Loading...
1// SPDX-License-Identifier: GPL-2.0
2#include "builtin.h"
3#include "perf-sys.h"
4
5#include "util/cpumap.h"
6#include "util/evlist.h"
7#include "util/evsel.h"
8#include "util/evsel_fprintf.h"
9#include "util/mutex.h"
10#include "util/symbol.h"
11#include "util/thread.h"
12#include "util/header.h"
13#include "util/session.h"
14#include "util/tool.h"
15#include "util/cloexec.h"
16#include "util/thread_map.h"
17#include "util/color.h"
18#include "util/stat.h"
19#include "util/string2.h"
20#include "util/callchain.h"
21#include "util/time-utils.h"
22
23#include <subcmd/pager.h>
24#include <subcmd/parse-options.h>
25#include "util/trace-event.h"
26
27#include "util/debug.h"
28#include "util/event.h"
29#include "util/util.h"
30
31#include <linux/kernel.h>
32#include <linux/log2.h>
33#include <linux/zalloc.h>
34#include <sys/prctl.h>
35#include <sys/resource.h>
36#include <inttypes.h>
37
38#include <errno.h>
39#include <semaphore.h>
40#include <pthread.h>
41#include <math.h>
42#include <api/fs/fs.h>
43#include <perf/cpumap.h>
44#include <linux/time64.h>
45#include <linux/err.h>
46
47#include <linux/ctype.h>
48
49#define PR_SET_NAME 15 /* Set process name */
50#define MAX_CPUS 4096
51#define COMM_LEN 20
52#define SYM_LEN 129
53#define MAX_PID 1024000
54
55static const char *cpu_list;
56static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
57
58struct sched_atom;
59
60struct task_desc {
61 unsigned long nr;
62 unsigned long pid;
63 char comm[COMM_LEN];
64
65 unsigned long nr_events;
66 unsigned long curr_event;
67 struct sched_atom **atoms;
68
69 pthread_t thread;
70 sem_t sleep_sem;
71
72 sem_t ready_for_work;
73 sem_t work_done_sem;
74
75 u64 cpu_usage;
76};
77
78enum sched_event_type {
79 SCHED_EVENT_RUN,
80 SCHED_EVENT_SLEEP,
81 SCHED_EVENT_WAKEUP,
82 SCHED_EVENT_MIGRATION,
83};
84
85struct sched_atom {
86 enum sched_event_type type;
87 int specific_wait;
88 u64 timestamp;
89 u64 duration;
90 unsigned long nr;
91 sem_t *wait_sem;
92 struct task_desc *wakee;
93};
94
95enum thread_state {
96 THREAD_SLEEPING = 0,
97 THREAD_WAIT_CPU,
98 THREAD_SCHED_IN,
99 THREAD_IGNORE
100};
101
102struct work_atom {
103 struct list_head list;
104 enum thread_state state;
105 u64 sched_out_time;
106 u64 wake_up_time;
107 u64 sched_in_time;
108 u64 runtime;
109};
110
111struct work_atoms {
112 struct list_head work_list;
113 struct thread *thread;
114 struct rb_node node;
115 u64 max_lat;
116 u64 max_lat_start;
117 u64 max_lat_end;
118 u64 total_lat;
119 u64 nb_atoms;
120 u64 total_runtime;
121 int num_merged;
122};
123
124typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
125
126struct perf_sched;
127
128struct trace_sched_handler {
129 int (*switch_event)(struct perf_sched *sched, struct evsel *evsel,
130 struct perf_sample *sample, struct machine *machine);
131
132 int (*runtime_event)(struct perf_sched *sched, struct evsel *evsel,
133 struct perf_sample *sample, struct machine *machine);
134
135 int (*wakeup_event)(struct perf_sched *sched, struct evsel *evsel,
136 struct perf_sample *sample, struct machine *machine);
137
138 /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
139 int (*fork_event)(struct perf_sched *sched, union perf_event *event,
140 struct machine *machine);
141
142 int (*migrate_task_event)(struct perf_sched *sched,
143 struct evsel *evsel,
144 struct perf_sample *sample,
145 struct machine *machine);
146};
147
148#define COLOR_PIDS PERF_COLOR_BLUE
149#define COLOR_CPUS PERF_COLOR_BG_RED
150
151struct perf_sched_map {
152 DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
153 struct perf_cpu *comp_cpus;
154 bool comp;
155 struct perf_thread_map *color_pids;
156 const char *color_pids_str;
157 struct perf_cpu_map *color_cpus;
158 const char *color_cpus_str;
159 struct perf_cpu_map *cpus;
160 const char *cpus_str;
161};
162
163struct perf_sched {
164 struct perf_tool tool;
165 const char *sort_order;
166 unsigned long nr_tasks;
167 struct task_desc **pid_to_task;
168 struct task_desc **tasks;
169 const struct trace_sched_handler *tp_handler;
170 struct mutex start_work_mutex;
171 struct mutex work_done_wait_mutex;
172 int profile_cpu;
173/*
174 * Track the current task - that way we can know whether there's any
175 * weird events, such as a task being switched away that is not current.
176 */
177 struct perf_cpu max_cpu;
178 u32 *curr_pid;
179 struct thread **curr_thread;
180 char next_shortname1;
181 char next_shortname2;
182 unsigned int replay_repeat;
183 unsigned long nr_run_events;
184 unsigned long nr_sleep_events;
185 unsigned long nr_wakeup_events;
186 unsigned long nr_sleep_corrections;
187 unsigned long nr_run_events_optimized;
188 unsigned long targetless_wakeups;
189 unsigned long multitarget_wakeups;
190 unsigned long nr_runs;
191 unsigned long nr_timestamps;
192 unsigned long nr_unordered_timestamps;
193 unsigned long nr_context_switch_bugs;
194 unsigned long nr_events;
195 unsigned long nr_lost_chunks;
196 unsigned long nr_lost_events;
197 u64 run_measurement_overhead;
198 u64 sleep_measurement_overhead;
199 u64 start_time;
200 u64 cpu_usage;
201 u64 runavg_cpu_usage;
202 u64 parent_cpu_usage;
203 u64 runavg_parent_cpu_usage;
204 u64 sum_runtime;
205 u64 sum_fluct;
206 u64 run_avg;
207 u64 all_runtime;
208 u64 all_count;
209 u64 *cpu_last_switched;
210 struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
211 struct list_head sort_list, cmp_pid;
212 bool force;
213 bool skip_merge;
214 struct perf_sched_map map;
215
216 /* options for timehist command */
217 bool summary;
218 bool summary_only;
219 bool idle_hist;
220 bool show_callchain;
221 unsigned int max_stack;
222 bool show_cpu_visual;
223 bool show_wakeups;
224 bool show_next;
225 bool show_migrations;
226 bool show_state;
227 u64 skipped_samples;
228 const char *time_str;
229 struct perf_time_interval ptime;
230 struct perf_time_interval hist_time;
231 volatile bool thread_funcs_exit;
232};
233
234/* per thread run time data */
235struct thread_runtime {
236 u64 last_time; /* time of previous sched in/out event */
237 u64 dt_run; /* run time */
238 u64 dt_sleep; /* time between CPU access by sleep (off cpu) */
239 u64 dt_iowait; /* time between CPU access by iowait (off cpu) */
240 u64 dt_preempt; /* time between CPU access by preempt (off cpu) */
241 u64 dt_delay; /* time between wakeup and sched-in */
242 u64 ready_to_run; /* time of wakeup */
243
244 struct stats run_stats;
245 u64 total_run_time;
246 u64 total_sleep_time;
247 u64 total_iowait_time;
248 u64 total_preempt_time;
249 u64 total_delay_time;
250
251 char last_state;
252
253 char shortname[3];
254 bool comm_changed;
255
256 u64 migrations;
257};
258
259/* per event run time data */
260struct evsel_runtime {
261 u64 *last_time; /* time this event was last seen per cpu */
262 u32 ncpu; /* highest cpu slot allocated */
263};
264
265/* per cpu idle time data */
266struct idle_thread_runtime {
267 struct thread_runtime tr;
268 struct thread *last_thread;
269 struct rb_root_cached sorted_root;
270 struct callchain_root callchain;
271 struct callchain_cursor cursor;
272};
273
274/* track idle times per cpu */
275static struct thread **idle_threads;
276static int idle_max_cpu;
277static char idle_comm[] = "<idle>";
278
279static u64 get_nsecs(void)
280{
281 struct timespec ts;
282
283 clock_gettime(CLOCK_MONOTONIC, &ts);
284
285 return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
286}
287
288static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
289{
290 u64 T0 = get_nsecs(), T1;
291
292 do {
293 T1 = get_nsecs();
294 } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
295}
296
297static void sleep_nsecs(u64 nsecs)
298{
299 struct timespec ts;
300
301 ts.tv_nsec = nsecs % 999999999;
302 ts.tv_sec = nsecs / 999999999;
303
304 nanosleep(&ts, NULL);
305}
306
307static void calibrate_run_measurement_overhead(struct perf_sched *sched)
308{
309 u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
310 int i;
311
312 for (i = 0; i < 10; i++) {
313 T0 = get_nsecs();
314 burn_nsecs(sched, 0);
315 T1 = get_nsecs();
316 delta = T1-T0;
317 min_delta = min(min_delta, delta);
318 }
319 sched->run_measurement_overhead = min_delta;
320
321 printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
322}
323
324static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
325{
326 u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
327 int i;
328
329 for (i = 0; i < 10; i++) {
330 T0 = get_nsecs();
331 sleep_nsecs(10000);
332 T1 = get_nsecs();
333 delta = T1-T0;
334 min_delta = min(min_delta, delta);
335 }
336 min_delta -= 10000;
337 sched->sleep_measurement_overhead = min_delta;
338
339 printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
340}
341
342static struct sched_atom *
343get_new_event(struct task_desc *task, u64 timestamp)
344{
345 struct sched_atom *event = zalloc(sizeof(*event));
346 unsigned long idx = task->nr_events;
347 size_t size;
348
349 event->timestamp = timestamp;
350 event->nr = idx;
351
352 task->nr_events++;
353 size = sizeof(struct sched_atom *) * task->nr_events;
354 task->atoms = realloc(task->atoms, size);
355 BUG_ON(!task->atoms);
356
357 task->atoms[idx] = event;
358
359 return event;
360}
361
362static struct sched_atom *last_event(struct task_desc *task)
363{
364 if (!task->nr_events)
365 return NULL;
366
367 return task->atoms[task->nr_events - 1];
368}
369
370static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
371 u64 timestamp, u64 duration)
372{
373 struct sched_atom *event, *curr_event = last_event(task);
374
375 /*
376 * optimize an existing RUN event by merging this one
377 * to it:
378 */
379 if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
380 sched->nr_run_events_optimized++;
381 curr_event->duration += duration;
382 return;
383 }
384
385 event = get_new_event(task, timestamp);
386
387 event->type = SCHED_EVENT_RUN;
388 event->duration = duration;
389
390 sched->nr_run_events++;
391}
392
393static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
394 u64 timestamp, struct task_desc *wakee)
395{
396 struct sched_atom *event, *wakee_event;
397
398 event = get_new_event(task, timestamp);
399 event->type = SCHED_EVENT_WAKEUP;
400 event->wakee = wakee;
401
402 wakee_event = last_event(wakee);
403 if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
404 sched->targetless_wakeups++;
405 return;
406 }
407 if (wakee_event->wait_sem) {
408 sched->multitarget_wakeups++;
409 return;
410 }
411
412 wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
413 sem_init(wakee_event->wait_sem, 0, 0);
414 wakee_event->specific_wait = 1;
415 event->wait_sem = wakee_event->wait_sem;
416
417 sched->nr_wakeup_events++;
418}
419
420static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
421 u64 timestamp, const char task_state __maybe_unused)
422{
423 struct sched_atom *event = get_new_event(task, timestamp);
424
425 event->type = SCHED_EVENT_SLEEP;
426
427 sched->nr_sleep_events++;
428}
429
430static struct task_desc *register_pid(struct perf_sched *sched,
431 unsigned long pid, const char *comm)
432{
433 struct task_desc *task;
434 static int pid_max;
435
436 if (sched->pid_to_task == NULL) {
437 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
438 pid_max = MAX_PID;
439 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
440 }
441 if (pid >= (unsigned long)pid_max) {
442 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
443 sizeof(struct task_desc *))) == NULL);
444 while (pid >= (unsigned long)pid_max)
445 sched->pid_to_task[pid_max++] = NULL;
446 }
447
448 task = sched->pid_to_task[pid];
449
450 if (task)
451 return task;
452
453 task = zalloc(sizeof(*task));
454 task->pid = pid;
455 task->nr = sched->nr_tasks;
456 strcpy(task->comm, comm);
457 /*
458 * every task starts in sleeping state - this gets ignored
459 * if there's no wakeup pointing to this sleep state:
460 */
461 add_sched_event_sleep(sched, task, 0, 0);
462
463 sched->pid_to_task[pid] = task;
464 sched->nr_tasks++;
465 sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
466 BUG_ON(!sched->tasks);
467 sched->tasks[task->nr] = task;
468
469 if (verbose > 0)
470 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
471
472 return task;
473}
474
475
476static void print_task_traces(struct perf_sched *sched)
477{
478 struct task_desc *task;
479 unsigned long i;
480
481 for (i = 0; i < sched->nr_tasks; i++) {
482 task = sched->tasks[i];
483 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
484 task->nr, task->comm, task->pid, task->nr_events);
485 }
486}
487
488static void add_cross_task_wakeups(struct perf_sched *sched)
489{
490 struct task_desc *task1, *task2;
491 unsigned long i, j;
492
493 for (i = 0; i < sched->nr_tasks; i++) {
494 task1 = sched->tasks[i];
495 j = i + 1;
496 if (j == sched->nr_tasks)
497 j = 0;
498 task2 = sched->tasks[j];
499 add_sched_event_wakeup(sched, task1, 0, task2);
500 }
501}
502
503static void perf_sched__process_event(struct perf_sched *sched,
504 struct sched_atom *atom)
505{
506 int ret = 0;
507
508 switch (atom->type) {
509 case SCHED_EVENT_RUN:
510 burn_nsecs(sched, atom->duration);
511 break;
512 case SCHED_EVENT_SLEEP:
513 if (atom->wait_sem)
514 ret = sem_wait(atom->wait_sem);
515 BUG_ON(ret);
516 break;
517 case SCHED_EVENT_WAKEUP:
518 if (atom->wait_sem)
519 ret = sem_post(atom->wait_sem);
520 BUG_ON(ret);
521 break;
522 case SCHED_EVENT_MIGRATION:
523 break;
524 default:
525 BUG_ON(1);
526 }
527}
528
529static u64 get_cpu_usage_nsec_parent(void)
530{
531 struct rusage ru;
532 u64 sum;
533 int err;
534
535 err = getrusage(RUSAGE_SELF, &ru);
536 BUG_ON(err);
537
538 sum = ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
539 sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
540
541 return sum;
542}
543
544static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
545{
546 struct perf_event_attr attr;
547 char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
548 int fd;
549 struct rlimit limit;
550 bool need_privilege = false;
551
552 memset(&attr, 0, sizeof(attr));
553
554 attr.type = PERF_TYPE_SOFTWARE;
555 attr.config = PERF_COUNT_SW_TASK_CLOCK;
556
557force_again:
558 fd = sys_perf_event_open(&attr, 0, -1, -1,
559 perf_event_open_cloexec_flag());
560
561 if (fd < 0) {
562 if (errno == EMFILE) {
563 if (sched->force) {
564 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
565 limit.rlim_cur += sched->nr_tasks - cur_task;
566 if (limit.rlim_cur > limit.rlim_max) {
567 limit.rlim_max = limit.rlim_cur;
568 need_privilege = true;
569 }
570 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
571 if (need_privilege && errno == EPERM)
572 strcpy(info, "Need privilege\n");
573 } else
574 goto force_again;
575 } else
576 strcpy(info, "Have a try with -f option\n");
577 }
578 pr_err("Error: sys_perf_event_open() syscall returned "
579 "with %d (%s)\n%s", fd,
580 str_error_r(errno, sbuf, sizeof(sbuf)), info);
581 exit(EXIT_FAILURE);
582 }
583 return fd;
584}
585
586static u64 get_cpu_usage_nsec_self(int fd)
587{
588 u64 runtime;
589 int ret;
590
591 ret = read(fd, &runtime, sizeof(runtime));
592 BUG_ON(ret != sizeof(runtime));
593
594 return runtime;
595}
596
597struct sched_thread_parms {
598 struct task_desc *task;
599 struct perf_sched *sched;
600 int fd;
601};
602
603static void *thread_func(void *ctx)
604{
605 struct sched_thread_parms *parms = ctx;
606 struct task_desc *this_task = parms->task;
607 struct perf_sched *sched = parms->sched;
608 u64 cpu_usage_0, cpu_usage_1;
609 unsigned long i, ret;
610 char comm2[22];
611 int fd = parms->fd;
612
613 zfree(&parms);
614
615 sprintf(comm2, ":%s", this_task->comm);
616 prctl(PR_SET_NAME, comm2);
617 if (fd < 0)
618 return NULL;
619
620 while (!sched->thread_funcs_exit) {
621 ret = sem_post(&this_task->ready_for_work);
622 BUG_ON(ret);
623 mutex_lock(&sched->start_work_mutex);
624 mutex_unlock(&sched->start_work_mutex);
625
626 cpu_usage_0 = get_cpu_usage_nsec_self(fd);
627
628 for (i = 0; i < this_task->nr_events; i++) {
629 this_task->curr_event = i;
630 perf_sched__process_event(sched, this_task->atoms[i]);
631 }
632
633 cpu_usage_1 = get_cpu_usage_nsec_self(fd);
634 this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
635 ret = sem_post(&this_task->work_done_sem);
636 BUG_ON(ret);
637
638 mutex_lock(&sched->work_done_wait_mutex);
639 mutex_unlock(&sched->work_done_wait_mutex);
640 }
641 return NULL;
642}
643
644static void create_tasks(struct perf_sched *sched)
645 EXCLUSIVE_LOCK_FUNCTION(sched->start_work_mutex)
646 EXCLUSIVE_LOCK_FUNCTION(sched->work_done_wait_mutex)
647{
648 struct task_desc *task;
649 pthread_attr_t attr;
650 unsigned long i;
651 int err;
652
653 err = pthread_attr_init(&attr);
654 BUG_ON(err);
655 err = pthread_attr_setstacksize(&attr,
656 (size_t) max(16 * 1024, (int)PTHREAD_STACK_MIN));
657 BUG_ON(err);
658 mutex_lock(&sched->start_work_mutex);
659 mutex_lock(&sched->work_done_wait_mutex);
660 for (i = 0; i < sched->nr_tasks; i++) {
661 struct sched_thread_parms *parms = malloc(sizeof(*parms));
662 BUG_ON(parms == NULL);
663 parms->task = task = sched->tasks[i];
664 parms->sched = sched;
665 parms->fd = self_open_counters(sched, i);
666 sem_init(&task->sleep_sem, 0, 0);
667 sem_init(&task->ready_for_work, 0, 0);
668 sem_init(&task->work_done_sem, 0, 0);
669 task->curr_event = 0;
670 err = pthread_create(&task->thread, &attr, thread_func, parms);
671 BUG_ON(err);
672 }
673}
674
675static void destroy_tasks(struct perf_sched *sched)
676 UNLOCK_FUNCTION(sched->start_work_mutex)
677 UNLOCK_FUNCTION(sched->work_done_wait_mutex)
678{
679 struct task_desc *task;
680 unsigned long i;
681 int err;
682
683 mutex_unlock(&sched->start_work_mutex);
684 mutex_unlock(&sched->work_done_wait_mutex);
685 /* Get rid of threads so they won't be upset by mutex destrunction */
686 for (i = 0; i < sched->nr_tasks; i++) {
687 task = sched->tasks[i];
688 err = pthread_join(task->thread, NULL);
689 BUG_ON(err);
690 sem_destroy(&task->sleep_sem);
691 sem_destroy(&task->ready_for_work);
692 sem_destroy(&task->work_done_sem);
693 }
694}
695
696static void wait_for_tasks(struct perf_sched *sched)
697 EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
698 EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
699{
700 u64 cpu_usage_0, cpu_usage_1;
701 struct task_desc *task;
702 unsigned long i, ret;
703
704 sched->start_time = get_nsecs();
705 sched->cpu_usage = 0;
706 mutex_unlock(&sched->work_done_wait_mutex);
707
708 for (i = 0; i < sched->nr_tasks; i++) {
709 task = sched->tasks[i];
710 ret = sem_wait(&task->ready_for_work);
711 BUG_ON(ret);
712 sem_init(&task->ready_for_work, 0, 0);
713 }
714 mutex_lock(&sched->work_done_wait_mutex);
715
716 cpu_usage_0 = get_cpu_usage_nsec_parent();
717
718 mutex_unlock(&sched->start_work_mutex);
719
720 for (i = 0; i < sched->nr_tasks; i++) {
721 task = sched->tasks[i];
722 ret = sem_wait(&task->work_done_sem);
723 BUG_ON(ret);
724 sem_init(&task->work_done_sem, 0, 0);
725 sched->cpu_usage += task->cpu_usage;
726 task->cpu_usage = 0;
727 }
728
729 cpu_usage_1 = get_cpu_usage_nsec_parent();
730 if (!sched->runavg_cpu_usage)
731 sched->runavg_cpu_usage = sched->cpu_usage;
732 sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
733
734 sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
735 if (!sched->runavg_parent_cpu_usage)
736 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
737 sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
738 sched->parent_cpu_usage)/sched->replay_repeat;
739
740 mutex_lock(&sched->start_work_mutex);
741
742 for (i = 0; i < sched->nr_tasks; i++) {
743 task = sched->tasks[i];
744 sem_init(&task->sleep_sem, 0, 0);
745 task->curr_event = 0;
746 }
747}
748
749static void run_one_test(struct perf_sched *sched)
750 EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
751 EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
752{
753 u64 T0, T1, delta, avg_delta, fluct;
754
755 T0 = get_nsecs();
756 wait_for_tasks(sched);
757 T1 = get_nsecs();
758
759 delta = T1 - T0;
760 sched->sum_runtime += delta;
761 sched->nr_runs++;
762
763 avg_delta = sched->sum_runtime / sched->nr_runs;
764 if (delta < avg_delta)
765 fluct = avg_delta - delta;
766 else
767 fluct = delta - avg_delta;
768 sched->sum_fluct += fluct;
769 if (!sched->run_avg)
770 sched->run_avg = delta;
771 sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
772
773 printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
774
775 printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
776
777 printf("cpu: %0.2f / %0.2f",
778 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
779
780#if 0
781 /*
782 * rusage statistics done by the parent, these are less
783 * accurate than the sched->sum_exec_runtime based statistics:
784 */
785 printf(" [%0.2f / %0.2f]",
786 (double)sched->parent_cpu_usage / NSEC_PER_MSEC,
787 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
788#endif
789
790 printf("\n");
791
792 if (sched->nr_sleep_corrections)
793 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
794 sched->nr_sleep_corrections = 0;
795}
796
797static void test_calibrations(struct perf_sched *sched)
798{
799 u64 T0, T1;
800
801 T0 = get_nsecs();
802 burn_nsecs(sched, NSEC_PER_MSEC);
803 T1 = get_nsecs();
804
805 printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
806
807 T0 = get_nsecs();
808 sleep_nsecs(NSEC_PER_MSEC);
809 T1 = get_nsecs();
810
811 printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
812}
813
814static int
815replay_wakeup_event(struct perf_sched *sched,
816 struct evsel *evsel, struct perf_sample *sample,
817 struct machine *machine __maybe_unused)
818{
819 const char *comm = evsel__strval(evsel, sample, "comm");
820 const u32 pid = evsel__intval(evsel, sample, "pid");
821 struct task_desc *waker, *wakee;
822
823 if (verbose > 0) {
824 printf("sched_wakeup event %p\n", evsel);
825
826 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
827 }
828
829 waker = register_pid(sched, sample->tid, "<unknown>");
830 wakee = register_pid(sched, pid, comm);
831
832 add_sched_event_wakeup(sched, waker, sample->time, wakee);
833 return 0;
834}
835
836static int replay_switch_event(struct perf_sched *sched,
837 struct evsel *evsel,
838 struct perf_sample *sample,
839 struct machine *machine __maybe_unused)
840{
841 const char *prev_comm = evsel__strval(evsel, sample, "prev_comm"),
842 *next_comm = evsel__strval(evsel, sample, "next_comm");
843 const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
844 next_pid = evsel__intval(evsel, sample, "next_pid");
845 const char prev_state = evsel__taskstate(evsel, sample, "prev_state");
846 struct task_desc *prev, __maybe_unused *next;
847 u64 timestamp0, timestamp = sample->time;
848 int cpu = sample->cpu;
849 s64 delta;
850
851 if (verbose > 0)
852 printf("sched_switch event %p\n", evsel);
853
854 if (cpu >= MAX_CPUS || cpu < 0)
855 return 0;
856
857 timestamp0 = sched->cpu_last_switched[cpu];
858 if (timestamp0)
859 delta = timestamp - timestamp0;
860 else
861 delta = 0;
862
863 if (delta < 0) {
864 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
865 return -1;
866 }
867
868 pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
869 prev_comm, prev_pid, next_comm, next_pid, delta);
870
871 prev = register_pid(sched, prev_pid, prev_comm);
872 next = register_pid(sched, next_pid, next_comm);
873
874 sched->cpu_last_switched[cpu] = timestamp;
875
876 add_sched_event_run(sched, prev, timestamp, delta);
877 add_sched_event_sleep(sched, prev, timestamp, prev_state);
878
879 return 0;
880}
881
882static int replay_fork_event(struct perf_sched *sched,
883 union perf_event *event,
884 struct machine *machine)
885{
886 struct thread *child, *parent;
887
888 child = machine__findnew_thread(machine, event->fork.pid,
889 event->fork.tid);
890 parent = machine__findnew_thread(machine, event->fork.ppid,
891 event->fork.ptid);
892
893 if (child == NULL || parent == NULL) {
894 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
895 child, parent);
896 goto out_put;
897 }
898
899 if (verbose > 0) {
900 printf("fork event\n");
901 printf("... parent: %s/%d\n", thread__comm_str(parent), thread__tid(parent));
902 printf("... child: %s/%d\n", thread__comm_str(child), thread__tid(child));
903 }
904
905 register_pid(sched, thread__tid(parent), thread__comm_str(parent));
906 register_pid(sched, thread__tid(child), thread__comm_str(child));
907out_put:
908 thread__put(child);
909 thread__put(parent);
910 return 0;
911}
912
913struct sort_dimension {
914 const char *name;
915 sort_fn_t cmp;
916 struct list_head list;
917};
918
919/*
920 * handle runtime stats saved per thread
921 */
922static struct thread_runtime *thread__init_runtime(struct thread *thread)
923{
924 struct thread_runtime *r;
925
926 r = zalloc(sizeof(struct thread_runtime));
927 if (!r)
928 return NULL;
929
930 init_stats(&r->run_stats);
931 thread__set_priv(thread, r);
932
933 return r;
934}
935
936static struct thread_runtime *thread__get_runtime(struct thread *thread)
937{
938 struct thread_runtime *tr;
939
940 tr = thread__priv(thread);
941 if (tr == NULL) {
942 tr = thread__init_runtime(thread);
943 if (tr == NULL)
944 pr_debug("Failed to malloc memory for runtime data.\n");
945 }
946
947 return tr;
948}
949
950static int
951thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
952{
953 struct sort_dimension *sort;
954 int ret = 0;
955
956 BUG_ON(list_empty(list));
957
958 list_for_each_entry(sort, list, list) {
959 ret = sort->cmp(l, r);
960 if (ret)
961 return ret;
962 }
963
964 return ret;
965}
966
967static struct work_atoms *
968thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
969 struct list_head *sort_list)
970{
971 struct rb_node *node = root->rb_root.rb_node;
972 struct work_atoms key = { .thread = thread };
973
974 while (node) {
975 struct work_atoms *atoms;
976 int cmp;
977
978 atoms = container_of(node, struct work_atoms, node);
979
980 cmp = thread_lat_cmp(sort_list, &key, atoms);
981 if (cmp > 0)
982 node = node->rb_left;
983 else if (cmp < 0)
984 node = node->rb_right;
985 else {
986 BUG_ON(thread != atoms->thread);
987 return atoms;
988 }
989 }
990 return NULL;
991}
992
993static void
994__thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
995 struct list_head *sort_list)
996{
997 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
998 bool leftmost = true;
999
1000 while (*new) {
1001 struct work_atoms *this;
1002 int cmp;
1003
1004 this = container_of(*new, struct work_atoms, node);
1005 parent = *new;
1006
1007 cmp = thread_lat_cmp(sort_list, data, this);
1008
1009 if (cmp > 0)
1010 new = &((*new)->rb_left);
1011 else {
1012 new = &((*new)->rb_right);
1013 leftmost = false;
1014 }
1015 }
1016
1017 rb_link_node(&data->node, parent, new);
1018 rb_insert_color_cached(&data->node, root, leftmost);
1019}
1020
1021static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1022{
1023 struct work_atoms *atoms = zalloc(sizeof(*atoms));
1024 if (!atoms) {
1025 pr_err("No memory at %s\n", __func__);
1026 return -1;
1027 }
1028
1029 atoms->thread = thread__get(thread);
1030 INIT_LIST_HEAD(&atoms->work_list);
1031 __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1032 return 0;
1033}
1034
1035static int
1036add_sched_out_event(struct work_atoms *atoms,
1037 char run_state,
1038 u64 timestamp)
1039{
1040 struct work_atom *atom = zalloc(sizeof(*atom));
1041 if (!atom) {
1042 pr_err("Non memory at %s", __func__);
1043 return -1;
1044 }
1045
1046 atom->sched_out_time = timestamp;
1047
1048 if (run_state == 'R') {
1049 atom->state = THREAD_WAIT_CPU;
1050 atom->wake_up_time = atom->sched_out_time;
1051 }
1052
1053 list_add_tail(&atom->list, &atoms->work_list);
1054 return 0;
1055}
1056
1057static void
1058add_runtime_event(struct work_atoms *atoms, u64 delta,
1059 u64 timestamp __maybe_unused)
1060{
1061 struct work_atom *atom;
1062
1063 BUG_ON(list_empty(&atoms->work_list));
1064
1065 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1066
1067 atom->runtime += delta;
1068 atoms->total_runtime += delta;
1069}
1070
1071static void
1072add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1073{
1074 struct work_atom *atom;
1075 u64 delta;
1076
1077 if (list_empty(&atoms->work_list))
1078 return;
1079
1080 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1081
1082 if (atom->state != THREAD_WAIT_CPU)
1083 return;
1084
1085 if (timestamp < atom->wake_up_time) {
1086 atom->state = THREAD_IGNORE;
1087 return;
1088 }
1089
1090 atom->state = THREAD_SCHED_IN;
1091 atom->sched_in_time = timestamp;
1092
1093 delta = atom->sched_in_time - atom->wake_up_time;
1094 atoms->total_lat += delta;
1095 if (delta > atoms->max_lat) {
1096 atoms->max_lat = delta;
1097 atoms->max_lat_start = atom->wake_up_time;
1098 atoms->max_lat_end = timestamp;
1099 }
1100 atoms->nb_atoms++;
1101}
1102
1103static int latency_switch_event(struct perf_sched *sched,
1104 struct evsel *evsel,
1105 struct perf_sample *sample,
1106 struct machine *machine)
1107{
1108 const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1109 next_pid = evsel__intval(evsel, sample, "next_pid");
1110 const char prev_state = evsel__taskstate(evsel, sample, "prev_state");
1111 struct work_atoms *out_events, *in_events;
1112 struct thread *sched_out, *sched_in;
1113 u64 timestamp0, timestamp = sample->time;
1114 int cpu = sample->cpu, err = -1;
1115 s64 delta;
1116
1117 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1118
1119 timestamp0 = sched->cpu_last_switched[cpu];
1120 sched->cpu_last_switched[cpu] = timestamp;
1121 if (timestamp0)
1122 delta = timestamp - timestamp0;
1123 else
1124 delta = 0;
1125
1126 if (delta < 0) {
1127 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1128 return -1;
1129 }
1130
1131 sched_out = machine__findnew_thread(machine, -1, prev_pid);
1132 sched_in = machine__findnew_thread(machine, -1, next_pid);
1133 if (sched_out == NULL || sched_in == NULL)
1134 goto out_put;
1135
1136 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1137 if (!out_events) {
1138 if (thread_atoms_insert(sched, sched_out))
1139 goto out_put;
1140 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1141 if (!out_events) {
1142 pr_err("out-event: Internal tree error");
1143 goto out_put;
1144 }
1145 }
1146 if (add_sched_out_event(out_events, prev_state, timestamp))
1147 return -1;
1148
1149 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1150 if (!in_events) {
1151 if (thread_atoms_insert(sched, sched_in))
1152 goto out_put;
1153 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1154 if (!in_events) {
1155 pr_err("in-event: Internal tree error");
1156 goto out_put;
1157 }
1158 /*
1159 * Take came in we have not heard about yet,
1160 * add in an initial atom in runnable state:
1161 */
1162 if (add_sched_out_event(in_events, 'R', timestamp))
1163 goto out_put;
1164 }
1165 add_sched_in_event(in_events, timestamp);
1166 err = 0;
1167out_put:
1168 thread__put(sched_out);
1169 thread__put(sched_in);
1170 return err;
1171}
1172
1173static int latency_runtime_event(struct perf_sched *sched,
1174 struct evsel *evsel,
1175 struct perf_sample *sample,
1176 struct machine *machine)
1177{
1178 const u32 pid = evsel__intval(evsel, sample, "pid");
1179 const u64 runtime = evsel__intval(evsel, sample, "runtime");
1180 struct thread *thread = machine__findnew_thread(machine, -1, pid);
1181 struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1182 u64 timestamp = sample->time;
1183 int cpu = sample->cpu, err = -1;
1184
1185 if (thread == NULL)
1186 return -1;
1187
1188 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1189 if (!atoms) {
1190 if (thread_atoms_insert(sched, thread))
1191 goto out_put;
1192 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1193 if (!atoms) {
1194 pr_err("in-event: Internal tree error");
1195 goto out_put;
1196 }
1197 if (add_sched_out_event(atoms, 'R', timestamp))
1198 goto out_put;
1199 }
1200
1201 add_runtime_event(atoms, runtime, timestamp);
1202 err = 0;
1203out_put:
1204 thread__put(thread);
1205 return err;
1206}
1207
1208static int latency_wakeup_event(struct perf_sched *sched,
1209 struct evsel *evsel,
1210 struct perf_sample *sample,
1211 struct machine *machine)
1212{
1213 const u32 pid = evsel__intval(evsel, sample, "pid");
1214 struct work_atoms *atoms;
1215 struct work_atom *atom;
1216 struct thread *wakee;
1217 u64 timestamp = sample->time;
1218 int err = -1;
1219
1220 wakee = machine__findnew_thread(machine, -1, pid);
1221 if (wakee == NULL)
1222 return -1;
1223 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1224 if (!atoms) {
1225 if (thread_atoms_insert(sched, wakee))
1226 goto out_put;
1227 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1228 if (!atoms) {
1229 pr_err("wakeup-event: Internal tree error");
1230 goto out_put;
1231 }
1232 if (add_sched_out_event(atoms, 'S', timestamp))
1233 goto out_put;
1234 }
1235
1236 BUG_ON(list_empty(&atoms->work_list));
1237
1238 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1239
1240 /*
1241 * As we do not guarantee the wakeup event happens when
1242 * task is out of run queue, also may happen when task is
1243 * on run queue and wakeup only change ->state to TASK_RUNNING,
1244 * then we should not set the ->wake_up_time when wake up a
1245 * task which is on run queue.
1246 *
1247 * You WILL be missing events if you've recorded only
1248 * one CPU, or are only looking at only one, so don't
1249 * skip in this case.
1250 */
1251 if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1252 goto out_ok;
1253
1254 sched->nr_timestamps++;
1255 if (atom->sched_out_time > timestamp) {
1256 sched->nr_unordered_timestamps++;
1257 goto out_ok;
1258 }
1259
1260 atom->state = THREAD_WAIT_CPU;
1261 atom->wake_up_time = timestamp;
1262out_ok:
1263 err = 0;
1264out_put:
1265 thread__put(wakee);
1266 return err;
1267}
1268
1269static int latency_migrate_task_event(struct perf_sched *sched,
1270 struct evsel *evsel,
1271 struct perf_sample *sample,
1272 struct machine *machine)
1273{
1274 const u32 pid = evsel__intval(evsel, sample, "pid");
1275 u64 timestamp = sample->time;
1276 struct work_atoms *atoms;
1277 struct work_atom *atom;
1278 struct thread *migrant;
1279 int err = -1;
1280
1281 /*
1282 * Only need to worry about migration when profiling one CPU.
1283 */
1284 if (sched->profile_cpu == -1)
1285 return 0;
1286
1287 migrant = machine__findnew_thread(machine, -1, pid);
1288 if (migrant == NULL)
1289 return -1;
1290 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1291 if (!atoms) {
1292 if (thread_atoms_insert(sched, migrant))
1293 goto out_put;
1294 register_pid(sched, thread__tid(migrant), thread__comm_str(migrant));
1295 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1296 if (!atoms) {
1297 pr_err("migration-event: Internal tree error");
1298 goto out_put;
1299 }
1300 if (add_sched_out_event(atoms, 'R', timestamp))
1301 goto out_put;
1302 }
1303
1304 BUG_ON(list_empty(&atoms->work_list));
1305
1306 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1307 atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1308
1309 sched->nr_timestamps++;
1310
1311 if (atom->sched_out_time > timestamp)
1312 sched->nr_unordered_timestamps++;
1313 err = 0;
1314out_put:
1315 thread__put(migrant);
1316 return err;
1317}
1318
1319static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1320{
1321 int i;
1322 int ret;
1323 u64 avg;
1324 char max_lat_start[32], max_lat_end[32];
1325
1326 if (!work_list->nb_atoms)
1327 return;
1328 /*
1329 * Ignore idle threads:
1330 */
1331 if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1332 return;
1333
1334 sched->all_runtime += work_list->total_runtime;
1335 sched->all_count += work_list->nb_atoms;
1336
1337 if (work_list->num_merged > 1) {
1338 ret = printf(" %s:(%d) ", thread__comm_str(work_list->thread),
1339 work_list->num_merged);
1340 } else {
1341 ret = printf(" %s:%d ", thread__comm_str(work_list->thread),
1342 thread__tid(work_list->thread));
1343 }
1344
1345 for (i = 0; i < 24 - ret; i++)
1346 printf(" ");
1347
1348 avg = work_list->total_lat / work_list->nb_atoms;
1349 timestamp__scnprintf_usec(work_list->max_lat_start, max_lat_start, sizeof(max_lat_start));
1350 timestamp__scnprintf_usec(work_list->max_lat_end, max_lat_end, sizeof(max_lat_end));
1351
1352 printf("|%11.3f ms |%9" PRIu64 " | avg:%8.3f ms | max:%8.3f ms | max start: %12s s | max end: %12s s\n",
1353 (double)work_list->total_runtime / NSEC_PER_MSEC,
1354 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1355 (double)work_list->max_lat / NSEC_PER_MSEC,
1356 max_lat_start, max_lat_end);
1357}
1358
1359static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1360{
1361 pid_t l_tid, r_tid;
1362
1363 if (RC_CHK_EQUAL(l->thread, r->thread))
1364 return 0;
1365 l_tid = thread__tid(l->thread);
1366 r_tid = thread__tid(r->thread);
1367 if (l_tid < r_tid)
1368 return -1;
1369 if (l_tid > r_tid)
1370 return 1;
1371 return (int)(RC_CHK_ACCESS(l->thread) - RC_CHK_ACCESS(r->thread));
1372}
1373
1374static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1375{
1376 u64 avgl, avgr;
1377
1378 if (!l->nb_atoms)
1379 return -1;
1380
1381 if (!r->nb_atoms)
1382 return 1;
1383
1384 avgl = l->total_lat / l->nb_atoms;
1385 avgr = r->total_lat / r->nb_atoms;
1386
1387 if (avgl < avgr)
1388 return -1;
1389 if (avgl > avgr)
1390 return 1;
1391
1392 return 0;
1393}
1394
1395static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1396{
1397 if (l->max_lat < r->max_lat)
1398 return -1;
1399 if (l->max_lat > r->max_lat)
1400 return 1;
1401
1402 return 0;
1403}
1404
1405static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1406{
1407 if (l->nb_atoms < r->nb_atoms)
1408 return -1;
1409 if (l->nb_atoms > r->nb_atoms)
1410 return 1;
1411
1412 return 0;
1413}
1414
1415static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1416{
1417 if (l->total_runtime < r->total_runtime)
1418 return -1;
1419 if (l->total_runtime > r->total_runtime)
1420 return 1;
1421
1422 return 0;
1423}
1424
1425static int sort_dimension__add(const char *tok, struct list_head *list)
1426{
1427 size_t i;
1428 static struct sort_dimension avg_sort_dimension = {
1429 .name = "avg",
1430 .cmp = avg_cmp,
1431 };
1432 static struct sort_dimension max_sort_dimension = {
1433 .name = "max",
1434 .cmp = max_cmp,
1435 };
1436 static struct sort_dimension pid_sort_dimension = {
1437 .name = "pid",
1438 .cmp = pid_cmp,
1439 };
1440 static struct sort_dimension runtime_sort_dimension = {
1441 .name = "runtime",
1442 .cmp = runtime_cmp,
1443 };
1444 static struct sort_dimension switch_sort_dimension = {
1445 .name = "switch",
1446 .cmp = switch_cmp,
1447 };
1448 struct sort_dimension *available_sorts[] = {
1449 &pid_sort_dimension,
1450 &avg_sort_dimension,
1451 &max_sort_dimension,
1452 &switch_sort_dimension,
1453 &runtime_sort_dimension,
1454 };
1455
1456 for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1457 if (!strcmp(available_sorts[i]->name, tok)) {
1458 list_add_tail(&available_sorts[i]->list, list);
1459
1460 return 0;
1461 }
1462 }
1463
1464 return -1;
1465}
1466
1467static void perf_sched__sort_lat(struct perf_sched *sched)
1468{
1469 struct rb_node *node;
1470 struct rb_root_cached *root = &sched->atom_root;
1471again:
1472 for (;;) {
1473 struct work_atoms *data;
1474 node = rb_first_cached(root);
1475 if (!node)
1476 break;
1477
1478 rb_erase_cached(node, root);
1479 data = rb_entry(node, struct work_atoms, node);
1480 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1481 }
1482 if (root == &sched->atom_root) {
1483 root = &sched->merged_atom_root;
1484 goto again;
1485 }
1486}
1487
1488static int process_sched_wakeup_event(struct perf_tool *tool,
1489 struct evsel *evsel,
1490 struct perf_sample *sample,
1491 struct machine *machine)
1492{
1493 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1494
1495 if (sched->tp_handler->wakeup_event)
1496 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1497
1498 return 0;
1499}
1500
1501static int process_sched_wakeup_ignore(struct perf_tool *tool __maybe_unused,
1502 struct evsel *evsel __maybe_unused,
1503 struct perf_sample *sample __maybe_unused,
1504 struct machine *machine __maybe_unused)
1505{
1506 return 0;
1507}
1508
1509union map_priv {
1510 void *ptr;
1511 bool color;
1512};
1513
1514static bool thread__has_color(struct thread *thread)
1515{
1516 union map_priv priv = {
1517 .ptr = thread__priv(thread),
1518 };
1519
1520 return priv.color;
1521}
1522
1523static struct thread*
1524map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1525{
1526 struct thread *thread = machine__findnew_thread(machine, pid, tid);
1527 union map_priv priv = {
1528 .color = false,
1529 };
1530
1531 if (!sched->map.color_pids || !thread || thread__priv(thread))
1532 return thread;
1533
1534 if (thread_map__has(sched->map.color_pids, tid))
1535 priv.color = true;
1536
1537 thread__set_priv(thread, priv.ptr);
1538 return thread;
1539}
1540
1541static int map_switch_event(struct perf_sched *sched, struct evsel *evsel,
1542 struct perf_sample *sample, struct machine *machine)
1543{
1544 const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
1545 struct thread *sched_in;
1546 struct thread_runtime *tr;
1547 int new_shortname;
1548 u64 timestamp0, timestamp = sample->time;
1549 s64 delta;
1550 int i;
1551 struct perf_cpu this_cpu = {
1552 .cpu = sample->cpu,
1553 };
1554 int cpus_nr;
1555 bool new_cpu = false;
1556 const char *color = PERF_COLOR_NORMAL;
1557 char stimestamp[32];
1558
1559 BUG_ON(this_cpu.cpu >= MAX_CPUS || this_cpu.cpu < 0);
1560
1561 if (this_cpu.cpu > sched->max_cpu.cpu)
1562 sched->max_cpu = this_cpu;
1563
1564 if (sched->map.comp) {
1565 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1566 if (!__test_and_set_bit(this_cpu.cpu, sched->map.comp_cpus_mask)) {
1567 sched->map.comp_cpus[cpus_nr++] = this_cpu;
1568 new_cpu = true;
1569 }
1570 } else
1571 cpus_nr = sched->max_cpu.cpu;
1572
1573 timestamp0 = sched->cpu_last_switched[this_cpu.cpu];
1574 sched->cpu_last_switched[this_cpu.cpu] = timestamp;
1575 if (timestamp0)
1576 delta = timestamp - timestamp0;
1577 else
1578 delta = 0;
1579
1580 if (delta < 0) {
1581 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1582 return -1;
1583 }
1584
1585 sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1586 if (sched_in == NULL)
1587 return -1;
1588
1589 tr = thread__get_runtime(sched_in);
1590 if (tr == NULL) {
1591 thread__put(sched_in);
1592 return -1;
1593 }
1594
1595 sched->curr_thread[this_cpu.cpu] = thread__get(sched_in);
1596
1597 printf(" ");
1598
1599 new_shortname = 0;
1600 if (!tr->shortname[0]) {
1601 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1602 /*
1603 * Don't allocate a letter-number for swapper:0
1604 * as a shortname. Instead, we use '.' for it.
1605 */
1606 tr->shortname[0] = '.';
1607 tr->shortname[1] = ' ';
1608 } else {
1609 tr->shortname[0] = sched->next_shortname1;
1610 tr->shortname[1] = sched->next_shortname2;
1611
1612 if (sched->next_shortname1 < 'Z') {
1613 sched->next_shortname1++;
1614 } else {
1615 sched->next_shortname1 = 'A';
1616 if (sched->next_shortname2 < '9')
1617 sched->next_shortname2++;
1618 else
1619 sched->next_shortname2 = '0';
1620 }
1621 }
1622 new_shortname = 1;
1623 }
1624
1625 for (i = 0; i < cpus_nr; i++) {
1626 struct perf_cpu cpu = {
1627 .cpu = sched->map.comp ? sched->map.comp_cpus[i].cpu : i,
1628 };
1629 struct thread *curr_thread = sched->curr_thread[cpu.cpu];
1630 struct thread_runtime *curr_tr;
1631 const char *pid_color = color;
1632 const char *cpu_color = color;
1633
1634 if (curr_thread && thread__has_color(curr_thread))
1635 pid_color = COLOR_PIDS;
1636
1637 if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, cpu))
1638 continue;
1639
1640 if (sched->map.color_cpus && perf_cpu_map__has(sched->map.color_cpus, cpu))
1641 cpu_color = COLOR_CPUS;
1642
1643 if (cpu.cpu != this_cpu.cpu)
1644 color_fprintf(stdout, color, " ");
1645 else
1646 color_fprintf(stdout, cpu_color, "*");
1647
1648 if (sched->curr_thread[cpu.cpu]) {
1649 curr_tr = thread__get_runtime(sched->curr_thread[cpu.cpu]);
1650 if (curr_tr == NULL) {
1651 thread__put(sched_in);
1652 return -1;
1653 }
1654 color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1655 } else
1656 color_fprintf(stdout, color, " ");
1657 }
1658
1659 if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, this_cpu))
1660 goto out;
1661
1662 timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1663 color_fprintf(stdout, color, " %12s secs ", stimestamp);
1664 if (new_shortname || tr->comm_changed || (verbose > 0 && thread__tid(sched_in))) {
1665 const char *pid_color = color;
1666
1667 if (thread__has_color(sched_in))
1668 pid_color = COLOR_PIDS;
1669
1670 color_fprintf(stdout, pid_color, "%s => %s:%d",
1671 tr->shortname, thread__comm_str(sched_in), thread__tid(sched_in));
1672 tr->comm_changed = false;
1673 }
1674
1675 if (sched->map.comp && new_cpu)
1676 color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1677
1678out:
1679 color_fprintf(stdout, color, "\n");
1680
1681 thread__put(sched_in);
1682
1683 return 0;
1684}
1685
1686static int process_sched_switch_event(struct perf_tool *tool,
1687 struct evsel *evsel,
1688 struct perf_sample *sample,
1689 struct machine *machine)
1690{
1691 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1692 int this_cpu = sample->cpu, err = 0;
1693 u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1694 next_pid = evsel__intval(evsel, sample, "next_pid");
1695
1696 if (sched->curr_pid[this_cpu] != (u32)-1) {
1697 /*
1698 * Are we trying to switch away a PID that is
1699 * not current?
1700 */
1701 if (sched->curr_pid[this_cpu] != prev_pid)
1702 sched->nr_context_switch_bugs++;
1703 }
1704
1705 if (sched->tp_handler->switch_event)
1706 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1707
1708 sched->curr_pid[this_cpu] = next_pid;
1709 return err;
1710}
1711
1712static int process_sched_runtime_event(struct perf_tool *tool,
1713 struct evsel *evsel,
1714 struct perf_sample *sample,
1715 struct machine *machine)
1716{
1717 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1718
1719 if (sched->tp_handler->runtime_event)
1720 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1721
1722 return 0;
1723}
1724
1725static int perf_sched__process_fork_event(struct perf_tool *tool,
1726 union perf_event *event,
1727 struct perf_sample *sample,
1728 struct machine *machine)
1729{
1730 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1731
1732 /* run the fork event through the perf machinery */
1733 perf_event__process_fork(tool, event, sample, machine);
1734
1735 /* and then run additional processing needed for this command */
1736 if (sched->tp_handler->fork_event)
1737 return sched->tp_handler->fork_event(sched, event, machine);
1738
1739 return 0;
1740}
1741
1742static int process_sched_migrate_task_event(struct perf_tool *tool,
1743 struct evsel *evsel,
1744 struct perf_sample *sample,
1745 struct machine *machine)
1746{
1747 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1748
1749 if (sched->tp_handler->migrate_task_event)
1750 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1751
1752 return 0;
1753}
1754
1755typedef int (*tracepoint_handler)(struct perf_tool *tool,
1756 struct evsel *evsel,
1757 struct perf_sample *sample,
1758 struct machine *machine);
1759
1760static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1761 union perf_event *event __maybe_unused,
1762 struct perf_sample *sample,
1763 struct evsel *evsel,
1764 struct machine *machine)
1765{
1766 int err = 0;
1767
1768 if (evsel->handler != NULL) {
1769 tracepoint_handler f = evsel->handler;
1770 err = f(tool, evsel, sample, machine);
1771 }
1772
1773 return err;
1774}
1775
1776static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1777 union perf_event *event,
1778 struct perf_sample *sample,
1779 struct machine *machine)
1780{
1781 struct thread *thread;
1782 struct thread_runtime *tr;
1783 int err;
1784
1785 err = perf_event__process_comm(tool, event, sample, machine);
1786 if (err)
1787 return err;
1788
1789 thread = machine__find_thread(machine, sample->pid, sample->tid);
1790 if (!thread) {
1791 pr_err("Internal error: can't find thread\n");
1792 return -1;
1793 }
1794
1795 tr = thread__get_runtime(thread);
1796 if (tr == NULL) {
1797 thread__put(thread);
1798 return -1;
1799 }
1800
1801 tr->comm_changed = true;
1802 thread__put(thread);
1803
1804 return 0;
1805}
1806
1807static int perf_sched__read_events(struct perf_sched *sched)
1808{
1809 struct evsel_str_handler handlers[] = {
1810 { "sched:sched_switch", process_sched_switch_event, },
1811 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1812 { "sched:sched_wakeup", process_sched_wakeup_event, },
1813 { "sched:sched_waking", process_sched_wakeup_event, },
1814 { "sched:sched_wakeup_new", process_sched_wakeup_event, },
1815 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1816 };
1817 struct perf_session *session;
1818 struct perf_data data = {
1819 .path = input_name,
1820 .mode = PERF_DATA_MODE_READ,
1821 .force = sched->force,
1822 };
1823 int rc = -1;
1824
1825 session = perf_session__new(&data, &sched->tool);
1826 if (IS_ERR(session)) {
1827 pr_debug("Error creating perf session");
1828 return PTR_ERR(session);
1829 }
1830
1831 symbol__init(&session->header.env);
1832
1833 /* prefer sched_waking if it is captured */
1834 if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
1835 handlers[2].handler = process_sched_wakeup_ignore;
1836
1837 if (perf_session__set_tracepoints_handlers(session, handlers))
1838 goto out_delete;
1839
1840 if (perf_session__has_traces(session, "record -R")) {
1841 int err = perf_session__process_events(session);
1842 if (err) {
1843 pr_err("Failed to process events, error %d", err);
1844 goto out_delete;
1845 }
1846
1847 sched->nr_events = session->evlist->stats.nr_events[0];
1848 sched->nr_lost_events = session->evlist->stats.total_lost;
1849 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1850 }
1851
1852 rc = 0;
1853out_delete:
1854 perf_session__delete(session);
1855 return rc;
1856}
1857
1858/*
1859 * scheduling times are printed as msec.usec
1860 */
1861static inline void print_sched_time(unsigned long long nsecs, int width)
1862{
1863 unsigned long msecs;
1864 unsigned long usecs;
1865
1866 msecs = nsecs / NSEC_PER_MSEC;
1867 nsecs -= msecs * NSEC_PER_MSEC;
1868 usecs = nsecs / NSEC_PER_USEC;
1869 printf("%*lu.%03lu ", width, msecs, usecs);
1870}
1871
1872/*
1873 * returns runtime data for event, allocating memory for it the
1874 * first time it is used.
1875 */
1876static struct evsel_runtime *evsel__get_runtime(struct evsel *evsel)
1877{
1878 struct evsel_runtime *r = evsel->priv;
1879
1880 if (r == NULL) {
1881 r = zalloc(sizeof(struct evsel_runtime));
1882 evsel->priv = r;
1883 }
1884
1885 return r;
1886}
1887
1888/*
1889 * save last time event was seen per cpu
1890 */
1891static void evsel__save_time(struct evsel *evsel, u64 timestamp, u32 cpu)
1892{
1893 struct evsel_runtime *r = evsel__get_runtime(evsel);
1894
1895 if (r == NULL)
1896 return;
1897
1898 if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1899 int i, n = __roundup_pow_of_two(cpu+1);
1900 void *p = r->last_time;
1901
1902 p = realloc(r->last_time, n * sizeof(u64));
1903 if (!p)
1904 return;
1905
1906 r->last_time = p;
1907 for (i = r->ncpu; i < n; ++i)
1908 r->last_time[i] = (u64) 0;
1909
1910 r->ncpu = n;
1911 }
1912
1913 r->last_time[cpu] = timestamp;
1914}
1915
1916/* returns last time this event was seen on the given cpu */
1917static u64 evsel__get_time(struct evsel *evsel, u32 cpu)
1918{
1919 struct evsel_runtime *r = evsel__get_runtime(evsel);
1920
1921 if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1922 return 0;
1923
1924 return r->last_time[cpu];
1925}
1926
1927static int comm_width = 30;
1928
1929static char *timehist_get_commstr(struct thread *thread)
1930{
1931 static char str[32];
1932 const char *comm = thread__comm_str(thread);
1933 pid_t tid = thread__tid(thread);
1934 pid_t pid = thread__pid(thread);
1935 int n;
1936
1937 if (pid == 0)
1938 n = scnprintf(str, sizeof(str), "%s", comm);
1939
1940 else if (tid != pid)
1941 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1942
1943 else
1944 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1945
1946 if (n > comm_width)
1947 comm_width = n;
1948
1949 return str;
1950}
1951
1952static void timehist_header(struct perf_sched *sched)
1953{
1954 u32 ncpus = sched->max_cpu.cpu + 1;
1955 u32 i, j;
1956
1957 printf("%15s %6s ", "time", "cpu");
1958
1959 if (sched->show_cpu_visual) {
1960 printf(" ");
1961 for (i = 0, j = 0; i < ncpus; ++i) {
1962 printf("%x", j++);
1963 if (j > 15)
1964 j = 0;
1965 }
1966 printf(" ");
1967 }
1968
1969 printf(" %-*s %9s %9s %9s", comm_width,
1970 "task name", "wait time", "sch delay", "run time");
1971
1972 if (sched->show_state)
1973 printf(" %s", "state");
1974
1975 printf("\n");
1976
1977 /*
1978 * units row
1979 */
1980 printf("%15s %-6s ", "", "");
1981
1982 if (sched->show_cpu_visual)
1983 printf(" %*s ", ncpus, "");
1984
1985 printf(" %-*s %9s %9s %9s", comm_width,
1986 "[tid/pid]", "(msec)", "(msec)", "(msec)");
1987
1988 if (sched->show_state)
1989 printf(" %5s", "");
1990
1991 printf("\n");
1992
1993 /*
1994 * separator
1995 */
1996 printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1997
1998 if (sched->show_cpu_visual)
1999 printf(" %.*s ", ncpus, graph_dotted_line);
2000
2001 printf(" %.*s %.9s %.9s %.9s", comm_width,
2002 graph_dotted_line, graph_dotted_line, graph_dotted_line,
2003 graph_dotted_line);
2004
2005 if (sched->show_state)
2006 printf(" %.5s", graph_dotted_line);
2007
2008 printf("\n");
2009}
2010
2011static void timehist_print_sample(struct perf_sched *sched,
2012 struct evsel *evsel,
2013 struct perf_sample *sample,
2014 struct addr_location *al,
2015 struct thread *thread,
2016 u64 t, const char state)
2017{
2018 struct thread_runtime *tr = thread__priv(thread);
2019 const char *next_comm = evsel__strval(evsel, sample, "next_comm");
2020 const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
2021 u32 max_cpus = sched->max_cpu.cpu + 1;
2022 char tstr[64];
2023 char nstr[30];
2024 u64 wait_time;
2025
2026 if (cpu_list && !test_bit(sample->cpu, cpu_bitmap))
2027 return;
2028
2029 timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2030 printf("%15s [%04d] ", tstr, sample->cpu);
2031
2032 if (sched->show_cpu_visual) {
2033 u32 i;
2034 char c;
2035
2036 printf(" ");
2037 for (i = 0; i < max_cpus; ++i) {
2038 /* flag idle times with 'i'; others are sched events */
2039 if (i == sample->cpu)
2040 c = (thread__tid(thread) == 0) ? 'i' : 's';
2041 else
2042 c = ' ';
2043 printf("%c", c);
2044 }
2045 printf(" ");
2046 }
2047
2048 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2049
2050 wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2051 print_sched_time(wait_time, 6);
2052
2053 print_sched_time(tr->dt_delay, 6);
2054 print_sched_time(tr->dt_run, 6);
2055
2056 if (sched->show_state)
2057 printf(" %5c ", thread__tid(thread) == 0 ? 'I' : state);
2058
2059 if (sched->show_next) {
2060 snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2061 printf(" %-*s", comm_width, nstr);
2062 }
2063
2064 if (sched->show_wakeups && !sched->show_next)
2065 printf(" %-*s", comm_width, "");
2066
2067 if (thread__tid(thread) == 0)
2068 goto out;
2069
2070 if (sched->show_callchain)
2071 printf(" ");
2072
2073 sample__fprintf_sym(sample, al, 0,
2074 EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2075 EVSEL__PRINT_CALLCHAIN_ARROW |
2076 EVSEL__PRINT_SKIP_IGNORED,
2077 get_tls_callchain_cursor(), symbol_conf.bt_stop_list, stdout);
2078
2079out:
2080 printf("\n");
2081}
2082
2083/*
2084 * Explanation of delta-time stats:
2085 *
2086 * t = time of current schedule out event
2087 * tprev = time of previous sched out event
2088 * also time of schedule-in event for current task
2089 * last_time = time of last sched change event for current task
2090 * (i.e, time process was last scheduled out)
2091 * ready_to_run = time of wakeup for current task
2092 *
2093 * -----|------------|------------|------------|------
2094 * last ready tprev t
2095 * time to run
2096 *
2097 * |-------- dt_wait --------|
2098 * |- dt_delay -|-- dt_run --|
2099 *
2100 * dt_run = run time of current task
2101 * dt_wait = time between last schedule out event for task and tprev
2102 * represents time spent off the cpu
2103 * dt_delay = time between wakeup and schedule-in of task
2104 */
2105
2106static void timehist_update_runtime_stats(struct thread_runtime *r,
2107 u64 t, u64 tprev)
2108{
2109 r->dt_delay = 0;
2110 r->dt_sleep = 0;
2111 r->dt_iowait = 0;
2112 r->dt_preempt = 0;
2113 r->dt_run = 0;
2114
2115 if (tprev) {
2116 r->dt_run = t - tprev;
2117 if (r->ready_to_run) {
2118 if (r->ready_to_run > tprev)
2119 pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2120 else
2121 r->dt_delay = tprev - r->ready_to_run;
2122 }
2123
2124 if (r->last_time > tprev)
2125 pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2126 else if (r->last_time) {
2127 u64 dt_wait = tprev - r->last_time;
2128
2129 if (r->last_state == 'R')
2130 r->dt_preempt = dt_wait;
2131 else if (r->last_state == 'D')
2132 r->dt_iowait = dt_wait;
2133 else
2134 r->dt_sleep = dt_wait;
2135 }
2136 }
2137
2138 update_stats(&r->run_stats, r->dt_run);
2139
2140 r->total_run_time += r->dt_run;
2141 r->total_delay_time += r->dt_delay;
2142 r->total_sleep_time += r->dt_sleep;
2143 r->total_iowait_time += r->dt_iowait;
2144 r->total_preempt_time += r->dt_preempt;
2145}
2146
2147static bool is_idle_sample(struct perf_sample *sample,
2148 struct evsel *evsel)
2149{
2150 /* pid 0 == swapper == idle task */
2151 if (strcmp(evsel__name(evsel), "sched:sched_switch") == 0)
2152 return evsel__intval(evsel, sample, "prev_pid") == 0;
2153
2154 return sample->pid == 0;
2155}
2156
2157static void save_task_callchain(struct perf_sched *sched,
2158 struct perf_sample *sample,
2159 struct evsel *evsel,
2160 struct machine *machine)
2161{
2162 struct callchain_cursor *cursor;
2163 struct thread *thread;
2164
2165 /* want main thread for process - has maps */
2166 thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2167 if (thread == NULL) {
2168 pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2169 return;
2170 }
2171
2172 if (!sched->show_callchain || sample->callchain == NULL)
2173 return;
2174
2175 cursor = get_tls_callchain_cursor();
2176
2177 if (thread__resolve_callchain(thread, cursor, evsel, sample,
2178 NULL, NULL, sched->max_stack + 2) != 0) {
2179 if (verbose > 0)
2180 pr_err("Failed to resolve callchain. Skipping\n");
2181
2182 return;
2183 }
2184
2185 callchain_cursor_commit(cursor);
2186
2187 while (true) {
2188 struct callchain_cursor_node *node;
2189 struct symbol *sym;
2190
2191 node = callchain_cursor_current(cursor);
2192 if (node == NULL)
2193 break;
2194
2195 sym = node->ms.sym;
2196 if (sym) {
2197 if (!strcmp(sym->name, "schedule") ||
2198 !strcmp(sym->name, "__schedule") ||
2199 !strcmp(sym->name, "preempt_schedule"))
2200 sym->ignore = 1;
2201 }
2202
2203 callchain_cursor_advance(cursor);
2204 }
2205}
2206
2207static int init_idle_thread(struct thread *thread)
2208{
2209 struct idle_thread_runtime *itr;
2210
2211 thread__set_comm(thread, idle_comm, 0);
2212
2213 itr = zalloc(sizeof(*itr));
2214 if (itr == NULL)
2215 return -ENOMEM;
2216
2217 init_stats(&itr->tr.run_stats);
2218 callchain_init(&itr->callchain);
2219 callchain_cursor_reset(&itr->cursor);
2220 thread__set_priv(thread, itr);
2221
2222 return 0;
2223}
2224
2225/*
2226 * Track idle stats per cpu by maintaining a local thread
2227 * struct for the idle task on each cpu.
2228 */
2229static int init_idle_threads(int ncpu)
2230{
2231 int i, ret;
2232
2233 idle_threads = zalloc(ncpu * sizeof(struct thread *));
2234 if (!idle_threads)
2235 return -ENOMEM;
2236
2237 idle_max_cpu = ncpu;
2238
2239 /* allocate the actual thread struct if needed */
2240 for (i = 0; i < ncpu; ++i) {
2241 idle_threads[i] = thread__new(0, 0);
2242 if (idle_threads[i] == NULL)
2243 return -ENOMEM;
2244
2245 ret = init_idle_thread(idle_threads[i]);
2246 if (ret < 0)
2247 return ret;
2248 }
2249
2250 return 0;
2251}
2252
2253static void free_idle_threads(void)
2254{
2255 int i;
2256
2257 if (idle_threads == NULL)
2258 return;
2259
2260 for (i = 0; i < idle_max_cpu; ++i) {
2261 if ((idle_threads[i]))
2262 thread__delete(idle_threads[i]);
2263 }
2264
2265 free(idle_threads);
2266}
2267
2268static struct thread *get_idle_thread(int cpu)
2269{
2270 /*
2271 * expand/allocate array of pointers to local thread
2272 * structs if needed
2273 */
2274 if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2275 int i, j = __roundup_pow_of_two(cpu+1);
2276 void *p;
2277
2278 p = realloc(idle_threads, j * sizeof(struct thread *));
2279 if (!p)
2280 return NULL;
2281
2282 idle_threads = (struct thread **) p;
2283 for (i = idle_max_cpu; i < j; ++i)
2284 idle_threads[i] = NULL;
2285
2286 idle_max_cpu = j;
2287 }
2288
2289 /* allocate a new thread struct if needed */
2290 if (idle_threads[cpu] == NULL) {
2291 idle_threads[cpu] = thread__new(0, 0);
2292
2293 if (idle_threads[cpu]) {
2294 if (init_idle_thread(idle_threads[cpu]) < 0)
2295 return NULL;
2296 }
2297 }
2298
2299 return idle_threads[cpu];
2300}
2301
2302static void save_idle_callchain(struct perf_sched *sched,
2303 struct idle_thread_runtime *itr,
2304 struct perf_sample *sample)
2305{
2306 struct callchain_cursor *cursor;
2307
2308 if (!sched->show_callchain || sample->callchain == NULL)
2309 return;
2310
2311 cursor = get_tls_callchain_cursor();
2312 if (cursor == NULL)
2313 return;
2314
2315 callchain_cursor__copy(&itr->cursor, cursor);
2316}
2317
2318static struct thread *timehist_get_thread(struct perf_sched *sched,
2319 struct perf_sample *sample,
2320 struct machine *machine,
2321 struct evsel *evsel)
2322{
2323 struct thread *thread;
2324
2325 if (is_idle_sample(sample, evsel)) {
2326 thread = get_idle_thread(sample->cpu);
2327 if (thread == NULL)
2328 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2329
2330 } else {
2331 /* there were samples with tid 0 but non-zero pid */
2332 thread = machine__findnew_thread(machine, sample->pid,
2333 sample->tid ?: sample->pid);
2334 if (thread == NULL) {
2335 pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2336 sample->tid);
2337 }
2338
2339 save_task_callchain(sched, sample, evsel, machine);
2340 if (sched->idle_hist) {
2341 struct thread *idle;
2342 struct idle_thread_runtime *itr;
2343
2344 idle = get_idle_thread(sample->cpu);
2345 if (idle == NULL) {
2346 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2347 return NULL;
2348 }
2349
2350 itr = thread__priv(idle);
2351 if (itr == NULL)
2352 return NULL;
2353
2354 itr->last_thread = thread;
2355
2356 /* copy task callchain when entering to idle */
2357 if (evsel__intval(evsel, sample, "next_pid") == 0)
2358 save_idle_callchain(sched, itr, sample);
2359 }
2360 }
2361
2362 return thread;
2363}
2364
2365static bool timehist_skip_sample(struct perf_sched *sched,
2366 struct thread *thread,
2367 struct evsel *evsel,
2368 struct perf_sample *sample)
2369{
2370 bool rc = false;
2371
2372 if (thread__is_filtered(thread)) {
2373 rc = true;
2374 sched->skipped_samples++;
2375 }
2376
2377 if (sched->idle_hist) {
2378 if (strcmp(evsel__name(evsel), "sched:sched_switch"))
2379 rc = true;
2380 else if (evsel__intval(evsel, sample, "prev_pid") != 0 &&
2381 evsel__intval(evsel, sample, "next_pid") != 0)
2382 rc = true;
2383 }
2384
2385 return rc;
2386}
2387
2388static void timehist_print_wakeup_event(struct perf_sched *sched,
2389 struct evsel *evsel,
2390 struct perf_sample *sample,
2391 struct machine *machine,
2392 struct thread *awakened)
2393{
2394 struct thread *thread;
2395 char tstr[64];
2396
2397 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2398 if (thread == NULL)
2399 return;
2400
2401 /* show wakeup unless both awakee and awaker are filtered */
2402 if (timehist_skip_sample(sched, thread, evsel, sample) &&
2403 timehist_skip_sample(sched, awakened, evsel, sample)) {
2404 return;
2405 }
2406
2407 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2408 printf("%15s [%04d] ", tstr, sample->cpu);
2409 if (sched->show_cpu_visual)
2410 printf(" %*s ", sched->max_cpu.cpu + 1, "");
2411
2412 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2413
2414 /* dt spacer */
2415 printf(" %9s %9s %9s ", "", "", "");
2416
2417 printf("awakened: %s", timehist_get_commstr(awakened));
2418
2419 printf("\n");
2420}
2421
2422static int timehist_sched_wakeup_ignore(struct perf_tool *tool __maybe_unused,
2423 union perf_event *event __maybe_unused,
2424 struct evsel *evsel __maybe_unused,
2425 struct perf_sample *sample __maybe_unused,
2426 struct machine *machine __maybe_unused)
2427{
2428 return 0;
2429}
2430
2431static int timehist_sched_wakeup_event(struct perf_tool *tool,
2432 union perf_event *event __maybe_unused,
2433 struct evsel *evsel,
2434 struct perf_sample *sample,
2435 struct machine *machine)
2436{
2437 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2438 struct thread *thread;
2439 struct thread_runtime *tr = NULL;
2440 /* want pid of awakened task not pid in sample */
2441 const u32 pid = evsel__intval(evsel, sample, "pid");
2442
2443 thread = machine__findnew_thread(machine, 0, pid);
2444 if (thread == NULL)
2445 return -1;
2446
2447 tr = thread__get_runtime(thread);
2448 if (tr == NULL)
2449 return -1;
2450
2451 if (tr->ready_to_run == 0)
2452 tr->ready_to_run = sample->time;
2453
2454 /* show wakeups if requested */
2455 if (sched->show_wakeups &&
2456 !perf_time__skip_sample(&sched->ptime, sample->time))
2457 timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2458
2459 return 0;
2460}
2461
2462static void timehist_print_migration_event(struct perf_sched *sched,
2463 struct evsel *evsel,
2464 struct perf_sample *sample,
2465 struct machine *machine,
2466 struct thread *migrated)
2467{
2468 struct thread *thread;
2469 char tstr[64];
2470 u32 max_cpus;
2471 u32 ocpu, dcpu;
2472
2473 if (sched->summary_only)
2474 return;
2475
2476 max_cpus = sched->max_cpu.cpu + 1;
2477 ocpu = evsel__intval(evsel, sample, "orig_cpu");
2478 dcpu = evsel__intval(evsel, sample, "dest_cpu");
2479
2480 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2481 if (thread == NULL)
2482 return;
2483
2484 if (timehist_skip_sample(sched, thread, evsel, sample) &&
2485 timehist_skip_sample(sched, migrated, evsel, sample)) {
2486 return;
2487 }
2488
2489 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2490 printf("%15s [%04d] ", tstr, sample->cpu);
2491
2492 if (sched->show_cpu_visual) {
2493 u32 i;
2494 char c;
2495
2496 printf(" ");
2497 for (i = 0; i < max_cpus; ++i) {
2498 c = (i == sample->cpu) ? 'm' : ' ';
2499 printf("%c", c);
2500 }
2501 printf(" ");
2502 }
2503
2504 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2505
2506 /* dt spacer */
2507 printf(" %9s %9s %9s ", "", "", "");
2508
2509 printf("migrated: %s", timehist_get_commstr(migrated));
2510 printf(" cpu %d => %d", ocpu, dcpu);
2511
2512 printf("\n");
2513}
2514
2515static int timehist_migrate_task_event(struct perf_tool *tool,
2516 union perf_event *event __maybe_unused,
2517 struct evsel *evsel,
2518 struct perf_sample *sample,
2519 struct machine *machine)
2520{
2521 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2522 struct thread *thread;
2523 struct thread_runtime *tr = NULL;
2524 /* want pid of migrated task not pid in sample */
2525 const u32 pid = evsel__intval(evsel, sample, "pid");
2526
2527 thread = machine__findnew_thread(machine, 0, pid);
2528 if (thread == NULL)
2529 return -1;
2530
2531 tr = thread__get_runtime(thread);
2532 if (tr == NULL)
2533 return -1;
2534
2535 tr->migrations++;
2536
2537 /* show migrations if requested */
2538 timehist_print_migration_event(sched, evsel, sample, machine, thread);
2539
2540 return 0;
2541}
2542
2543static int timehist_sched_change_event(struct perf_tool *tool,
2544 union perf_event *event,
2545 struct evsel *evsel,
2546 struct perf_sample *sample,
2547 struct machine *machine)
2548{
2549 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2550 struct perf_time_interval *ptime = &sched->ptime;
2551 struct addr_location al;
2552 struct thread *thread;
2553 struct thread_runtime *tr = NULL;
2554 u64 tprev, t = sample->time;
2555 int rc = 0;
2556 const char state = evsel__taskstate(evsel, sample, "prev_state");
2557
2558 addr_location__init(&al);
2559 if (machine__resolve(machine, &al, sample) < 0) {
2560 pr_err("problem processing %d event. skipping it\n",
2561 event->header.type);
2562 rc = -1;
2563 goto out;
2564 }
2565
2566 thread = timehist_get_thread(sched, sample, machine, evsel);
2567 if (thread == NULL) {
2568 rc = -1;
2569 goto out;
2570 }
2571
2572 if (timehist_skip_sample(sched, thread, evsel, sample))
2573 goto out;
2574
2575 tr = thread__get_runtime(thread);
2576 if (tr == NULL) {
2577 rc = -1;
2578 goto out;
2579 }
2580
2581 tprev = evsel__get_time(evsel, sample->cpu);
2582
2583 /*
2584 * If start time given:
2585 * - sample time is under window user cares about - skip sample
2586 * - tprev is under window user cares about - reset to start of window
2587 */
2588 if (ptime->start && ptime->start > t)
2589 goto out;
2590
2591 if (tprev && ptime->start > tprev)
2592 tprev = ptime->start;
2593
2594 /*
2595 * If end time given:
2596 * - previous sched event is out of window - we are done
2597 * - sample time is beyond window user cares about - reset it
2598 * to close out stats for time window interest
2599 */
2600 if (ptime->end) {
2601 if (tprev > ptime->end)
2602 goto out;
2603
2604 if (t > ptime->end)
2605 t = ptime->end;
2606 }
2607
2608 if (!sched->idle_hist || thread__tid(thread) == 0) {
2609 if (!cpu_list || test_bit(sample->cpu, cpu_bitmap))
2610 timehist_update_runtime_stats(tr, t, tprev);
2611
2612 if (sched->idle_hist) {
2613 struct idle_thread_runtime *itr = (void *)tr;
2614 struct thread_runtime *last_tr;
2615
2616 BUG_ON(thread__tid(thread) != 0);
2617
2618 if (itr->last_thread == NULL)
2619 goto out;
2620
2621 /* add current idle time as last thread's runtime */
2622 last_tr = thread__get_runtime(itr->last_thread);
2623 if (last_tr == NULL)
2624 goto out;
2625
2626 timehist_update_runtime_stats(last_tr, t, tprev);
2627 /*
2628 * remove delta time of last thread as it's not updated
2629 * and otherwise it will show an invalid value next
2630 * time. we only care total run time and run stat.
2631 */
2632 last_tr->dt_run = 0;
2633 last_tr->dt_delay = 0;
2634 last_tr->dt_sleep = 0;
2635 last_tr->dt_iowait = 0;
2636 last_tr->dt_preempt = 0;
2637
2638 if (itr->cursor.nr)
2639 callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2640
2641 itr->last_thread = NULL;
2642 }
2643 }
2644
2645 if (!sched->summary_only)
2646 timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2647
2648out:
2649 if (sched->hist_time.start == 0 && t >= ptime->start)
2650 sched->hist_time.start = t;
2651 if (ptime->end == 0 || t <= ptime->end)
2652 sched->hist_time.end = t;
2653
2654 if (tr) {
2655 /* time of this sched_switch event becomes last time task seen */
2656 tr->last_time = sample->time;
2657
2658 /* last state is used to determine where to account wait time */
2659 tr->last_state = state;
2660
2661 /* sched out event for task so reset ready to run time */
2662 tr->ready_to_run = 0;
2663 }
2664
2665 evsel__save_time(evsel, sample->time, sample->cpu);
2666
2667 addr_location__exit(&al);
2668 return rc;
2669}
2670
2671static int timehist_sched_switch_event(struct perf_tool *tool,
2672 union perf_event *event,
2673 struct evsel *evsel,
2674 struct perf_sample *sample,
2675 struct machine *machine __maybe_unused)
2676{
2677 return timehist_sched_change_event(tool, event, evsel, sample, machine);
2678}
2679
2680static int process_lost(struct perf_tool *tool __maybe_unused,
2681 union perf_event *event,
2682 struct perf_sample *sample,
2683 struct machine *machine __maybe_unused)
2684{
2685 char tstr[64];
2686
2687 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2688 printf("%15s ", tstr);
2689 printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2690
2691 return 0;
2692}
2693
2694
2695static void print_thread_runtime(struct thread *t,
2696 struct thread_runtime *r)
2697{
2698 double mean = avg_stats(&r->run_stats);
2699 float stddev;
2700
2701 printf("%*s %5d %9" PRIu64 " ",
2702 comm_width, timehist_get_commstr(t), thread__ppid(t),
2703 (u64) r->run_stats.n);
2704
2705 print_sched_time(r->total_run_time, 8);
2706 stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2707 print_sched_time(r->run_stats.min, 6);
2708 printf(" ");
2709 print_sched_time((u64) mean, 6);
2710 printf(" ");
2711 print_sched_time(r->run_stats.max, 6);
2712 printf(" ");
2713 printf("%5.2f", stddev);
2714 printf(" %5" PRIu64, r->migrations);
2715 printf("\n");
2716}
2717
2718static void print_thread_waittime(struct thread *t,
2719 struct thread_runtime *r)
2720{
2721 printf("%*s %5d %9" PRIu64 " ",
2722 comm_width, timehist_get_commstr(t), thread__ppid(t),
2723 (u64) r->run_stats.n);
2724
2725 print_sched_time(r->total_run_time, 8);
2726 print_sched_time(r->total_sleep_time, 6);
2727 printf(" ");
2728 print_sched_time(r->total_iowait_time, 6);
2729 printf(" ");
2730 print_sched_time(r->total_preempt_time, 6);
2731 printf(" ");
2732 print_sched_time(r->total_delay_time, 6);
2733 printf("\n");
2734}
2735
2736struct total_run_stats {
2737 struct perf_sched *sched;
2738 u64 sched_count;
2739 u64 task_count;
2740 u64 total_run_time;
2741};
2742
2743static int show_thread_runtime(struct thread *t, void *priv)
2744{
2745 struct total_run_stats *stats = priv;
2746 struct thread_runtime *r;
2747
2748 if (thread__is_filtered(t))
2749 return 0;
2750
2751 r = thread__priv(t);
2752 if (r && r->run_stats.n) {
2753 stats->task_count++;
2754 stats->sched_count += r->run_stats.n;
2755 stats->total_run_time += r->total_run_time;
2756
2757 if (stats->sched->show_state)
2758 print_thread_waittime(t, r);
2759 else
2760 print_thread_runtime(t, r);
2761 }
2762
2763 return 0;
2764}
2765
2766static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2767{
2768 const char *sep = " <- ";
2769 struct callchain_list *chain;
2770 size_t ret = 0;
2771 char bf[1024];
2772 bool first;
2773
2774 if (node == NULL)
2775 return 0;
2776
2777 ret = callchain__fprintf_folded(fp, node->parent);
2778 first = (ret == 0);
2779
2780 list_for_each_entry(chain, &node->val, list) {
2781 if (chain->ip >= PERF_CONTEXT_MAX)
2782 continue;
2783 if (chain->ms.sym && chain->ms.sym->ignore)
2784 continue;
2785 ret += fprintf(fp, "%s%s", first ? "" : sep,
2786 callchain_list__sym_name(chain, bf, sizeof(bf),
2787 false));
2788 first = false;
2789 }
2790
2791 return ret;
2792}
2793
2794static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2795{
2796 size_t ret = 0;
2797 FILE *fp = stdout;
2798 struct callchain_node *chain;
2799 struct rb_node *rb_node = rb_first_cached(root);
2800
2801 printf(" %16s %8s %s\n", "Idle time (msec)", "Count", "Callchains");
2802 printf(" %.16s %.8s %.50s\n", graph_dotted_line, graph_dotted_line,
2803 graph_dotted_line);
2804
2805 while (rb_node) {
2806 chain = rb_entry(rb_node, struct callchain_node, rb_node);
2807 rb_node = rb_next(rb_node);
2808
2809 ret += fprintf(fp, " ");
2810 print_sched_time(chain->hit, 12);
2811 ret += 16; /* print_sched_time returns 2nd arg + 4 */
2812 ret += fprintf(fp, " %8d ", chain->count);
2813 ret += callchain__fprintf_folded(fp, chain);
2814 ret += fprintf(fp, "\n");
2815 }
2816
2817 return ret;
2818}
2819
2820static void timehist_print_summary(struct perf_sched *sched,
2821 struct perf_session *session)
2822{
2823 struct machine *m = &session->machines.host;
2824 struct total_run_stats totals;
2825 u64 task_count;
2826 struct thread *t;
2827 struct thread_runtime *r;
2828 int i;
2829 u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2830
2831 memset(&totals, 0, sizeof(totals));
2832 totals.sched = sched;
2833
2834 if (sched->idle_hist) {
2835 printf("\nIdle-time summary\n");
2836 printf("%*s parent sched-out ", comm_width, "comm");
2837 printf(" idle-time min-idle avg-idle max-idle stddev migrations\n");
2838 } else if (sched->show_state) {
2839 printf("\nWait-time summary\n");
2840 printf("%*s parent sched-in ", comm_width, "comm");
2841 printf(" run-time sleep iowait preempt delay\n");
2842 } else {
2843 printf("\nRuntime summary\n");
2844 printf("%*s parent sched-in ", comm_width, "comm");
2845 printf(" run-time min-run avg-run max-run stddev migrations\n");
2846 }
2847 printf("%*s (count) ", comm_width, "");
2848 printf(" (msec) (msec) (msec) (msec) %s\n",
2849 sched->show_state ? "(msec)" : "%");
2850 printf("%.117s\n", graph_dotted_line);
2851
2852 machine__for_each_thread(m, show_thread_runtime, &totals);
2853 task_count = totals.task_count;
2854 if (!task_count)
2855 printf("<no still running tasks>\n");
2856
2857 /* CPU idle stats not tracked when samples were skipped */
2858 if (sched->skipped_samples && !sched->idle_hist)
2859 return;
2860
2861 printf("\nIdle stats:\n");
2862 for (i = 0; i < idle_max_cpu; ++i) {
2863 if (cpu_list && !test_bit(i, cpu_bitmap))
2864 continue;
2865
2866 t = idle_threads[i];
2867 if (!t)
2868 continue;
2869
2870 r = thread__priv(t);
2871 if (r && r->run_stats.n) {
2872 totals.sched_count += r->run_stats.n;
2873 printf(" CPU %2d idle for ", i);
2874 print_sched_time(r->total_run_time, 6);
2875 printf(" msec (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2876 } else
2877 printf(" CPU %2d idle entire time window\n", i);
2878 }
2879
2880 if (sched->idle_hist && sched->show_callchain) {
2881 callchain_param.mode = CHAIN_FOLDED;
2882 callchain_param.value = CCVAL_PERIOD;
2883
2884 callchain_register_param(&callchain_param);
2885
2886 printf("\nIdle stats by callchain:\n");
2887 for (i = 0; i < idle_max_cpu; ++i) {
2888 struct idle_thread_runtime *itr;
2889
2890 t = idle_threads[i];
2891 if (!t)
2892 continue;
2893
2894 itr = thread__priv(t);
2895 if (itr == NULL)
2896 continue;
2897
2898 callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
2899 0, &callchain_param);
2900
2901 printf(" CPU %2d:", i);
2902 print_sched_time(itr->tr.total_run_time, 6);
2903 printf(" msec\n");
2904 timehist_print_idlehist_callchain(&itr->sorted_root);
2905 printf("\n");
2906 }
2907 }
2908
2909 printf("\n"
2910 " Total number of unique tasks: %" PRIu64 "\n"
2911 "Total number of context switches: %" PRIu64 "\n",
2912 totals.task_count, totals.sched_count);
2913
2914 printf(" Total run time (msec): ");
2915 print_sched_time(totals.total_run_time, 2);
2916 printf("\n");
2917
2918 printf(" Total scheduling time (msec): ");
2919 print_sched_time(hist_time, 2);
2920 printf(" (x %d)\n", sched->max_cpu.cpu);
2921}
2922
2923typedef int (*sched_handler)(struct perf_tool *tool,
2924 union perf_event *event,
2925 struct evsel *evsel,
2926 struct perf_sample *sample,
2927 struct machine *machine);
2928
2929static int perf_timehist__process_sample(struct perf_tool *tool,
2930 union perf_event *event,
2931 struct perf_sample *sample,
2932 struct evsel *evsel,
2933 struct machine *machine)
2934{
2935 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2936 int err = 0;
2937 struct perf_cpu this_cpu = {
2938 .cpu = sample->cpu,
2939 };
2940
2941 if (this_cpu.cpu > sched->max_cpu.cpu)
2942 sched->max_cpu = this_cpu;
2943
2944 if (evsel->handler != NULL) {
2945 sched_handler f = evsel->handler;
2946
2947 err = f(tool, event, evsel, sample, machine);
2948 }
2949
2950 return err;
2951}
2952
2953static int timehist_check_attr(struct perf_sched *sched,
2954 struct evlist *evlist)
2955{
2956 struct evsel *evsel;
2957 struct evsel_runtime *er;
2958
2959 list_for_each_entry(evsel, &evlist->core.entries, core.node) {
2960 er = evsel__get_runtime(evsel);
2961 if (er == NULL) {
2962 pr_err("Failed to allocate memory for evsel runtime data\n");
2963 return -1;
2964 }
2965
2966 /* only need to save callchain related to sched_switch event */
2967 if (sched->show_callchain &&
2968 evsel__name_is(evsel, "sched:sched_switch") &&
2969 !evsel__has_callchain(evsel)) {
2970 pr_info("Samples of sched_switch event do not have callchains.\n");
2971 sched->show_callchain = 0;
2972 symbol_conf.use_callchain = 0;
2973 }
2974 }
2975
2976 return 0;
2977}
2978
2979static int perf_sched__timehist(struct perf_sched *sched)
2980{
2981 struct evsel_str_handler handlers[] = {
2982 { "sched:sched_switch", timehist_sched_switch_event, },
2983 { "sched:sched_wakeup", timehist_sched_wakeup_event, },
2984 { "sched:sched_waking", timehist_sched_wakeup_event, },
2985 { "sched:sched_wakeup_new", timehist_sched_wakeup_event, },
2986 };
2987 const struct evsel_str_handler migrate_handlers[] = {
2988 { "sched:sched_migrate_task", timehist_migrate_task_event, },
2989 };
2990 struct perf_data data = {
2991 .path = input_name,
2992 .mode = PERF_DATA_MODE_READ,
2993 .force = sched->force,
2994 };
2995
2996 struct perf_session *session;
2997 struct evlist *evlist;
2998 int err = -1;
2999
3000 /*
3001 * event handlers for timehist option
3002 */
3003 sched->tool.sample = perf_timehist__process_sample;
3004 sched->tool.mmap = perf_event__process_mmap;
3005 sched->tool.comm = perf_event__process_comm;
3006 sched->tool.exit = perf_event__process_exit;
3007 sched->tool.fork = perf_event__process_fork;
3008 sched->tool.lost = process_lost;
3009 sched->tool.attr = perf_event__process_attr;
3010 sched->tool.tracing_data = perf_event__process_tracing_data;
3011 sched->tool.build_id = perf_event__process_build_id;
3012
3013 sched->tool.ordered_events = true;
3014 sched->tool.ordering_requires_timestamps = true;
3015
3016 symbol_conf.use_callchain = sched->show_callchain;
3017
3018 session = perf_session__new(&data, &sched->tool);
3019 if (IS_ERR(session))
3020 return PTR_ERR(session);
3021
3022 if (cpu_list) {
3023 err = perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap);
3024 if (err < 0)
3025 goto out;
3026 }
3027
3028 evlist = session->evlist;
3029
3030 symbol__init(&session->header.env);
3031
3032 if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
3033 pr_err("Invalid time string\n");
3034 return -EINVAL;
3035 }
3036
3037 if (timehist_check_attr(sched, evlist) != 0)
3038 goto out;
3039
3040 setup_pager();
3041
3042 /* prefer sched_waking if it is captured */
3043 if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
3044 handlers[1].handler = timehist_sched_wakeup_ignore;
3045
3046 /* setup per-evsel handlers */
3047 if (perf_session__set_tracepoints_handlers(session, handlers))
3048 goto out;
3049
3050 /* sched_switch event at a minimum needs to exist */
3051 if (!evlist__find_tracepoint_by_name(session->evlist, "sched:sched_switch")) {
3052 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3053 goto out;
3054 }
3055
3056 if (sched->show_migrations &&
3057 perf_session__set_tracepoints_handlers(session, migrate_handlers))
3058 goto out;
3059
3060 /* pre-allocate struct for per-CPU idle stats */
3061 sched->max_cpu.cpu = session->header.env.nr_cpus_online;
3062 if (sched->max_cpu.cpu == 0)
3063 sched->max_cpu.cpu = 4;
3064 if (init_idle_threads(sched->max_cpu.cpu))
3065 goto out;
3066
3067 /* summary_only implies summary option, but don't overwrite summary if set */
3068 if (sched->summary_only)
3069 sched->summary = sched->summary_only;
3070
3071 if (!sched->summary_only)
3072 timehist_header(sched);
3073
3074 err = perf_session__process_events(session);
3075 if (err) {
3076 pr_err("Failed to process events, error %d", err);
3077 goto out;
3078 }
3079
3080 sched->nr_events = evlist->stats.nr_events[0];
3081 sched->nr_lost_events = evlist->stats.total_lost;
3082 sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3083
3084 if (sched->summary)
3085 timehist_print_summary(sched, session);
3086
3087out:
3088 free_idle_threads();
3089 perf_session__delete(session);
3090
3091 return err;
3092}
3093
3094
3095static void print_bad_events(struct perf_sched *sched)
3096{
3097 if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3098 printf(" INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3099 (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3100 sched->nr_unordered_timestamps, sched->nr_timestamps);
3101 }
3102 if (sched->nr_lost_events && sched->nr_events) {
3103 printf(" INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3104 (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3105 sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3106 }
3107 if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3108 printf(" INFO: %.3f%% context switch bugs (%ld out of %ld)",
3109 (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3110 sched->nr_context_switch_bugs, sched->nr_timestamps);
3111 if (sched->nr_lost_events)
3112 printf(" (due to lost events?)");
3113 printf("\n");
3114 }
3115}
3116
3117static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3118{
3119 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3120 struct work_atoms *this;
3121 const char *comm = thread__comm_str(data->thread), *this_comm;
3122 bool leftmost = true;
3123
3124 while (*new) {
3125 int cmp;
3126
3127 this = container_of(*new, struct work_atoms, node);
3128 parent = *new;
3129
3130 this_comm = thread__comm_str(this->thread);
3131 cmp = strcmp(comm, this_comm);
3132 if (cmp > 0) {
3133 new = &((*new)->rb_left);
3134 } else if (cmp < 0) {
3135 new = &((*new)->rb_right);
3136 leftmost = false;
3137 } else {
3138 this->num_merged++;
3139 this->total_runtime += data->total_runtime;
3140 this->nb_atoms += data->nb_atoms;
3141 this->total_lat += data->total_lat;
3142 list_splice(&data->work_list, &this->work_list);
3143 if (this->max_lat < data->max_lat) {
3144 this->max_lat = data->max_lat;
3145 this->max_lat_start = data->max_lat_start;
3146 this->max_lat_end = data->max_lat_end;
3147 }
3148 zfree(&data);
3149 return;
3150 }
3151 }
3152
3153 data->num_merged++;
3154 rb_link_node(&data->node, parent, new);
3155 rb_insert_color_cached(&data->node, root, leftmost);
3156}
3157
3158static void perf_sched__merge_lat(struct perf_sched *sched)
3159{
3160 struct work_atoms *data;
3161 struct rb_node *node;
3162
3163 if (sched->skip_merge)
3164 return;
3165
3166 while ((node = rb_first_cached(&sched->atom_root))) {
3167 rb_erase_cached(node, &sched->atom_root);
3168 data = rb_entry(node, struct work_atoms, node);
3169 __merge_work_atoms(&sched->merged_atom_root, data);
3170 }
3171}
3172
3173static int setup_cpus_switch_event(struct perf_sched *sched)
3174{
3175 unsigned int i;
3176
3177 sched->cpu_last_switched = calloc(MAX_CPUS, sizeof(*(sched->cpu_last_switched)));
3178 if (!sched->cpu_last_switched)
3179 return -1;
3180
3181 sched->curr_pid = malloc(MAX_CPUS * sizeof(*(sched->curr_pid)));
3182 if (!sched->curr_pid) {
3183 zfree(&sched->cpu_last_switched);
3184 return -1;
3185 }
3186
3187 for (i = 0; i < MAX_CPUS; i++)
3188 sched->curr_pid[i] = -1;
3189
3190 return 0;
3191}
3192
3193static void free_cpus_switch_event(struct perf_sched *sched)
3194{
3195 zfree(&sched->curr_pid);
3196 zfree(&sched->cpu_last_switched);
3197}
3198
3199static int perf_sched__lat(struct perf_sched *sched)
3200{
3201 int rc = -1;
3202 struct rb_node *next;
3203
3204 setup_pager();
3205
3206 if (setup_cpus_switch_event(sched))
3207 return rc;
3208
3209 if (perf_sched__read_events(sched))
3210 goto out_free_cpus_switch_event;
3211
3212 perf_sched__merge_lat(sched);
3213 perf_sched__sort_lat(sched);
3214
3215 printf("\n -------------------------------------------------------------------------------------------------------------------------------------------\n");
3216 printf(" Task | Runtime ms | Switches | Avg delay ms | Max delay ms | Max delay start | Max delay end |\n");
3217 printf(" -------------------------------------------------------------------------------------------------------------------------------------------\n");
3218
3219 next = rb_first_cached(&sched->sorted_atom_root);
3220
3221 while (next) {
3222 struct work_atoms *work_list;
3223
3224 work_list = rb_entry(next, struct work_atoms, node);
3225 output_lat_thread(sched, work_list);
3226 next = rb_next(next);
3227 thread__zput(work_list->thread);
3228 }
3229
3230 printf(" -----------------------------------------------------------------------------------------------------------------\n");
3231 printf(" TOTAL: |%11.3f ms |%9" PRIu64 " |\n",
3232 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3233
3234 printf(" ---------------------------------------------------\n");
3235
3236 print_bad_events(sched);
3237 printf("\n");
3238
3239 rc = 0;
3240
3241out_free_cpus_switch_event:
3242 free_cpus_switch_event(sched);
3243 return rc;
3244}
3245
3246static int setup_map_cpus(struct perf_sched *sched)
3247{
3248 sched->max_cpu.cpu = sysconf(_SC_NPROCESSORS_CONF);
3249
3250 if (sched->map.comp) {
3251 sched->map.comp_cpus = zalloc(sched->max_cpu.cpu * sizeof(int));
3252 if (!sched->map.comp_cpus)
3253 return -1;
3254 }
3255
3256 if (sched->map.cpus_str) {
3257 sched->map.cpus = perf_cpu_map__new(sched->map.cpus_str);
3258 if (!sched->map.cpus) {
3259 pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3260 zfree(&sched->map.comp_cpus);
3261 return -1;
3262 }
3263 }
3264
3265 return 0;
3266}
3267
3268static int setup_color_pids(struct perf_sched *sched)
3269{
3270 struct perf_thread_map *map;
3271
3272 if (!sched->map.color_pids_str)
3273 return 0;
3274
3275 map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3276 if (!map) {
3277 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3278 return -1;
3279 }
3280
3281 sched->map.color_pids = map;
3282 return 0;
3283}
3284
3285static int setup_color_cpus(struct perf_sched *sched)
3286{
3287 struct perf_cpu_map *map;
3288
3289 if (!sched->map.color_cpus_str)
3290 return 0;
3291
3292 map = perf_cpu_map__new(sched->map.color_cpus_str);
3293 if (!map) {
3294 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3295 return -1;
3296 }
3297
3298 sched->map.color_cpus = map;
3299 return 0;
3300}
3301
3302static int perf_sched__map(struct perf_sched *sched)
3303{
3304 int rc = -1;
3305
3306 sched->curr_thread = calloc(MAX_CPUS, sizeof(*(sched->curr_thread)));
3307 if (!sched->curr_thread)
3308 return rc;
3309
3310 if (setup_cpus_switch_event(sched))
3311 goto out_free_curr_thread;
3312
3313 if (setup_map_cpus(sched))
3314 goto out_free_cpus_switch_event;
3315
3316 if (setup_color_pids(sched))
3317 goto out_put_map_cpus;
3318
3319 if (setup_color_cpus(sched))
3320 goto out_put_color_pids;
3321
3322 setup_pager();
3323 if (perf_sched__read_events(sched))
3324 goto out_put_color_cpus;
3325
3326 rc = 0;
3327 print_bad_events(sched);
3328
3329out_put_color_cpus:
3330 perf_cpu_map__put(sched->map.color_cpus);
3331
3332out_put_color_pids:
3333 perf_thread_map__put(sched->map.color_pids);
3334
3335out_put_map_cpus:
3336 zfree(&sched->map.comp_cpus);
3337 perf_cpu_map__put(sched->map.cpus);
3338
3339out_free_cpus_switch_event:
3340 free_cpus_switch_event(sched);
3341
3342out_free_curr_thread:
3343 zfree(&sched->curr_thread);
3344 return rc;
3345}
3346
3347static int perf_sched__replay(struct perf_sched *sched)
3348{
3349 int ret;
3350 unsigned long i;
3351
3352 mutex_init(&sched->start_work_mutex);
3353 mutex_init(&sched->work_done_wait_mutex);
3354
3355 ret = setup_cpus_switch_event(sched);
3356 if (ret)
3357 goto out_mutex_destroy;
3358
3359 calibrate_run_measurement_overhead(sched);
3360 calibrate_sleep_measurement_overhead(sched);
3361
3362 test_calibrations(sched);
3363
3364 ret = perf_sched__read_events(sched);
3365 if (ret)
3366 goto out_free_cpus_switch_event;
3367
3368 printf("nr_run_events: %ld\n", sched->nr_run_events);
3369 printf("nr_sleep_events: %ld\n", sched->nr_sleep_events);
3370 printf("nr_wakeup_events: %ld\n", sched->nr_wakeup_events);
3371
3372 if (sched->targetless_wakeups)
3373 printf("target-less wakeups: %ld\n", sched->targetless_wakeups);
3374 if (sched->multitarget_wakeups)
3375 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3376 if (sched->nr_run_events_optimized)
3377 printf("run atoms optimized: %ld\n",
3378 sched->nr_run_events_optimized);
3379
3380 print_task_traces(sched);
3381 add_cross_task_wakeups(sched);
3382
3383 sched->thread_funcs_exit = false;
3384 create_tasks(sched);
3385 printf("------------------------------------------------------------\n");
3386 for (i = 0; i < sched->replay_repeat; i++)
3387 run_one_test(sched);
3388
3389 sched->thread_funcs_exit = true;
3390 destroy_tasks(sched);
3391
3392out_free_cpus_switch_event:
3393 free_cpus_switch_event(sched);
3394
3395out_mutex_destroy:
3396 mutex_destroy(&sched->start_work_mutex);
3397 mutex_destroy(&sched->work_done_wait_mutex);
3398 return ret;
3399}
3400
3401static void setup_sorting(struct perf_sched *sched, const struct option *options,
3402 const char * const usage_msg[])
3403{
3404 char *tmp, *tok, *str = strdup(sched->sort_order);
3405
3406 for (tok = strtok_r(str, ", ", &tmp);
3407 tok; tok = strtok_r(NULL, ", ", &tmp)) {
3408 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3409 usage_with_options_msg(usage_msg, options,
3410 "Unknown --sort key: `%s'", tok);
3411 }
3412 }
3413
3414 free(str);
3415
3416 sort_dimension__add("pid", &sched->cmp_pid);
3417}
3418
3419static bool schedstat_events_exposed(void)
3420{
3421 /*
3422 * Select "sched:sched_stat_wait" event to check
3423 * whether schedstat tracepoints are exposed.
3424 */
3425 return IS_ERR(trace_event__tp_format("sched", "sched_stat_wait")) ?
3426 false : true;
3427}
3428
3429static int __cmd_record(int argc, const char **argv)
3430{
3431 unsigned int rec_argc, i, j;
3432 char **rec_argv;
3433 const char **rec_argv_copy;
3434 const char * const record_args[] = {
3435 "record",
3436 "-a",
3437 "-R",
3438 "-m", "1024",
3439 "-c", "1",
3440 "-e", "sched:sched_switch",
3441 "-e", "sched:sched_stat_runtime",
3442 "-e", "sched:sched_process_fork",
3443 "-e", "sched:sched_wakeup_new",
3444 "-e", "sched:sched_migrate_task",
3445 };
3446
3447 /*
3448 * The tracepoints trace_sched_stat_{wait, sleep, iowait}
3449 * are not exposed to user if CONFIG_SCHEDSTATS is not set,
3450 * to prevent "perf sched record" execution failure, determine
3451 * whether to record schedstat events according to actual situation.
3452 */
3453 const char * const schedstat_args[] = {
3454 "-e", "sched:sched_stat_wait",
3455 "-e", "sched:sched_stat_sleep",
3456 "-e", "sched:sched_stat_iowait",
3457 };
3458 unsigned int schedstat_argc = schedstat_events_exposed() ?
3459 ARRAY_SIZE(schedstat_args) : 0;
3460
3461 struct tep_event *waking_event;
3462 int ret;
3463
3464 /*
3465 * +2 for either "-e", "sched:sched_wakeup" or
3466 * "-e", "sched:sched_waking"
3467 */
3468 rec_argc = ARRAY_SIZE(record_args) + 2 + schedstat_argc + argc - 1;
3469 rec_argv = calloc(rec_argc + 1, sizeof(char *));
3470 if (rec_argv == NULL)
3471 return -ENOMEM;
3472 rec_argv_copy = calloc(rec_argc + 1, sizeof(char *));
3473 if (rec_argv_copy == NULL) {
3474 free(rec_argv);
3475 return -ENOMEM;
3476 }
3477
3478 for (i = 0; i < ARRAY_SIZE(record_args); i++)
3479 rec_argv[i] = strdup(record_args[i]);
3480
3481 rec_argv[i++] = strdup("-e");
3482 waking_event = trace_event__tp_format("sched", "sched_waking");
3483 if (!IS_ERR(waking_event))
3484 rec_argv[i++] = strdup("sched:sched_waking");
3485 else
3486 rec_argv[i++] = strdup("sched:sched_wakeup");
3487
3488 for (j = 0; j < schedstat_argc; j++)
3489 rec_argv[i++] = strdup(schedstat_args[j]);
3490
3491 for (j = 1; j < (unsigned int)argc; j++, i++)
3492 rec_argv[i] = strdup(argv[j]);
3493
3494 BUG_ON(i != rec_argc);
3495
3496 memcpy(rec_argv_copy, rec_argv, sizeof(char *) * rec_argc);
3497 ret = cmd_record(rec_argc, rec_argv_copy);
3498
3499 for (i = 0; i < rec_argc; i++)
3500 free(rec_argv[i]);
3501 free(rec_argv);
3502 free(rec_argv_copy);
3503
3504 return ret;
3505}
3506
3507int cmd_sched(int argc, const char **argv)
3508{
3509 static const char default_sort_order[] = "avg, max, switch, runtime";
3510 struct perf_sched sched = {
3511 .tool = {
3512 .sample = perf_sched__process_tracepoint_sample,
3513 .comm = perf_sched__process_comm,
3514 .namespaces = perf_event__process_namespaces,
3515 .lost = perf_event__process_lost,
3516 .fork = perf_sched__process_fork_event,
3517 .ordered_events = true,
3518 },
3519 .cmp_pid = LIST_HEAD_INIT(sched.cmp_pid),
3520 .sort_list = LIST_HEAD_INIT(sched.sort_list),
3521 .sort_order = default_sort_order,
3522 .replay_repeat = 10,
3523 .profile_cpu = -1,
3524 .next_shortname1 = 'A',
3525 .next_shortname2 = '0',
3526 .skip_merge = 0,
3527 .show_callchain = 1,
3528 .max_stack = 5,
3529 };
3530 const struct option sched_options[] = {
3531 OPT_STRING('i', "input", &input_name, "file",
3532 "input file name"),
3533 OPT_INCR('v', "verbose", &verbose,
3534 "be more verbose (show symbol address, etc)"),
3535 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3536 "dump raw trace in ASCII"),
3537 OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3538 OPT_END()
3539 };
3540 const struct option latency_options[] = {
3541 OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3542 "sort by key(s): runtime, switch, avg, max"),
3543 OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3544 "CPU to profile on"),
3545 OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3546 "latency stats per pid instead of per comm"),
3547 OPT_PARENT(sched_options)
3548 };
3549 const struct option replay_options[] = {
3550 OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3551 "repeat the workload replay N times (-1: infinite)"),
3552 OPT_PARENT(sched_options)
3553 };
3554 const struct option map_options[] = {
3555 OPT_BOOLEAN(0, "compact", &sched.map.comp,
3556 "map output in compact mode"),
3557 OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3558 "highlight given pids in map"),
3559 OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3560 "highlight given CPUs in map"),
3561 OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3562 "display given CPUs in map"),
3563 OPT_PARENT(sched_options)
3564 };
3565 const struct option timehist_options[] = {
3566 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3567 "file", "vmlinux pathname"),
3568 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3569 "file", "kallsyms pathname"),
3570 OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3571 "Display call chains if present (default on)"),
3572 OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3573 "Maximum number of functions to display backtrace."),
3574 OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3575 "Look for files with symbols relative to this directory"),
3576 OPT_BOOLEAN('s', "summary", &sched.summary_only,
3577 "Show only syscall summary with statistics"),
3578 OPT_BOOLEAN('S', "with-summary", &sched.summary,
3579 "Show all syscalls and summary with statistics"),
3580 OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3581 OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3582 OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3583 OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3584 OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3585 OPT_STRING(0, "time", &sched.time_str, "str",
3586 "Time span for analysis (start,stop)"),
3587 OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3588 OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3589 "analyze events only for given process id(s)"),
3590 OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3591 "analyze events only for given thread id(s)"),
3592 OPT_STRING('C', "cpu", &cpu_list, "cpu", "list of cpus to profile"),
3593 OPT_PARENT(sched_options)
3594 };
3595
3596 const char * const latency_usage[] = {
3597 "perf sched latency [<options>]",
3598 NULL
3599 };
3600 const char * const replay_usage[] = {
3601 "perf sched replay [<options>]",
3602 NULL
3603 };
3604 const char * const map_usage[] = {
3605 "perf sched map [<options>]",
3606 NULL
3607 };
3608 const char * const timehist_usage[] = {
3609 "perf sched timehist [<options>]",
3610 NULL
3611 };
3612 const char *const sched_subcommands[] = { "record", "latency", "map",
3613 "replay", "script",
3614 "timehist", NULL };
3615 const char *sched_usage[] = {
3616 NULL,
3617 NULL
3618 };
3619 struct trace_sched_handler lat_ops = {
3620 .wakeup_event = latency_wakeup_event,
3621 .switch_event = latency_switch_event,
3622 .runtime_event = latency_runtime_event,
3623 .migrate_task_event = latency_migrate_task_event,
3624 };
3625 struct trace_sched_handler map_ops = {
3626 .switch_event = map_switch_event,
3627 };
3628 struct trace_sched_handler replay_ops = {
3629 .wakeup_event = replay_wakeup_event,
3630 .switch_event = replay_switch_event,
3631 .fork_event = replay_fork_event,
3632 };
3633 int ret;
3634
3635 argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3636 sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3637 if (!argc)
3638 usage_with_options(sched_usage, sched_options);
3639
3640 /*
3641 * Aliased to 'perf script' for now:
3642 */
3643 if (!strcmp(argv[0], "script")) {
3644 return cmd_script(argc, argv);
3645 } else if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
3646 return __cmd_record(argc, argv);
3647 } else if (strlen(argv[0]) > 2 && strstarts("latency", argv[0])) {
3648 sched.tp_handler = &lat_ops;
3649 if (argc > 1) {
3650 argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3651 if (argc)
3652 usage_with_options(latency_usage, latency_options);
3653 }
3654 setup_sorting(&sched, latency_options, latency_usage);
3655 return perf_sched__lat(&sched);
3656 } else if (!strcmp(argv[0], "map")) {
3657 if (argc) {
3658 argc = parse_options(argc, argv, map_options, map_usage, 0);
3659 if (argc)
3660 usage_with_options(map_usage, map_options);
3661 }
3662 sched.tp_handler = &map_ops;
3663 setup_sorting(&sched, latency_options, latency_usage);
3664 return perf_sched__map(&sched);
3665 } else if (strlen(argv[0]) > 2 && strstarts("replay", argv[0])) {
3666 sched.tp_handler = &replay_ops;
3667 if (argc) {
3668 argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3669 if (argc)
3670 usage_with_options(replay_usage, replay_options);
3671 }
3672 return perf_sched__replay(&sched);
3673 } else if (!strcmp(argv[0], "timehist")) {
3674 if (argc) {
3675 argc = parse_options(argc, argv, timehist_options,
3676 timehist_usage, 0);
3677 if (argc)
3678 usage_with_options(timehist_usage, timehist_options);
3679 }
3680 if ((sched.show_wakeups || sched.show_next) &&
3681 sched.summary_only) {
3682 pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3683 parse_options_usage(timehist_usage, timehist_options, "s", true);
3684 if (sched.show_wakeups)
3685 parse_options_usage(NULL, timehist_options, "w", true);
3686 if (sched.show_next)
3687 parse_options_usage(NULL, timehist_options, "n", true);
3688 return -EINVAL;
3689 }
3690 ret = symbol__validate_sym_arguments();
3691 if (ret)
3692 return ret;
3693
3694 return perf_sched__timehist(&sched);
3695 } else {
3696 usage_with_options(sched_usage, sched_options);
3697 }
3698
3699 return 0;
3700}
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
14#include "util/parse-options.h"
15#include "util/trace-event.h"
16
17#include "util/debug.h"
18
19#include <sys/prctl.h>
20#include <sys/resource.h>
21
22#include <semaphore.h>
23#include <pthread.h>
24#include <math.h>
25
26static const char *input_name;
27
28static char default_sort_order[] = "avg, max, switch, runtime";
29static const char *sort_order = default_sort_order;
30
31static int profile_cpu = -1;
32
33#define PR_SET_NAME 15 /* Set process name */
34#define MAX_CPUS 4096
35
36static u64 run_measurement_overhead;
37static u64 sleep_measurement_overhead;
38
39#define COMM_LEN 20
40#define SYM_LEN 129
41
42#define MAX_PID 65536
43
44static unsigned long nr_tasks;
45
46struct sched_atom;
47
48struct task_desc {
49 unsigned long nr;
50 unsigned long pid;
51 char comm[COMM_LEN];
52
53 unsigned long nr_events;
54 unsigned long curr_event;
55 struct sched_atom **atoms;
56
57 pthread_t thread;
58 sem_t sleep_sem;
59
60 sem_t ready_for_work;
61 sem_t work_done_sem;
62
63 u64 cpu_usage;
64};
65
66enum sched_event_type {
67 SCHED_EVENT_RUN,
68 SCHED_EVENT_SLEEP,
69 SCHED_EVENT_WAKEUP,
70 SCHED_EVENT_MIGRATION,
71};
72
73struct sched_atom {
74 enum sched_event_type type;
75 int specific_wait;
76 u64 timestamp;
77 u64 duration;
78 unsigned long nr;
79 sem_t *wait_sem;
80 struct task_desc *wakee;
81};
82
83static struct task_desc *pid_to_task[MAX_PID];
84
85static struct task_desc **tasks;
86
87static pthread_mutex_t start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
88static u64 start_time;
89
90static pthread_mutex_t work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
91
92static unsigned long nr_run_events;
93static unsigned long nr_sleep_events;
94static unsigned long nr_wakeup_events;
95
96static unsigned long nr_sleep_corrections;
97static unsigned long nr_run_events_optimized;
98
99static unsigned long targetless_wakeups;
100static unsigned long multitarget_wakeups;
101
102static u64 cpu_usage;
103static u64 runavg_cpu_usage;
104static u64 parent_cpu_usage;
105static u64 runavg_parent_cpu_usage;
106
107static unsigned long nr_runs;
108static u64 sum_runtime;
109static u64 sum_fluct;
110static u64 run_avg;
111
112static unsigned int replay_repeat = 10;
113static unsigned long nr_timestamps;
114static unsigned long nr_unordered_timestamps;
115static unsigned long nr_state_machine_bugs;
116static unsigned long nr_context_switch_bugs;
117static unsigned long nr_events;
118static unsigned long nr_lost_chunks;
119static unsigned long nr_lost_events;
120
121#define TASK_STATE_TO_CHAR_STR "RSDTtZX"
122
123enum thread_state {
124 THREAD_SLEEPING = 0,
125 THREAD_WAIT_CPU,
126 THREAD_SCHED_IN,
127 THREAD_IGNORE
128};
129
130struct work_atom {
131 struct list_head list;
132 enum thread_state state;
133 u64 sched_out_time;
134 u64 wake_up_time;
135 u64 sched_in_time;
136 u64 runtime;
137};
138
139struct work_atoms {
140 struct list_head work_list;
141 struct thread *thread;
142 struct rb_node node;
143 u64 max_lat;
144 u64 max_lat_at;
145 u64 total_lat;
146 u64 nb_atoms;
147 u64 total_runtime;
148};
149
150typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
151
152static struct rb_root atom_root, sorted_atom_root;
153
154static u64 all_runtime;
155static u64 all_count;
156
157
158static u64 get_nsecs(void)
159{
160 struct timespec ts;
161
162 clock_gettime(CLOCK_MONOTONIC, &ts);
163
164 return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
165}
166
167static void burn_nsecs(u64 nsecs)
168{
169 u64 T0 = get_nsecs(), T1;
170
171 do {
172 T1 = get_nsecs();
173 } while (T1 + run_measurement_overhead < T0 + nsecs);
174}
175
176static void sleep_nsecs(u64 nsecs)
177{
178 struct timespec ts;
179
180 ts.tv_nsec = nsecs % 999999999;
181 ts.tv_sec = nsecs / 999999999;
182
183 nanosleep(&ts, NULL);
184}
185
186static void calibrate_run_measurement_overhead(void)
187{
188 u64 T0, T1, delta, min_delta = 1000000000ULL;
189 int i;
190
191 for (i = 0; i < 10; i++) {
192 T0 = get_nsecs();
193 burn_nsecs(0);
194 T1 = get_nsecs();
195 delta = T1-T0;
196 min_delta = min(min_delta, delta);
197 }
198 run_measurement_overhead = min_delta;
199
200 printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
201}
202
203static void calibrate_sleep_measurement_overhead(void)
204{
205 u64 T0, T1, delta, min_delta = 1000000000ULL;
206 int i;
207
208 for (i = 0; i < 10; i++) {
209 T0 = get_nsecs();
210 sleep_nsecs(10000);
211 T1 = get_nsecs();
212 delta = T1-T0;
213 min_delta = min(min_delta, delta);
214 }
215 min_delta -= 10000;
216 sleep_measurement_overhead = min_delta;
217
218 printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
219}
220
221static struct sched_atom *
222get_new_event(struct task_desc *task, u64 timestamp)
223{
224 struct sched_atom *event = zalloc(sizeof(*event));
225 unsigned long idx = task->nr_events;
226 size_t size;
227
228 event->timestamp = timestamp;
229 event->nr = idx;
230
231 task->nr_events++;
232 size = sizeof(struct sched_atom *) * task->nr_events;
233 task->atoms = realloc(task->atoms, size);
234 BUG_ON(!task->atoms);
235
236 task->atoms[idx] = event;
237
238 return event;
239}
240
241static struct sched_atom *last_event(struct task_desc *task)
242{
243 if (!task->nr_events)
244 return NULL;
245
246 return task->atoms[task->nr_events - 1];
247}
248
249static void
250add_sched_event_run(struct task_desc *task, u64 timestamp, u64 duration)
251{
252 struct sched_atom *event, *curr_event = last_event(task);
253
254 /*
255 * optimize an existing RUN event by merging this one
256 * to it:
257 */
258 if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
259 nr_run_events_optimized++;
260 curr_event->duration += duration;
261 return;
262 }
263
264 event = get_new_event(task, timestamp);
265
266 event->type = SCHED_EVENT_RUN;
267 event->duration = duration;
268
269 nr_run_events++;
270}
271
272static void
273add_sched_event_wakeup(struct task_desc *task, u64 timestamp,
274 struct task_desc *wakee)
275{
276 struct sched_atom *event, *wakee_event;
277
278 event = get_new_event(task, timestamp);
279 event->type = SCHED_EVENT_WAKEUP;
280 event->wakee = wakee;
281
282 wakee_event = last_event(wakee);
283 if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
284 targetless_wakeups++;
285 return;
286 }
287 if (wakee_event->wait_sem) {
288 multitarget_wakeups++;
289 return;
290 }
291
292 wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
293 sem_init(wakee_event->wait_sem, 0, 0);
294 wakee_event->specific_wait = 1;
295 event->wait_sem = wakee_event->wait_sem;
296
297 nr_wakeup_events++;
298}
299
300static void
301add_sched_event_sleep(struct task_desc *task, u64 timestamp,
302 u64 task_state __used)
303{
304 struct sched_atom *event = get_new_event(task, timestamp);
305
306 event->type = SCHED_EVENT_SLEEP;
307
308 nr_sleep_events++;
309}
310
311static struct task_desc *register_pid(unsigned long pid, const char *comm)
312{
313 struct task_desc *task;
314
315 BUG_ON(pid >= MAX_PID);
316
317 task = pid_to_task[pid];
318
319 if (task)
320 return task;
321
322 task = zalloc(sizeof(*task));
323 task->pid = pid;
324 task->nr = nr_tasks;
325 strcpy(task->comm, comm);
326 /*
327 * every task starts in sleeping state - this gets ignored
328 * if there's no wakeup pointing to this sleep state:
329 */
330 add_sched_event_sleep(task, 0, 0);
331
332 pid_to_task[pid] = task;
333 nr_tasks++;
334 tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
335 BUG_ON(!tasks);
336 tasks[task->nr] = task;
337
338 if (verbose)
339 printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
340
341 return task;
342}
343
344
345static void print_task_traces(void)
346{
347 struct task_desc *task;
348 unsigned long i;
349
350 for (i = 0; i < nr_tasks; i++) {
351 task = tasks[i];
352 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
353 task->nr, task->comm, task->pid, task->nr_events);
354 }
355}
356
357static void add_cross_task_wakeups(void)
358{
359 struct task_desc *task1, *task2;
360 unsigned long i, j;
361
362 for (i = 0; i < nr_tasks; i++) {
363 task1 = tasks[i];
364 j = i + 1;
365 if (j == nr_tasks)
366 j = 0;
367 task2 = tasks[j];
368 add_sched_event_wakeup(task1, 0, task2);
369 }
370}
371
372static void
373process_sched_event(struct task_desc *this_task __used, struct sched_atom *atom)
374{
375 int ret = 0;
376
377 switch (atom->type) {
378 case SCHED_EVENT_RUN:
379 burn_nsecs(atom->duration);
380 break;
381 case SCHED_EVENT_SLEEP:
382 if (atom->wait_sem)
383 ret = sem_wait(atom->wait_sem);
384 BUG_ON(ret);
385 break;
386 case SCHED_EVENT_WAKEUP:
387 if (atom->wait_sem)
388 ret = sem_post(atom->wait_sem);
389 BUG_ON(ret);
390 break;
391 case SCHED_EVENT_MIGRATION:
392 break;
393 default:
394 BUG_ON(1);
395 }
396}
397
398static u64 get_cpu_usage_nsec_parent(void)
399{
400 struct rusage ru;
401 u64 sum;
402 int err;
403
404 err = getrusage(RUSAGE_SELF, &ru);
405 BUG_ON(err);
406
407 sum = ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
408 sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
409
410 return sum;
411}
412
413static int self_open_counters(void)
414{
415 struct perf_event_attr attr;
416 int fd;
417
418 memset(&attr, 0, sizeof(attr));
419
420 attr.type = PERF_TYPE_SOFTWARE;
421 attr.config = PERF_COUNT_SW_TASK_CLOCK;
422
423 fd = sys_perf_event_open(&attr, 0, -1, -1, 0);
424
425 if (fd < 0)
426 die("Error: sys_perf_event_open() syscall returned"
427 "with %d (%s)\n", fd, strerror(errno));
428 return fd;
429}
430
431static u64 get_cpu_usage_nsec_self(int fd)
432{
433 u64 runtime;
434 int ret;
435
436 ret = read(fd, &runtime, sizeof(runtime));
437 BUG_ON(ret != sizeof(runtime));
438
439 return runtime;
440}
441
442static void *thread_func(void *ctx)
443{
444 struct task_desc *this_task = ctx;
445 u64 cpu_usage_0, cpu_usage_1;
446 unsigned long i, ret;
447 char comm2[22];
448 int fd;
449
450 sprintf(comm2, ":%s", this_task->comm);
451 prctl(PR_SET_NAME, comm2);
452 fd = self_open_counters();
453
454again:
455 ret = sem_post(&this_task->ready_for_work);
456 BUG_ON(ret);
457 ret = pthread_mutex_lock(&start_work_mutex);
458 BUG_ON(ret);
459 ret = pthread_mutex_unlock(&start_work_mutex);
460 BUG_ON(ret);
461
462 cpu_usage_0 = get_cpu_usage_nsec_self(fd);
463
464 for (i = 0; i < this_task->nr_events; i++) {
465 this_task->curr_event = i;
466 process_sched_event(this_task, this_task->atoms[i]);
467 }
468
469 cpu_usage_1 = get_cpu_usage_nsec_self(fd);
470 this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
471 ret = sem_post(&this_task->work_done_sem);
472 BUG_ON(ret);
473
474 ret = pthread_mutex_lock(&work_done_wait_mutex);
475 BUG_ON(ret);
476 ret = pthread_mutex_unlock(&work_done_wait_mutex);
477 BUG_ON(ret);
478
479 goto again;
480}
481
482static void create_tasks(void)
483{
484 struct task_desc *task;
485 pthread_attr_t attr;
486 unsigned long i;
487 int err;
488
489 err = pthread_attr_init(&attr);
490 BUG_ON(err);
491 err = pthread_attr_setstacksize(&attr,
492 (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
493 BUG_ON(err);
494 err = pthread_mutex_lock(&start_work_mutex);
495 BUG_ON(err);
496 err = pthread_mutex_lock(&work_done_wait_mutex);
497 BUG_ON(err);
498 for (i = 0; i < nr_tasks; i++) {
499 task = tasks[i];
500 sem_init(&task->sleep_sem, 0, 0);
501 sem_init(&task->ready_for_work, 0, 0);
502 sem_init(&task->work_done_sem, 0, 0);
503 task->curr_event = 0;
504 err = pthread_create(&task->thread, &attr, thread_func, task);
505 BUG_ON(err);
506 }
507}
508
509static void wait_for_tasks(void)
510{
511 u64 cpu_usage_0, cpu_usage_1;
512 struct task_desc *task;
513 unsigned long i, ret;
514
515 start_time = get_nsecs();
516 cpu_usage = 0;
517 pthread_mutex_unlock(&work_done_wait_mutex);
518
519 for (i = 0; i < nr_tasks; i++) {
520 task = tasks[i];
521 ret = sem_wait(&task->ready_for_work);
522 BUG_ON(ret);
523 sem_init(&task->ready_for_work, 0, 0);
524 }
525 ret = pthread_mutex_lock(&work_done_wait_mutex);
526 BUG_ON(ret);
527
528 cpu_usage_0 = get_cpu_usage_nsec_parent();
529
530 pthread_mutex_unlock(&start_work_mutex);
531
532 for (i = 0; i < nr_tasks; i++) {
533 task = tasks[i];
534 ret = sem_wait(&task->work_done_sem);
535 BUG_ON(ret);
536 sem_init(&task->work_done_sem, 0, 0);
537 cpu_usage += task->cpu_usage;
538 task->cpu_usage = 0;
539 }
540
541 cpu_usage_1 = get_cpu_usage_nsec_parent();
542 if (!runavg_cpu_usage)
543 runavg_cpu_usage = cpu_usage;
544 runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
545
546 parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
547 if (!runavg_parent_cpu_usage)
548 runavg_parent_cpu_usage = parent_cpu_usage;
549 runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
550 parent_cpu_usage)/10;
551
552 ret = pthread_mutex_lock(&start_work_mutex);
553 BUG_ON(ret);
554
555 for (i = 0; i < nr_tasks; i++) {
556 task = tasks[i];
557 sem_init(&task->sleep_sem, 0, 0);
558 task->curr_event = 0;
559 }
560}
561
562static void run_one_test(void)
563{
564 u64 T0, T1, delta, avg_delta, fluct;
565
566 T0 = get_nsecs();
567 wait_for_tasks();
568 T1 = get_nsecs();
569
570 delta = T1 - T0;
571 sum_runtime += delta;
572 nr_runs++;
573
574 avg_delta = sum_runtime / nr_runs;
575 if (delta < avg_delta)
576 fluct = avg_delta - delta;
577 else
578 fluct = delta - avg_delta;
579 sum_fluct += fluct;
580 if (!run_avg)
581 run_avg = delta;
582 run_avg = (run_avg*9 + delta)/10;
583
584 printf("#%-3ld: %0.3f, ",
585 nr_runs, (double)delta/1000000.0);
586
587 printf("ravg: %0.2f, ",
588 (double)run_avg/1e6);
589
590 printf("cpu: %0.2f / %0.2f",
591 (double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6);
592
593#if 0
594 /*
595 * rusage statistics done by the parent, these are less
596 * accurate than the sum_exec_runtime based statistics:
597 */
598 printf(" [%0.2f / %0.2f]",
599 (double)parent_cpu_usage/1e6,
600 (double)runavg_parent_cpu_usage/1e6);
601#endif
602
603 printf("\n");
604
605 if (nr_sleep_corrections)
606 printf(" (%ld sleep corrections)\n", nr_sleep_corrections);
607 nr_sleep_corrections = 0;
608}
609
610static void test_calibrations(void)
611{
612 u64 T0, T1;
613
614 T0 = get_nsecs();
615 burn_nsecs(1e6);
616 T1 = get_nsecs();
617
618 printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
619
620 T0 = get_nsecs();
621 sleep_nsecs(1e6);
622 T1 = get_nsecs();
623
624 printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
625}
626
627#define FILL_FIELD(ptr, field, event, data) \
628 ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
629
630#define FILL_ARRAY(ptr, array, event, data) \
631do { \
632 void *__array = raw_field_ptr(event, #array, data); \
633 memcpy(ptr.array, __array, sizeof(ptr.array)); \
634} while(0)
635
636#define FILL_COMMON_FIELDS(ptr, event, data) \
637do { \
638 FILL_FIELD(ptr, common_type, event, data); \
639 FILL_FIELD(ptr, common_flags, event, data); \
640 FILL_FIELD(ptr, common_preempt_count, event, data); \
641 FILL_FIELD(ptr, common_pid, event, data); \
642 FILL_FIELD(ptr, common_tgid, event, data); \
643} while (0)
644
645
646
647struct trace_switch_event {
648 u32 size;
649
650 u16 common_type;
651 u8 common_flags;
652 u8 common_preempt_count;
653 u32 common_pid;
654 u32 common_tgid;
655
656 char prev_comm[16];
657 u32 prev_pid;
658 u32 prev_prio;
659 u64 prev_state;
660 char next_comm[16];
661 u32 next_pid;
662 u32 next_prio;
663};
664
665struct trace_runtime_event {
666 u32 size;
667
668 u16 common_type;
669 u8 common_flags;
670 u8 common_preempt_count;
671 u32 common_pid;
672 u32 common_tgid;
673
674 char comm[16];
675 u32 pid;
676 u64 runtime;
677 u64 vruntime;
678};
679
680struct trace_wakeup_event {
681 u32 size;
682
683 u16 common_type;
684 u8 common_flags;
685 u8 common_preempt_count;
686 u32 common_pid;
687 u32 common_tgid;
688
689 char comm[16];
690 u32 pid;
691
692 u32 prio;
693 u32 success;
694 u32 cpu;
695};
696
697struct trace_fork_event {
698 u32 size;
699
700 u16 common_type;
701 u8 common_flags;
702 u8 common_preempt_count;
703 u32 common_pid;
704 u32 common_tgid;
705
706 char parent_comm[16];
707 u32 parent_pid;
708 char child_comm[16];
709 u32 child_pid;
710};
711
712struct trace_migrate_task_event {
713 u32 size;
714
715 u16 common_type;
716 u8 common_flags;
717 u8 common_preempt_count;
718 u32 common_pid;
719 u32 common_tgid;
720
721 char comm[16];
722 u32 pid;
723
724 u32 prio;
725 u32 cpu;
726};
727
728struct trace_sched_handler {
729 void (*switch_event)(struct trace_switch_event *,
730 struct machine *,
731 struct event_format *,
732 int cpu,
733 u64 timestamp,
734 struct thread *thread);
735
736 void (*runtime_event)(struct trace_runtime_event *,
737 struct machine *,
738 struct event_format *,
739 int cpu,
740 u64 timestamp,
741 struct thread *thread);
742
743 void (*wakeup_event)(struct trace_wakeup_event *,
744 struct machine *,
745 struct event_format *,
746 int cpu,
747 u64 timestamp,
748 struct thread *thread);
749
750 void (*fork_event)(struct trace_fork_event *,
751 struct event_format *,
752 int cpu,
753 u64 timestamp,
754 struct thread *thread);
755
756 void (*migrate_task_event)(struct trace_migrate_task_event *,
757 struct machine *machine,
758 struct event_format *,
759 int cpu,
760 u64 timestamp,
761 struct thread *thread);
762};
763
764
765static void
766replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
767 struct machine *machine __used,
768 struct event_format *event,
769 int cpu __used,
770 u64 timestamp __used,
771 struct thread *thread __used)
772{
773 struct task_desc *waker, *wakee;
774
775 if (verbose) {
776 printf("sched_wakeup event %p\n", event);
777
778 printf(" ... pid %d woke up %s/%d\n",
779 wakeup_event->common_pid,
780 wakeup_event->comm,
781 wakeup_event->pid);
782 }
783
784 waker = register_pid(wakeup_event->common_pid, "<unknown>");
785 wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
786
787 add_sched_event_wakeup(waker, timestamp, wakee);
788}
789
790static u64 cpu_last_switched[MAX_CPUS];
791
792static void
793replay_switch_event(struct trace_switch_event *switch_event,
794 struct machine *machine __used,
795 struct event_format *event,
796 int cpu,
797 u64 timestamp,
798 struct thread *thread __used)
799{
800 struct task_desc *prev, __used *next;
801 u64 timestamp0;
802 s64 delta;
803
804 if (verbose)
805 printf("sched_switch event %p\n", event);
806
807 if (cpu >= MAX_CPUS || cpu < 0)
808 return;
809
810 timestamp0 = cpu_last_switched[cpu];
811 if (timestamp0)
812 delta = timestamp - timestamp0;
813 else
814 delta = 0;
815
816 if (delta < 0)
817 die("hm, delta: %" PRIu64 " < 0 ?\n", delta);
818
819 if (verbose) {
820 printf(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
821 switch_event->prev_comm, switch_event->prev_pid,
822 switch_event->next_comm, switch_event->next_pid,
823 delta);
824 }
825
826 prev = register_pid(switch_event->prev_pid, switch_event->prev_comm);
827 next = register_pid(switch_event->next_pid, switch_event->next_comm);
828
829 cpu_last_switched[cpu] = timestamp;
830
831 add_sched_event_run(prev, timestamp, delta);
832 add_sched_event_sleep(prev, timestamp, switch_event->prev_state);
833}
834
835
836static void
837replay_fork_event(struct trace_fork_event *fork_event,
838 struct event_format *event,
839 int cpu __used,
840 u64 timestamp __used,
841 struct thread *thread __used)
842{
843 if (verbose) {
844 printf("sched_fork event %p\n", event);
845 printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid);
846 printf("... child: %s/%d\n", fork_event->child_comm, fork_event->child_pid);
847 }
848 register_pid(fork_event->parent_pid, fork_event->parent_comm);
849 register_pid(fork_event->child_pid, fork_event->child_comm);
850}
851
852static struct trace_sched_handler replay_ops = {
853 .wakeup_event = replay_wakeup_event,
854 .switch_event = replay_switch_event,
855 .fork_event = replay_fork_event,
856};
857
858struct sort_dimension {
859 const char *name;
860 sort_fn_t cmp;
861 struct list_head list;
862};
863
864static LIST_HEAD(cmp_pid);
865
866static int
867thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
868{
869 struct sort_dimension *sort;
870 int ret = 0;
871
872 BUG_ON(list_empty(list));
873
874 list_for_each_entry(sort, list, list) {
875 ret = sort->cmp(l, r);
876 if (ret)
877 return ret;
878 }
879
880 return ret;
881}
882
883static struct work_atoms *
884thread_atoms_search(struct rb_root *root, struct thread *thread,
885 struct list_head *sort_list)
886{
887 struct rb_node *node = root->rb_node;
888 struct work_atoms key = { .thread = thread };
889
890 while (node) {
891 struct work_atoms *atoms;
892 int cmp;
893
894 atoms = container_of(node, struct work_atoms, node);
895
896 cmp = thread_lat_cmp(sort_list, &key, atoms);
897 if (cmp > 0)
898 node = node->rb_left;
899 else if (cmp < 0)
900 node = node->rb_right;
901 else {
902 BUG_ON(thread != atoms->thread);
903 return atoms;
904 }
905 }
906 return NULL;
907}
908
909static void
910__thread_latency_insert(struct rb_root *root, struct work_atoms *data,
911 struct list_head *sort_list)
912{
913 struct rb_node **new = &(root->rb_node), *parent = NULL;
914
915 while (*new) {
916 struct work_atoms *this;
917 int cmp;
918
919 this = container_of(*new, struct work_atoms, node);
920 parent = *new;
921
922 cmp = thread_lat_cmp(sort_list, data, this);
923
924 if (cmp > 0)
925 new = &((*new)->rb_left);
926 else
927 new = &((*new)->rb_right);
928 }
929
930 rb_link_node(&data->node, parent, new);
931 rb_insert_color(&data->node, root);
932}
933
934static void thread_atoms_insert(struct thread *thread)
935{
936 struct work_atoms *atoms = zalloc(sizeof(*atoms));
937 if (!atoms)
938 die("No memory");
939
940 atoms->thread = thread;
941 INIT_LIST_HEAD(&atoms->work_list);
942 __thread_latency_insert(&atom_root, atoms, &cmp_pid);
943}
944
945static void
946latency_fork_event(struct trace_fork_event *fork_event __used,
947 struct event_format *event __used,
948 int cpu __used,
949 u64 timestamp __used,
950 struct thread *thread __used)
951{
952 /* should insert the newcomer */
953}
954
955__used
956static char sched_out_state(struct trace_switch_event *switch_event)
957{
958 const char *str = TASK_STATE_TO_CHAR_STR;
959
960 return str[switch_event->prev_state];
961}
962
963static void
964add_sched_out_event(struct work_atoms *atoms,
965 char run_state,
966 u64 timestamp)
967{
968 struct work_atom *atom = zalloc(sizeof(*atom));
969 if (!atom)
970 die("Non memory");
971
972 atom->sched_out_time = timestamp;
973
974 if (run_state == 'R') {
975 atom->state = THREAD_WAIT_CPU;
976 atom->wake_up_time = atom->sched_out_time;
977 }
978
979 list_add_tail(&atom->list, &atoms->work_list);
980}
981
982static void
983add_runtime_event(struct work_atoms *atoms, u64 delta, u64 timestamp __used)
984{
985 struct work_atom *atom;
986
987 BUG_ON(list_empty(&atoms->work_list));
988
989 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
990
991 atom->runtime += delta;
992 atoms->total_runtime += delta;
993}
994
995static void
996add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
997{
998 struct work_atom *atom;
999 u64 delta;
1000
1001 if (list_empty(&atoms->work_list))
1002 return;
1003
1004 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1005
1006 if (atom->state != THREAD_WAIT_CPU)
1007 return;
1008
1009 if (timestamp < atom->wake_up_time) {
1010 atom->state = THREAD_IGNORE;
1011 return;
1012 }
1013
1014 atom->state = THREAD_SCHED_IN;
1015 atom->sched_in_time = timestamp;
1016
1017 delta = atom->sched_in_time - atom->wake_up_time;
1018 atoms->total_lat += delta;
1019 if (delta > atoms->max_lat) {
1020 atoms->max_lat = delta;
1021 atoms->max_lat_at = timestamp;
1022 }
1023 atoms->nb_atoms++;
1024}
1025
1026static void
1027latency_switch_event(struct trace_switch_event *switch_event,
1028 struct machine *machine,
1029 struct event_format *event __used,
1030 int cpu,
1031 u64 timestamp,
1032 struct thread *thread __used)
1033{
1034 struct work_atoms *out_events, *in_events;
1035 struct thread *sched_out, *sched_in;
1036 u64 timestamp0;
1037 s64 delta;
1038
1039 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1040
1041 timestamp0 = cpu_last_switched[cpu];
1042 cpu_last_switched[cpu] = timestamp;
1043 if (timestamp0)
1044 delta = timestamp - timestamp0;
1045 else
1046 delta = 0;
1047
1048 if (delta < 0)
1049 die("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1050
1051
1052 sched_out = machine__findnew_thread(machine, switch_event->prev_pid);
1053 sched_in = machine__findnew_thread(machine, switch_event->next_pid);
1054
1055 out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1056 if (!out_events) {
1057 thread_atoms_insert(sched_out);
1058 out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1059 if (!out_events)
1060 die("out-event: Internal tree error");
1061 }
1062 add_sched_out_event(out_events, sched_out_state(switch_event), timestamp);
1063
1064 in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1065 if (!in_events) {
1066 thread_atoms_insert(sched_in);
1067 in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1068 if (!in_events)
1069 die("in-event: Internal tree error");
1070 /*
1071 * Take came in we have not heard about yet,
1072 * add in an initial atom in runnable state:
1073 */
1074 add_sched_out_event(in_events, 'R', timestamp);
1075 }
1076 add_sched_in_event(in_events, timestamp);
1077}
1078
1079static void
1080latency_runtime_event(struct trace_runtime_event *runtime_event,
1081 struct machine *machine,
1082 struct event_format *event __used,
1083 int cpu,
1084 u64 timestamp,
1085 struct thread *this_thread __used)
1086{
1087 struct thread *thread = machine__findnew_thread(machine, runtime_event->pid);
1088 struct work_atoms *atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1089
1090 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1091 if (!atoms) {
1092 thread_atoms_insert(thread);
1093 atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1094 if (!atoms)
1095 die("in-event: Internal tree error");
1096 add_sched_out_event(atoms, 'R', timestamp);
1097 }
1098
1099 add_runtime_event(atoms, runtime_event->runtime, timestamp);
1100}
1101
1102static void
1103latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
1104 struct machine *machine,
1105 struct event_format *__event __used,
1106 int cpu __used,
1107 u64 timestamp,
1108 struct thread *thread __used)
1109{
1110 struct work_atoms *atoms;
1111 struct work_atom *atom;
1112 struct thread *wakee;
1113
1114 /* Note for later, it may be interesting to observe the failing cases */
1115 if (!wakeup_event->success)
1116 return;
1117
1118 wakee = machine__findnew_thread(machine, wakeup_event->pid);
1119 atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1120 if (!atoms) {
1121 thread_atoms_insert(wakee);
1122 atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1123 if (!atoms)
1124 die("wakeup-event: Internal tree error");
1125 add_sched_out_event(atoms, 'S', timestamp);
1126 }
1127
1128 BUG_ON(list_empty(&atoms->work_list));
1129
1130 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1131
1132 /*
1133 * You WILL be missing events if you've recorded only
1134 * one CPU, or are only looking at only one, so don't
1135 * make useless noise.
1136 */
1137 if (profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1138 nr_state_machine_bugs++;
1139
1140 nr_timestamps++;
1141 if (atom->sched_out_time > timestamp) {
1142 nr_unordered_timestamps++;
1143 return;
1144 }
1145
1146 atom->state = THREAD_WAIT_CPU;
1147 atom->wake_up_time = timestamp;
1148}
1149
1150static void
1151latency_migrate_task_event(struct trace_migrate_task_event *migrate_task_event,
1152 struct machine *machine,
1153 struct event_format *__event __used,
1154 int cpu __used,
1155 u64 timestamp,
1156 struct thread *thread __used)
1157{
1158 struct work_atoms *atoms;
1159 struct work_atom *atom;
1160 struct thread *migrant;
1161
1162 /*
1163 * Only need to worry about migration when profiling one CPU.
1164 */
1165 if (profile_cpu == -1)
1166 return;
1167
1168 migrant = machine__findnew_thread(machine, migrate_task_event->pid);
1169 atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1170 if (!atoms) {
1171 thread_atoms_insert(migrant);
1172 register_pid(migrant->pid, migrant->comm);
1173 atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1174 if (!atoms)
1175 die("migration-event: Internal tree error");
1176 add_sched_out_event(atoms, 'R', timestamp);
1177 }
1178
1179 BUG_ON(list_empty(&atoms->work_list));
1180
1181 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1182 atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1183
1184 nr_timestamps++;
1185
1186 if (atom->sched_out_time > timestamp)
1187 nr_unordered_timestamps++;
1188}
1189
1190static struct trace_sched_handler lat_ops = {
1191 .wakeup_event = latency_wakeup_event,
1192 .switch_event = latency_switch_event,
1193 .runtime_event = latency_runtime_event,
1194 .fork_event = latency_fork_event,
1195 .migrate_task_event = latency_migrate_task_event,
1196};
1197
1198static void output_lat_thread(struct work_atoms *work_list)
1199{
1200 int i;
1201 int ret;
1202 u64 avg;
1203
1204 if (!work_list->nb_atoms)
1205 return;
1206 /*
1207 * Ignore idle threads:
1208 */
1209 if (!strcmp(work_list->thread->comm, "swapper"))
1210 return;
1211
1212 all_runtime += work_list->total_runtime;
1213 all_count += work_list->nb_atoms;
1214
1215 ret = printf(" %s:%d ", work_list->thread->comm, work_list->thread->pid);
1216
1217 for (i = 0; i < 24 - ret; i++)
1218 printf(" ");
1219
1220 avg = work_list->total_lat / work_list->nb_atoms;
1221
1222 printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %9.6f s\n",
1223 (double)work_list->total_runtime / 1e6,
1224 work_list->nb_atoms, (double)avg / 1e6,
1225 (double)work_list->max_lat / 1e6,
1226 (double)work_list->max_lat_at / 1e9);
1227}
1228
1229static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1230{
1231 if (l->thread->pid < r->thread->pid)
1232 return -1;
1233 if (l->thread->pid > r->thread->pid)
1234 return 1;
1235
1236 return 0;
1237}
1238
1239static struct sort_dimension pid_sort_dimension = {
1240 .name = "pid",
1241 .cmp = pid_cmp,
1242};
1243
1244static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1245{
1246 u64 avgl, avgr;
1247
1248 if (!l->nb_atoms)
1249 return -1;
1250
1251 if (!r->nb_atoms)
1252 return 1;
1253
1254 avgl = l->total_lat / l->nb_atoms;
1255 avgr = r->total_lat / r->nb_atoms;
1256
1257 if (avgl < avgr)
1258 return -1;
1259 if (avgl > avgr)
1260 return 1;
1261
1262 return 0;
1263}
1264
1265static struct sort_dimension avg_sort_dimension = {
1266 .name = "avg",
1267 .cmp = avg_cmp,
1268};
1269
1270static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1271{
1272 if (l->max_lat < r->max_lat)
1273 return -1;
1274 if (l->max_lat > r->max_lat)
1275 return 1;
1276
1277 return 0;
1278}
1279
1280static struct sort_dimension max_sort_dimension = {
1281 .name = "max",
1282 .cmp = max_cmp,
1283};
1284
1285static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1286{
1287 if (l->nb_atoms < r->nb_atoms)
1288 return -1;
1289 if (l->nb_atoms > r->nb_atoms)
1290 return 1;
1291
1292 return 0;
1293}
1294
1295static struct sort_dimension switch_sort_dimension = {
1296 .name = "switch",
1297 .cmp = switch_cmp,
1298};
1299
1300static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1301{
1302 if (l->total_runtime < r->total_runtime)
1303 return -1;
1304 if (l->total_runtime > r->total_runtime)
1305 return 1;
1306
1307 return 0;
1308}
1309
1310static struct sort_dimension runtime_sort_dimension = {
1311 .name = "runtime",
1312 .cmp = runtime_cmp,
1313};
1314
1315static struct sort_dimension *available_sorts[] = {
1316 &pid_sort_dimension,
1317 &avg_sort_dimension,
1318 &max_sort_dimension,
1319 &switch_sort_dimension,
1320 &runtime_sort_dimension,
1321};
1322
1323#define NB_AVAILABLE_SORTS (int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
1324
1325static LIST_HEAD(sort_list);
1326
1327static int sort_dimension__add(const char *tok, struct list_head *list)
1328{
1329 int i;
1330
1331 for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
1332 if (!strcmp(available_sorts[i]->name, tok)) {
1333 list_add_tail(&available_sorts[i]->list, list);
1334
1335 return 0;
1336 }
1337 }
1338
1339 return -1;
1340}
1341
1342static void setup_sorting(void);
1343
1344static void sort_lat(void)
1345{
1346 struct rb_node *node;
1347
1348 for (;;) {
1349 struct work_atoms *data;
1350 node = rb_first(&atom_root);
1351 if (!node)
1352 break;
1353
1354 rb_erase(node, &atom_root);
1355 data = rb_entry(node, struct work_atoms, node);
1356 __thread_latency_insert(&sorted_atom_root, data, &sort_list);
1357 }
1358}
1359
1360static struct trace_sched_handler *trace_handler;
1361
1362static void
1363process_sched_wakeup_event(struct perf_tool *tool __used,
1364 struct event_format *event,
1365 struct perf_sample *sample,
1366 struct machine *machine,
1367 struct thread *thread)
1368{
1369 void *data = sample->raw_data;
1370 struct trace_wakeup_event wakeup_event;
1371
1372 FILL_COMMON_FIELDS(wakeup_event, event, data);
1373
1374 FILL_ARRAY(wakeup_event, comm, event, data);
1375 FILL_FIELD(wakeup_event, pid, event, data);
1376 FILL_FIELD(wakeup_event, prio, event, data);
1377 FILL_FIELD(wakeup_event, success, event, data);
1378 FILL_FIELD(wakeup_event, cpu, event, data);
1379
1380 if (trace_handler->wakeup_event)
1381 trace_handler->wakeup_event(&wakeup_event, machine, event,
1382 sample->cpu, sample->time, thread);
1383}
1384
1385/*
1386 * Track the current task - that way we can know whether there's any
1387 * weird events, such as a task being switched away that is not current.
1388 */
1389static int max_cpu;
1390
1391static u32 curr_pid[MAX_CPUS] = { [0 ... MAX_CPUS-1] = -1 };
1392
1393static struct thread *curr_thread[MAX_CPUS];
1394
1395static char next_shortname1 = 'A';
1396static char next_shortname2 = '0';
1397
1398static void
1399map_switch_event(struct trace_switch_event *switch_event,
1400 struct machine *machine,
1401 struct event_format *event __used,
1402 int this_cpu,
1403 u64 timestamp,
1404 struct thread *thread __used)
1405{
1406 struct thread *sched_out __used, *sched_in;
1407 int new_shortname;
1408 u64 timestamp0;
1409 s64 delta;
1410 int cpu;
1411
1412 BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1413
1414 if (this_cpu > max_cpu)
1415 max_cpu = this_cpu;
1416
1417 timestamp0 = cpu_last_switched[this_cpu];
1418 cpu_last_switched[this_cpu] = timestamp;
1419 if (timestamp0)
1420 delta = timestamp - timestamp0;
1421 else
1422 delta = 0;
1423
1424 if (delta < 0)
1425 die("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1426
1427
1428 sched_out = machine__findnew_thread(machine, switch_event->prev_pid);
1429 sched_in = machine__findnew_thread(machine, switch_event->next_pid);
1430
1431 curr_thread[this_cpu] = sched_in;
1432
1433 printf(" ");
1434
1435 new_shortname = 0;
1436 if (!sched_in->shortname[0]) {
1437 sched_in->shortname[0] = next_shortname1;
1438 sched_in->shortname[1] = next_shortname2;
1439
1440 if (next_shortname1 < 'Z') {
1441 next_shortname1++;
1442 } else {
1443 next_shortname1='A';
1444 if (next_shortname2 < '9') {
1445 next_shortname2++;
1446 } else {
1447 next_shortname2='0';
1448 }
1449 }
1450 new_shortname = 1;
1451 }
1452
1453 for (cpu = 0; cpu <= max_cpu; cpu++) {
1454 if (cpu != this_cpu)
1455 printf(" ");
1456 else
1457 printf("*");
1458
1459 if (curr_thread[cpu]) {
1460 if (curr_thread[cpu]->pid)
1461 printf("%2s ", curr_thread[cpu]->shortname);
1462 else
1463 printf(". ");
1464 } else
1465 printf(" ");
1466 }
1467
1468 printf(" %12.6f secs ", (double)timestamp/1e9);
1469 if (new_shortname) {
1470 printf("%s => %s:%d\n",
1471 sched_in->shortname, sched_in->comm, sched_in->pid);
1472 } else {
1473 printf("\n");
1474 }
1475}
1476
1477static void
1478process_sched_switch_event(struct perf_tool *tool __used,
1479 struct event_format *event,
1480 struct perf_sample *sample,
1481 struct machine *machine,
1482 struct thread *thread)
1483{
1484 int this_cpu = sample->cpu;
1485 void *data = sample->raw_data;
1486 struct trace_switch_event switch_event;
1487
1488 FILL_COMMON_FIELDS(switch_event, event, data);
1489
1490 FILL_ARRAY(switch_event, prev_comm, event, data);
1491 FILL_FIELD(switch_event, prev_pid, event, data);
1492 FILL_FIELD(switch_event, prev_prio, event, data);
1493 FILL_FIELD(switch_event, prev_state, event, data);
1494 FILL_ARRAY(switch_event, next_comm, event, data);
1495 FILL_FIELD(switch_event, next_pid, event, data);
1496 FILL_FIELD(switch_event, next_prio, event, data);
1497
1498 if (curr_pid[this_cpu] != (u32)-1) {
1499 /*
1500 * Are we trying to switch away a PID that is
1501 * not current?
1502 */
1503 if (curr_pid[this_cpu] != switch_event.prev_pid)
1504 nr_context_switch_bugs++;
1505 }
1506 if (trace_handler->switch_event)
1507 trace_handler->switch_event(&switch_event, machine, event,
1508 this_cpu, sample->time, thread);
1509
1510 curr_pid[this_cpu] = switch_event.next_pid;
1511}
1512
1513static void
1514process_sched_runtime_event(struct perf_tool *tool __used,
1515 struct event_format *event,
1516 struct perf_sample *sample,
1517 struct machine *machine,
1518 struct thread *thread)
1519{
1520 void *data = sample->raw_data;
1521 struct trace_runtime_event runtime_event;
1522
1523 FILL_ARRAY(runtime_event, comm, event, data);
1524 FILL_FIELD(runtime_event, pid, event, data);
1525 FILL_FIELD(runtime_event, runtime, event, data);
1526 FILL_FIELD(runtime_event, vruntime, event, data);
1527
1528 if (trace_handler->runtime_event)
1529 trace_handler->runtime_event(&runtime_event, machine, event,
1530 sample->cpu, sample->time, thread);
1531}
1532
1533static void
1534process_sched_fork_event(struct perf_tool *tool __used,
1535 struct event_format *event,
1536 struct perf_sample *sample,
1537 struct machine *machine __used,
1538 struct thread *thread)
1539{
1540 void *data = sample->raw_data;
1541 struct trace_fork_event fork_event;
1542
1543 FILL_COMMON_FIELDS(fork_event, event, data);
1544
1545 FILL_ARRAY(fork_event, parent_comm, event, data);
1546 FILL_FIELD(fork_event, parent_pid, event, data);
1547 FILL_ARRAY(fork_event, child_comm, event, data);
1548 FILL_FIELD(fork_event, child_pid, event, data);
1549
1550 if (trace_handler->fork_event)
1551 trace_handler->fork_event(&fork_event, event,
1552 sample->cpu, sample->time, thread);
1553}
1554
1555static void
1556process_sched_exit_event(struct perf_tool *tool __used,
1557 struct event_format *event,
1558 struct perf_sample *sample __used,
1559 struct machine *machine __used,
1560 struct thread *thread __used)
1561{
1562 if (verbose)
1563 printf("sched_exit event %p\n", event);
1564}
1565
1566static void
1567process_sched_migrate_task_event(struct perf_tool *tool __used,
1568 struct event_format *event,
1569 struct perf_sample *sample,
1570 struct machine *machine,
1571 struct thread *thread)
1572{
1573 void *data = sample->raw_data;
1574 struct trace_migrate_task_event migrate_task_event;
1575
1576 FILL_COMMON_FIELDS(migrate_task_event, event, data);
1577
1578 FILL_ARRAY(migrate_task_event, comm, event, data);
1579 FILL_FIELD(migrate_task_event, pid, event, data);
1580 FILL_FIELD(migrate_task_event, prio, event, data);
1581 FILL_FIELD(migrate_task_event, cpu, event, data);
1582
1583 if (trace_handler->migrate_task_event)
1584 trace_handler->migrate_task_event(&migrate_task_event, machine,
1585 event, sample->cpu,
1586 sample->time, thread);
1587}
1588
1589typedef void (*tracepoint_handler)(struct perf_tool *tool, struct event_format *event,
1590 struct perf_sample *sample,
1591 struct machine *machine,
1592 struct thread *thread);
1593
1594static int perf_sched__process_tracepoint_sample(struct perf_tool *tool,
1595 union perf_event *event __used,
1596 struct perf_sample *sample,
1597 struct perf_evsel *evsel,
1598 struct machine *machine)
1599{
1600 struct thread *thread = machine__findnew_thread(machine, sample->pid);
1601
1602 if (thread == NULL) {
1603 pr_debug("problem processing %s event, skipping it.\n",
1604 evsel->name);
1605 return -1;
1606 }
1607
1608 evsel->hists.stats.total_period += sample->period;
1609 hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
1610
1611 if (evsel->handler.func != NULL) {
1612 tracepoint_handler f = evsel->handler.func;
1613
1614 if (evsel->handler.data == NULL)
1615 evsel->handler.data = trace_find_event(evsel->attr.config);
1616
1617 f(tool, evsel->handler.data, sample, machine, thread);
1618 }
1619
1620 return 0;
1621}
1622
1623static struct perf_tool perf_sched = {
1624 .sample = perf_sched__process_tracepoint_sample,
1625 .comm = perf_event__process_comm,
1626 .lost = perf_event__process_lost,
1627 .fork = perf_event__process_task,
1628 .ordered_samples = true,
1629};
1630
1631static void read_events(bool destroy, struct perf_session **psession)
1632{
1633 int err = -EINVAL;
1634 const struct perf_evsel_str_handler handlers[] = {
1635 { "sched:sched_switch", process_sched_switch_event, },
1636 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1637 { "sched:sched_wakeup", process_sched_wakeup_event, },
1638 { "sched:sched_wakeup_new", process_sched_wakeup_event, },
1639 { "sched:sched_process_fork", process_sched_fork_event, },
1640 { "sched:sched_process_exit", process_sched_exit_event, },
1641 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1642 };
1643 struct perf_session *session = perf_session__new(input_name, O_RDONLY,
1644 0, false, &perf_sched);
1645 if (session == NULL)
1646 die("No Memory");
1647
1648 err = perf_evlist__set_tracepoints_handlers_array(session->evlist, handlers);
1649 assert(err == 0);
1650
1651 if (perf_session__has_traces(session, "record -R")) {
1652 err = perf_session__process_events(session, &perf_sched);
1653 if (err)
1654 die("Failed to process events, error %d", err);
1655
1656 nr_events = session->hists.stats.nr_events[0];
1657 nr_lost_events = session->hists.stats.total_lost;
1658 nr_lost_chunks = session->hists.stats.nr_events[PERF_RECORD_LOST];
1659 }
1660
1661 if (destroy)
1662 perf_session__delete(session);
1663
1664 if (psession)
1665 *psession = session;
1666}
1667
1668static void print_bad_events(void)
1669{
1670 if (nr_unordered_timestamps && nr_timestamps) {
1671 printf(" INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1672 (double)nr_unordered_timestamps/(double)nr_timestamps*100.0,
1673 nr_unordered_timestamps, nr_timestamps);
1674 }
1675 if (nr_lost_events && nr_events) {
1676 printf(" INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1677 (double)nr_lost_events/(double)nr_events*100.0,
1678 nr_lost_events, nr_events, nr_lost_chunks);
1679 }
1680 if (nr_state_machine_bugs && nr_timestamps) {
1681 printf(" INFO: %.3f%% state machine bugs (%ld out of %ld)",
1682 (double)nr_state_machine_bugs/(double)nr_timestamps*100.0,
1683 nr_state_machine_bugs, nr_timestamps);
1684 if (nr_lost_events)
1685 printf(" (due to lost events?)");
1686 printf("\n");
1687 }
1688 if (nr_context_switch_bugs && nr_timestamps) {
1689 printf(" INFO: %.3f%% context switch bugs (%ld out of %ld)",
1690 (double)nr_context_switch_bugs/(double)nr_timestamps*100.0,
1691 nr_context_switch_bugs, nr_timestamps);
1692 if (nr_lost_events)
1693 printf(" (due to lost events?)");
1694 printf("\n");
1695 }
1696}
1697
1698static void __cmd_lat(void)
1699{
1700 struct rb_node *next;
1701 struct perf_session *session;
1702
1703 setup_pager();
1704 read_events(false, &session);
1705 sort_lat();
1706
1707 printf("\n ---------------------------------------------------------------------------------------------------------------\n");
1708 printf(" Task | Runtime ms | Switches | Average delay ms | Maximum delay ms | Maximum delay at |\n");
1709 printf(" ---------------------------------------------------------------------------------------------------------------\n");
1710
1711 next = rb_first(&sorted_atom_root);
1712
1713 while (next) {
1714 struct work_atoms *work_list;
1715
1716 work_list = rb_entry(next, struct work_atoms, node);
1717 output_lat_thread(work_list);
1718 next = rb_next(next);
1719 }
1720
1721 printf(" -----------------------------------------------------------------------------------------\n");
1722 printf(" TOTAL: |%11.3f ms |%9" PRIu64 " |\n",
1723 (double)all_runtime/1e6, all_count);
1724
1725 printf(" ---------------------------------------------------\n");
1726
1727 print_bad_events();
1728 printf("\n");
1729
1730 perf_session__delete(session);
1731}
1732
1733static struct trace_sched_handler map_ops = {
1734 .wakeup_event = NULL,
1735 .switch_event = map_switch_event,
1736 .runtime_event = NULL,
1737 .fork_event = NULL,
1738};
1739
1740static void __cmd_map(void)
1741{
1742 max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1743
1744 setup_pager();
1745 read_events(true, NULL);
1746 print_bad_events();
1747}
1748
1749static void __cmd_replay(void)
1750{
1751 unsigned long i;
1752
1753 calibrate_run_measurement_overhead();
1754 calibrate_sleep_measurement_overhead();
1755
1756 test_calibrations();
1757
1758 read_events(true, NULL);
1759
1760 printf("nr_run_events: %ld\n", nr_run_events);
1761 printf("nr_sleep_events: %ld\n", nr_sleep_events);
1762 printf("nr_wakeup_events: %ld\n", nr_wakeup_events);
1763
1764 if (targetless_wakeups)
1765 printf("target-less wakeups: %ld\n", targetless_wakeups);
1766 if (multitarget_wakeups)
1767 printf("multi-target wakeups: %ld\n", multitarget_wakeups);
1768 if (nr_run_events_optimized)
1769 printf("run atoms optimized: %ld\n",
1770 nr_run_events_optimized);
1771
1772 print_task_traces();
1773 add_cross_task_wakeups();
1774
1775 create_tasks();
1776 printf("------------------------------------------------------------\n");
1777 for (i = 0; i < replay_repeat; i++)
1778 run_one_test();
1779}
1780
1781
1782static const char * const sched_usage[] = {
1783 "perf sched [<options>] {record|latency|map|replay|script}",
1784 NULL
1785};
1786
1787static const struct option sched_options[] = {
1788 OPT_STRING('i', "input", &input_name, "file",
1789 "input file name"),
1790 OPT_INCR('v', "verbose", &verbose,
1791 "be more verbose (show symbol address, etc)"),
1792 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1793 "dump raw trace in ASCII"),
1794 OPT_END()
1795};
1796
1797static const char * const latency_usage[] = {
1798 "perf sched latency [<options>]",
1799 NULL
1800};
1801
1802static const struct option latency_options[] = {
1803 OPT_STRING('s', "sort", &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", &profile_cpu,
1808 "CPU to profile on"),
1809 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1810 "dump raw trace in ASCII"),
1811 OPT_END()
1812};
1813
1814static const char * const replay_usage[] = {
1815 "perf sched replay [<options>]",
1816 NULL
1817};
1818
1819static const struct option replay_options[] = {
1820 OPT_UINTEGER('r', "repeat", &replay_repeat,
1821 "repeat the workload replay N times (-1: infinite)"),
1822 OPT_INCR('v', "verbose", &verbose,
1823 "be more verbose (show symbol address, etc)"),
1824 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1825 "dump raw trace in ASCII"),
1826 OPT_END()
1827};
1828
1829static void setup_sorting(void)
1830{
1831 char *tmp, *tok, *str = strdup(sort_order);
1832
1833 for (tok = strtok_r(str, ", ", &tmp);
1834 tok; tok = strtok_r(NULL, ", ", &tmp)) {
1835 if (sort_dimension__add(tok, &sort_list) < 0) {
1836 error("Unknown --sort key: `%s'", tok);
1837 usage_with_options(latency_usage, latency_options);
1838 }
1839 }
1840
1841 free(str);
1842
1843 sort_dimension__add("pid", &cmp_pid);
1844}
1845
1846static const char *record_args[] = {
1847 "record",
1848 "-a",
1849 "-R",
1850 "-f",
1851 "-m", "1024",
1852 "-c", "1",
1853 "-e", "sched:sched_switch",
1854 "-e", "sched:sched_stat_wait",
1855 "-e", "sched:sched_stat_sleep",
1856 "-e", "sched:sched_stat_iowait",
1857 "-e", "sched:sched_stat_runtime",
1858 "-e", "sched:sched_process_exit",
1859 "-e", "sched:sched_process_fork",
1860 "-e", "sched:sched_wakeup",
1861 "-e", "sched:sched_migrate_task",
1862};
1863
1864static int __cmd_record(int argc, const char **argv)
1865{
1866 unsigned int rec_argc, i, j;
1867 const char **rec_argv;
1868
1869 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1870 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1871
1872 if (rec_argv == NULL)
1873 return -ENOMEM;
1874
1875 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1876 rec_argv[i] = strdup(record_args[i]);
1877
1878 for (j = 1; j < (unsigned int)argc; j++, i++)
1879 rec_argv[i] = argv[j];
1880
1881 BUG_ON(i != rec_argc);
1882
1883 return cmd_record(i, rec_argv, NULL);
1884}
1885
1886int cmd_sched(int argc, const char **argv, const char *prefix __used)
1887{
1888 argc = parse_options(argc, argv, sched_options, sched_usage,
1889 PARSE_OPT_STOP_AT_NON_OPTION);
1890 if (!argc)
1891 usage_with_options(sched_usage, sched_options);
1892
1893 /*
1894 * Aliased to 'perf script' for now:
1895 */
1896 if (!strcmp(argv[0], "script"))
1897 return cmd_script(argc, argv, prefix);
1898
1899 symbol__init();
1900 if (!strncmp(argv[0], "rec", 3)) {
1901 return __cmd_record(argc, argv);
1902 } else if (!strncmp(argv[0], "lat", 3)) {
1903 trace_handler = &lat_ops;
1904 if (argc > 1) {
1905 argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1906 if (argc)
1907 usage_with_options(latency_usage, latency_options);
1908 }
1909 setup_sorting();
1910 __cmd_lat();
1911 } else if (!strcmp(argv[0], "map")) {
1912 trace_handler = &map_ops;
1913 setup_sorting();
1914 __cmd_map();
1915 } else if (!strncmp(argv[0], "rep", 3)) {
1916 trace_handler = &replay_ops;
1917 if (argc) {
1918 argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1919 if (argc)
1920 usage_with_options(replay_usage, replay_options);
1921 }
1922 __cmd_replay();
1923 } else {
1924 usage_with_options(sched_usage, sched_options);
1925 }
1926
1927 return 0;
1928}