i++)
1193 printf(" ");
1194
1195 avg = work_list->total_lat / work_list->nb_atoms;
1196
1197 printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13.6f s\n",
1198 (double)work_list->total_runtime / 1e6,
1199 work_list->nb_atoms, (double)avg / 1e6,
1200 (double)work_list->max_lat / 1e6,
1201 (double)work_list->max_lat_at / 1e9);
1202}
1203
1204static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1205{
1206 if (l->thread == r->thread)
1207 return 0;
1208 if (l->thread->tid < r->thread->tid)
1209 return -1;
1210 if (l->thread->tid > r->thread->tid)
1211 return 1;
1212 return (int)(l->thread - r->thread);
1213}
1214
1215static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1216{
1217 u64 avgl, avgr;
1218
1219 if (!l->nb_atoms)
1220 return -1;
1221
1222 if (!r->nb_atoms)
1223 return 1;
1224
1225 avgl = l->total_lat / l->nb_atoms;
1226 avgr = r->total_lat / r->nb_atoms;
1227
1228 if (avgl < avgr)
1229 return -1;
1230 if (avgl > avgr)
1231 return 1;
1232
1233 return 0;
1234}
1235
1236static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1237{
1238 if (l->max_lat < r->max_lat)
1239 return -1;
1240 if (l->max_lat > r->max_lat)
1241 return 1;
1242
1243 return 0;
1244}
1245
1246static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1247{
1248 if (l->nb_atoms < r->nb_atoms)
1249 return -1;
1250 if (l->nb_atoms > r->nb_atoms)
1251 return 1;
1252
1253 return 0;
1254}
1255
1256static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1257{
1258 if (l->total_runtime < r->total_runtime)
1259 return -1;
1260 if (l->total_runtime > r->total_runtime)
1261 return 1;
1262
1263 return 0;
1264}
1265
1266static int sort_dimension__add(const char *tok, struct list_head *list)
1267{
1268 size_t i;
1269 static struct sort_dimension avg_sort_dimension = {
1270 .name = "avg",
1271 .cmp = avg_cmp,
1272 };
1273 static struct sort_dimension max_sort_dimension = {
1274 .name = "max",
1275 .cmp = max_cmp,
1276 };
1277 static struct sort_dimension pid_sort_dimension = {
1278 .name = "pid",
1279 .cmp = pid_cmp,
1280 };
1281 static struct sort_dimension runtime_sort_dimension = {
1282 .name = "runtime",
1283 .cmp = runtime_cmp,
1284 };
1285 static struct sort_dimension switch_sort_dimension = {
1286 .name = "switch",
1287 .cmp = switch_cmp,
1288 };
1289 struct sort_dimension *available_sorts[] = {
1290 &pid_sort_dimension,
1291 &avg_sort_dimension,
1292 &max_sort_dimension,
1293 &switch_sort_dimension,
1294 &runtime_sort_dimension,
1295 };
1296
1297 for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1298 if (!strcmp(available_sorts[i]->name, tok)) {
1299 list_add_tail(&available_sorts[i]->list, list);
1300
1301 return 0;
1302 }
1303 }
1304
1305 return -1;
1306}
1307
1308static void perf_sched__sort_lat(struct perf_sched *sched)
1309{
1310 struct rb_node *node;
1311 struct rb_root *root = &sched->atom_root;
1312again:
1313 for (;;) {
1314 struct work_atoms *data;
1315 node = rb_first(root);
1316 if (!node)
1317 break;
1318
1319 rb_erase(node, root);
1320 data = rb_entry(node, struct work_atoms, node);
1321 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1322 }
1323 if (root == &sched->atom_root) {
1324 root = &sched->merged_atom_root;
1325 goto again;
1326 }
1327}
1328
1329static int process_sched_wakeup_event(struct perf_tool *tool,
1330 struct perf_evsel *evsel,
1331 struct perf_sample *sample,
1332 struct machine *machine)
1333{
1334 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1335
1336 if (sched->tp_handler->wakeup_event)
1337 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1338
1339 return 0;
1340}
1341
1342static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1343 struct perf_sample *sample, struct machine *machine)
1344{
1345 const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1346 struct thread *sched_in;
1347 int new_shortname;
1348 u64 timestamp0, timestamp = sample->time;
1349 s64 delta;
1350 int cpu, this_cpu = sample->cpu;
1351
1352 BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1353
1354 if (this_cpu > sched->max_cpu)
1355 sched->max_cpu = this_cpu;
1356
1357 timestamp0 = sched->cpu_last_switched[this_cpu];
1358 sched->cpu_last_switched[this_cpu] = timestamp;
1359 if (timestamp0)
1360 delta = timestamp - timestamp0;
1361 else
1362 delta = 0;
1363
1364 if (delta < 0) {
1365 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1366 return -1;
1367 }
1368
1369 sched_in = machine__findnew_thread(machine, -1, next_pid);
1370 if (sched_in == NULL)
1371 return -1;
1372
1373 sched->curr_thread[this_cpu] = thread__get(sched_in);
1374
1375 printf(" ");
1376
1377 new_shortname = 0;
1378 if (!sched_in->shortname[0]) {
1379 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1380 /*
1381 * Don't allocate a letter-number for swapper:0
1382 * as a shortname. Instead, we use '.' for it.
1383 */
1384 sched_in->shortname[0] = '.';
1385 sched_in->shortname[1] = ' ';
1386 } else {
1387 sched_in->shortname[0] = sched->next_shortname1;
1388 sched_in->shortname[1] = sched->next_shortname2;
1389
1390 if (sched->next_shortname1 < 'Z') {
1391 sched->next_shortname1++;
1392 } else {
1393 sched->next_shortname1 = 'A';
1394 if (sched->next_shortname2 < '9')
1395 sched->next_shortname2++;
1396 else
1397 sched->next_shortname2 = '0';
1398 }
1399 }
1400 new_shortname = 1;
1401 }
1402
1403 for (cpu = 0; cpu <= sched->max_cpu; cpu++) {
1404 if (cpu != this_cpu)
1405 printf(" ");
1406 else
1407 printf("*");
1408
1409 if (sched->curr_thread[cpu])
1410 printf("%2s ", sched->curr_thread[cpu]->shortname);
1411 else
1412 printf(" ");
1413 }
1414
1415 printf(" %12.6f secs ", (double)timestamp/1e9);
1416 if (new_shortname) {
1417 printf("%s => %s:%d\n",
1418 sched_in->shortname, thread__comm_str(sched_in), sched_in->tid);
1419 } else {
1420 printf("\n");
1421 }
1422
1423 thread__put(sched_in);
1424
1425 return 0;
1426}
1427
1428static int process_sched_switch_event(struct perf_tool *tool,
1429 struct perf_evsel *evsel,
1430 struct perf_sample *sample,
1431 struct machine *machine)
1432{
1433 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1434 int this_cpu = sample->cpu, err = 0;
1435 u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1436 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1437
1438 if (sched->curr_pid[this_cpu] != (u32)-1) {
1439 /*
1440 * Are we trying to switch away a PID that is
1441 * not current?
1442 */
1443 if (sched->curr_pid[this_cpu] != prev_pid)
1444 sched->nr_context_switch_bugs++;
1445 }
1446
1447 if (sched->tp_handler->switch_event)
1448 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1449
1450 sched->curr_pid[this_cpu] = next_pid;
1451 return err;
1452}
1453
1454static int process_sched_runtime_event(struct perf_tool *tool,
1455 struct perf_evsel *evsel,
1456 struct perf_sample *sample,
1457 struct machine *machine)
1458{
1459 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1460
1461 if (sched->tp_handler->runtime_event)
1462 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1463
1464 return 0;
1465}
1466
1467static int perf_sched__process_fork_event(struct perf_tool *tool,
1468 union perf_event *event,
1469 struct perf_sample *sample,
1470 struct machine *machine)
1471{
1472 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1473
1474 /* run the fork event through the perf machineruy */
1475 perf_event__process_fork(tool, event, sample, machine);
1476
1477 /* and then run additional processing needed for this command */
1478 if (sched->tp_handler->fork_event)
1479 return sched->tp_handler->fork_event(sched, event, machine);
1480
1481 return 0;
1482}
1483
1484static int process_sched_migrate_task_event(struct perf_tool *tool,
1485 struct perf_evsel *evsel,
1486 struct perf_sample *sample,
1487 struct machine *machine)
1488{
1489 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1490
1491 if (sched->tp_handler->migrate_task_event)
1492 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1493
1494 return 0;
1495}
1496
1497typedef int (*tracepoint_handler)(struct perf_tool *tool,
1498 struct perf_evsel *evsel,
1499 struct perf_sample *sample,
1500 struct machine *machine);
1501
1502static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1503 union perf_event *event __maybe_unused,
1504 struct perf_sample *sample,
1505 struct perf_evsel *evsel,
1506 struct machine *machine)
1507{
1508 int err = 0;
1509
1510 if (evsel->handler != NULL) {
1511 tracepoint_handler f = evsel->handler;
1512 err = f(tool, evsel, sample, machine);
1513 }
1514
1515 return err;
1516}
1517
1518static int perf_sched__read_events(struct perf_sched *sched)
1519{
1520 const struct perf_evsel_str_handler handlers[] = {
1521 { "sched:sched_switch", process_sched_switch_event, },
1522 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1523 { "sched:sched_wakeup", process_sched_wakeup_event, },
1524 { "sched:sched_wakeup_new", process_sched_wakeup_event, },
1525 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1526 };
1527 struct perf_session *session;
1528 struct perf_data_file file = {
1529 .path = input_name,
1530 .mode = PERF_DATA_MODE_READ,
1531 .force = sched->force,
1532 };
1533 int rc = -1;
1534
1535 session = perf_session__new(&file, false, &sched->tool);
1536 if (session == NULL) {
1537 pr_debug("No Memory for session\n");
1538 return -1;
1539 }
1540
1541 symbol__init(&session->header.env);
1542
1543 if (perf_session__set_tracepoints_handlers(session, handlers))
1544 goto out_delete;
1545
1546 if (perf_session__has_traces(session, "record -R")) {
1547 int err = perf_session__process_events(session);
1548 if (err) {
1549 pr_err("Failed to process events, error %d", err);
1550 goto out_delete;
1551 }
1552
1553 sched->nr_events = session->evlist->stats.nr_events[0];
1554 sched->nr_lost_events = session->evlist->stats.total_lost;
1555 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1556 }
1557
1558 rc = 0;
1559out_delete:
1560 perf_session__delete(session);
1561 return rc;
1562}
1563
1564static void print_bad_events(struct perf_sched *sched)
1565{
1566 if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
1567 printf(" INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1568 (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
1569 sched->nr_unordered_timestamps, sched->nr_timestamps);
1570 }
1571 if (sched->nr_lost_events && sched->nr_events) {
1572 printf(" INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1573 (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
1574 sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
1575 }
1576 if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
1577 printf(" INFO: %.3f%% context switch bugs (%ld out of %ld)",
1578 (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
1579 sched->nr_context_switch_bugs, sched->nr_timestamps);
1580 if (sched->nr_lost_events)
1581 printf(" (due to lost events?)");
1582 printf("\n");
1583 }
1584}
1585
1586static void __merge_work_atoms(struct rb_root *root, struct work_atoms *data)
1587{
1588 struct rb_node **new = &(root->rb_node), *parent = NULL;
1589 struct work_atoms *this;
1590 const char *comm = thread__comm_str(data->thread), *this_comm;
1591
1592 while (*new) {
1593 int cmp;
1594
1595 this = container_of(*new, struct work_atoms, node);
1596 parent = *new;
1597
1598 this_comm = thread__comm_str(this->thread);
1599 cmp = strcmp(comm, this_comm);
1600 if (cmp > 0) {
1601 new = &((*new)->rb_left);
1602 } else if (cmp < 0) {
1603 new = &((*new)->rb_right);
1604 } else {
1605 this->num_merged++;
1606 this->total_runtime += data->total_runtime;
1607 this->nb_atoms += data->nb_atoms;
1608 this->total_lat += data->total_lat;
1609 list_splice(&data->work_list, &this->work_list);
1610 if (this->max_lat < data->max_lat) {
1611 this->max_lat = data->max_lat;
1612 this->max_lat_at = data->max_lat_at;
1613 }
1614 zfree(&data);
1615 return;
1616 }
1617 }
1618
1619 data->num_merged++;
1620 rb_link_node(&data->node, parent, new);
1621 rb_insert_color(&data->node, root);
1622}
1623
1624static void perf_sched__merge_lat(struct perf_sched *sched)
1625{
1626 struct work_atoms *data;
1627 struct rb_node *node;
1628
1629 if (sched->skip_merge)
1630 return;
1631
1632 while ((node = rb_first(&sched->atom_root))) {
1633 rb_erase(node, &sched->atom_root);
1634 data = rb_entry(node, struct work_atoms, node);
1635 __merge_work_atoms(&sched->merged_atom_root, data);
1636 }
1637}
1638
1639static int perf_sched__lat(struct perf_sched *sched)
1640{
1641 struct rb_node *next;
1642
1643 setup_pager();
1644
1645 if (perf_sched__read_events(sched))
1646 return -1;
1647
1648 perf_sched__merge_lat(sched);
1649 perf_sched__sort_lat(sched);
1650
1651 printf("\n -----------------------------------------------------------------------------------------------------------------\n");
1652 printf(" Task | Runtime ms | Switches | Average delay ms | Maximum delay ms | Maximum delay at |\n");
1653 printf(" -----------------------------------------------------------------------------------------------------------------\n");
1654
1655 next = rb_first(&sched->sorted_atom_root);
1656
1657 while (next) {
1658 struct work_atoms *work_list;
1659
1660 work_list = rb_entry(next, struct work_atoms, node);
1661 output_lat_thread(sched, work_list);
1662 next = rb_next(next);
1663 thread__zput(work_list->thread);
1664 }
1665
1666 printf(" -----------------------------------------------------------------------------------------------------------------\n");
1667 printf(" TOTAL: |%11.3f ms |%9" PRIu64 " |\n",
1668 (double)sched->all_runtime / 1e6, sched->all_count);
1669
1670 printf(" ---------------------------------------------------\n");
1671
1672 print_bad_events(sched);
1673 printf("\n");
1674
1675 return 0;
1676}
1677
1678static int perf_sched__map(struct perf_sched *sched)
1679{
1680 sched->max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1681
1682 setup_pager();
1683 if (perf_sched__read_events(sched))
1684 return -1;
1685 print_bad_events(sched);
1686 return 0;
1687}
1688
1689static int perf_sched__replay(struct perf_sched *sched)
1690{
1691 unsigned long i;
1692
1693 calibrate_run_measurement_overhead(sched);
1694 calibrate_sleep_measurement_overhead(sched);
1695
1696 test_calibrations(sched);
1697
1698 if (perf_sched__read_events(sched))
1699 return -1;
1700
1701 printf("nr_run_events: %ld\n", sched->nr_run_events);
1702 printf("nr_sleep_events: %ld\n", sched->nr_sleep_events);
1703 printf("nr_wakeup_events: %ld\n", sched->nr_wakeup_events);
1704
1705 if (sched->targetless_wakeups)
1706 printf("target-less wakeups: %ld\n", sched->targetless_wakeups);
1707 if (sched->multitarget_wakeups)
1708 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
1709 if (sched->nr_run_events_optimized)
1710 printf("run atoms optimized: %ld\n",
1711 sched->nr_run_events_optimized);
1712
1713 print_task_traces(sched);
1714 add_cross_task_wakeups(sched);
1715
1716 create_tasks(sched);
1717 printf("------------------------------------------------------------\n");
1718 for (i = 0; i < sched->replay_repeat; i++)
1719 run_one_test(sched);
1720
1721 return 0;
1722}
1723
1724static void setup_sorting(struct perf_sched *sched, const struct option *options,
1725 const char * const usage_msg[])
1726{
1727 char *tmp, *tok, *str = strdup(sched->sort_order);
1728
1729 for (tok = strtok_r(str, ", ", &tmp);
1730 tok; tok = strtok_r(NULL, ", ", &tmp)) {
1731 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
1732 usage_with_options_msg(usage_msg, options,
1733 "Unknown --sort key: `%s'", tok);
1734 }
1735 }
1736
1737 free(str);
1738
1739 sort_dimension__add("pid", &sched->cmp_pid);
1740}
1741
1742static int __cmd_record(int argc, const char **argv)
1743{
1744 unsigned int rec_argc, i, j;
1745 const char **rec_argv;
1746 const char * const record_args[] = {
1747 "record",
1748 "-a",
1749 "-R",
1750 "-m", "1024",
1751 "-c", "1",
1752 "-e", "sched:sched_switch",
1753 "-e", "sched:sched_stat_wait",
1754 "-e", "sched:sched_stat_sleep",
1755 "-e", "sched:sched_stat_iowait",
1756 "-e", "sched:sched_stat_runtime",
1757 "-e", "sched:sched_process_fork",
1758 "-e", "sched:sched_wakeup",
1759 "-e", "sched:sched_wakeup_new",
1760 "-e", "sched:sched_migrate_task",
1761 };
1762
1763 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1764 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1765
1766 if (rec_argv == NULL)
1767 return -ENOMEM;
1768
1769 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1770 rec_argv[i] = strdup(record_args[i]);
1771
1772 for (j = 1; j < (unsigned int)argc; j++, i++)
1773 rec_argv[i] = argv[j];
1774
1775 BUG_ON(i != rec_argc);
1776
1777 return cmd_record(i, rec_argv, NULL);
1778}
1779
1780int cmd_sched(int argc, const char **argv, const char *prefix __maybe_unused)
1781{
1782 const char default_sort_order[] = "avg, max, switch, runtime";
1783 struct perf_sched sched = {
1784 .tool = {
1785 .sample = perf_sched__process_tracepoint_sample,
1786 .comm = perf_event__process_comm,
1787 .lost = perf_event__process_lost,
1788 .fork = perf_sched__process_fork_event,
1789 .ordered_events = true,
1790 },
1791 .cmp_pid = LIST_HEAD_INIT(sched.cmp_pid),
1792 .sort_list = LIST_HEAD_INIT(sched.sort_list),
1793 .start_work_mutex = PTHREAD_MUTEX_INITIALIZER,
1794 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
1795 .sort_order = default_sort_order,
1796 .replay_repeat = 10,
1797 .profile_cpu = -1,
1798 .next_shortname1 = 'A',
1799 .next_shortname2 = '0',
1800 .skip_merge = 0,
1801 };
1802 const struct option latency_options[] = {
1803 OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
1804 "sort by key(s): runtime, switch, avg, max"),
1805 OPT_INCR('v', "verbose", &verbose,
1806 "be more verbose (show symbol address, etc)"),
1807 OPT_INTEGER('C', "CPU", &sched.profile_cpu,
1808 "CPU to profile on"),
1809 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1810 "dump raw trace in ASCII"),
1811 OPT_BOOLEAN('p', "pids", &sched.skip_merge,
1812 "latency stats per pid instead of per comm"),
1813 OPT_END()
1814 };
1815 const struct option replay_options[] = {
1816 OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
1817 "repeat the workload replay N times (-1: infinite)"),
1818 OPT_INCR('v', "verbose", &verbose,
1819 "be more verbose (show symbol address, etc)"),
1820 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1821 "dump raw trace in ASCII"),
1822 OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
1823 OPT_END()
1824 };
1825 const struct option sched_options[] = {
1826 OPT_STRING('i', "input", &input_name, "file",
1827 "input file name"),
1828 OPT_INCR('v', "verbose", &verbose,
1829 "be more verbose (show symbol address, etc)"),
1830 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1831 "dump raw trace in ASCII"),
1832 OPT_END()
1833 };
1834 const char * const latency_usage[] = {
1835 "perf sched latency [<options>]",
1836 NULL
1837 };
1838 const char * const replay_usage[] = {
1839 "perf sched replay [<options>]",
1840 NULL
1841 };
1842 const char *const sched_subcommands[] = { "record", "latency", "map",
1843 "replay", "script", NULL };
1844 const char *sched_usage[] = {
1845 NULL,
1846 NULL
1847 };
1848 struct trace_sched_handler lat_ops = {
1849 .wakeup_event = latency_wakeup_event,
1850 .switch_event = latency_switch_event,
1851 .runtime_event = latency_runtime_event,
1852 .migrate_task_event = latency_migrate_task_event,
1853 };
1854 struct trace_sched_handler map_ops = {
1855 .switch_event = map_switch_event,
1856 };
1857 struct trace_sched_handler replay_ops = {
1858 .wakeup_event = replay_wakeup_event,
1859 .switch_event = replay_switch_event,
1860 .fork_event = replay_fork_event,
1861 };
1862 unsigned int i;
1863
1864 for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
1865 sched.curr_pid[i] = -1;
1866
1867 argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
1868 sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
1869 if (!argc)
1870 usage_with_options(sched_usage, sched_options);
1871
1872 /*
1873 * Aliased to 'perf script' for now:
1874 */
1875 if (!strcmp(argv[0], "script"))
1876 return cmd_script(argc, argv, prefix);
1877
1878 if (!strncmp(argv[0], "rec", 3)) {
1879 return __cmd_record(argc, argv);
1880 } else if (!strncmp(argv[0], "lat", 3)) {
1881 sched.tp_handler = &lat_ops;
1882 if (argc > 1) {
1883 argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1884 if (argc)
1885 usage_with_options(latency_usage, latency_options);
1886 }
1887 setup_sorting(&sched, latency_options, latency_usage);
1888 return perf_sched__lat(&sched);
1889 } else if (!strcmp(argv[0], "map")) {
1890 sched.tp_handler = &map_ops;
1891 setup_sorting(&sched, latency_options, latency_usage);
1892 return perf_sched__map(&sched);
1893 } else if (!strncmp(argv[0], "rep", 3)) {
1894 sched.tp_handler = &replay_ops;
1895 if (argc) {
1896 argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1897 if (argc)
1898 usage_with_options(replay_usage, replay_options);
1899 }
1900 return perf_sched__replay(&sched);
1901 } else {
1902 usage_with_options(sched_usage, sched_options);
1903 }
1904
1905 return 0;
1906}