projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
perf/core: Remove unused perf_flags
[linux-2.6-block.git] / tools / perf / builtin-sched.c
CommitLineData
b2441318 1// SPDX-License-Identifier: GPL-2.0
0a02ad93 2#include "builtin.h"
b1ffe8f3 3#include "perf.h"
0a02ad93
IM		 4
5#include "util/util.h"
ee29be62 6#include "util/evlist.h"
0a02ad93 7#include "util/cache.h"
e3f42609 8#include "util/evsel.h"
0a02ad93
IM		 9#include "util/symbol.h"
10#include "util/thread.h"
11#include "util/header.h"
94c744b6 12#include "util/session.h"
45694aa7 13#include "util/tool.h"
57480d2c 14#include "util/cloexec.h"
a151a37a 15#include "util/thread_map.h"
8cd91195 16#include "util/color.h"
49394a2a 17#include "util/stat.h"
6c973c90 18#include "util/callchain.h"
853b7407 19#include "util/time-utils.h"
0a02ad93 20
4b6ab94e 21#include <subcmd/parse-options.h>
b1ffe8f3 22#include "util/trace-event.h"
0a02ad93 23
0a02ad93
IM		 24#include "util/debug.h"
25
877a7a11 26#include <linux/kernel.h>
49394a2a 27#include <linux/log2.h>
b1ffe8f3 28#include <sys/prctl.h>
7b78f136 29#include <sys/resource.h>
fd20e811 30#include <inttypes.h>
0a02ad93 31
a43783ae 32#include <errno.h>
b1ffe8f3
IM		 33#include <semaphore.h>
34#include <pthread.h>
35#include <math.h>
cb06ac25 36#include <api/fs/fs.h>
4fc76e49 37#include <linux/time64.h>
419ab0d6 38
3d689ed6
ACM		 39#include "sane_ctype.h"
40
b1ffe8f3
IM		 41#define PR_SET_NAME 15 /* Set process name */
42#define MAX_CPUS 4096
b1ffe8f3
IM		 43#define COMM_LEN 20
44#define SYM_LEN 129
a35e27d0 45#define MAX_PID 1024000
ec156764 46
39aeb52f 47struct sched_atom;
ec156764 48
b1ffe8f3
IM		 49struct task_desc {
50 unsigned long nr;
51 unsigned long pid;
52 char comm[COMM_LEN];
ec156764 53
b1ffe8f3
IM		 54 unsigned long nr_events;
55 unsigned long curr_event;
39aeb52f 56 struct sched_atom **atoms;
b1ffe8f3
IM		 57
58 pthread_t thread;
59 sem_t sleep_sem;
ec156764 60
b1ffe8f3
IM		 61 sem_t ready_for_work;
62 sem_t work_done_sem;
63
64 u64 cpu_usage;
65};
66
67enum sched_event_type {
68 SCHED_EVENT_RUN,
69 SCHED_EVENT_SLEEP,
70 SCHED_EVENT_WAKEUP,
55ffb7a6 71 SCHED_EVENT_MIGRATION,
b1ffe8f3
IM		 72};
73
39aeb52f 74struct sched_atom {
b1ffe8f3 75 enum sched_event_type type;
eed05fe7 76 int specific_wait;
b1ffe8f3
IM		 77 u64 timestamp;
78 u64 duration;
79 unsigned long nr;
b1ffe8f3
IM		 80 sem_t *wait_sem;
81 struct task_desc *wakee;
82};
83
e936e8e4 84#define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
b1ffe8f3 85
941bdea7
NK		 86/* task state bitmask, copied from include/linux/sched.h */
87#define TASK_RUNNING 0
88#define TASK_INTERRUPTIBLE 1
89#define TASK_UNINTERRUPTIBLE 2
90#define __TASK_STOPPED 4
91#define __TASK_TRACED 8
92/* in tsk->exit_state */
93#define EXIT_DEAD 16
94#define EXIT_ZOMBIE 32
95#define EXIT_TRACE (EXIT_ZOMBIE | EXIT_DEAD)
96/* in tsk->state again */
97#define TASK_DEAD 64
98#define TASK_WAKEKILL 128
99#define TASK_WAKING 256
100#define TASK_PARKED 512
101
b1ffe8f3
IM		 102enum thread_state {
103 THREAD_SLEEPING = 0,
104 THREAD_WAIT_CPU,
105 THREAD_SCHED_IN,
106 THREAD_IGNORE
107};
108
109struct work_atom {
110 struct list_head list;
111 enum thread_state state;
aa1ab9d2 112 u64 sched_out_time;
b1ffe8f3
IM		 113 u64 wake_up_time;
114 u64 sched_in_time;
115 u64 runtime;
116};
117
39aeb52f 118struct work_atoms {
119 struct list_head work_list;
b1ffe8f3
IM		 120 struct thread *thread;
121 struct rb_node node;
122 u64 max_lat;
3786310a 123 u64 max_lat_at;
b1ffe8f3
IM		 124 u64 total_lat;
125 u64 nb_atoms;
126 u64 total_runtime;
2f80dd44 127 int num_merged;
b1ffe8f3
IM		 128};
129
39aeb52f 130typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
b1ffe8f3 131
9ec3f4e4 132struct perf_sched;
0e9b07e5 133
9ec3f4e4
ACM		 134struct trace_sched_handler {
135 int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
136 struct perf_sample *sample, struct machine *machine);
0e9b07e5 137
9ec3f4e4
ACM		 138 int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
139 struct perf_sample *sample, struct machine *machine);
0e9b07e5 140
9ec3f4e4
ACM		 141 int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
142 struct perf_sample *sample, struct machine *machine);
0e9b07e5 143
cb627505
DA		 144 /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
145 int (*fork_event)(struct perf_sched *sched, union perf_event *event,
146 struct machine *machine);
0e9b07e5
ACM		 147
148 int (*migrate_task_event)(struct perf_sched *sched,
9ec3f4e4
ACM		 149 struct perf_evsel *evsel,
150 struct perf_sample *sample,
151 struct machine *machine);
0e9b07e5
ACM		 152};
153
a151a37a 154#define COLOR_PIDS PERF_COLOR_BLUE
cf294f24 155#define COLOR_CPUS PERF_COLOR_BG_RED
a151a37a 156
99623c62
JO		 157struct perf_sched_map {
158 DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
159 int *comp_cpus;
160 bool comp;
a151a37a
JO		 161 struct thread_map *color_pids;
162 const char *color_pids_str;
cf294f24
JO		 163 struct cpu_map *color_cpus;
164 const char *color_cpus_str;
73643bb6
JO		 165 struct cpu_map *cpus;
166 const char *cpus_str;
99623c62
JO		 167};
168
0e9b07e5
ACM		 169struct perf_sched {
170 struct perf_tool tool;
0e9b07e5
ACM		 171 const char *sort_order;
172 unsigned long nr_tasks;
cb06ac25 173 struct task_desc **pid_to_task;
0e9b07e5
ACM		 174 struct task_desc **tasks;
175 const struct trace_sched_handler *tp_handler;
176 pthread_mutex_t start_work_mutex;
177 pthread_mutex_t work_done_wait_mutex;
178 int profile_cpu;
179/*
180 * Track the current task - that way we can know whether there's any
181 * weird events, such as a task being switched away that is not current.
182 */
183 int max_cpu;
184 u32 curr_pid[MAX_CPUS];
185 struct thread *curr_thread[MAX_CPUS];
186 char next_shortname1;
187 char next_shortname2;
188 unsigned int replay_repeat;
189 unsigned long nr_run_events;
190 unsigned long nr_sleep_events;
191 unsigned long nr_wakeup_events;
192 unsigned long nr_sleep_corrections;
193 unsigned long nr_run_events_optimized;
194 unsigned long targetless_wakeups;
195 unsigned long multitarget_wakeups;
196 unsigned long nr_runs;
197 unsigned long nr_timestamps;
198 unsigned long nr_unordered_timestamps;
0e9b07e5
ACM		 199 unsigned long nr_context_switch_bugs;
200 unsigned long nr_events;
201 unsigned long nr_lost_chunks;
202 unsigned long nr_lost_events;
203 u64 run_measurement_overhead;
204 u64 sleep_measurement_overhead;
205 u64 start_time;
206 u64 cpu_usage;
207 u64 runavg_cpu_usage;
208 u64 parent_cpu_usage;
209 u64 runavg_parent_cpu_usage;
210 u64 sum_runtime;
211 u64 sum_fluct;
212 u64 run_avg;
213 u64 all_runtime;
214 u64 all_count;
215 u64 cpu_last_switched[MAX_CPUS];
2f80dd44 216 struct rb_root atom_root, sorted_atom_root, merged_atom_root;
0e9b07e5 217 struct list_head sort_list, cmp_pid;
939cda52 218 bool force;
2f80dd44 219 bool skip_merge;
99623c62 220 struct perf_sched_map map;
52df138c
DA		 221
222 /* options for timehist command */
223 bool summary;
224 bool summary_only;
699b5b92 225 bool idle_hist;
6c973c90
DA		 226 bool show_callchain;
227 unsigned int max_stack;
a407b067 228 bool show_cpu_visual;
fc1469f1 229 bool show_wakeups;
292c4a8f 230 bool show_next;
350f54fa 231 bool show_migrations;
414e050c 232 bool show_state;
52df138c 233 u64 skipped_samples;
853b7407
DA		 234 const char *time_str;
235 struct perf_time_interval ptime;
9396c9cb 236 struct perf_time_interval hist_time;
0e9b07e5 237};
b1ffe8f3 238
49394a2a
DA		 239/* per thread run time data */
240struct thread_runtime {
241 u64 last_time; /* time of previous sched in/out event */
242 u64 dt_run; /* run time */
941bdea7
NK		 243 u64 dt_sleep; /* time between CPU access by sleep (off cpu) */
244 u64 dt_iowait; /* time between CPU access by iowait (off cpu) */
245 u64 dt_preempt; /* time between CPU access by preempt (off cpu) */
49394a2a
DA		 246 u64 dt_delay; /* time between wakeup and sched-in */
247 u64 ready_to_run; /* time of wakeup */
248
249 struct stats run_stats;
250 u64 total_run_time;
587782c5
NK		 251 u64 total_sleep_time;
252 u64 total_iowait_time;
253 u64 total_preempt_time;
254 u64 total_delay_time;
350f54fa 255
941bdea7 256 int last_state;
8640da9f
CD		 257
258 char shortname[3];
99a3c3a9
CD		 259 bool comm_changed;
260
350f54fa 261 u64 migrations;
49394a2a
DA		 262};
263
264/* per event run time data */
265struct evsel_runtime {
266 u64 *last_time; /* time this event was last seen per cpu */
267 u32 ncpu; /* highest cpu slot allocated */
268};
269
3bc2fa9c
NK		 270/* per cpu idle time data */
271struct idle_thread_runtime {
272 struct thread_runtime tr;
273 struct thread *last_thread;
274 struct rb_root sorted_root;
275 struct callchain_root callchain;
276 struct callchain_cursor cursor;
277};
278
49394a2a
DA		 279/* track idle times per cpu */
280static struct thread **idle_threads;
281static int idle_max_cpu;
282static char idle_comm[] = "<idle>";
283
b1ffe8f3 284static u64 get_nsecs(void)
ec156764
IM		 285{
286 struct timespec ts;
287
288 clock_gettime(CLOCK_MONOTONIC, &ts);
289
4fc76e49 290 return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
ec156764
IM		 291}
292
0e9b07e5 293static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
ec156764 294{
b1ffe8f3 295 u64 T0 = get_nsecs(), T1;
ec156764
IM		 296
297 do {
298 T1 = get_nsecs();
0e9b07e5 299 } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
ec156764
IM		 300}
301
b1ffe8f3 302static void sleep_nsecs(u64 nsecs)
ec156764
IM		 303{
304 struct timespec ts;
305
306 ts.tv_nsec = nsecs % 999999999;
307 ts.tv_sec = nsecs / 999999999;
308
309 nanosleep(&ts, NULL);
310}
311
0e9b07e5 312static void calibrate_run_measurement_overhead(struct perf_sched *sched)
ec156764 313{
4fc76e49 314 u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
ec156764
IM		 315 int i;
316
317 for (i = 0; i < 10; i++) {
318 T0 = get_nsecs();
0e9b07e5 319 burn_nsecs(sched, 0);
ec156764
IM		 320 T1 = get_nsecs();
321 delta = T1-T0;
322 min_delta = min(min_delta, delta);
323 }
0e9b07e5 324 sched->run_measurement_overhead = min_delta;
ec156764 325
9486aa38 326 printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
ec156764
IM		 327}
328
0e9b07e5 329static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
ec156764 330{
4fc76e49 331 u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
ec156764
IM		 332 int i;
333
334 for (i = 0; i < 10; i++) {
335 T0 = get_nsecs();
336 sleep_nsecs(10000);
337 T1 = get_nsecs();
338 delta = T1-T0;
339 min_delta = min(min_delta, delta);
340 }
341 min_delta -= 10000;
0e9b07e5 342 sched->sleep_measurement_overhead = min_delta;
ec156764 343
9486aa38 344 printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
ec156764
IM		 345}
346
39aeb52f 347static struct sched_atom *
b1ffe8f3 348get_new_event(struct task_desc *task, u64 timestamp)
ec156764 349{
36479484 350 struct sched_atom *event = zalloc(sizeof(*event));
ec156764
IM		 351 unsigned long idx = task->nr_events;
352 size_t size;
353
354 event->timestamp = timestamp;
355 event->nr = idx;
356
357 task->nr_events++;
39aeb52f 358 size = sizeof(struct sched_atom *) * task->nr_events;
359 task->atoms = realloc(task->atoms, size);
360 BUG_ON(!task->atoms);
ec156764 361
39aeb52f 362 task->atoms[idx] = event;
ec156764
IM		 363
364 return event;
365}
366
39aeb52f 367static struct sched_atom *last_event(struct task_desc *task)
ec156764
IM		 368{
369 if (!task->nr_events)
370 return NULL;
371
39aeb52f 372 return task->atoms[task->nr_events - 1];
ec156764
IM		 373}
374
0e9b07e5
ACM		 375static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
376 u64 timestamp, u64 duration)
ec156764 377{
39aeb52f 378 struct sched_atom *event, *curr_event = last_event(task);
ec156764
IM		 379
380 /*
fbf94829
IM		 381 * optimize an existing RUN event by merging this one
382 * to it:
383 */
ec156764 384 if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
0e9b07e5 385 sched->nr_run_events_optimized++;
ec156764
IM		 386 curr_event->duration += duration;
387 return;
388 }
389
390 event = get_new_event(task, timestamp);
391
392 event->type = SCHED_EVENT_RUN;
393 event->duration = duration;
394
0e9b07e5 395 sched->nr_run_events++;
ec156764
IM		 396}
397
0e9b07e5
ACM		 398static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
399 u64 timestamp, struct task_desc *wakee)
ec156764 400{
39aeb52f 401 struct sched_atom *event, *wakee_event;
ec156764
IM		 402
403 event = get_new_event(task, timestamp);
404 event->type = SCHED_EVENT_WAKEUP;
405 event->wakee = wakee;
406
407 wakee_event = last_event(wakee);
408 if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
0e9b07e5 409 sched->targetless_wakeups++;
ec156764
IM		 410 return;
411 }
412 if (wakee_event->wait_sem) {
0e9b07e5 413 sched->multitarget_wakeups++;
ec156764
IM		 414 return;
415 }
416
36479484 417 wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
ec156764
IM		 418 sem_init(wakee_event->wait_sem, 0, 0);
419 wakee_event->specific_wait = 1;
420 event->wait_sem = wakee_event->wait_sem;
421
0e9b07e5 422 sched->nr_wakeup_events++;
ec156764
IM		 423}
424
0e9b07e5
ACM		 425static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
426 u64 timestamp, u64 task_state __maybe_unused)
ec156764 427{
39aeb52f 428 struct sched_atom *event = get_new_event(task, timestamp);
ec156764
IM		 429
430 event->type = SCHED_EVENT_SLEEP;
431
0e9b07e5 432 sched->nr_sleep_events++;
ec156764
IM		 433}
434
0e9b07e5
ACM		 435static struct task_desc *register_pid(struct perf_sched *sched,
436 unsigned long pid, const char *comm)
ec156764
IM		 437{
438 struct task_desc *task;
cb06ac25 439 static int pid_max;
ec156764 440
cb06ac25
YS		 441 if (sched->pid_to_task == NULL) {
442 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
443 pid_max = MAX_PID;
444 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
445 }
3a423a5c
YS		 446 if (pid >= (unsigned long)pid_max) {
447 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
448 sizeof(struct task_desc *))) == NULL);
449 while (pid >= (unsigned long)pid_max)
450 sched->pid_to_task[pid_max++] = NULL;
451 }
ec156764 452
0e9b07e5 453 task = sched->pid_to_task[pid];
ec156764
IM		 454
455 if (task)
456 return task;
457
36479484 458 task = zalloc(sizeof(*task));
ec156764 459 task->pid = pid;
0e9b07e5 460 task->nr = sched->nr_tasks;
ec156764
IM		 461 strcpy(task->comm, comm);
462 /*
463 * every task starts in sleeping state - this gets ignored
464 * if there's no wakeup pointing to this sleep state:
465 */
0e9b07e5 466 add_sched_event_sleep(sched, task, 0, 0);
ec156764 467
0e9b07e5
ACM		 468 sched->pid_to_task[pid] = task;
469 sched->nr_tasks++;
0755bc4d 470 sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
0e9b07e5
ACM		 471 BUG_ON(!sched->tasks);
472 sched->tasks[task->nr] = task;
ec156764 473
bb963e16 474 if (verbose > 0)
0e9b07e5 475 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
ec156764
IM		 476
477 return task;
478}
479
480
0e9b07e5 481static void print_task_traces(struct perf_sched *sched)
ec156764
IM		 482{
483 struct task_desc *task;
484 unsigned long i;
485
0e9b07e5
ACM		 486 for (i = 0; i < sched->nr_tasks; i++) {
487 task = sched->tasks[i];
ad236fd2 488 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
ec156764
IM		 489 task->nr, task->comm, task->pid, task->nr_events);
490 }
491}
492
0e9b07e5 493static void add_cross_task_wakeups(struct perf_sched *sched)
ec156764
IM		 494{
495 struct task_desc *task1, *task2;
496 unsigned long i, j;
497
0e9b07e5
ACM		 498 for (i = 0; i < sched->nr_tasks; i++) {
499 task1 = sched->tasks[i];
ec156764 500 j = i + 1;
0e9b07e5 501 if (j == sched->nr_tasks)
ec156764 502 j = 0;
0e9b07e5
ACM		 503 task2 = sched->tasks[j];
504 add_sched_event_wakeup(sched, task1, 0, task2);
ec156764
IM		 505 }
506}
507
0e9b07e5
ACM		 508static void perf_sched__process_event(struct perf_sched *sched,
509 struct sched_atom *atom)
ec156764
IM		 510{
511 int ret = 0;
ec156764 512
39aeb52f 513 switch (atom->type) {
ec156764 514 case SCHED_EVENT_RUN:
0e9b07e5 515 burn_nsecs(sched, atom->duration);
ec156764
IM		 516 break;
517 case SCHED_EVENT_SLEEP:
39aeb52f 518 if (atom->wait_sem)
519 ret = sem_wait(atom->wait_sem);
ec156764
IM		 520 BUG_ON(ret);
521 break;
522 case SCHED_EVENT_WAKEUP:
39aeb52f 523 if (atom->wait_sem)
524 ret = sem_post(atom->wait_sem);
ec156764
IM		 525 BUG_ON(ret);
526 break;
55ffb7a6
MG		 527 case SCHED_EVENT_MIGRATION:
528 break;
ec156764
IM		 529 default:
530 BUG_ON(1);
531 }
532}
533
b1ffe8f3 534static u64 get_cpu_usage_nsec_parent(void)
ec156764
IM		 535{
536 struct rusage ru;
b1ffe8f3 537 u64 sum;
ec156764
IM		 538 int err;
539
540 err = getrusage(RUSAGE_SELF, &ru);
541 BUG_ON(err);
542
4fc76e49
ACM		 543 sum = ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
544 sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
ec156764
IM		 545
546 return sum;
547}
548
939cda52 549static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
ec156764 550{
c0c9e721 551 struct perf_event_attr attr;
939cda52 552 char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
c0c9e721 553 int fd;
939cda52
YS		 554 struct rlimit limit;
555 bool need_privilege = false;
ec156764 556
c0c9e721 557 memset(&attr, 0, sizeof(attr));
ec156764 558
c0c9e721
XG		 559 attr.type = PERF_TYPE_SOFTWARE;
560 attr.config = PERF_COUNT_SW_TASK_CLOCK;
ec156764 561
939cda52 562force_again:
57480d2c
YD		 563 fd = sys_perf_event_open(&attr, 0, -1, -1,
564 perf_event_open_cloexec_flag());
c0c9e721 565
1aff59be 566 if (fd < 0) {
939cda52
YS		 567 if (errno == EMFILE) {
568 if (sched->force) {
569 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
570 limit.rlim_cur += sched->nr_tasks - cur_task;
571 if (limit.rlim_cur > limit.rlim_max) {
572 limit.rlim_max = limit.rlim_cur;
573 need_privilege = true;
574 }
575 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
576 if (need_privilege && errno == EPERM)
577 strcpy(info, "Need privilege\n");
578 } else
579 goto force_again;
580 } else
581 strcpy(info, "Have a try with -f option\n");
582 }
60b7d14a 583 pr_err("Error: sys_perf_event_open() syscall returned "
939cda52 584 "with %d (%s)\n%s", fd,
c8b5f2c9 585 str_error_r(errno, sbuf, sizeof(sbuf)), info);
1aff59be
YS		 586 exit(EXIT_FAILURE);
587 }
c0c9e721
XG		 588 return fd;
589}
590
591static u64 get_cpu_usage_nsec_self(int fd)
592{
593 u64 runtime;
594 int ret;
595
596 ret = read(fd, &runtime, sizeof(runtime));
597 BUG_ON(ret != sizeof(runtime));
598
599 return runtime;
ec156764
IM		 600}
601
0e9b07e5
ACM		 602struct sched_thread_parms {
603 struct task_desc *task;
604 struct perf_sched *sched;
08097abc 605 int fd;
0e9b07e5
ACM		 606};
607
ec156764
IM		 608static void *thread_func(void *ctx)
609{
0e9b07e5
ACM		 610 struct sched_thread_parms *parms = ctx;
611 struct task_desc *this_task = parms->task;
612 struct perf_sched *sched = parms->sched;
b1ffe8f3 613 u64 cpu_usage_0, cpu_usage_1;
ec156764
IM		 614 unsigned long i, ret;
615 char comm2[22];
08097abc 616 int fd = parms->fd;
ec156764 617
74cf249d 618 zfree(&parms);
0e9b07e5 619
ec156764
IM		 620 sprintf(comm2, ":%s", this_task->comm);
621 prctl(PR_SET_NAME, comm2);
a116e05d
ACM		 622 if (fd < 0)
623 return NULL;
ec156764
IM		 624again:
625 ret = sem_post(&this_task->ready_for_work);
626 BUG_ON(ret);
0e9b07e5 627 ret = pthread_mutex_lock(&sched->start_work_mutex);
ec156764 628 BUG_ON(ret);
0e9b07e5 629 ret = pthread_mutex_unlock(&sched->start_work_mutex);
ec156764 630 BUG_ON(ret);
ec156764 631
c0c9e721 632 cpu_usage_0 = get_cpu_usage_nsec_self(fd);
ec156764
IM		 633
634 for (i = 0; i < this_task->nr_events; i++) {
635 this_task->curr_event = i;
0e9b07e5 636 perf_sched__process_event(sched, this_task->atoms[i]);
ec156764
IM		 637 }
638
c0c9e721 639 cpu_usage_1 = get_cpu_usage_nsec_self(fd);
ec156764 640 this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
ec156764
IM		 641 ret = sem_post(&this_task->work_done_sem);
642 BUG_ON(ret);
ec156764 643
0e9b07e5 644 ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
ec156764 645 BUG_ON(ret);
0e9b07e5 646 ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
ec156764 647 BUG_ON(ret);
ec156764
IM		 648
649 goto again;
650}
651
0e9b07e5 652static void create_tasks(struct perf_sched *sched)
ec156764
IM		 653{
654 struct task_desc *task;
655 pthread_attr_t attr;
656 unsigned long i;
657 int err;
658
659 err = pthread_attr_init(&attr);
660 BUG_ON(err);
12f7e036
JP		 661 err = pthread_attr_setstacksize(&attr,
662 (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
ec156764 663 BUG_ON(err);
0e9b07e5 664 err = pthread_mutex_lock(&sched->start_work_mutex);
ec156764 665 BUG_ON(err);
0e9b07e5 666 err = pthread_mutex_lock(&sched->work_done_wait_mutex);
ec156764 667 BUG_ON(err);
0e9b07e5
ACM		 668 for (i = 0; i < sched->nr_tasks; i++) {
669 struct sched_thread_parms *parms = malloc(sizeof(*parms));
670 BUG_ON(parms == NULL);
671 parms->task = task = sched->tasks[i];
672 parms->sched = sched;
939cda52 673 parms->fd = self_open_counters(sched, i);
ec156764
IM		 674 sem_init(&task->sleep_sem, 0, 0);
675 sem_init(&task->ready_for_work, 0, 0);
676 sem_init(&task->work_done_sem, 0, 0);
677 task->curr_event = 0;
0e9b07e5 678 err = pthread_create(&task->thread, &attr, thread_func, parms);
ec156764
IM		 679 BUG_ON(err);
680 }
681}
682
0e9b07e5 683static void wait_for_tasks(struct perf_sched *sched)
ec156764 684{
b1ffe8f3 685 u64 cpu_usage_0, cpu_usage_1;
ec156764
IM		 686 struct task_desc *task;
687 unsigned long i, ret;
688
0e9b07e5
ACM		 689 sched->start_time = get_nsecs();
690 sched->cpu_usage = 0;
691 pthread_mutex_unlock(&sched->work_done_wait_mutex);
ec156764 692
0e9b07e5
ACM		 693 for (i = 0; i < sched->nr_tasks; i++) {
694 task = sched->tasks[i];
ec156764
IM		 695 ret = sem_wait(&task->ready_for_work);
696 BUG_ON(ret);
697 sem_init(&task->ready_for_work, 0, 0);
698 }
0e9b07e5 699 ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
ec156764
IM		 700 BUG_ON(ret);
701
702 cpu_usage_0 = get_cpu_usage_nsec_parent();
703
0e9b07e5 704 pthread_mutex_unlock(&sched->start_work_mutex);
ec156764 705
0e9b07e5
ACM		 706 for (i = 0; i < sched->nr_tasks; i++) {
707 task = sched->tasks[i];
ec156764
IM		 708 ret = sem_wait(&task->work_done_sem);
709 BUG_ON(ret);
710 sem_init(&task->work_done_sem, 0, 0);
0e9b07e5 711 sched->cpu_usage += task->cpu_usage;
ec156764
IM		 712 task->cpu_usage = 0;
713 }
714
715 cpu_usage_1 = get_cpu_usage_nsec_parent();
0e9b07e5
ACM		 716 if (!sched->runavg_cpu_usage)
717 sched->runavg_cpu_usage = sched->cpu_usage;
ff5f3bbd 718 sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
ec156764 719
0e9b07e5
ACM		 720 sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
721 if (!sched->runavg_parent_cpu_usage)
722 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
ff5f3bbd
YS		 723 sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
724 sched->parent_cpu_usage)/sched->replay_repeat;
ec156764 725
0e9b07e5 726 ret = pthread_mutex_lock(&sched->start_work_mutex);
ec156764
IM		 727 BUG_ON(ret);
728
0e9b07e5
ACM		 729 for (i = 0; i < sched->nr_tasks; i++) {
730 task = sched->tasks[i];
ec156764
IM		 731 sem_init(&task->sleep_sem, 0, 0);
732 task->curr_event = 0;
733 }
734}
735
0e9b07e5 736static void run_one_test(struct perf_sched *sched)
ec156764 737{
fb7d0b3c 738 u64 T0, T1, delta, avg_delta, fluct;
ec156764
IM		 739
740 T0 = get_nsecs();
0e9b07e5 741 wait_for_tasks(sched);
ec156764
IM		 742 T1 = get_nsecs();
743
744 delta = T1 - T0;
0e9b07e5
ACM		 745 sched->sum_runtime += delta;
746 sched->nr_runs++;
ec156764 747
0e9b07e5 748 avg_delta = sched->sum_runtime / sched->nr_runs;
ec156764
IM		 749 if (delta < avg_delta)
750 fluct = avg_delta - delta;
751 else
752 fluct = delta - avg_delta;
0e9b07e5
ACM		 753 sched->sum_fluct += fluct;
754 if (!sched->run_avg)
755 sched->run_avg = delta;
ff5f3bbd 756 sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
ec156764 757
4fc76e49 758 printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
ec156764 759
4fc76e49 760 printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
ec156764 761
ad236fd2 762 printf("cpu: %0.2f / %0.2f",
4fc76e49 763 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
ec156764
IM		 764
765#if 0
766 /*
fbf94829 767 * rusage statistics done by the parent, these are less
0e9b07e5 768 * accurate than the sched->sum_exec_runtime based statistics:
fbf94829 769 */
ad236fd2 770 printf(" [%0.2f / %0.2f]",
4fc76e49
ACM		 771 (double)sched->parent_cpu_usage / NSEC_PER_MSEC,
772 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
ec156764
IM		 773#endif
774
ad236fd2 775 printf("\n");
ec156764 776
0e9b07e5
ACM		 777 if (sched->nr_sleep_corrections)
778 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
779 sched->nr_sleep_corrections = 0;
ec156764
IM		 780}
781
0e9b07e5 782static void test_calibrations(struct perf_sched *sched)
ec156764 783{
b1ffe8f3 784 u64 T0, T1;
ec156764
IM		 785
786 T0 = get_nsecs();
4fc76e49 787 burn_nsecs(sched, NSEC_PER_MSEC);
ec156764
IM		 788 T1 = get_nsecs();
789
9486aa38 790 printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
ec156764
IM		 791
792 T0 = get_nsecs();
4fc76e49 793 sleep_nsecs(NSEC_PER_MSEC);
ec156764
IM		 794 T1 = get_nsecs();
795
9486aa38 796 printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
ec156764
IM		 797}
798
a116e05d 799static int
0e9b07e5 800replay_wakeup_event(struct perf_sched *sched,
9ec3f4e4
ACM		 801 struct perf_evsel *evsel, struct perf_sample *sample,
802 struct machine *machine __maybe_unused)
419ab0d6 803{
9ec3f4e4
ACM		 804 const char *comm = perf_evsel__strval(evsel, sample, "comm");
805 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
419ab0d6 806 struct task_desc *waker, *wakee;
fbf94829 807
bb963e16 808 if (verbose > 0) {
2b7fcbc5 809 printf("sched_wakeup event %p\n", evsel);
fbf94829 810
9ec3f4e4 811 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
ad236fd2 812 }
fbf94829 813
2b7fcbc5 814 waker = register_pid(sched, sample->tid, "<unknown>");
9ec3f4e4 815 wakee = register_pid(sched, pid, comm);
fbf94829 816
0e9b07e5 817 add_sched_event_wakeup(sched, waker, sample->time, wakee);
a116e05d 818 return 0;
ec156764
IM		 819}
820
9ec3f4e4
ACM		 821static int replay_switch_event(struct perf_sched *sched,
822 struct perf_evsel *evsel,
823 struct perf_sample *sample,
824 struct machine *machine __maybe_unused)
ec156764 825{
9ec3f4e4
ACM		 826 const char *prev_comm = perf_evsel__strval(evsel, sample, "prev_comm"),
827 *next_comm = perf_evsel__strval(evsel, sample, "next_comm");
828 const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
829 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
830 const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
1d037ca1 831 struct task_desc *prev, __maybe_unused *next;
7f7f8d0b
ACM		 832 u64 timestamp0, timestamp = sample->time;
833 int cpu = sample->cpu;
fbf94829
IM		 834 s64 delta;
835
bb963e16 836 if (verbose > 0)
2b7fcbc5 837 printf("sched_switch event %p\n", evsel);
ad236fd2 838
fbf94829 839 if (cpu >= MAX_CPUS || cpu < 0)
a116e05d 840 return 0;
fbf94829 841
0e9b07e5 842 timestamp0 = sched->cpu_last_switched[cpu];
fbf94829
IM		 843 if (timestamp0)
844 delta = timestamp - timestamp0;
845 else
846 delta = 0;
847
a116e05d 848 if (delta < 0) {
60b7d14a 849 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
a116e05d
ACM		 850 return -1;
851 }
fbf94829 852
9ec3f4e4
ACM		 853 pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
854 prev_comm, prev_pid, next_comm, next_pid, delta);
fbf94829 855
9ec3f4e4
ACM		 856 prev = register_pid(sched, prev_pid, prev_comm);
857 next = register_pid(sched, next_pid, next_comm);
fbf94829 858
0e9b07e5 859 sched->cpu_last_switched[cpu] = timestamp;
fbf94829 860
0e9b07e5 861 add_sched_event_run(sched, prev, timestamp, delta);
9ec3f4e4 862 add_sched_event_sleep(sched, prev, timestamp, prev_state);
a116e05d
ACM		 863
864 return 0;
fbf94829
IM		 865}
866
cb627505
DA		 867static int replay_fork_event(struct perf_sched *sched,
868 union perf_event *event,
869 struct machine *machine)
419ab0d6 870{
cb627505
DA		 871 struct thread *child, *parent;
872
314add6b
AH		 873 child = machine__findnew_thread(machine, event->fork.pid,
874 event->fork.tid);
875 parent = machine__findnew_thread(machine, event->fork.ppid,
876 event->fork.ptid);
cb627505
DA		 877
878 if (child == NULL || parent == NULL) {
879 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
880 child, parent);
b91fc39f 881 goto out_put;
cb627505 882 }
9ec3f4e4 883
bb963e16 884 if (verbose > 0) {
cb627505 885 printf("fork event\n");
b9c5143a
FW		 886 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
887 printf("... child: %s/%d\n", thread__comm_str(child), child->tid);
419ab0d6 888 }
9ec3f4e4 889
b9c5143a
FW		 890 register_pid(sched, parent->tid, thread__comm_str(parent));
891 register_pid(sched, child->tid, thread__comm_str(child));
b91fc39f
ACM		 892out_put:
893 thread__put(child);
894 thread__put(parent);
a116e05d 895 return 0;
419ab0d6 896}
fbf94829 897
b1ffe8f3
IM		 898struct sort_dimension {
899 const char *name;
b5fae128 900 sort_fn_t cmp;
b1ffe8f3
IM		 901 struct list_head list;
902};
903
8640da9f
CD		 904/*
905 * handle runtime stats saved per thread
906 */
907static struct thread_runtime *thread__init_runtime(struct thread *thread)
908{
909 struct thread_runtime *r;
910
911 r = zalloc(sizeof(struct thread_runtime));
912 if (!r)
913 return NULL;
914
915 init_stats(&r->run_stats);
916 thread__set_priv(thread, r);
917
918 return r;
919}
920
921static struct thread_runtime *thread__get_runtime(struct thread *thread)
922{
923 struct thread_runtime *tr;
924
925 tr = thread__priv(thread);
926 if (tr == NULL) {
927 tr = thread__init_runtime(thread);
928 if (tr == NULL)
929 pr_debug("Failed to malloc memory for runtime data.\n");
930 }
931
932 return tr;
933}
934
daa1d7a5 935static int
39aeb52f 936thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
daa1d7a5
FW		 937{
938 struct sort_dimension *sort;
939 int ret = 0;
940
b5fae128
IM		 941 BUG_ON(list_empty(list));
942
daa1d7a5
FW		 943 list_for_each_entry(sort, list, list) {
944 ret = sort->cmp(l, r);
945 if (ret)
946 return ret;
947 }
948
949 return ret;
950}
951
39aeb52f 952static struct work_atoms *
b5fae128
IM		 953thread_atoms_search(struct rb_root *root, struct thread *thread,
954 struct list_head *sort_list)
955{
956 struct rb_node *node = root->rb_node;
39aeb52f 957 struct work_atoms key = { .thread = thread };
b5fae128
IM		 958
959 while (node) {
39aeb52f 960 struct work_atoms *atoms;
b5fae128
IM		 961 int cmp;
962
39aeb52f 963 atoms = container_of(node, struct work_atoms, node);
b5fae128
IM		 964
965 cmp = thread_lat_cmp(sort_list, &key, atoms);
966 if (cmp > 0)
967 node = node->rb_left;
968 else if (cmp < 0)
969 node = node->rb_right;
970 else {
971 BUG_ON(thread != atoms->thread);
972 return atoms;
973 }
974 }
975 return NULL;
976}
977
cdce9d73 978static void
39aeb52f 979__thread_latency_insert(struct rb_root *root, struct work_atoms *data,
daa1d7a5 980 struct list_head *sort_list)
cdce9d73
FW		 981{
982 struct rb_node **new = &(root->rb_node), *parent = NULL;
983
984 while (*new) {
39aeb52f 985 struct work_atoms *this;
daa1d7a5 986 int cmp;
cdce9d73 987
39aeb52f 988 this = container_of(*new, struct work_atoms, node);
cdce9d73 989 parent = *new;
daa1d7a5
FW		 990
991 cmp = thread_lat_cmp(sort_list, data, this);
992
993 if (cmp > 0)
cdce9d73 994 new = &((*new)->rb_left);
cdce9d73 995 else
daa1d7a5 996 new = &((*new)->rb_right);
cdce9d73
FW		 997 }
998
999 rb_link_node(&data->node, parent, new);
1000 rb_insert_color(&data->node, root);
1001}
1002
0e9b07e5 1003static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
cdce9d73 1004{
36479484 1005 struct work_atoms *atoms = zalloc(sizeof(*atoms));
a116e05d
ACM		 1006 if (!atoms) {
1007 pr_err("No memory at %s\n", __func__);
1008 return -1;
1009 }
cdce9d73 1010
f3b623b8 1011 atoms->thread = thread__get(thread);
39aeb52f 1012 INIT_LIST_HEAD(&atoms->work_list);
0e9b07e5 1013 __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
a116e05d 1014 return 0;
cdce9d73
FW		 1015}
1016
9ec3f4e4 1017static char sched_out_state(u64 prev_state)
cdce9d73
FW		 1018{
1019 const char *str = TASK_STATE_TO_CHAR_STR;
1020
9ec3f4e4 1021 return str[prev_state];
cdce9d73
FW		 1022}
1023
a116e05d 1024static int
39aeb52f 1025add_sched_out_event(struct work_atoms *atoms,
1026 char run_state,
1027 u64 timestamp)
cdce9d73 1028{
36479484 1029 struct work_atom *atom = zalloc(sizeof(*atom));
a116e05d
ACM		 1030 if (!atom) {
1031 pr_err("Non memory at %s", __func__);
1032 return -1;
1033 }
cdce9d73 1034
aa1ab9d2
FW		 1035 atom->sched_out_time = timestamp;
1036
39aeb52f 1037 if (run_state == 'R') {
b1ffe8f3 1038 atom->state = THREAD_WAIT_CPU;
aa1ab9d2 1039 atom->wake_up_time = atom->sched_out_time;
c6ced611
FW		 1040 }
1041
39aeb52f 1042 list_add_tail(&atom->list, &atoms->work_list);
a116e05d 1043 return 0;
cdce9d73
FW		 1044}
1045
1046static void
1d037ca1
IT		 1047add_runtime_event(struct work_atoms *atoms, u64 delta,
1048 u64 timestamp __maybe_unused)
39aeb52f 1049{
1050 struct work_atom *atom;
1051
1052 BUG_ON(list_empty(&atoms->work_list));
1053
1054 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1055
1056 atom->runtime += delta;
1057 atoms->total_runtime += delta;
1058}
1059
1060static void
1061add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
cdce9d73 1062{
b1ffe8f3 1063 struct work_atom *atom;
66685678 1064 u64 delta;
cdce9d73 1065
39aeb52f 1066 if (list_empty(&atoms->work_list))
cdce9d73
FW		 1067 return;
1068
39aeb52f 1069 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
cdce9d73 1070
b1ffe8f3 1071 if (atom->state != THREAD_WAIT_CPU)
cdce9d73
FW		 1072 return;
1073
b1ffe8f3
IM		 1074 if (timestamp < atom->wake_up_time) {
1075 atom->state = THREAD_IGNORE;
cdce9d73
FW		 1076 return;
1077 }
1078
b1ffe8f3
IM		 1079 atom->state = THREAD_SCHED_IN;
1080 atom->sched_in_time = timestamp;
66685678 1081
b1ffe8f3 1082 delta = atom->sched_in_time - atom->wake_up_time;
66685678 1083 atoms->total_lat += delta;
3786310a 1084 if (delta > atoms->max_lat) {
66685678 1085 atoms->max_lat = delta;
3786310a
FW		 1086 atoms->max_lat_at = timestamp;
1087 }
66685678 1088 atoms->nb_atoms++;
cdce9d73
FW		 1089}
1090
9ec3f4e4
ACM		 1091static int latency_switch_event(struct perf_sched *sched,
1092 struct perf_evsel *evsel,
1093 struct perf_sample *sample,
1094 struct machine *machine)
cdce9d73 1095{
9ec3f4e4
ACM		 1096 const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1097 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1098 const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
39aeb52f 1099 struct work_atoms *out_events, *in_events;
cdce9d73 1100 struct thread *sched_out, *sched_in;
7f7f8d0b 1101 u64 timestamp0, timestamp = sample->time;
b91fc39f 1102 int cpu = sample->cpu, err = -1;
ea92ed5a
IM		 1103 s64 delta;
1104
39aeb52f 1105 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
ea92ed5a 1106
0e9b07e5
ACM		 1107 timestamp0 = sched->cpu_last_switched[cpu];
1108 sched->cpu_last_switched[cpu] = timestamp;
ea92ed5a
IM		 1109 if (timestamp0)
1110 delta = timestamp - timestamp0;
1111 else
1112 delta = 0;
1113
a116e05d
ACM		 1114 if (delta < 0) {
1115 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1116 return -1;
1117 }
cdce9d73 1118
1fcb8768
AH		 1119 sched_out = machine__findnew_thread(machine, -1, prev_pid);
1120 sched_in = machine__findnew_thread(machine, -1, next_pid);
b91fc39f
ACM		 1121 if (sched_out == NULL || sched_in == NULL)
1122 goto out_put;
cdce9d73 1123
0e9b07e5 1124 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
39aeb52f 1125 if (!out_events) {
0e9b07e5 1126 if (thread_atoms_insert(sched, sched_out))
b91fc39f 1127 goto out_put;
0e9b07e5 1128 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
a116e05d
ACM		 1129 if (!out_events) {
1130 pr_err("out-event: Internal tree error");
b91fc39f 1131 goto out_put;
a116e05d 1132 }
39aeb52f 1133 }
9ec3f4e4 1134 if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
a116e05d 1135 return -1;
39aeb52f 1136
0e9b07e5 1137 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
39aeb52f 1138 if (!in_events) {
0e9b07e5 1139 if (thread_atoms_insert(sched, sched_in))
b91fc39f 1140 goto out_put;
0e9b07e5 1141 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
a116e05d
ACM		 1142 if (!in_events) {
1143 pr_err("in-event: Internal tree error");
b91fc39f 1144 goto out_put;
a116e05d 1145 }
39aeb52f 1146 /*
1147 * Take came in we have not heard about yet,
1148 * add in an initial atom in runnable state:
1149 */
a116e05d 1150 if (add_sched_out_event(in_events, 'R', timestamp))
b91fc39f 1151 goto out_put;
cdce9d73 1152 }
39aeb52f 1153 add_sched_in_event(in_events, timestamp);
b91fc39f
ACM		 1154 err = 0;
1155out_put:
1156 thread__put(sched_out);
1157 thread__put(sched_in);
1158 return err;
39aeb52f 1159}
cdce9d73 1160
9ec3f4e4
ACM		 1161static int latency_runtime_event(struct perf_sched *sched,
1162 struct perf_evsel *evsel,
1163 struct perf_sample *sample,
1164 struct machine *machine)
39aeb52f 1165{
9ec3f4e4
ACM		 1166 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1167 const u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
1fcb8768 1168 struct thread *thread = machine__findnew_thread(machine, -1, pid);
0e9b07e5 1169 struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
7f7f8d0b 1170 u64 timestamp = sample->time;
b91fc39f
ACM		 1171 int cpu = sample->cpu, err = -1;
1172
1173 if (thread == NULL)
1174 return -1;
39aeb52f 1175
1176 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
39aeb52f 1177 if (!atoms) {
0e9b07e5 1178 if (thread_atoms_insert(sched, thread))
b91fc39f 1179 goto out_put;
0e9b07e5 1180 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
a116e05d 1181 if (!atoms) {
60b7d14a 1182 pr_err("in-event: Internal tree error");
b91fc39f 1183 goto out_put;
a116e05d
ACM		 1184 }
1185 if (add_sched_out_event(atoms, 'R', timestamp))
b91fc39f 1186 goto out_put;
cdce9d73
FW		 1187 }
1188
9ec3f4e4 1189 add_runtime_event(atoms, runtime, timestamp);
b91fc39f
ACM		 1190 err = 0;
1191out_put:
1192 thread__put(thread);
1193 return err;
cdce9d73
FW		 1194}
1195
9ec3f4e4
ACM		 1196static int latency_wakeup_event(struct perf_sched *sched,
1197 struct perf_evsel *evsel,
1198 struct perf_sample *sample,
1199 struct machine *machine)
cdce9d73 1200{
0680ee7d 1201 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
39aeb52f 1202 struct work_atoms *atoms;
b1ffe8f3 1203 struct work_atom *atom;
cdce9d73 1204 struct thread *wakee;
7f7f8d0b 1205 u64 timestamp = sample->time;
b91fc39f 1206 int err = -1;
cdce9d73 1207
1fcb8768 1208 wakee = machine__findnew_thread(machine, -1, pid);
b91fc39f
ACM		 1209 if (wakee == NULL)
1210 return -1;
0e9b07e5 1211 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
17562205 1212 if (!atoms) {
0e9b07e5 1213 if (thread_atoms_insert(sched, wakee))
b91fc39f 1214 goto out_put;
0e9b07e5 1215 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
a116e05d 1216 if (!atoms) {
60b7d14a 1217 pr_err("wakeup-event: Internal tree error");
b91fc39f 1218 goto out_put;
a116e05d
ACM		 1219 }
1220 if (add_sched_out_event(atoms, 'S', timestamp))
b91fc39f 1221 goto out_put;
cdce9d73
FW		 1222 }
1223
39aeb52f 1224 BUG_ON(list_empty(&atoms->work_list));
cdce9d73 1225
39aeb52f 1226 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
cdce9d73 1227
55ffb7a6 1228 /*
67d6259d
DY		 1229 * As we do not guarantee the wakeup event happens when
1230 * task is out of run queue, also may happen when task is
1231 * on run queue and wakeup only change ->state to TASK_RUNNING,
1232 * then we should not set the ->wake_up_time when wake up a
1233 * task which is on run queue.
1234 *
55ffb7a6
MG		 1235 * You WILL be missing events if you've recorded only
1236 * one CPU, or are only looking at only one, so don't
67d6259d 1237 * skip in this case.
55ffb7a6 1238 */
0e9b07e5 1239 if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
b91fc39f 1240 goto out_ok;
cdce9d73 1241
0e9b07e5 1242 sched->nr_timestamps++;
ea57c4f5 1243 if (atom->sched_out_time > timestamp) {
0e9b07e5 1244 sched->nr_unordered_timestamps++;
b91fc39f 1245 goto out_ok;
ea57c4f5 1246 }
aa1ab9d2 1247
b1ffe8f3
IM		 1248 atom->state = THREAD_WAIT_CPU;
1249 atom->wake_up_time = timestamp;
b91fc39f
ACM		 1250out_ok:
1251 err = 0;
1252out_put:
1253 thread__put(wakee);
1254 return err;
cdce9d73
FW		 1255}
1256
9ec3f4e4
ACM		 1257static int latency_migrate_task_event(struct perf_sched *sched,
1258 struct perf_evsel *evsel,
1259 struct perf_sample *sample,
1260 struct machine *machine)
55ffb7a6 1261{
9ec3f4e4 1262 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
7f7f8d0b 1263 u64 timestamp = sample->time;
55ffb7a6
MG		 1264 struct work_atoms *atoms;
1265 struct work_atom *atom;
1266 struct thread *migrant;
b91fc39f 1267 int err = -1;
55ffb7a6
MG		 1268
1269 /*
1270 * Only need to worry about migration when profiling one CPU.
1271 */
0e9b07e5 1272 if (sched->profile_cpu == -1)
a116e05d 1273 return 0;
55ffb7a6 1274
1fcb8768 1275 migrant = machine__findnew_thread(machine, -1, pid);
b91fc39f
ACM		 1276 if (migrant == NULL)
1277 return -1;
0e9b07e5 1278 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
55ffb7a6 1279 if (!atoms) {
0e9b07e5 1280 if (thread_atoms_insert(sched, migrant))
b91fc39f 1281 goto out_put;
b9c5143a 1282 register_pid(sched, migrant->tid, thread__comm_str(migrant));
0e9b07e5 1283 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
a116e05d 1284 if (!atoms) {
60b7d14a 1285 pr_err("migration-event: Internal tree error");
b91fc39f 1286 goto out_put;
a116e05d
ACM		 1287 }
1288 if (add_sched_out_event(atoms, 'R', timestamp))
b91fc39f 1289 goto out_put;
55ffb7a6
MG		 1290 }
1291
1292 BUG_ON(list_empty(&atoms->work_list));
1293
1294 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1295 atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1296
0e9b07e5 1297 sched->nr_timestamps++;
55ffb7a6
MG		 1298
1299 if (atom->sched_out_time > timestamp)
0e9b07e5 1300 sched->nr_unordered_timestamps++;
b91fc39f
ACM		 1301 err = 0;
1302out_put:
1303 thread__put(migrant);
1304 return err;
55ffb7a6
MG		 1305}
1306
0e9b07e5 1307static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
cdce9d73 1308{
cdce9d73
FW		 1309 int i;
1310 int ret;
66685678 1311 u64 avg;
99620a5d 1312 char max_lat_at[32];
cdce9d73 1313
39aeb52f 1314 if (!work_list->nb_atoms)
cdce9d73 1315 return;
ea57c4f5
IM		 1316 /*
1317 * Ignore idle threads:
1318 */
b9c5143a 1319 if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
ea57c4f5 1320 return;
cdce9d73 1321
0e9b07e5
ACM		 1322 sched->all_runtime += work_list->total_runtime;
1323 sched->all_count += work_list->nb_atoms;
66685678 1324
2f80dd44
JB		 1325 if (work_list->num_merged > 1)
1326 ret = printf(" %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1327 else
1328 ret = printf(" %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
cdce9d73 1329
08f69e6c 1330 for (i = 0; i < 24 - ret; i++)
cdce9d73
FW		 1331 printf(" ");
1332
39aeb52f 1333 avg = work_list->total_lat / work_list->nb_atoms;
99620a5d 1334 timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at));
cdce9d73 1335
99620a5d 1336 printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n",
4fc76e49
ACM		 1337 (double)work_list->total_runtime / NSEC_PER_MSEC,
1338 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1339 (double)work_list->max_lat / NSEC_PER_MSEC,
99620a5d 1340 max_lat_at);
cdce9d73
FW		 1341}
1342
39aeb52f 1343static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
daa1d7a5 1344{
0014de17
JO		 1345 if (l->thread == r->thread)
1346 return 0;
38051234 1347 if (l->thread->tid < r->thread->tid)
daa1d7a5 1348 return -1;
38051234 1349 if (l->thread->tid > r->thread->tid)
daa1d7a5 1350 return 1;
0014de17 1351 return (int)(l->thread - r->thread);
daa1d7a5
FW		 1352}
1353
39aeb52f 1354static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
daa1d7a5
FW		 1355{
1356 u64 avgl, avgr;
1357
1358 if (!l->nb_atoms)
1359 return -1;
1360
1361 if (!r->nb_atoms)
1362 return 1;
1363
1364 avgl = l->total_lat / l->nb_atoms;
1365 avgr = r->total_lat / r->nb_atoms;
1366
1367 if (avgl < avgr)
1368 return -1;
1369 if (avgl > avgr)
1370 return 1;
1371
1372 return 0;
1373}
1374
39aeb52f 1375static int max_cmp(struct work_atoms *l, struct work_atoms *r)
daa1d7a5
FW		 1376{
1377 if (l->max_lat < r->max_lat)
1378 return -1;
1379 if (l->max_lat > r->max_lat)
1380 return 1;
1381
1382 return 0;
1383}
1384
39aeb52f 1385static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
daa1d7a5
FW		 1386{
1387 if (l->nb_atoms < r->nb_atoms)
1388 return -1;
1389 if (l->nb_atoms > r->nb_atoms)
1390 return 1;
1391
1392 return 0;
1393}
1394
39aeb52f 1395static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
daa1d7a5
FW		 1396{
1397 if (l->total_runtime < r->total_runtime)
1398 return -1;
1399 if (l->total_runtime > r->total_runtime)
1400 return 1;
1401
1402 return 0;
1403}
1404
cbef79a8 1405static int sort_dimension__add(const char *tok, struct list_head *list)
daa1d7a5 1406{
0e9b07e5
ACM		 1407 size_t i;
1408 static struct sort_dimension avg_sort_dimension = {
1409 .name = "avg",
1410 .cmp = avg_cmp,
1411 };
1412 static struct sort_dimension max_sort_dimension = {
1413 .name = "max",
1414 .cmp = max_cmp,
1415 };
1416 static struct sort_dimension pid_sort_dimension = {
1417 .name = "pid",
1418 .cmp = pid_cmp,
1419 };
1420 static struct sort_dimension runtime_sort_dimension = {
1421 .name = "runtime",
1422 .cmp = runtime_cmp,
1423 };
1424 static struct sort_dimension switch_sort_dimension = {
1425 .name = "switch",
1426 .cmp = switch_cmp,
1427 };
1428 struct sort_dimension *available_sorts[] = {
1429 &pid_sort_dimension,
1430 &avg_sort_dimension,
1431 &max_sort_dimension,
1432 &switch_sort_dimension,
1433 &runtime_sort_dimension,
1434 };
daa1d7a5 1435
0e9b07e5 1436 for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
daa1d7a5
FW		 1437 if (!strcmp(available_sorts[i]->name, tok)) {
1438 list_add_tail(&available_sorts[i]->list, list);
1439
1440 return 0;
1441 }
1442 }
1443
1444 return -1;
1445}
1446
0e9b07e5 1447static void perf_sched__sort_lat(struct perf_sched *sched)
daa1d7a5
FW		 1448{
1449 struct rb_node *node;
2f80dd44
JB		 1450 struct rb_root *root = &sched->atom_root;
1451again:
daa1d7a5 1452 for (;;) {
39aeb52f 1453 struct work_atoms *data;
2f80dd44 1454 node = rb_first(root);
daa1d7a5
FW		 1455 if (!node)
1456 break;
1457
2f80dd44 1458 rb_erase(node, root);
39aeb52f 1459 data = rb_entry(node, struct work_atoms, node);
0e9b07e5 1460 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
daa1d7a5 1461 }
2f80dd44
JB		 1462 if (root == &sched->atom_root) {
1463 root = &sched->merged_atom_root;
1464 goto again;
1465 }
daa1d7a5
FW		 1466}
1467
0e9b07e5 1468static int process_sched_wakeup_event(struct perf_tool *tool,
2b7fcbc5 1469 struct perf_evsel *evsel,
1d037ca1 1470 struct perf_sample *sample,
4218e673 1471 struct machine *machine)
419ab0d6 1472{
0e9b07e5 1473 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
419ab0d6 1474
9ec3f4e4
ACM		 1475 if (sched->tp_handler->wakeup_event)
1476 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
a116e05d 1477
2b7fcbc5 1478 return 0;
419ab0d6
FW		 1479}
1480
a151a37a
JO		 1481union map_priv {
1482 void *ptr;
1483 bool color;
1484};
1485
1486static bool thread__has_color(struct thread *thread)
1487{
1488 union map_priv priv = {
1489 .ptr = thread__priv(thread),
1490 };
1491
1492 return priv.color;
1493}
1494
1495static struct thread*
1496map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1497{
1498 struct thread *thread = machine__findnew_thread(machine, pid, tid);
1499 union map_priv priv = {
1500 .color = false,
1501 };
1502
1503 if (!sched->map.color_pids || !thread || thread__priv(thread))
1504 return thread;
1505
1506 if (thread_map__has(sched->map.color_pids, tid))
1507 priv.color = true;
1508
1509 thread__set_priv(thread, priv.ptr);
1510 return thread;
1511}
1512
9ec3f4e4
ACM		 1513static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1514 struct perf_sample *sample, struct machine *machine)
0ec04e16 1515{
9d372ca5
DY		 1516 const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1517 struct thread *sched_in;
8640da9f 1518 struct thread_runtime *tr;
0ec04e16 1519 int new_shortname;
7f7f8d0b 1520 u64 timestamp0, timestamp = sample->time;
0ec04e16 1521 s64 delta;
99623c62
JO		 1522 int i, this_cpu = sample->cpu;
1523 int cpus_nr;
1524 bool new_cpu = false;
8cd91195 1525 const char *color = PERF_COLOR_NORMAL;
99620a5d 1526 char stimestamp[32];
0ec04e16
IM		 1527
1528 BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1529
0e9b07e5
ACM		 1530 if (this_cpu > sched->max_cpu)
1531 sched->max_cpu = this_cpu;
0ec04e16 1532
99623c62
JO		 1533 if (sched->map.comp) {
1534 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1535 if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1536 sched->map.comp_cpus[cpus_nr++] = this_cpu;
1537 new_cpu = true;
1538 }
1539 } else
1540 cpus_nr = sched->max_cpu;
1541
0e9b07e5
ACM		 1542 timestamp0 = sched->cpu_last_switched[this_cpu];
1543 sched->cpu_last_switched[this_cpu] = timestamp;
0ec04e16
IM		 1544 if (timestamp0)
1545 delta = timestamp - timestamp0;
1546 else
1547 delta = 0;
1548
a116e05d 1549 if (delta < 0) {
60b7d14a 1550 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
a116e05d
ACM		 1551 return -1;
1552 }
0ec04e16 1553
a151a37a 1554 sched_in = map__findnew_thread(sched, machine, -1, next_pid);
b91fc39f
ACM		 1555 if (sched_in == NULL)
1556 return -1;
0ec04e16 1557
8640da9f
CD		 1558 tr = thread__get_runtime(sched_in);
1559 if (tr == NULL) {
1560 thread__put(sched_in);
1561 return -1;
1562 }
1563
b91fc39f 1564 sched->curr_thread[this_cpu] = thread__get(sched_in);
0ec04e16
IM		 1565
1566 printf(" ");
1567
1568 new_shortname = 0;
8640da9f 1569 if (!tr->shortname[0]) {
6bcab4e1
D		 1570 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1571 /*
1572 * Don't allocate a letter-number for swapper:0
1573 * as a shortname. Instead, we use '.' for it.
1574 */
8640da9f
CD		 1575 tr->shortname[0] = '.';
1576 tr->shortname[1] = ' ';
0ec04e16 1577 } else {
8640da9f
CD		 1578 tr->shortname[0] = sched->next_shortname1;
1579 tr->shortname[1] = sched->next_shortname2;
6bcab4e1
D		 1580
1581 if (sched->next_shortname1 < 'Z') {
1582 sched->next_shortname1++;
0ec04e16 1583 } else {
6bcab4e1
D		 1584 sched->next_shortname1 = 'A';
1585 if (sched->next_shortname2 < '9')
1586 sched->next_shortname2++;
1587 else
1588 sched->next_shortname2 = '0';
0ec04e16
IM		 1589 }
1590 }
1591 new_shortname = 1;
1592 }
1593
99623c62
JO		 1594 for (i = 0; i < cpus_nr; i++) {
1595 int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
a151a37a 1596 struct thread *curr_thread = sched->curr_thread[cpu];
8640da9f 1597 struct thread_runtime *curr_tr;
a151a37a 1598 const char *pid_color = color;
cf294f24 1599 const char *cpu_color = color;
a151a37a
JO		 1600
1601 if (curr_thread && thread__has_color(curr_thread))
1602 pid_color = COLOR_PIDS;
99623c62 1603
73643bb6
JO		 1604 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1605 continue;
1606
cf294f24
JO		 1607 if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1608 cpu_color = COLOR_CPUS;
1609
0ec04e16 1610 if (cpu != this_cpu)
1208bb27 1611 color_fprintf(stdout, color, " ");
0ec04e16 1612 else
cf294f24 1613 color_fprintf(stdout, cpu_color, "*");
0ec04e16 1614
8640da9f
CD		 1615 if (sched->curr_thread[cpu]) {
1616 curr_tr = thread__get_runtime(sched->curr_thread[cpu]);
1617 if (curr_tr == NULL) {
1618 thread__put(sched_in);
1619 return -1;
1620 }
1621 color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1622 } else
8cd91195 1623 color_fprintf(stdout, color, " ");
0ec04e16
IM		 1624 }
1625
73643bb6
JO		 1626 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1627 goto out;
1628
99620a5d
NK		 1629 timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1630 color_fprintf(stdout, color, " %12s secs ", stimestamp);
99a3c3a9 1631 if (new_shortname || tr->comm_changed || (verbose > 0 && sched_in->tid)) {
a151a37a
JO		 1632 const char *pid_color = color;
1633
1634 if (thread__has_color(sched_in))
1635 pid_color = COLOR_PIDS;
1636
1637 color_fprintf(stdout, pid_color, "%s => %s:%d",
8640da9f 1638 tr->shortname, thread__comm_str(sched_in), sched_in->tid);
99a3c3a9 1639 tr->comm_changed = false;
0ec04e16 1640 }
a116e05d 1641
99623c62 1642 if (sched->map.comp && new_cpu)
8cd91195 1643 color_fprintf(stdout, color, " (CPU %d)", this_cpu);
99623c62 1644
73643bb6 1645out:
8cd91195 1646 color_fprintf(stdout, color, "\n");
99623c62 1647
b91fc39f
ACM		 1648 thread__put(sched_in);
1649
a116e05d 1650 return 0;
0ec04e16
IM		 1651}
1652
0e9b07e5 1653static int process_sched_switch_event(struct perf_tool *tool,
2b7fcbc5 1654 struct perf_evsel *evsel,
1d037ca1 1655 struct perf_sample *sample,
4218e673 1656 struct machine *machine)
419ab0d6 1657{
0e9b07e5 1658 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
a116e05d 1659 int this_cpu = sample->cpu, err = 0;
2b7fcbc5
ACM		 1660 u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1661 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
419ab0d6 1662
0e9b07e5 1663 if (sched->curr_pid[this_cpu] != (u32)-1) {
c8a37751
IM		 1664 /*
1665 * Are we trying to switch away a PID that is
1666 * not current?
1667 */
2b7fcbc5 1668 if (sched->curr_pid[this_cpu] != prev_pid)
0e9b07e5 1669 sched->nr_context_switch_bugs++;
c8a37751 1670 }
c8a37751 1671
9ec3f4e4
ACM		 1672 if (sched->tp_handler->switch_event)
1673 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
2b7fcbc5
ACM		 1674
1675 sched->curr_pid[this_cpu] = next_pid;
a116e05d 1676 return err;
419ab0d6
FW		 1677}
1678
0e9b07e5 1679static int process_sched_runtime_event(struct perf_tool *tool,
2b7fcbc5 1680 struct perf_evsel *evsel,
1d037ca1 1681 struct perf_sample *sample,
4218e673 1682 struct machine *machine)
39aeb52f 1683{
0e9b07e5 1684 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
39aeb52f 1685
9ec3f4e4
ACM		 1686 if (sched->tp_handler->runtime_event)
1687 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
a116e05d 1688
2b7fcbc5 1689 return 0;
39aeb52f 1690}
1691
cb627505
DA		 1692static int perf_sched__process_fork_event(struct perf_tool *tool,
1693 union perf_event *event,
1694 struct perf_sample *sample,
1695 struct machine *machine)
fbf94829 1696{
0e9b07e5 1697 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
46538818 1698
cb627505
DA		 1699 /* run the fork event through the perf machineruy */
1700 perf_event__process_fork(tool, event, sample, machine);
1701
1702 /* and then run additional processing needed for this command */
9ec3f4e4 1703 if (sched->tp_handler->fork_event)
cb627505 1704 return sched->tp_handler->fork_event(sched, event, machine);
a116e05d 1705
2b7fcbc5 1706 return 0;
fbf94829
IM		 1707}
1708
0e9b07e5 1709static int process_sched_migrate_task_event(struct perf_tool *tool,
2b7fcbc5 1710 struct perf_evsel *evsel,
1d037ca1 1711 struct perf_sample *sample,
4218e673 1712 struct machine *machine)
55ffb7a6 1713{
0e9b07e5 1714 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
55ffb7a6 1715
9ec3f4e4
ACM		 1716 if (sched->tp_handler->migrate_task_event)
1717 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
a116e05d 1718
2b7fcbc5 1719 return 0;
55ffb7a6
MG		 1720}
1721
a116e05d 1722typedef int (*tracepoint_handler)(struct perf_tool *tool,
2b7fcbc5 1723 struct perf_evsel *evsel,
a116e05d 1724 struct perf_sample *sample,
4218e673 1725 struct machine *machine);
ec156764 1726
1d037ca1
IT		 1727static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1728 union perf_event *event __maybe_unused,
ee29be62
ACM		 1729 struct perf_sample *sample,
1730 struct perf_evsel *evsel,
1731 struct machine *machine)
0a02ad93 1732{
a116e05d 1733 int err = 0;
0a02ad93 1734
744a9719
ACM		 1735 if (evsel->handler != NULL) {
1736 tracepoint_handler f = evsel->handler;
2b7fcbc5 1737 err = f(tool, evsel, sample, machine);
ee29be62 1738 }
0a02ad93 1739
a116e05d 1740 return err;
0a02ad93
IM		 1741}
1742
99a3c3a9
CD		 1743static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1744 union perf_event *event,
1745 struct perf_sample *sample,
1746 struct machine *machine)
1747{
1748 struct thread *thread;
1749 struct thread_runtime *tr;
1750 int err;
1751
1752 err = perf_event__process_comm(tool, event, sample, machine);
1753 if (err)
1754 return err;
1755
1756 thread = machine__find_thread(machine, sample->pid, sample->tid);
1757 if (!thread) {
1758 pr_err("Internal error: can't find thread\n");
1759 return -1;
1760 }
1761
1762 tr = thread__get_runtime(thread);
1763 if (tr == NULL) {
1764 thread__put(thread);
1765 return -1;
1766 }
1767
1768 tr->comm_changed = true;
1769 thread__put(thread);
1770
1771 return 0;
1772}
1773
ae536acf 1774static int perf_sched__read_events(struct perf_sched *sched)
0a02ad93 1775{
ee29be62
ACM		 1776 const struct perf_evsel_str_handler handlers[] = {
1777 { "sched:sched_switch", process_sched_switch_event, },
1778 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1779 { "sched:sched_wakeup", process_sched_wakeup_event, },
1780 { "sched:sched_wakeup_new", process_sched_wakeup_event, },
ee29be62
ACM		 1781 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1782 };
da378962 1783 struct perf_session *session;
8ceb41d7 1784 struct perf_data data = {
eae8ad80
JO		 1785 .file = {
1786 .path = input_name,
1787 },
1788 .mode = PERF_DATA_MODE_READ,
1789 .force = sched->force,
f5fc1412 1790 };
ae536acf 1791 int rc = -1;
da378962 1792
8ceb41d7 1793 session = perf_session__new(&data, false, &sched->tool);
a116e05d
ACM		 1794 if (session == NULL) {
1795 pr_debug("No Memory for session\n");
1796 return -1;
1797 }
94c744b6 1798
0a7e6d1b 1799 symbol__init(&session->header.env);
04934106 1800
a116e05d
ACM		 1801 if (perf_session__set_tracepoints_handlers(session, handlers))
1802 goto out_delete;
ee29be62 1803
cee75ac7 1804 if (perf_session__has_traces(session, "record -R")) {
b7b61cbe 1805 int err = perf_session__process_events(session);
a116e05d
ACM		 1806 if (err) {
1807 pr_err("Failed to process events, error %d", err);
1808 goto out_delete;
1809 }
4c09bafa 1810
75be989a
ACM		 1811 sched->nr_events = session->evlist->stats.nr_events[0];
1812 sched->nr_lost_events = session->evlist->stats.total_lost;
1813 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
cee75ac7 1814 }
d549c769 1815
ae536acf 1816 rc = 0;
a116e05d
ACM		 1817out_delete:
1818 perf_session__delete(session);
ae536acf 1819 return rc;
0a02ad93
IM		 1820}
1821
49394a2a
DA		 1822/*
1823 * scheduling times are printed as msec.usec
1824 */
1825static inline void print_sched_time(unsigned long long nsecs, int width)
1826{
1827 unsigned long msecs;
1828 unsigned long usecs;
1829
1830 msecs = nsecs / NSEC_PER_MSEC;
1831 nsecs -= msecs * NSEC_PER_MSEC;
1832 usecs = nsecs / NSEC_PER_USEC;
1833 printf("%*lu.%03lu ", width, msecs, usecs);
1834}
1835
1836/*
1837 * returns runtime data for event, allocating memory for it the
1838 * first time it is used.
1839 */
1840static struct evsel_runtime *perf_evsel__get_runtime(struct perf_evsel *evsel)
1841{
1842 struct evsel_runtime *r = evsel->priv;
1843
1844 if (r == NULL) {
1845 r = zalloc(sizeof(struct evsel_runtime));
1846 evsel->priv = r;
1847 }
1848
1849 return r;
1850}
1851
1852/*
1853 * save last time event was seen per cpu
1854 */
1855static void perf_evsel__save_time(struct perf_evsel *evsel,
1856 u64 timestamp, u32 cpu)
1857{
1858 struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1859
1860 if (r == NULL)
1861 return;
1862
1863 if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1864 int i, n = __roundup_pow_of_two(cpu+1);
1865 void *p = r->last_time;
1866
1867 p = realloc(r->last_time, n * sizeof(u64));
1868 if (!p)
1869 return;
1870
1871 r->last_time = p;
1872 for (i = r->ncpu; i < n; ++i)
1873 r->last_time[i] = (u64) 0;
1874
1875 r->ncpu = n;
1876 }
1877
1878 r->last_time[cpu] = timestamp;
1879}
1880
1881/* returns last time this event was seen on the given cpu */
1882static u64 perf_evsel__get_time(struct perf_evsel *evsel, u32 cpu)
1883{
1884 struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1885
1886 if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1887 return 0;
1888
1889 return r->last_time[cpu];
1890}
1891
9b8087d7 1892static int comm_width = 30;
49394a2a
DA		 1893
1894static char *timehist_get_commstr(struct thread *thread)
1895{
1896 static char str[32];
1897 const char *comm = thread__comm_str(thread);
1898 pid_t tid = thread->tid;
1899 pid_t pid = thread->pid_;
1900 int n;
1901
1902 if (pid == 0)
1903 n = scnprintf(str, sizeof(str), "%s", comm);
1904
1905 else if (tid != pid)
1906 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1907
1908 else
1909 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1910
1911 if (n > comm_width)
1912 comm_width = n;
1913
1914 return str;
1915}
1916
a407b067 1917static void timehist_header(struct perf_sched *sched)
49394a2a 1918{
a407b067
DA		 1919 u32 ncpus = sched->max_cpu + 1;
1920 u32 i, j;
1921
49394a2a
DA		 1922 printf("%15s %6s ", "time", "cpu");
1923
a407b067
DA		 1924 if (sched->show_cpu_visual) {
1925 printf(" ");
1926 for (i = 0, j = 0; i < ncpus; ++i) {
1927 printf("%x", j++);
1928 if (j > 15)
1929 j = 0;
1930 }
1931 printf(" ");
1932 }
1933
0e6758e8 1934 printf(" %-*s %9s %9s %9s", comm_width,
49394a2a
DA		 1935 "task name", "wait time", "sch delay", "run time");
1936
414e050c
NK		 1937 if (sched->show_state)
1938 printf(" %s", "state");
1939
49394a2a
DA		 1940 printf("\n");
1941
1942 /*
1943 * units row
1944 */
1945 printf("%15s %-6s ", "", "");
1946
a407b067
DA		 1947 if (sched->show_cpu_visual)
1948 printf(" %*s ", ncpus, "");
1949
414e050c 1950 printf(" %-*s %9s %9s %9s", comm_width,
0e6758e8 1951 "[tid/pid]", "(msec)", "(msec)", "(msec)");
49394a2a 1952
414e050c
NK		 1953 if (sched->show_state)
1954 printf(" %5s", "");
1955
1956 printf("\n");
1957
49394a2a
DA		 1958 /*
1959 * separator
1960 */
1961 printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1962
a407b067
DA		 1963 if (sched->show_cpu_visual)
1964 printf(" %.*s ", ncpus, graph_dotted_line);
1965
0e6758e8 1966 printf(" %.*s %.9s %.9s %.9s", comm_width,
49394a2a
DA		 1967 graph_dotted_line, graph_dotted_line, graph_dotted_line,
1968 graph_dotted_line);
1969
414e050c
NK		 1970 if (sched->show_state)
1971 printf(" %.5s", graph_dotted_line);
1972
49394a2a
DA		 1973 printf("\n");
1974}
1975
414e050c
NK		 1976static char task_state_char(struct thread *thread, int state)
1977{
1978 static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
1979 unsigned bit = state ? ffs(state) : 0;
1980
1981 /* 'I' for idle */
1982 if (thread->tid == 0)
1983 return 'I';
1984
1985 return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
1986}
1987
fc1469f1 1988static void timehist_print_sample(struct perf_sched *sched,
292c4a8f 1989 struct perf_evsel *evsel,
fc1469f1 1990 struct perf_sample *sample,
6c973c90 1991 struct addr_location *al,
853b7407 1992 struct thread *thread,
414e050c 1993 u64 t, int state)
49394a2a
DA		 1994{
1995 struct thread_runtime *tr = thread__priv(thread);
292c4a8f
BG		 1996 const char *next_comm = perf_evsel__strval(evsel, sample, "next_comm");
1997 const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
a407b067 1998 u32 max_cpus = sched->max_cpu + 1;
49394a2a 1999 char tstr[64];
292c4a8f 2000 char nstr[30];
941bdea7 2001 u64 wait_time;
49394a2a 2002
853b7407 2003 timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
49394a2a
DA		 2004 printf("%15s [%04d] ", tstr, sample->cpu);
2005
a407b067
DA		 2006 if (sched->show_cpu_visual) {
2007 u32 i;
2008 char c;
2009
2010 printf(" ");
2011 for (i = 0; i < max_cpus; ++i) {
2012 /* flag idle times with 'i'; others are sched events */
2013 if (i == sample->cpu)
2014 c = (thread->tid == 0) ? 'i' : 's';
2015 else
2016 c = ' ';
2017 printf("%c", c);
2018 }
2019 printf(" ");
2020 }
2021
49394a2a
DA		 2022 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2023
941bdea7
NK		 2024 wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2025 print_sched_time(wait_time, 6);
2026
49394a2a
DA		 2027 print_sched_time(tr->dt_delay, 6);
2028 print_sched_time(tr->dt_run, 6);
fc1469f1 2029
414e050c
NK		 2030 if (sched->show_state)
2031 printf(" %5c ", task_state_char(thread, state));
2032
292c4a8f
BG		 2033 if (sched->show_next) {
2034 snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2035 printf(" %-*s", comm_width, nstr);
2036 }
2037
2038 if (sched->show_wakeups && !sched->show_next)
fc1469f1
DA		 2039 printf(" %-*s", comm_width, "");
2040
6c973c90
DA		 2041 if (thread->tid == 0)
2042 goto out;
2043
2044 if (sched->show_callchain)
2045 printf(" ");
2046
2047 sample__fprintf_sym(sample, al, 0,
2048 EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2d9bbf6e
NK		 2049 EVSEL__PRINT_CALLCHAIN_ARROW |
2050 EVSEL__PRINT_SKIP_IGNORED,
6c973c90
DA		 2051 &callchain_cursor, stdout);
2052
2053out:
49394a2a
DA		 2054 printf("\n");
2055}
2056
2057/*
2058 * Explanation of delta-time stats:
2059 *
2060 * t = time of current schedule out event
2061 * tprev = time of previous sched out event
2062 * also time of schedule-in event for current task
2063 * last_time = time of last sched change event for current task
2064 * (i.e, time process was last scheduled out)
2065 * ready_to_run = time of wakeup for current task
2066 *
2067 * -----|------------|------------|------------|------
2068 * last ready tprev t
2069 * time to run
2070 *
2071 * |-------- dt_wait --------|
2072 * |- dt_delay -|-- dt_run --|
2073 *
2074 * dt_run = run time of current task
2075 * dt_wait = time between last schedule out event for task and tprev
2076 * represents time spent off the cpu
2077 * dt_delay = time between wakeup and schedule-in of task
2078 */
2079
2080static void timehist_update_runtime_stats(struct thread_runtime *r,
2081 u64 t, u64 tprev)
2082{
2083 r->dt_delay = 0;
941bdea7
NK		 2084 r->dt_sleep = 0;
2085 r->dt_iowait = 0;
2086 r->dt_preempt = 0;
49394a2a 2087 r->dt_run = 0;
941bdea7 2088
49394a2a
DA		 2089 if (tprev) {
2090 r->dt_run = t - tprev;
2091 if (r->ready_to_run) {
2092 if (r->ready_to_run > tprev)
2093 pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2094 else
2095 r->dt_delay = tprev - r->ready_to_run;
2096 }
2097
2098 if (r->last_time > tprev)
2099 pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
941bdea7
NK		 2100 else if (r->last_time) {
2101 u64 dt_wait = tprev - r->last_time;
2102
2103 if (r->last_state == TASK_RUNNING)
2104 r->dt_preempt = dt_wait;
2105 else if (r->last_state == TASK_UNINTERRUPTIBLE)
2106 r->dt_iowait = dt_wait;
2107 else
2108 r->dt_sleep = dt_wait;
2109 }
49394a2a
DA		 2110 }
2111
2112 update_stats(&r->run_stats, r->dt_run);
587782c5
NK		 2113
2114 r->total_run_time += r->dt_run;
2115 r->total_delay_time += r->dt_delay;
2116 r->total_sleep_time += r->dt_sleep;
2117 r->total_iowait_time += r->dt_iowait;
2118 r->total_preempt_time += r->dt_preempt;
49394a2a
DA		 2119}
2120
96039c7c
NK		 2121static bool is_idle_sample(struct perf_sample *sample,
2122 struct perf_evsel *evsel)
49394a2a
DA		 2123{
2124 /* pid 0 == swapper == idle task */
96039c7c
NK		 2125 if (strcmp(perf_evsel__name(evsel), "sched:sched_switch") == 0)
2126 return perf_evsel__intval(evsel, sample, "prev_pid") == 0;
49394a2a 2127
96039c7c
NK		 2128 return sample->pid == 0;
2129}
2130
2131static void save_task_callchain(struct perf_sched *sched,
2132 struct perf_sample *sample,
2133 struct perf_evsel *evsel,
2134 struct machine *machine)
2135{
2136 struct callchain_cursor *cursor = &callchain_cursor;
2137 struct thread *thread;
6c973c90
DA		 2138
2139 /* want main thread for process - has maps */
2140 thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2141 if (thread == NULL) {
2142 pr_debug("Failed to get thread for pid %d.\n", sample->pid);
96039c7c 2143 return;
6c973c90
DA		 2144 }
2145
4c50563d 2146 if (!sched->show_callchain || sample->callchain == NULL)
96039c7c 2147 return;
6c973c90
DA		 2148
2149 if (thread__resolve_callchain(thread, cursor, evsel, sample,
8388deb3 2150 NULL, NULL, sched->max_stack + 2) != 0) {
bb963e16 2151 if (verbose > 0)
62d94b00 2152 pr_err("Failed to resolve callchain. Skipping\n");
6c973c90 2153
96039c7c 2154 return;
6c973c90 2155 }
cdeb01bf 2156
6c973c90 2157 callchain_cursor_commit(cursor);
cdeb01bf
NK		 2158
2159 while (true) {
2160 struct callchain_cursor_node *node;
2161 struct symbol *sym;
2162
2163 node = callchain_cursor_current(cursor);
2164 if (node == NULL)
2165 break;
2166
2167 sym = node->sym;
a7c3899c 2168 if (sym) {
cdeb01bf
NK		 2169 if (!strcmp(sym->name, "schedule") ||
2170 !strcmp(sym->name, "__schedule") ||
2171 !strcmp(sym->name, "preempt_schedule"))
2172 sym->ignore = 1;
2173 }
2174
2175 callchain_cursor_advance(cursor);
2176 }
49394a2a
DA		 2177}
2178
3bc2fa9c
NK		 2179static int init_idle_thread(struct thread *thread)
2180{
2181 struct idle_thread_runtime *itr;
2182
2183 thread__set_comm(thread, idle_comm, 0);
2184
2185 itr = zalloc(sizeof(*itr));
2186 if (itr == NULL)
2187 return -ENOMEM;
2188
2189 init_stats(&itr->tr.run_stats);
2190 callchain_init(&itr->callchain);
2191 callchain_cursor_reset(&itr->cursor);
2192 thread__set_priv(thread, itr);
2193
2194 return 0;
2195}
2196
49394a2a
DA		 2197/*
2198 * Track idle stats per cpu by maintaining a local thread
2199 * struct for the idle task on each cpu.
2200 */
2201static int init_idle_threads(int ncpu)
2202{
3bc2fa9c 2203 int i, ret;
49394a2a
DA		 2204
2205 idle_threads = zalloc(ncpu * sizeof(struct thread *));
2206 if (!idle_threads)
2207 return -ENOMEM;
2208
b336352b 2209 idle_max_cpu = ncpu;
49394a2a
DA		 2210
2211 /* allocate the actual thread struct if needed */
2212 for (i = 0; i < ncpu; ++i) {
2213 idle_threads[i] = thread__new(0, 0);
2214 if (idle_threads[i] == NULL)
2215 return -ENOMEM;
2216
3bc2fa9c
NK		 2217 ret = init_idle_thread(idle_threads[i]);
2218 if (ret < 0)
2219 return ret;
49394a2a
DA		 2220 }
2221
2222 return 0;
2223}
2224
2225static void free_idle_threads(void)
2226{
2227 int i;
2228
2229 if (idle_threads == NULL)
2230 return;
2231
b336352b 2232 for (i = 0; i < idle_max_cpu; ++i) {
49394a2a
DA		 2233 if ((idle_threads[i]))
2234 thread__delete(idle_threads[i]);
2235 }
2236
2237 free(idle_threads);
2238}
2239
2240static struct thread *get_idle_thread(int cpu)
2241{
2242 /*
2243 * expand/allocate array of pointers to local thread
2244 * structs if needed
2245 */
2246 if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2247 int i, j = __roundup_pow_of_two(cpu+1);
2248 void *p;
2249
2250 p = realloc(idle_threads, j * sizeof(struct thread *));
2251 if (!p)
2252 return NULL;
2253
2254 idle_threads = (struct thread **) p;
b336352b 2255 for (i = idle_max_cpu; i < j; ++i)
49394a2a
DA		 2256 idle_threads[i] = NULL;
2257
2258 idle_max_cpu = j;
2259 }
2260
2261 /* allocate a new thread struct if needed */
2262 if (idle_threads[cpu] == NULL) {
2263 idle_threads[cpu] = thread__new(0, 0);
2264
2265 if (idle_threads[cpu]) {
3bc2fa9c
NK		 2266 if (init_idle_thread(idle_threads[cpu]) < 0)
2267 return NULL;
49394a2a
DA		 2268 }
2269 }
2270
2271 return idle_threads[cpu];
2272}
2273
4c50563d
ACM		 2274static void save_idle_callchain(struct perf_sched *sched,
2275 struct idle_thread_runtime *itr,
699b5b92
NK		 2276 struct perf_sample *sample)
2277{
4c50563d 2278 if (!sched->show_callchain || sample->callchain == NULL)
699b5b92
NK		 2279 return;
2280
2281 callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2282}
2283
6c973c90
DA		 2284static struct thread *timehist_get_thread(struct perf_sched *sched,
2285 struct perf_sample *sample,
49394a2a
DA		 2286 struct machine *machine,
2287 struct perf_evsel *evsel)
2288{
2289 struct thread *thread;
2290
96039c7c 2291 if (is_idle_sample(sample, evsel)) {
49394a2a
DA		 2292 thread = get_idle_thread(sample->cpu);
2293 if (thread == NULL)
2294 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2295
2296 } else {
5d92d96a
NK		 2297 /* there were samples with tid 0 but non-zero pid */
2298 thread = machine__findnew_thread(machine, sample->pid,
2299 sample->tid ?: sample->pid);
49394a2a
DA		 2300 if (thread == NULL) {
2301 pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2302 sample->tid);
2303 }
96039c7c
NK		 2304
2305 save_task_callchain(sched, sample, evsel, machine);
699b5b92
NK		 2306 if (sched->idle_hist) {
2307 struct thread *idle;
2308 struct idle_thread_runtime *itr;
2309
2310 idle = get_idle_thread(sample->cpu);
2311 if (idle == NULL) {
2312 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2313 return NULL;
2314 }
2315
2316 itr = thread__priv(idle);
2317 if (itr == NULL)
2318 return NULL;
2319
2320 itr->last_thread = thread;
2321
2322 /* copy task callchain when entering to idle */
2323 if (perf_evsel__intval(evsel, sample, "next_pid") == 0)
4c50563d 2324 save_idle_callchain(sched, itr, sample);
699b5b92 2325 }
49394a2a
DA		 2326 }
2327
2328 return thread;
2329}
2330
52df138c 2331static bool timehist_skip_sample(struct perf_sched *sched,
a4b2b6f5
NK		 2332 struct thread *thread,
2333 struct perf_evsel *evsel,
2334 struct perf_sample *sample)
49394a2a
DA		 2335{
2336 bool rc = false;
2337
52df138c 2338 if (thread__is_filtered(thread)) {
49394a2a 2339 rc = true;
52df138c
DA		 2340 sched->skipped_samples++;
2341 }
49394a2a 2342
a4b2b6f5
NK		 2343 if (sched->idle_hist) {
2344 if (strcmp(perf_evsel__name(evsel), "sched:sched_switch"))
2345 rc = true;
2346 else if (perf_evsel__intval(evsel, sample, "prev_pid") != 0 &&
2347 perf_evsel__intval(evsel, sample, "next_pid") != 0)
2348 rc = true;
2349 }
2350
49394a2a
DA		 2351 return rc;
2352}
2353
fc1469f1 2354static void timehist_print_wakeup_event(struct perf_sched *sched,
a4b2b6f5 2355 struct perf_evsel *evsel,
fc1469f1
DA		 2356 struct perf_sample *sample,
2357 struct machine *machine,
2358 struct thread *awakened)
2359{
2360 struct thread *thread;
2361 char tstr[64];
2362
2363 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2364 if (thread == NULL)
2365 return;
2366
2367 /* show wakeup unless both awakee and awaker are filtered */
a4b2b6f5
NK		 2368 if (timehist_skip_sample(sched, thread, evsel, sample) &&
2369 timehist_skip_sample(sched, awakened, evsel, sample)) {
fc1469f1
DA		 2370 return;
2371 }
2372
2373 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2374 printf("%15s [%04d] ", tstr, sample->cpu);
a407b067
DA		 2375 if (sched->show_cpu_visual)
2376 printf(" %*s ", sched->max_cpu + 1, "");
fc1469f1
DA		 2377
2378 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2379
2380 /* dt spacer */
2381 printf(" %9s %9s %9s ", "", "", "");
2382
2383 printf("awakened: %s", timehist_get_commstr(awakened));
2384
2385 printf("\n");
2386}
2387
2388static int timehist_sched_wakeup_event(struct perf_tool *tool,
49394a2a
DA		 2389 union perf_event *event __maybe_unused,
2390 struct perf_evsel *evsel,
2391 struct perf_sample *sample,
2392 struct machine *machine)
2393{
fc1469f1 2394 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
49394a2a
DA		 2395 struct thread *thread;
2396 struct thread_runtime *tr = NULL;
2397 /* want pid of awakened task not pid in sample */
2398 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2399
2400 thread = machine__findnew_thread(machine, 0, pid);
2401 if (thread == NULL)
2402 return -1;
2403
2404 tr = thread__get_runtime(thread);
2405 if (tr == NULL)
2406 return -1;
2407
2408 if (tr->ready_to_run == 0)
2409 tr->ready_to_run = sample->time;
2410
fc1469f1 2411 /* show wakeups if requested */
853b7407
DA		 2412 if (sched->show_wakeups &&
2413 !perf_time__skip_sample(&sched->ptime, sample->time))
a4b2b6f5 2414 timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
fc1469f1 2415
49394a2a
DA		 2416 return 0;
2417}
2418
350f54fa
DA		 2419static void timehist_print_migration_event(struct perf_sched *sched,
2420 struct perf_evsel *evsel,
2421 struct perf_sample *sample,
2422 struct machine *machine,
2423 struct thread *migrated)
2424{
2425 struct thread *thread;
2426 char tstr[64];
2427 u32 max_cpus = sched->max_cpu + 1;
2428 u32 ocpu, dcpu;
2429
2430 if (sched->summary_only)
2431 return;
2432
2433 max_cpus = sched->max_cpu + 1;
2434 ocpu = perf_evsel__intval(evsel, sample, "orig_cpu");
2435 dcpu = perf_evsel__intval(evsel, sample, "dest_cpu");
2436
2437 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2438 if (thread == NULL)
2439 return;
2440
a4b2b6f5
NK		 2441 if (timehist_skip_sample(sched, thread, evsel, sample) &&
2442 timehist_skip_sample(sched, migrated, evsel, sample)) {
350f54fa
DA		 2443 return;
2444 }
2445
2446 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2447 printf("%15s [%04d] ", tstr, sample->cpu);
2448
2449 if (sched->show_cpu_visual) {
2450 u32 i;
2451 char c;
2452
2453 printf(" ");
2454 for (i = 0; i < max_cpus; ++i) {
2455 c = (i == sample->cpu) ? 'm' : ' ';
2456 printf("%c", c);
2457 }
2458 printf(" ");
2459 }
2460
2461 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2462
2463 /* dt spacer */
2464 printf(" %9s %9s %9s ", "", "", "");
2465
2466 printf("migrated: %s", timehist_get_commstr(migrated));
2467 printf(" cpu %d => %d", ocpu, dcpu);
2468
2469 printf("\n");
2470}
2471
2472static int timehist_migrate_task_event(struct perf_tool *tool,
2473 union perf_event *event __maybe_unused,
2474 struct perf_evsel *evsel,
2475 struct perf_sample *sample,
2476 struct machine *machine)
2477{
2478 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2479 struct thread *thread;
2480 struct thread_runtime *tr = NULL;
2481 /* want pid of migrated task not pid in sample */
2482 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2483
2484 thread = machine__findnew_thread(machine, 0, pid);
2485 if (thread == NULL)
2486 return -1;
2487
2488 tr = thread__get_runtime(thread);
2489 if (tr == NULL)
2490 return -1;
2491
2492 tr->migrations++;
2493
2494 /* show migrations if requested */
2495 timehist_print_migration_event(sched, evsel, sample, machine, thread);
2496
2497 return 0;
2498}
2499
52df138c 2500static int timehist_sched_change_event(struct perf_tool *tool,
49394a2a
DA		 2501 union perf_event *event,
2502 struct perf_evsel *evsel,
2503 struct perf_sample *sample,
2504 struct machine *machine)
2505{
fc1469f1 2506 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
853b7407 2507 struct perf_time_interval *ptime = &sched->ptime;
49394a2a
DA		 2508 struct addr_location al;
2509 struct thread *thread;
2510 struct thread_runtime *tr = NULL;
853b7407 2511 u64 tprev, t = sample->time;
49394a2a 2512 int rc = 0;
414e050c
NK		 2513 int state = perf_evsel__intval(evsel, sample, "prev_state");
2514
49394a2a
DA		 2515
2516 if (machine__resolve(machine, &al, sample) < 0) {
2517 pr_err("problem processing %d event. skipping it\n",
2518 event->header.type);
2519 rc = -1;
2520 goto out;
2521 }
2522
6c973c90 2523 thread = timehist_get_thread(sched, sample, machine, evsel);
49394a2a
DA		 2524 if (thread == NULL) {
2525 rc = -1;
2526 goto out;
2527 }
2528
a4b2b6f5 2529 if (timehist_skip_sample(sched, thread, evsel, sample))
49394a2a
DA		 2530 goto out;
2531
2532 tr = thread__get_runtime(thread);
2533 if (tr == NULL) {
2534 rc = -1;
2535 goto out;
2536 }
2537
2538 tprev = perf_evsel__get_time(evsel, sample->cpu);
2539
853b7407
DA		 2540 /*
2541 * If start time given:
2542 * - sample time is under window user cares about - skip sample
2543 * - tprev is under window user cares about - reset to start of window
2544 */
2545 if (ptime->start && ptime->start > t)
2546 goto out;
2547
bdd75729 2548 if (tprev && ptime->start > tprev)
853b7407
DA		 2549 tprev = ptime->start;
2550
2551 /*
2552 * If end time given:
2553 * - previous sched event is out of window - we are done
2554 * - sample time is beyond window user cares about - reset it
2555 * to close out stats for time window interest
2556 */
2557 if (ptime->end) {
2558 if (tprev > ptime->end)
2559 goto out;
2560
2561 if (t > ptime->end)
2562 t = ptime->end;
2563 }
2564
07235f84
NK		 2565 if (!sched->idle_hist || thread->tid == 0) {
2566 timehist_update_runtime_stats(tr, t, tprev);
2567
2568 if (sched->idle_hist) {
2569 struct idle_thread_runtime *itr = (void *)tr;
2570 struct thread_runtime *last_tr;
2571
2572 BUG_ON(thread->tid != 0);
2573
2574 if (itr->last_thread == NULL)
2575 goto out;
2576
2577 /* add current idle time as last thread's runtime */
2578 last_tr = thread__get_runtime(itr->last_thread);
2579 if (last_tr == NULL)
2580 goto out;
2581
2582 timehist_update_runtime_stats(last_tr, t, tprev);
2583 /*
2584 * remove delta time of last thread as it's not updated
2585 * and otherwise it will show an invalid value next
2586 * time. we only care total run time and run stat.
2587 */
2588 last_tr->dt_run = 0;
07235f84 2589 last_tr->dt_delay = 0;
941bdea7
NK		 2590 last_tr->dt_sleep = 0;
2591 last_tr->dt_iowait = 0;
2592 last_tr->dt_preempt = 0;
07235f84 2593
ba957ebb
NK		 2594 if (itr->cursor.nr)
2595 callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2596
07235f84
NK		 2597 itr->last_thread = NULL;
2598 }
2599 }
853b7407 2600
52df138c 2601 if (!sched->summary_only)
292c4a8f 2602 timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
49394a2a
DA		 2603
2604out:
9396c9cb
NK		 2605 if (sched->hist_time.start == 0 && t >= ptime->start)
2606 sched->hist_time.start = t;
2607 if (ptime->end == 0 || t <= ptime->end)
2608 sched->hist_time.end = t;
2609
49394a2a
DA		 2610 if (tr) {
2611 /* time of this sched_switch event becomes last time task seen */
2612 tr->last_time = sample->time;
2613
941bdea7 2614 /* last state is used to determine where to account wait time */
414e050c 2615 tr->last_state = state;
941bdea7 2616
49394a2a
DA		 2617 /* sched out event for task so reset ready to run time */
2618 tr->ready_to_run = 0;
2619 }
2620
2621 perf_evsel__save_time(evsel, sample->time, sample->cpu);
2622
2623 return rc;
2624}
2625
2626static int timehist_sched_switch_event(struct perf_tool *tool,
2627 union perf_event *event,
2628 struct perf_evsel *evsel,
2629 struct perf_sample *sample,
2630 struct machine *machine __maybe_unused)
2631{
2632 return timehist_sched_change_event(tool, event, evsel, sample, machine);
2633}
2634
2635static int process_lost(struct perf_tool *tool __maybe_unused,
2636 union perf_event *event,
2637 struct perf_sample *sample,
2638 struct machine *machine __maybe_unused)
2639{
2640 char tstr[64];
2641
2642 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2643 printf("%15s ", tstr);
2644 printf("lost %" PRIu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2645
2646 return 0;
2647}
2648
2649
52df138c
DA		 2650static void print_thread_runtime(struct thread *t,
2651 struct thread_runtime *r)
2652{
2653 double mean = avg_stats(&r->run_stats);
2654 float stddev;
2655
2656 printf("%*s %5d %9" PRIu64 " ",
2657 comm_width, timehist_get_commstr(t), t->ppid,
2658 (u64) r->run_stats.n);
2659
2660 print_sched_time(r->total_run_time, 8);
2661 stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2662 print_sched_time(r->run_stats.min, 6);
2663 printf(" ");
2664 print_sched_time((u64) mean, 6);
2665 printf(" ");
2666 print_sched_time(r->run_stats.max, 6);
2667 printf(" ");
2668 printf("%5.2f", stddev);
350f54fa 2669 printf(" %5" PRIu64, r->migrations);
52df138c
DA		 2670 printf("\n");
2671}
2672
587782c5
NK		 2673static void print_thread_waittime(struct thread *t,
2674 struct thread_runtime *r)
2675{
2676 printf("%*s %5d %9" PRIu64 " ",
2677 comm_width, timehist_get_commstr(t), t->ppid,
2678 (u64) r->run_stats.n);
2679
2680 print_sched_time(r->total_run_time, 8);
2681 print_sched_time(r->total_sleep_time, 6);
2682 printf(" ");
2683 print_sched_time(r->total_iowait_time, 6);
2684 printf(" ");
2685 print_sched_time(r->total_preempt_time, 6);
2686 printf(" ");
2687 print_sched_time(r->total_delay_time, 6);
2688 printf("\n");
2689}
2690
52df138c 2691struct total_run_stats {
587782c5 2692 struct perf_sched *sched;
52df138c
DA		 2693 u64 sched_count;
2694 u64 task_count;
2695 u64 total_run_time;
2696};
2697
2698static int __show_thread_runtime(struct thread *t, void *priv)
2699{
2700 struct total_run_stats *stats = priv;
2701 struct thread_runtime *r;
2702
2703 if (thread__is_filtered(t))
2704 return 0;
2705
2706 r = thread__priv(t);
2707 if (r && r->run_stats.n) {
2708 stats->task_count++;
2709 stats->sched_count += r->run_stats.n;
2710 stats->total_run_time += r->total_run_time;
587782c5
NK		 2711
2712 if (stats->sched->show_state)
2713 print_thread_waittime(t, r);
2714 else
2715 print_thread_runtime(t, r);
52df138c
DA		 2716 }
2717
2718 return 0;
2719}
2720
2721static int show_thread_runtime(struct thread *t, void *priv)
2722{
2723 if (t->dead)
2724 return 0;
2725
2726 return __show_thread_runtime(t, priv);
2727}
2728
2729static int show_deadthread_runtime(struct thread *t, void *priv)
2730{
2731 if (!t->dead)
2732 return 0;
2733
2734 return __show_thread_runtime(t, priv);
2735}
2736
ba957ebb
NK		 2737static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2738{
2739 const char *sep = " <- ";
2740 struct callchain_list *chain;
2741 size_t ret = 0;
2742 char bf[1024];
2743 bool first;
2744
2745 if (node == NULL)
2746 return 0;
2747
2748 ret = callchain__fprintf_folded(fp, node->parent);
2749 first = (ret == 0);
2750
2751 list_for_each_entry(chain, &node->val, list) {
2752 if (chain->ip >= PERF_CONTEXT_MAX)
2753 continue;
2754 if (chain->ms.sym && chain->ms.sym->ignore)
2755 continue;
2756 ret += fprintf(fp, "%s%s", first ? "" : sep,
2757 callchain_list__sym_name(chain, bf, sizeof(bf),
2758 false));
2759 first = false;
2760 }
2761
2762 return ret;
2763}
2764
2765static size_t timehist_print_idlehist_callchain(struct rb_root *root)
2766{
2767 size_t ret = 0;
2768 FILE *fp = stdout;
2769 struct callchain_node *chain;
2770 struct rb_node *rb_node = rb_first(root);
2771
2772 printf(" %16s %8s %s\n", "Idle time (msec)", "Count", "Callchains");
2773 printf(" %.16s %.8s %.50s\n", graph_dotted_line, graph_dotted_line,
2774 graph_dotted_line);
2775
2776 while (rb_node) {
2777 chain = rb_entry(rb_node, struct callchain_node, rb_node);
2778 rb_node = rb_next(rb_node);
2779
2780 ret += fprintf(fp, " ");
2781 print_sched_time(chain->hit, 12);
2782 ret += 16; /* print_sched_time returns 2nd arg + 4 */
2783 ret += fprintf(fp, " %8d ", chain->count);
2784 ret += callchain__fprintf_folded(fp, chain);
2785 ret += fprintf(fp, "\n");
2786 }
2787
2788 return ret;
2789}
2790
52df138c
DA		 2791static void timehist_print_summary(struct perf_sched *sched,
2792 struct perf_session *session)
2793{
2794 struct machine *m = &session->machines.host;
2795 struct total_run_stats totals;
2796 u64 task_count;
2797 struct thread *t;
2798 struct thread_runtime *r;
2799 int i;
9396c9cb 2800 u64 hist_time = sched->hist_time.end - sched->hist_time.start;
52df138c
DA		 2801
2802 memset(&totals, 0, sizeof(totals));
587782c5 2803 totals.sched = sched;
52df138c 2804
07235f84
NK		 2805 if (sched->idle_hist) {
2806 printf("\nIdle-time summary\n");
2807 printf("%*s parent sched-out ", comm_width, "comm");
2808 printf(" idle-time min-idle avg-idle max-idle stddev migrations\n");
587782c5
NK		 2809 } else if (sched->show_state) {
2810 printf("\nWait-time summary\n");
2811 printf("%*s parent sched-in ", comm_width, "comm");
2812 printf(" run-time sleep iowait preempt delay\n");
07235f84
NK		 2813 } else {
2814 printf("\nRuntime summary\n");
2815 printf("%*s parent sched-in ", comm_width, "comm");
2816 printf(" run-time min-run avg-run max-run stddev migrations\n");
2817 }
52df138c 2818 printf("%*s (count) ", comm_width, "");
587782c5
NK		 2819 printf(" (msec) (msec) (msec) (msec) %s\n",
2820 sched->show_state ? "(msec)" : "%");
350f54fa 2821 printf("%.117s\n", graph_dotted_line);
52df138c
DA		 2822
2823 machine__for_each_thread(m, show_thread_runtime, &totals);
2824 task_count = totals.task_count;
2825 if (!task_count)
2826 printf("<no still running tasks>\n");
2827
2828 printf("\nTerminated tasks:\n");
2829 machine__for_each_thread(m, show_deadthread_runtime, &totals);
2830 if (task_count == totals.task_count)
2831 printf("<no terminated tasks>\n");
2832
2833 /* CPU idle stats not tracked when samples were skipped */
07235f84 2834 if (sched->skipped_samples && !sched->idle_hist)
52df138c
DA		 2835 return;
2836
2837 printf("\nIdle stats:\n");
b336352b 2838 for (i = 0; i < idle_max_cpu; ++i) {
52df138c
DA		 2839 t = idle_threads[i];
2840 if (!t)
2841 continue;
2842
2843 r = thread__priv(t);
2844 if (r && r->run_stats.n) {
2845 totals.sched_count += r->run_stats.n;
2846 printf(" CPU %2d idle for ", i);
2847 print_sched_time(r->total_run_time, 6);
9396c9cb 2848 printf(" msec (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
52df138c
DA		 2849 } else
2850 printf(" CPU %2d idle entire time window\n", i);
2851 }
2852
4c50563d 2853 if (sched->idle_hist && sched->show_callchain) {
ba957ebb
NK		 2854 callchain_param.mode = CHAIN_FOLDED;
2855 callchain_param.value = CCVAL_PERIOD;
2856
2857 callchain_register_param(&callchain_param);
2858
2859 printf("\nIdle stats by callchain:\n");
2860 for (i = 0; i < idle_max_cpu; ++i) {
2861 struct idle_thread_runtime *itr;
2862
2863 t = idle_threads[i];
2864 if (!t)
2865 continue;
2866
2867 itr = thread__priv(t);
2868 if (itr == NULL)
2869 continue;
2870
2871 callchain_param.sort(&itr->sorted_root, &itr->callchain,
2872 0, &callchain_param);
2873
2874 printf(" CPU %2d:", i);
2875 print_sched_time(itr->tr.total_run_time, 6);
2876 printf(" msec\n");
2877 timehist_print_idlehist_callchain(&itr->sorted_root);
2878 printf("\n");
2879 }
2880 }
2881
52df138c
DA		 2882 printf("\n"
2883 " Total number of unique tasks: %" PRIu64 "\n"
9396c9cb 2884 "Total number of context switches: %" PRIu64 "\n",
52df138c
DA		 2885 totals.task_count, totals.sched_count);
2886
9396c9cb 2887 printf(" Total run time (msec): ");
52df138c
DA		 2888 print_sched_time(totals.total_run_time, 2);
2889 printf("\n");
9396c9cb
NK		 2890
2891 printf(" Total scheduling time (msec): ");
2892 print_sched_time(hist_time, 2);
2893 printf(" (x %d)\n", sched->max_cpu);
52df138c
DA		 2894}
2895
49394a2a
DA		 2896typedef int (*sched_handler)(struct perf_tool *tool,
2897 union perf_event *event,
2898 struct perf_evsel *evsel,
2899 struct perf_sample *sample,
2900 struct machine *machine);
2901
2902static int perf_timehist__process_sample(struct perf_tool *tool,
2903 union perf_event *event,
2904 struct perf_sample *sample,
2905 struct perf_evsel *evsel,
2906 struct machine *machine)
2907{
2908 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2909 int err = 0;
2910 int this_cpu = sample->cpu;
2911
2912 if (this_cpu > sched->max_cpu)
2913 sched->max_cpu = this_cpu;
2914
2915 if (evsel->handler != NULL) {
2916 sched_handler f = evsel->handler;
2917
2918 err = f(tool, event, evsel, sample, machine);
2919 }
2920
2921 return err;
2922}
2923
6c973c90
DA		 2924static int timehist_check_attr(struct perf_sched *sched,
2925 struct perf_evlist *evlist)
2926{
2927 struct perf_evsel *evsel;
2928 struct evsel_runtime *er;
2929
2930 list_for_each_entry(evsel, &evlist->entries, node) {
2931 er = perf_evsel__get_runtime(evsel);
2932 if (er == NULL) {
2933 pr_err("Failed to allocate memory for evsel runtime data\n");
2934 return -1;
2935 }
2936
27de9b2b 2937 if (sched->show_callchain && !evsel__has_callchain(evsel)) {
6c973c90
DA		 2938 pr_info("Samples do not have callchains.\n");
2939 sched->show_callchain = 0;
2940 symbol_conf.use_callchain = 0;
2941 }
2942 }
2943
2944 return 0;
2945}
2946
49394a2a
DA		 2947static int perf_sched__timehist(struct perf_sched *sched)
2948{
2949 const struct perf_evsel_str_handler handlers[] = {
2950 { "sched:sched_switch", timehist_sched_switch_event, },
2951 { "sched:sched_wakeup", timehist_sched_wakeup_event, },
2952 { "sched:sched_wakeup_new", timehist_sched_wakeup_event, },
2953 };
350f54fa
DA		 2954 const struct perf_evsel_str_handler migrate_handlers[] = {
2955 { "sched:sched_migrate_task", timehist_migrate_task_event, },
2956 };
8ceb41d7 2957 struct perf_data data = {
eae8ad80
JO		 2958 .file = {
2959 .path = input_name,
2960 },
2961 .mode = PERF_DATA_MODE_READ,
2962 .force = sched->force,
49394a2a
DA		 2963 };
2964
2965 struct perf_session *session;
52df138c 2966 struct perf_evlist *evlist;
49394a2a
DA		 2967 int err = -1;
2968
2969 /*
2970 * event handlers for timehist option
2971 */
2972 sched->tool.sample = perf_timehist__process_sample;
2973 sched->tool.mmap = perf_event__process_mmap;
2974 sched->tool.comm = perf_event__process_comm;
2975 sched->tool.exit = perf_event__process_exit;
2976 sched->tool.fork = perf_event__process_fork;
2977 sched->tool.lost = process_lost;
2978 sched->tool.attr = perf_event__process_attr;
2979 sched->tool.tracing_data = perf_event__process_tracing_data;
2980 sched->tool.build_id = perf_event__process_build_id;
2981
2982 sched->tool.ordered_events = true;
2983 sched->tool.ordering_requires_timestamps = true;
2984
6c973c90
DA		 2985 symbol_conf.use_callchain = sched->show_callchain;
2986
8ceb41d7 2987 session = perf_session__new(&data, false, &sched->tool);
49394a2a
DA		 2988 if (session == NULL)
2989 return -ENOMEM;
2990
52df138c
DA		 2991 evlist = session->evlist;
2992
49394a2a
DA		 2993 symbol__init(&session->header.env);
2994
853b7407
DA		 2995 if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
2996 pr_err("Invalid time string\n");
2997 return -EINVAL;
2998 }
2999
6c973c90
DA		 3000 if (timehist_check_attr(sched, evlist) != 0)
3001 goto out;
3002
49394a2a
DA		 3003 setup_pager();
3004
3005 /* setup per-evsel handlers */
3006 if (perf_session__set_tracepoints_handlers(session, handlers))
3007 goto out;
3008
f45bf8d3
DA		 3009 /* sched_switch event at a minimum needs to exist */
3010 if (!perf_evlist__find_tracepoint_by_name(session->evlist,
3011 "sched:sched_switch")) {
3012 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
49394a2a 3013 goto out;
f45bf8d3 3014 }
49394a2a 3015
350f54fa
DA		 3016 if (sched->show_migrations &&
3017 perf_session__set_tracepoints_handlers(session, migrate_handlers))
3018 goto out;
3019
49394a2a
DA		 3020 /* pre-allocate struct for per-CPU idle stats */
3021 sched->max_cpu = session->header.env.nr_cpus_online;
3022 if (sched->max_cpu == 0)
3023 sched->max_cpu = 4;
3024 if (init_idle_threads(sched->max_cpu))
3025 goto out;
3026
52df138c
DA		 3027 /* summary_only implies summary option, but don't overwrite summary if set */
3028 if (sched->summary_only)
3029 sched->summary = sched->summary_only;
3030
3031 if (!sched->summary_only)
a407b067 3032 timehist_header(sched);
49394a2a
DA		 3033
3034 err = perf_session__process_events(session);
3035 if (err) {
3036 pr_err("Failed to process events, error %d", err);
3037 goto out;
3038 }
3039
52df138c
DA		 3040 sched->nr_events = evlist->stats.nr_events[0];
3041 sched->nr_lost_events = evlist->stats.total_lost;
3042 sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3043
3044 if (sched->summary)
3045 timehist_print_summary(sched, session);
3046
49394a2a
DA		 3047out:
3048 free_idle_threads();
3049 perf_session__delete(session);
3050
3051 return err;
3052}
3053
3054
0e9b07e5 3055static void print_bad_events(struct perf_sched *sched)
0ec04e16 3056{
0e9b07e5 3057 if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
0ec04e16 3058 printf(" INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
0e9b07e5
ACM		 3059 (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3060 sched->nr_unordered_timestamps, sched->nr_timestamps);
0ec04e16 3061 }
0e9b07e5 3062 if (sched->nr_lost_events && sched->nr_events) {
0ec04e16 3063 printf(" INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
0e9b07e5
ACM		 3064 (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3065 sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
0ec04e16 3066 }
0e9b07e5 3067 if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
0ec04e16 3068 printf(" INFO: %.3f%% context switch bugs (%ld out of %ld)",
0e9b07e5
ACM		 3069 (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3070 sched->nr_context_switch_bugs, sched->nr_timestamps);
3071 if (sched->nr_lost_events)
0ec04e16
IM		 3072 printf(" (due to lost events?)");
3073 printf("\n");
3074 }
3075}
3076
2f80dd44
JB		 3077static void __merge_work_atoms(struct rb_root *root, struct work_atoms *data)
3078{
3079 struct rb_node **new = &(root->rb_node), *parent = NULL;
3080 struct work_atoms *this;
3081 const char *comm = thread__comm_str(data->thread), *this_comm;
3082
3083 while (*new) {
3084 int cmp;
3085
3086 this = container_of(*new, struct work_atoms, node);
3087 parent = *new;
3088
3089 this_comm = thread__comm_str(this->thread);
3090 cmp = strcmp(comm, this_comm);
3091 if (cmp > 0) {
3092 new = &((*new)->rb_left);
3093 } else if (cmp < 0) {
3094 new = &((*new)->rb_right);
3095 } else {
3096 this->num_merged++;
3097 this->total_runtime += data->total_runtime;
3098 this->nb_atoms += data->nb_atoms;
3099 this->total_lat += data->total_lat;
3100 list_splice(&data->work_list, &this->work_list);
3101 if (this->max_lat < data->max_lat) {
3102 this->max_lat = data->max_lat;
3103 this->max_lat_at = data->max_lat_at;
3104 }
3105 zfree(&data);
3106 return;
3107 }
3108 }
3109
3110 data->num_merged++;
3111 rb_link_node(&data->node, parent, new);
3112 rb_insert_color(&data->node, root);
3113}
3114
3115static void perf_sched__merge_lat(struct perf_sched *sched)
3116{
3117 struct work_atoms *data;
3118 struct rb_node *node;
3119
3120 if (sched->skip_merge)
3121 return;
3122
3123 while ((node = rb_first(&sched->atom_root))) {
3124 rb_erase(node, &sched->atom_root);
3125 data = rb_entry(node, struct work_atoms, node);
3126 __merge_work_atoms(&sched->merged_atom_root, data);
3127 }
3128}
3129
0e9b07e5 3130static int perf_sched__lat(struct perf_sched *sched)
0ec04e16
IM		 3131{
3132 struct rb_node *next;
3133
3134 setup_pager();
ad9def7c 3135
ae536acf 3136 if (perf_sched__read_events(sched))
a116e05d 3137 return -1;
ad9def7c 3138
2f80dd44 3139 perf_sched__merge_lat(sched);
0e9b07e5 3140 perf_sched__sort_lat(sched);
0ec04e16 3141
80790e0b
RR		 3142 printf("\n -----------------------------------------------------------------------------------------------------------------\n");
3143 printf(" Task | Runtime ms | Switches | Average delay ms | Maximum delay ms | Maximum delay at |\n");
3144 printf(" -----------------------------------------------------------------------------------------------------------------\n");
0ec04e16 3145
0e9b07e5 3146 next = rb_first(&sched->sorted_atom_root);
0ec04e16
IM		 3147
3148 while (next) {
3149 struct work_atoms *work_list;
3150
3151 work_list = rb_entry(next, struct work_atoms, node);
0e9b07e5 3152 output_lat_thread(sched, work_list);
0ec04e16 3153 next = rb_next(next);
ae536acf 3154 thread__zput(work_list->thread);
0ec04e16
IM		 3155 }
3156
80790e0b 3157 printf(" -----------------------------------------------------------------------------------------------------------------\n");
9486aa38 3158 printf(" TOTAL: |%11.3f ms |%9" PRIu64 " |\n",
4fc76e49 3159 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
0ec04e16
IM		 3160
3161 printf(" ---------------------------------------------------\n");
3162
0e9b07e5 3163 print_bad_events(sched);
0ec04e16
IM		 3164 printf("\n");
3165
a116e05d 3166 return 0;
0ec04e16
IM		 3167}
3168
99623c62
JO		 3169static int setup_map_cpus(struct perf_sched *sched)
3170{
73643bb6
JO		 3171 struct cpu_map *map;
3172
99623c62
JO		 3173 sched->max_cpu = sysconf(_SC_NPROCESSORS_CONF);
3174
3175 if (sched->map.comp) {
3176 sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
cf294f24
JO		 3177 if (!sched->map.comp_cpus)
3178 return -1;
99623c62
JO		 3179 }
3180
73643bb6
JO		 3181 if (!sched->map.cpus_str)
3182 return 0;
3183
3184 map = cpu_map__new(sched->map.cpus_str);
3185 if (!map) {
3186 pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3187 return -1;
3188 }
3189
3190 sched->map.cpus = map;
99623c62
JO		 3191 return 0;
3192}
3193
a151a37a
JO		 3194static int setup_color_pids(struct perf_sched *sched)
3195{
3196 struct thread_map *map;
3197
3198 if (!sched->map.color_pids_str)
3199 return 0;
3200
3201 map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3202 if (!map) {
3203 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3204 return -1;
3205 }
3206
3207 sched->map.color_pids = map;
3208 return 0;
3209}
3210
cf294f24
JO		 3211static int setup_color_cpus(struct perf_sched *sched)
3212{
3213 struct cpu_map *map;
3214
3215 if (!sched->map.color_cpus_str)
3216 return 0;
3217
3218 map = cpu_map__new(sched->map.color_cpus_str);
3219 if (!map) {
3220 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3221 return -1;
3222 }
3223
3224 sched->map.color_cpus = map;
3225 return 0;
3226}
3227
0e9b07e5 3228static int perf_sched__map(struct perf_sched *sched)
0ec04e16 3229{
99623c62
JO		 3230 if (setup_map_cpus(sched))
3231 return -1;
40749d0f 3232
a151a37a
JO		 3233 if (setup_color_pids(sched))
3234 return -1;
3235
cf294f24
JO		 3236 if (setup_color_cpus(sched))
3237 return -1;
3238
0ec04e16 3239 setup_pager();
ae536acf 3240 if (perf_sched__read_events(sched))
a116e05d 3241 return -1;
0e9b07e5 3242 print_bad_events(sched);
a116e05d 3243 return 0;
0ec04e16
IM		 3244}
3245
0e9b07e5 3246static int perf_sched__replay(struct perf_sched *sched)
0ec04e16
IM		 3247{
3248 unsigned long i;
3249
0e9b07e5
ACM		 3250 calibrate_run_measurement_overhead(sched);
3251 calibrate_sleep_measurement_overhead(sched);
0ec04e16 3252
0e9b07e5 3253 test_calibrations(sched);
0ec04e16 3254
ae536acf 3255 if (perf_sched__read_events(sched))
a116e05d 3256 return -1;
0ec04e16 3257
0e9b07e5
ACM		 3258 printf("nr_run_events: %ld\n", sched->nr_run_events);
3259 printf("nr_sleep_events: %ld\n", sched->nr_sleep_events);
3260 printf("nr_wakeup_events: %ld\n", sched->nr_wakeup_events);
0ec04e16 3261
0e9b07e5
ACM		 3262 if (sched->targetless_wakeups)
3263 printf("target-less wakeups: %ld\n", sched->targetless_wakeups);
3264 if (sched->multitarget_wakeups)
3265 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3266 if (sched->nr_run_events_optimized)
0ec04e16 3267 printf("run atoms optimized: %ld\n",
0e9b07e5 3268 sched->nr_run_events_optimized);
0ec04e16 3269
0e9b07e5
ACM		 3270 print_task_traces(sched);
3271 add_cross_task_wakeups(sched);
0ec04e16 3272
0e9b07e5 3273 create_tasks(sched);
0ec04e16 3274 printf("------------------------------------------------------------\n");
0e9b07e5
ACM		 3275 for (i = 0; i < sched->replay_repeat; i++)
3276 run_one_test(sched);
a116e05d
ACM		 3277
3278 return 0;
0ec04e16
IM		 3279}
3280
0e9b07e5
ACM		 3281static void setup_sorting(struct perf_sched *sched, const struct option *options,
3282 const char * const usage_msg[])
daa1d7a5 3283{
0e9b07e5 3284 char *tmp, *tok, *str = strdup(sched->sort_order);
daa1d7a5
FW		 3285
3286 for (tok = strtok_r(str, ", ", &tmp);
3287 tok; tok = strtok_r(NULL, ", ", &tmp)) {
0e9b07e5 3288 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
c7118369
NK		 3289 usage_with_options_msg(usage_msg, options,
3290 "Unknown --sort key: `%s'", tok);
daa1d7a5
FW		 3291 }
3292 }
3293
3294 free(str);
3295
0e9b07e5 3296 sort_dimension__add("pid", &sched->cmp_pid);
daa1d7a5
FW		 3297}
3298
1fc35b29
IM		 3299static int __cmd_record(int argc, const char **argv)
3300{
3301 unsigned int rec_argc, i, j;
3302 const char **rec_argv;
0e9b07e5
ACM		 3303 const char * const record_args[] = {
3304 "record",
3305 "-a",
3306 "-R",
0e9b07e5
ACM		 3307 "-m", "1024",
3308 "-c", "1",
3309 "-e", "sched:sched_switch",
3310 "-e", "sched:sched_stat_wait",
3311 "-e", "sched:sched_stat_sleep",
3312 "-e", "sched:sched_stat_iowait",
3313 "-e", "sched:sched_stat_runtime",
0e9b07e5
ACM		 3314 "-e", "sched:sched_process_fork",
3315 "-e", "sched:sched_wakeup",
7fff9597 3316 "-e", "sched:sched_wakeup_new",
0e9b07e5
ACM		 3317 "-e", "sched:sched_migrate_task",
3318 };
1fc35b29
IM		 3319
3320 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
3321 rec_argv = calloc(rec_argc + 1, sizeof(char *));
3322
e462dc55 3323 if (rec_argv == NULL)
ce47dc56
CS		 3324 return -ENOMEM;
3325
1fc35b29
IM		 3326 for (i = 0; i < ARRAY_SIZE(record_args); i++)
3327 rec_argv[i] = strdup(record_args[i]);
3328
3329 for (j = 1; j < (unsigned int)argc; j++, i++)
3330 rec_argv[i] = argv[j];
3331
3332 BUG_ON(i != rec_argc);
3333
b0ad8ea6 3334 return cmd_record(i, rec_argv);
1fc35b29
IM		 3335}
3336
b0ad8ea6 3337int cmd_sched(int argc, const char **argv)
0a02ad93 3338{
8a39df8f
AH		 3339 const char default_sort_order[] = "avg, max, switch, runtime";
3340 struct perf_sched sched = {
3341 .tool = {
3342 .sample = perf_sched__process_tracepoint_sample,
99a3c3a9 3343 .comm = perf_sched__process_comm,
f3b3614a 3344 .namespaces = perf_event__process_namespaces,
8a39df8f
AH		 3345 .lost = perf_event__process_lost,
3346 .fork = perf_sched__process_fork_event,
0a8cb85c 3347 .ordered_events = true,
8a39df8f
AH		 3348 },
3349 .cmp_pid = LIST_HEAD_INIT(sched.cmp_pid),
3350 .sort_list = LIST_HEAD_INIT(sched.sort_list),
3351 .start_work_mutex = PTHREAD_MUTEX_INITIALIZER,
3352 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
8a39df8f
AH		 3353 .sort_order = default_sort_order,
3354 .replay_repeat = 10,
3355 .profile_cpu = -1,
3356 .next_shortname1 = 'A',
3357 .next_shortname2 = '0',
2f80dd44 3358 .skip_merge = 0,
6c973c90
DA		 3359 .show_callchain = 1,
3360 .max_stack = 5,
8a39df8f 3361 };
77f02f44
NK		 3362 const struct option sched_options[] = {
3363 OPT_STRING('i', "input", &input_name, "file",
3364 "input file name"),
3365 OPT_INCR('v', "verbose", &verbose,
3366 "be more verbose (show symbol address, etc)"),
3367 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3368 "dump raw trace in ASCII"),
6fa94258 3369 OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
77f02f44
NK		 3370 OPT_END()
3371 };
0e9b07e5
ACM		 3372 const struct option latency_options[] = {
3373 OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3374 "sort by key(s): runtime, switch, avg, max"),
0e9b07e5
ACM		 3375 OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3376 "CPU to profile on"),
2f80dd44
JB		 3377 OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3378 "latency stats per pid instead of per comm"),
77f02f44 3379 OPT_PARENT(sched_options)
0e9b07e5
ACM		 3380 };
3381 const struct option replay_options[] = {
3382 OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3383 "repeat the workload replay N times (-1: infinite)"),
77f02f44 3384 OPT_PARENT(sched_options)
0e9b07e5 3385 };
99623c62
JO		 3386 const struct option map_options[] = {
3387 OPT_BOOLEAN(0, "compact", &sched.map.comp,
3388 "map output in compact mode"),
a151a37a
JO		 3389 OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3390 "highlight given pids in map"),
cf294f24
JO		 3391 OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3392 "highlight given CPUs in map"),
73643bb6
JO		 3393 OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3394 "display given CPUs in map"),
77f02f44 3395 OPT_PARENT(sched_options)
99623c62 3396 };
49394a2a
DA		 3397 const struct option timehist_options[] = {
3398 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3399 "file", "vmlinux pathname"),
3400 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3401 "file", "kallsyms pathname"),
6c973c90
DA		 3402 OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3403 "Display call chains if present (default on)"),
3404 OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3405 "Maximum number of functions to display backtrace."),
49394a2a
DA		 3406 OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3407 "Look for files with symbols relative to this directory"),
52df138c
DA		 3408 OPT_BOOLEAN('s', "summary", &sched.summary_only,
3409 "Show only syscall summary with statistics"),
3410 OPT_BOOLEAN('S', "with-summary", &sched.summary,
3411 "Show all syscalls and summary with statistics"),
fc1469f1 3412 OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
292c4a8f 3413 OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
350f54fa 3414 OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
a407b067 3415 OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
07235f84 3416 OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
853b7407
DA		 3417 OPT_STRING(0, "time", &sched.time_str, "str",
3418 "Time span for analysis (start,stop)"),
414e050c 3419 OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
0f59d7a3
DA		 3420 OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3421 "analyze events only for given process id(s)"),
3422 OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3423 "analyze events only for given thread id(s)"),
49394a2a
DA		 3424 OPT_PARENT(sched_options)
3425 };
3426
0e9b07e5
ACM		 3427 const char * const latency_usage[] = {
3428 "perf sched latency [<options>]",
3429 NULL
3430 };
3431 const char * const replay_usage[] = {
3432 "perf sched replay [<options>]",
3433 NULL
3434 };
99623c62
JO		 3435 const char * const map_usage[] = {
3436 "perf sched map [<options>]",
3437 NULL
3438 };
49394a2a
DA		 3439 const char * const timehist_usage[] = {
3440 "perf sched timehist [<options>]",
3441 NULL
3442 };
a83edb2d 3443 const char *const sched_subcommands[] = { "record", "latency", "map",
49394a2a
DA		 3444 "replay", "script",
3445 "timehist", NULL };
a83edb2d
RR		 3446 const char *sched_usage[] = {
3447 NULL,
0e9b07e5
ACM		 3448 NULL
3449 };
3450 struct trace_sched_handler lat_ops = {
3451 .wakeup_event = latency_wakeup_event,
3452 .switch_event = latency_switch_event,
3453 .runtime_event = latency_runtime_event,
0e9b07e5
ACM		 3454 .migrate_task_event = latency_migrate_task_event,
3455 };
3456 struct trace_sched_handler map_ops = {
3457 .switch_event = map_switch_event,
3458 };
3459 struct trace_sched_handler replay_ops = {
3460 .wakeup_event = replay_wakeup_event,
3461 .switch_event = replay_switch_event,
3462 .fork_event = replay_fork_event,
3463 };
156a2b02
AH		 3464 unsigned int i;
3465
3466 for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3467 sched.curr_pid[i] = -1;
0e9b07e5 3468
a83edb2d
RR		 3469 argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3470 sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
f2858d8a
IM		 3471 if (!argc)
3472 usage_with_options(sched_usage, sched_options);
0a02ad93 3473
c0777c5a 3474 /*
133dc4c3 3475 * Aliased to 'perf script' for now:
c0777c5a 3476 */
133dc4c3 3477 if (!strcmp(argv[0], "script"))
b0ad8ea6 3478 return cmd_script(argc, argv);
c0777c5a 3479
1fc35b29
IM		 3480 if (!strncmp(argv[0], "rec", 3)) {
3481 return __cmd_record(argc, argv);
3482 } else if (!strncmp(argv[0], "lat", 3)) {
0e9b07e5 3483 sched.tp_handler = &lat_ops;
f2858d8a
IM		 3484 if (argc > 1) {
3485 argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3486 if (argc)
3487 usage_with_options(latency_usage, latency_options);
f2858d8a 3488 }
0e9b07e5
ACM		 3489 setup_sorting(&sched, latency_options, latency_usage);
3490 return perf_sched__lat(&sched);
0ec04e16 3491 } else if (!strcmp(argv[0], "map")) {
99623c62 3492 if (argc) {
a151a37a 3493 argc = parse_options(argc, argv, map_options, map_usage, 0);
99623c62
JO		 3494 if (argc)
3495 usage_with_options(map_usage, map_options);
3496 }
0e9b07e5
ACM		 3497 sched.tp_handler = &map_ops;
3498 setup_sorting(&sched, latency_options, latency_usage);
3499 return perf_sched__map(&sched);
f2858d8a 3500 } else if (!strncmp(argv[0], "rep", 3)) {
0e9b07e5 3501 sched.tp_handler = &replay_ops;
f2858d8a
IM		 3502 if (argc) {
3503 argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3504 if (argc)
3505 usage_with_options(replay_usage, replay_options);
3506 }
0e9b07e5 3507 return perf_sched__replay(&sched);
49394a2a
DA		 3508 } else if (!strcmp(argv[0], "timehist")) {
3509 if (argc) {
3510 argc = parse_options(argc, argv, timehist_options,
3511 timehist_usage, 0);
3512 if (argc)
3513 usage_with_options(timehist_usage, timehist_options);
3514 }
292c4a8f
BG		 3515 if ((sched.show_wakeups || sched.show_next) &&
3516 sched.summary_only) {
3517 pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
fc1469f1 3518 parse_options_usage(timehist_usage, timehist_options, "s", true);
292c4a8f
BG		 3519 if (sched.show_wakeups)
3520 parse_options_usage(NULL, timehist_options, "w", true);
3521 if (sched.show_next)
3522 parse_options_usage(NULL, timehist_options, "n", true);
fc1469f1
DA		 3523 return -EINVAL;
3524 }
3525
49394a2a 3526 return perf_sched__timehist(&sched);
f2858d8a
IM		 3527 } else {
3528 usage_with_options(sched_usage, sched_options);
3529 }
3530
ec156764 3531 return 0;
0a02ad93 3532}
Linux 6.x block layer and io_uring tree(s)