projects / linux-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
License cleanup: add SPDX GPL-2.0 license identifier to files with no license
[linux-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;
350f54fa 257 u64 migrations;
49394a2a
DA		 258};
259
260/* per event run time data */
261struct evsel_runtime {
262 u64 *last_time; /* time this event was last seen per cpu */
263 u32 ncpu; /* highest cpu slot allocated */
264};
265
3bc2fa9c
NK		 266/* per cpu idle time data */
267struct idle_thread_runtime {
268 struct thread_runtime tr;
269 struct thread *last_thread;
270 struct rb_root sorted_root;
271 struct callchain_root callchain;
272 struct callchain_cursor cursor;
273};
274
49394a2a
DA		 275/* track idle times per cpu */
276static struct thread **idle_threads;
277static int idle_max_cpu;
278static char idle_comm[] = "<idle>";
279
b1ffe8f3 280static u64 get_nsecs(void)
ec156764
IM		 281{
282 struct timespec ts;
283
284 clock_gettime(CLOCK_MONOTONIC, &ts);
285
4fc76e49 286 return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
ec156764
IM		 287}
288
0e9b07e5 289static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
ec156764 290{
b1ffe8f3 291 u64 T0 = get_nsecs(), T1;
ec156764
IM		 292
293 do {
294 T1 = get_nsecs();
0e9b07e5 295 } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
ec156764
IM		 296}
297
b1ffe8f3 298static void sleep_nsecs(u64 nsecs)
ec156764
IM		 299{
300 struct timespec ts;
301
302 ts.tv_nsec = nsecs % 999999999;
303 ts.tv_sec = nsecs / 999999999;
304
305 nanosleep(&ts, NULL);
306}
307
0e9b07e5 308static void calibrate_run_measurement_overhead(struct perf_sched *sched)
ec156764 309{
4fc76e49 310 u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
ec156764
IM		 311 int i;
312
313 for (i = 0; i < 10; i++) {
314 T0 = get_nsecs();
0e9b07e5 315 burn_nsecs(sched, 0);
ec156764
IM		 316 T1 = get_nsecs();
317 delta = T1-T0;
318 min_delta = min(min_delta, delta);
319 }
0e9b07e5 320 sched->run_measurement_overhead = min_delta;
ec156764 321
9486aa38 322 printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
ec156764
IM		 323}
324
0e9b07e5 325static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
ec156764 326{
4fc76e49 327 u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
ec156764
IM		 328 int i;
329
330 for (i = 0; i < 10; i++) {
331 T0 = get_nsecs();
332 sleep_nsecs(10000);
333 T1 = get_nsecs();
334 delta = T1-T0;
335 min_delta = min(min_delta, delta);
336 }
337 min_delta -= 10000;
0e9b07e5 338 sched->sleep_measurement_overhead = min_delta;
ec156764 339
9486aa38 340 printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
ec156764
IM		 341}
342
39aeb52f 343static struct sched_atom *
b1ffe8f3 344get_new_event(struct task_desc *task, u64 timestamp)
ec156764 345{
36479484 346 struct sched_atom *event = zalloc(sizeof(*event));
ec156764
IM		 347 unsigned long idx = task->nr_events;
348 size_t size;
349
350 event->timestamp = timestamp;
351 event->nr = idx;
352
353 task->nr_events++;
39aeb52f 354 size = sizeof(struct sched_atom *) * task->nr_events;
355 task->atoms = realloc(task->atoms, size);
356 BUG_ON(!task->atoms);
ec156764 357
39aeb52f 358 task->atoms[idx] = event;
ec156764
IM		 359
360 return event;
361}
362
39aeb52f 363static struct sched_atom *last_event(struct task_desc *task)
ec156764
IM		 364{
365 if (!task->nr_events)
366 return NULL;
367
39aeb52f 368 return task->atoms[task->nr_events - 1];
ec156764
IM		 369}
370
0e9b07e5
ACM		 371static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
372 u64 timestamp, u64 duration)
ec156764 373{
39aeb52f 374 struct sched_atom *event, *curr_event = last_event(task);
ec156764
IM		 375
376 /*
fbf94829
IM		 377 * optimize an existing RUN event by merging this one
378 * to it:
379 */
ec156764 380 if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
0e9b07e5 381 sched->nr_run_events_optimized++;
ec156764
IM		 382 curr_event->duration += duration;
383 return;
384 }
385
386 event = get_new_event(task, timestamp);
387
388 event->type = SCHED_EVENT_RUN;
389 event->duration = duration;
390
0e9b07e5 391 sched->nr_run_events++;
ec156764
IM		 392}
393
0e9b07e5
ACM		 394static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
395 u64 timestamp, struct task_desc *wakee)
ec156764 396{
39aeb52f 397 struct sched_atom *event, *wakee_event;
ec156764
IM		 398
399 event = get_new_event(task, timestamp);
400 event->type = SCHED_EVENT_WAKEUP;
401 event->wakee = wakee;
402
403 wakee_event = last_event(wakee);
404 if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
0e9b07e5 405 sched->targetless_wakeups++;
ec156764
IM		 406 return;
407 }
408 if (wakee_event->wait_sem) {
0e9b07e5 409 sched->multitarget_wakeups++;
ec156764
IM		 410 return;
411 }
412
36479484 413 wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
ec156764
IM		 414 sem_init(wakee_event->wait_sem, 0, 0);
415 wakee_event->specific_wait = 1;
416 event->wait_sem = wakee_event->wait_sem;
417
0e9b07e5 418 sched->nr_wakeup_events++;
ec156764
IM		 419}
420
0e9b07e5
ACM		 421static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
422 u64 timestamp, u64 task_state __maybe_unused)
ec156764 423{
39aeb52f 424 struct sched_atom *event = get_new_event(task, timestamp);
ec156764
IM		 425
426 event->type = SCHED_EVENT_SLEEP;
427
0e9b07e5 428 sched->nr_sleep_events++;
ec156764
IM		 429}
430
0e9b07e5
ACM		 431static struct task_desc *register_pid(struct perf_sched *sched,
432 unsigned long pid, const char *comm)
ec156764
IM		 433{
434 struct task_desc *task;
cb06ac25 435 static int pid_max;
ec156764 436
cb06ac25
YS		 437 if (sched->pid_to_task == NULL) {
438 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
439 pid_max = MAX_PID;
440 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
441 }
3a423a5c
YS		 442 if (pid >= (unsigned long)pid_max) {
443 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
444 sizeof(struct task_desc *))) == NULL);
445 while (pid >= (unsigned long)pid_max)
446 sched->pid_to_task[pid_max++] = NULL;
447 }
ec156764 448
0e9b07e5 449 task = sched->pid_to_task[pid];
ec156764
IM		 450
451 if (task)
452 return task;
453
36479484 454 task = zalloc(sizeof(*task));
ec156764 455 task->pid = pid;
0e9b07e5 456 task->nr = sched->nr_tasks;
ec156764
IM		 457 strcpy(task->comm, comm);
458 /*
459 * every task starts in sleeping state - this gets ignored
460 * if there's no wakeup pointing to this sleep state:
461 */
0e9b07e5 462 add_sched_event_sleep(sched, task, 0, 0);
ec156764 463
0e9b07e5
ACM		 464 sched->pid_to_task[pid] = task;
465 sched->nr_tasks++;
0755bc4d 466 sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
0e9b07e5
ACM		 467 BUG_ON(!sched->tasks);
468 sched->tasks[task->nr] = task;
ec156764 469
bb963e16 470 if (verbose > 0)
0e9b07e5 471 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
ec156764
IM		 472
473 return task;
474}
475
476
0e9b07e5 477static void print_task_traces(struct perf_sched *sched)
ec156764
IM		 478{
479 struct task_desc *task;
480 unsigned long i;
481
0e9b07e5
ACM		 482 for (i = 0; i < sched->nr_tasks; i++) {
483 task = sched->tasks[i];
ad236fd2 484 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
ec156764
IM		 485 task->nr, task->comm, task->pid, task->nr_events);
486 }
487}
488
0e9b07e5 489static void add_cross_task_wakeups(struct perf_sched *sched)
ec156764
IM		 490{
491 struct task_desc *task1, *task2;
492 unsigned long i, j;
493
0e9b07e5
ACM		 494 for (i = 0; i < sched->nr_tasks; i++) {
495 task1 = sched->tasks[i];
ec156764 496 j = i + 1;
0e9b07e5 497 if (j == sched->nr_tasks)
ec156764 498 j = 0;
0e9b07e5
ACM		 499 task2 = sched->tasks[j];
500 add_sched_event_wakeup(sched, task1, 0, task2);
ec156764
IM		 501 }
502}
503
0e9b07e5
ACM		 504static void perf_sched__process_event(struct perf_sched *sched,
505 struct sched_atom *atom)
ec156764
IM		 506{
507 int ret = 0;
ec156764 508
39aeb52f 509 switch (atom->type) {
ec156764 510 case SCHED_EVENT_RUN:
0e9b07e5 511 burn_nsecs(sched, atom->duration);
ec156764
IM		 512 break;
513 case SCHED_EVENT_SLEEP:
39aeb52f 514 if (atom->wait_sem)
515 ret = sem_wait(atom->wait_sem);
ec156764
IM		 516 BUG_ON(ret);
517 break;
518 case SCHED_EVENT_WAKEUP:
39aeb52f 519 if (atom->wait_sem)
520 ret = sem_post(atom->wait_sem);
ec156764
IM		 521 BUG_ON(ret);
522 break;
55ffb7a6
MG		 523 case SCHED_EVENT_MIGRATION:
524 break;
ec156764
IM		 525 default:
526 BUG_ON(1);
527 }
528}
529
b1ffe8f3 530static u64 get_cpu_usage_nsec_parent(void)
ec156764
IM		 531{
532 struct rusage ru;
b1ffe8f3 533 u64 sum;
ec156764
IM		 534 int err;
535
536 err = getrusage(RUSAGE_SELF, &ru);
537 BUG_ON(err);
538
4fc76e49
ACM		 539 sum = ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
540 sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
ec156764
IM		 541
542 return sum;
543}
544
939cda52 545static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
ec156764 546{
c0c9e721 547 struct perf_event_attr attr;
939cda52 548 char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
c0c9e721 549 int fd;
939cda52
YS		 550 struct rlimit limit;
551 bool need_privilege = false;
ec156764 552
c0c9e721 553 memset(&attr, 0, sizeof(attr));
ec156764 554
c0c9e721
XG		 555 attr.type = PERF_TYPE_SOFTWARE;
556 attr.config = PERF_COUNT_SW_TASK_CLOCK;
ec156764 557
939cda52 558force_again:
57480d2c
YD		 559 fd = sys_perf_event_open(&attr, 0, -1, -1,
560 perf_event_open_cloexec_flag());
c0c9e721 561
1aff59be 562 if (fd < 0) {
939cda52
YS		 563 if (errno == EMFILE) {
564 if (sched->force) {
565 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
566 limit.rlim_cur += sched->nr_tasks - cur_task;
567 if (limit.rlim_cur > limit.rlim_max) {
568 limit.rlim_max = limit.rlim_cur;
569 need_privilege = true;
570 }
571 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
572 if (need_privilege && errno == EPERM)
573 strcpy(info, "Need privilege\n");
574 } else
575 goto force_again;
576 } else
577 strcpy(info, "Have a try with -f option\n");
578 }
60b7d14a 579 pr_err("Error: sys_perf_event_open() syscall returned "
939cda52 580 "with %d (%s)\n%s", fd,
c8b5f2c9 581 str_error_r(errno, sbuf, sizeof(sbuf)), info);
1aff59be
YS		 582 exit(EXIT_FAILURE);
583 }
c0c9e721
XG		 584 return fd;
585}
586
587static u64 get_cpu_usage_nsec_self(int fd)
588{
589 u64 runtime;
590 int ret;
591
592 ret = read(fd, &runtime, sizeof(runtime));
593 BUG_ON(ret != sizeof(runtime));
594
595 return runtime;
ec156764
IM		 596}
597
0e9b07e5
ACM		 598struct sched_thread_parms {
599 struct task_desc *task;
600 struct perf_sched *sched;
08097abc 601 int fd;
0e9b07e5
ACM		 602};
603
ec156764
IM		 604static void *thread_func(void *ctx)
605{
0e9b07e5
ACM		 606 struct sched_thread_parms *parms = ctx;
607 struct task_desc *this_task = parms->task;
608 struct perf_sched *sched = parms->sched;
b1ffe8f3 609 u64 cpu_usage_0, cpu_usage_1;
ec156764
IM		 610 unsigned long i, ret;
611 char comm2[22];
08097abc 612 int fd = parms->fd;
ec156764 613
74cf249d 614 zfree(&parms);
0e9b07e5 615
ec156764
IM		 616 sprintf(comm2, ":%s", this_task->comm);
617 prctl(PR_SET_NAME, comm2);
a116e05d
ACM		 618 if (fd < 0)
619 return NULL;
ec156764
IM		 620again:
621 ret = sem_post(&this_task->ready_for_work);
622 BUG_ON(ret);
0e9b07e5 623 ret = pthread_mutex_lock(&sched->start_work_mutex);
ec156764 624 BUG_ON(ret);
0e9b07e5 625 ret = pthread_mutex_unlock(&sched->start_work_mutex);
ec156764 626 BUG_ON(ret);
ec156764 627
c0c9e721 628 cpu_usage_0 = get_cpu_usage_nsec_self(fd);
ec156764
IM		 629
630 for (i = 0; i < this_task->nr_events; i++) {
631 this_task->curr_event = i;
0e9b07e5 632 perf_sched__process_event(sched, this_task->atoms[i]);
ec156764
IM		 633 }
634
c0c9e721 635 cpu_usage_1 = get_cpu_usage_nsec_self(fd);
ec156764 636 this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
ec156764
IM		 637 ret = sem_post(&this_task->work_done_sem);
638 BUG_ON(ret);
ec156764 639
0e9b07e5 640 ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
ec156764 641 BUG_ON(ret);
0e9b07e5 642 ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
ec156764 643 BUG_ON(ret);
ec156764
IM		 644
645 goto again;
646}
647
0e9b07e5 648static void create_tasks(struct perf_sched *sched)
ec156764
IM		 649{
650 struct task_desc *task;
651 pthread_attr_t attr;
652 unsigned long i;
653 int err;
654
655 err = pthread_attr_init(&attr);
656 BUG_ON(err);
12f7e036
JP		 657 err = pthread_attr_setstacksize(&attr,
658 (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
ec156764 659 BUG_ON(err);
0e9b07e5 660 err = pthread_mutex_lock(&sched->start_work_mutex);
ec156764 661 BUG_ON(err);
0e9b07e5 662 err = pthread_mutex_lock(&sched->work_done_wait_mutex);
ec156764 663 BUG_ON(err);
0e9b07e5
ACM		 664 for (i = 0; i < sched->nr_tasks; i++) {
665 struct sched_thread_parms *parms = malloc(sizeof(*parms));
666 BUG_ON(parms == NULL);
667 parms->task = task = sched->tasks[i];
668 parms->sched = sched;
939cda52 669 parms->fd = self_open_counters(sched, i);
ec156764
IM		 670 sem_init(&task->sleep_sem, 0, 0);
671 sem_init(&task->ready_for_work, 0, 0);
672 sem_init(&task->work_done_sem, 0, 0);
673 task->curr_event = 0;
0e9b07e5 674 err = pthread_create(&task->thread, &attr, thread_func, parms);
ec156764
IM		 675 BUG_ON(err);
676 }
677}
678
0e9b07e5 679static void wait_for_tasks(struct perf_sched *sched)
ec156764 680{
b1ffe8f3 681 u64 cpu_usage_0, cpu_usage_1;
ec156764
IM		 682 struct task_desc *task;
683 unsigned long i, ret;
684
0e9b07e5
ACM		 685 sched->start_time = get_nsecs();
686 sched->cpu_usage = 0;
687 pthread_mutex_unlock(&sched->work_done_wait_mutex);
ec156764 688
0e9b07e5
ACM		 689 for (i = 0; i < sched->nr_tasks; i++) {
690 task = sched->tasks[i];
ec156764
IM		 691 ret = sem_wait(&task->ready_for_work);
692 BUG_ON(ret);
693 sem_init(&task->ready_for_work, 0, 0);
694 }
0e9b07e5 695 ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
ec156764
IM		 696 BUG_ON(ret);
697
698 cpu_usage_0 = get_cpu_usage_nsec_parent();
699
0e9b07e5 700 pthread_mutex_unlock(&sched->start_work_mutex);
ec156764 701
0e9b07e5
ACM		 702 for (i = 0; i < sched->nr_tasks; i++) {
703 task = sched->tasks[i];
ec156764
IM		 704 ret = sem_wait(&task->work_done_sem);
705 BUG_ON(ret);
706 sem_init(&task->work_done_sem, 0, 0);
0e9b07e5 707 sched->cpu_usage += task->cpu_usage;
ec156764
IM		 708 task->cpu_usage = 0;
709 }
710
711 cpu_usage_1 = get_cpu_usage_nsec_parent();
0e9b07e5
ACM		 712 if (!sched->runavg_cpu_usage)
713 sched->runavg_cpu_usage = sched->cpu_usage;
ff5f3bbd 714 sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
ec156764 715
0e9b07e5
ACM		 716 sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
717 if (!sched->runavg_parent_cpu_usage)
718 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
ff5f3bbd
YS		 719 sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
720 sched->parent_cpu_usage)/sched->replay_repeat;
ec156764 721
0e9b07e5 722 ret = pthread_mutex_lock(&sched->start_work_mutex);
ec156764
IM		 723 BUG_ON(ret);
724
0e9b07e5
ACM		 725 for (i = 0; i < sched->nr_tasks; i++) {
726 task = sched->tasks[i];
ec156764
IM		 727 sem_init(&task->sleep_sem, 0, 0);
728 task->curr_event = 0;
729 }
730}
731
0e9b07e5 732static void run_one_test(struct perf_sched *sched)
ec156764 733{
fb7d0b3c 734 u64 T0, T1, delta, avg_delta, fluct;
ec156764
IM		 735
736 T0 = get_nsecs();
0e9b07e5 737 wait_for_tasks(sched);
ec156764
IM		 738 T1 = get_nsecs();
739
740 delta = T1 - T0;
0e9b07e5
ACM		 741 sched->sum_runtime += delta;
742 sched->nr_runs++;
ec156764 743
0e9b07e5 744 avg_delta = sched->sum_runtime / sched->nr_runs;
ec156764
IM		 745 if (delta < avg_delta)
746 fluct = avg_delta - delta;
747 else
748 fluct = delta - avg_delta;
0e9b07e5
ACM		 749 sched->sum_fluct += fluct;
750 if (!sched->run_avg)
751 sched->run_avg = delta;
ff5f3bbd 752 sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
ec156764 753
4fc76e49 754 printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
ec156764 755
4fc76e49 756 printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
ec156764 757
ad236fd2 758 printf("cpu: %0.2f / %0.2f",
4fc76e49 759 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
ec156764
IM		 760
761#if 0
762 /*
fbf94829 763 * rusage statistics done by the parent, these are less
0e9b07e5 764 * accurate than the sched->sum_exec_runtime based statistics:
fbf94829 765 */
ad236fd2 766 printf(" [%0.2f / %0.2f]",
4fc76e49
ACM		 767 (double)sched->parent_cpu_usage / NSEC_PER_MSEC,
768 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
ec156764
IM		 769#endif
770
ad236fd2 771 printf("\n");
ec156764 772
0e9b07e5
ACM		 773 if (sched->nr_sleep_corrections)
774 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
775 sched->nr_sleep_corrections = 0;
ec156764
IM		 776}
777
0e9b07e5 778static void test_calibrations(struct perf_sched *sched)
ec156764 779{
b1ffe8f3 780 u64 T0, T1;
ec156764
IM		 781
782 T0 = get_nsecs();
4fc76e49 783 burn_nsecs(sched, NSEC_PER_MSEC);
ec156764
IM		 784 T1 = get_nsecs();
785
9486aa38 786 printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
ec156764
IM		 787
788 T0 = get_nsecs();
4fc76e49 789 sleep_nsecs(NSEC_PER_MSEC);
ec156764
IM		 790 T1 = get_nsecs();
791
9486aa38 792 printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
ec156764
IM		 793}
794
a116e05d 795static int
0e9b07e5 796replay_wakeup_event(struct perf_sched *sched,
9ec3f4e4
ACM		 797 struct perf_evsel *evsel, struct perf_sample *sample,
798 struct machine *machine __maybe_unused)
419ab0d6 799{
9ec3f4e4
ACM		 800 const char *comm = perf_evsel__strval(evsel, sample, "comm");
801 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
419ab0d6 802 struct task_desc *waker, *wakee;
fbf94829 803
bb963e16 804 if (verbose > 0) {
2b7fcbc5 805 printf("sched_wakeup event %p\n", evsel);
fbf94829 806
9ec3f4e4 807 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
ad236fd2 808 }
fbf94829 809
2b7fcbc5 810 waker = register_pid(sched, sample->tid, "<unknown>");
9ec3f4e4 811 wakee = register_pid(sched, pid, comm);
fbf94829 812
0e9b07e5 813 add_sched_event_wakeup(sched, waker, sample->time, wakee);
a116e05d 814 return 0;
ec156764
IM		 815}
816
9ec3f4e4
ACM		 817static int replay_switch_event(struct perf_sched *sched,
818 struct perf_evsel *evsel,
819 struct perf_sample *sample,
820 struct machine *machine __maybe_unused)
ec156764 821{
9ec3f4e4
ACM		 822 const char *prev_comm = perf_evsel__strval(evsel, sample, "prev_comm"),
823 *next_comm = perf_evsel__strval(evsel, sample, "next_comm");
824 const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
825 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
826 const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
1d037ca1 827 struct task_desc *prev, __maybe_unused *next;
7f7f8d0b
ACM		 828 u64 timestamp0, timestamp = sample->time;
829 int cpu = sample->cpu;
fbf94829
IM		 830 s64 delta;
831
bb963e16 832 if (verbose > 0)
2b7fcbc5 833 printf("sched_switch event %p\n", evsel);
ad236fd2 834
fbf94829 835 if (cpu >= MAX_CPUS || cpu < 0)
a116e05d 836 return 0;
fbf94829 837
0e9b07e5 838 timestamp0 = sched->cpu_last_switched[cpu];
fbf94829
IM		 839 if (timestamp0)
840 delta = timestamp - timestamp0;
841 else
842 delta = 0;
843
a116e05d 844 if (delta < 0) {
60b7d14a 845 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
a116e05d
ACM		 846 return -1;
847 }
fbf94829 848
9ec3f4e4
ACM		 849 pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
850 prev_comm, prev_pid, next_comm, next_pid, delta);
fbf94829 851
9ec3f4e4
ACM		 852 prev = register_pid(sched, prev_pid, prev_comm);
853 next = register_pid(sched, next_pid, next_comm);
fbf94829 854
0e9b07e5 855 sched->cpu_last_switched[cpu] = timestamp;
fbf94829 856
0e9b07e5 857 add_sched_event_run(sched, prev, timestamp, delta);
9ec3f4e4 858 add_sched_event_sleep(sched, prev, timestamp, prev_state);
a116e05d
ACM		 859
860 return 0;
fbf94829
IM		 861}
862
cb627505
DA		 863static int replay_fork_event(struct perf_sched *sched,
864 union perf_event *event,
865 struct machine *machine)
419ab0d6 866{
cb627505
DA		 867 struct thread *child, *parent;
868
314add6b
AH		 869 child = machine__findnew_thread(machine, event->fork.pid,
870 event->fork.tid);
871 parent = machine__findnew_thread(machine, event->fork.ppid,
872 event->fork.ptid);
cb627505
DA		 873
874 if (child == NULL || parent == NULL) {
875 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
876 child, parent);
b91fc39f 877 goto out_put;
cb627505 878 }
9ec3f4e4 879
bb963e16 880 if (verbose > 0) {
cb627505 881 printf("fork event\n");
b9c5143a
FW		 882 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
883 printf("... child: %s/%d\n", thread__comm_str(child), child->tid);
419ab0d6 884 }
9ec3f4e4 885
b9c5143a
FW		 886 register_pid(sched, parent->tid, thread__comm_str(parent));
887 register_pid(sched, child->tid, thread__comm_str(child));
b91fc39f
ACM		 888out_put:
889 thread__put(child);
890 thread__put(parent);
a116e05d 891 return 0;
419ab0d6 892}
fbf94829 893
b1ffe8f3
IM		 894struct sort_dimension {
895 const char *name;
b5fae128 896 sort_fn_t cmp;
b1ffe8f3
IM		 897 struct list_head list;
898};
899
daa1d7a5 900static int
39aeb52f 901thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
daa1d7a5
FW		 902{
903 struct sort_dimension *sort;
904 int ret = 0;
905
b5fae128
IM		 906 BUG_ON(list_empty(list));
907
daa1d7a5
FW		 908 list_for_each_entry(sort, list, list) {
909 ret = sort->cmp(l, r);
910 if (ret)
911 return ret;
912 }
913
914 return ret;
915}
916
39aeb52f 917static struct work_atoms *
b5fae128
IM		 918thread_atoms_search(struct rb_root *root, struct thread *thread,
919 struct list_head *sort_list)
920{
921 struct rb_node *node = root->rb_node;
39aeb52f 922 struct work_atoms key = { .thread = thread };
b5fae128
IM		 923
924 while (node) {
39aeb52f 925 struct work_atoms *atoms;
b5fae128
IM		 926 int cmp;
927
39aeb52f 928 atoms = container_of(node, struct work_atoms, node);
b5fae128
IM		 929
930 cmp = thread_lat_cmp(sort_list, &key, atoms);
931 if (cmp > 0)
932 node = node->rb_left;
933 else if (cmp < 0)
934 node = node->rb_right;
935 else {
936 BUG_ON(thread != atoms->thread);
937 return atoms;
938 }
939 }
940 return NULL;
941}
942
cdce9d73 943static void
39aeb52f 944__thread_latency_insert(struct rb_root *root, struct work_atoms *data,
daa1d7a5 945 struct list_head *sort_list)
cdce9d73
FW		 946{
947 struct rb_node **new = &(root->rb_node), *parent = NULL;
948
949 while (*new) {
39aeb52f 950 struct work_atoms *this;
daa1d7a5 951 int cmp;
cdce9d73 952
39aeb52f 953 this = container_of(*new, struct work_atoms, node);
cdce9d73 954 parent = *new;
daa1d7a5
FW		 955
956 cmp = thread_lat_cmp(sort_list, data, this);
957
958 if (cmp > 0)
cdce9d73 959 new = &((*new)->rb_left);
cdce9d73 960 else
daa1d7a5 961 new = &((*new)->rb_right);
cdce9d73
FW		 962 }
963
964 rb_link_node(&data->node, parent, new);
965 rb_insert_color(&data->node, root);
966}
967
0e9b07e5 968static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
cdce9d73 969{
36479484 970 struct work_atoms *atoms = zalloc(sizeof(*atoms));
a116e05d
ACM		 971 if (!atoms) {
972 pr_err("No memory at %s\n", __func__);
973 return -1;
974 }
cdce9d73 975
f3b623b8 976 atoms->thread = thread__get(thread);
39aeb52f 977 INIT_LIST_HEAD(&atoms->work_list);
0e9b07e5 978 __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
a116e05d 979 return 0;
cdce9d73
FW		 980}
981
9ec3f4e4 982static char sched_out_state(u64 prev_state)
cdce9d73
FW		 983{
984 const char *str = TASK_STATE_TO_CHAR_STR;
985
9ec3f4e4 986 return str[prev_state];
cdce9d73
FW		 987}
988
a116e05d 989static int
39aeb52f 990add_sched_out_event(struct work_atoms *atoms,
991 char run_state,
992 u64 timestamp)
cdce9d73 993{
36479484 994 struct work_atom *atom = zalloc(sizeof(*atom));
a116e05d
ACM		 995 if (!atom) {
996 pr_err("Non memory at %s", __func__);
997 return -1;
998 }
cdce9d73 999
aa1ab9d2
FW		 1000 atom->sched_out_time = timestamp;
1001
39aeb52f 1002 if (run_state == 'R') {
b1ffe8f3 1003 atom->state = THREAD_WAIT_CPU;
aa1ab9d2 1004 atom->wake_up_time = atom->sched_out_time;
c6ced611
FW		 1005 }
1006
39aeb52f 1007 list_add_tail(&atom->list, &atoms->work_list);
a116e05d 1008 return 0;
cdce9d73
FW		 1009}
1010
1011static void
1d037ca1
IT		 1012add_runtime_event(struct work_atoms *atoms, u64 delta,
1013 u64 timestamp __maybe_unused)
39aeb52f 1014{
1015 struct work_atom *atom;
1016
1017 BUG_ON(list_empty(&atoms->work_list));
1018
1019 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1020
1021 atom->runtime += delta;
1022 atoms->total_runtime += delta;
1023}
1024
1025static void
1026add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
cdce9d73 1027{
b1ffe8f3 1028 struct work_atom *atom;
66685678 1029 u64 delta;
cdce9d73 1030
39aeb52f 1031 if (list_empty(&atoms->work_list))
cdce9d73
FW		 1032 return;
1033
39aeb52f 1034 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
cdce9d73 1035
b1ffe8f3 1036 if (atom->state != THREAD_WAIT_CPU)
cdce9d73
FW		 1037 return;
1038
b1ffe8f3
IM		 1039 if (timestamp < atom->wake_up_time) {
1040 atom->state = THREAD_IGNORE;
cdce9d73
FW		 1041 return;
1042 }
1043
b1ffe8f3
IM		 1044 atom->state = THREAD_SCHED_IN;
1045 atom->sched_in_time = timestamp;
66685678 1046
b1ffe8f3 1047 delta = atom->sched_in_time - atom->wake_up_time;
66685678 1048 atoms->total_lat += delta;
3786310a 1049 if (delta > atoms->max_lat) {
66685678 1050 atoms->max_lat = delta;
3786310a
FW		 1051 atoms->max_lat_at = timestamp;
1052 }
66685678 1053 atoms->nb_atoms++;
cdce9d73
FW		 1054}
1055
9ec3f4e4
ACM		 1056static int latency_switch_event(struct perf_sched *sched,
1057 struct perf_evsel *evsel,
1058 struct perf_sample *sample,
1059 struct machine *machine)
cdce9d73 1060{
9ec3f4e4
ACM		 1061 const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1062 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1063 const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
39aeb52f 1064 struct work_atoms *out_events, *in_events;
cdce9d73 1065 struct thread *sched_out, *sched_in;
7f7f8d0b 1066 u64 timestamp0, timestamp = sample->time;
b91fc39f 1067 int cpu = sample->cpu, err = -1;
ea92ed5a
IM		 1068 s64 delta;
1069
39aeb52f 1070 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
ea92ed5a 1071
0e9b07e5
ACM		 1072 timestamp0 = sched->cpu_last_switched[cpu];
1073 sched->cpu_last_switched[cpu] = timestamp;
ea92ed5a
IM		 1074 if (timestamp0)
1075 delta = timestamp - timestamp0;
1076 else
1077 delta = 0;
1078
a116e05d
ACM		 1079 if (delta < 0) {
1080 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1081 return -1;
1082 }
cdce9d73 1083
1fcb8768
AH		 1084 sched_out = machine__findnew_thread(machine, -1, prev_pid);
1085 sched_in = machine__findnew_thread(machine, -1, next_pid);
b91fc39f
ACM		 1086 if (sched_out == NULL || sched_in == NULL)
1087 goto out_put;
cdce9d73 1088
0e9b07e5 1089 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
39aeb52f 1090 if (!out_events) {
0e9b07e5 1091 if (thread_atoms_insert(sched, sched_out))
b91fc39f 1092 goto out_put;
0e9b07e5 1093 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
a116e05d
ACM		 1094 if (!out_events) {
1095 pr_err("out-event: Internal tree error");
b91fc39f 1096 goto out_put;
a116e05d 1097 }
39aeb52f 1098 }
9ec3f4e4 1099 if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
a116e05d 1100 return -1;
39aeb52f 1101
0e9b07e5 1102 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
39aeb52f 1103 if (!in_events) {
0e9b07e5 1104 if (thread_atoms_insert(sched, sched_in))
b91fc39f 1105 goto out_put;
0e9b07e5 1106 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
a116e05d
ACM		 1107 if (!in_events) {
1108 pr_err("in-event: Internal tree error");
b91fc39f 1109 goto out_put;
a116e05d 1110 }
39aeb52f 1111 /*
1112 * Take came in we have not heard about yet,
1113 * add in an initial atom in runnable state:
1114 */
a116e05d 1115 if (add_sched_out_event(in_events, 'R', timestamp))
b91fc39f 1116 goto out_put;
cdce9d73 1117 }
39aeb52f 1118 add_sched_in_event(in_events, timestamp);
b91fc39f
ACM		 1119 err = 0;
1120out_put:
1121 thread__put(sched_out);
1122 thread__put(sched_in);
1123 return err;
39aeb52f 1124}
cdce9d73 1125
9ec3f4e4
ACM		 1126static int latency_runtime_event(struct perf_sched *sched,
1127 struct perf_evsel *evsel,
1128 struct perf_sample *sample,
1129 struct machine *machine)
39aeb52f 1130{
9ec3f4e4
ACM		 1131 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1132 const u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
1fcb8768 1133 struct thread *thread = machine__findnew_thread(machine, -1, pid);
0e9b07e5 1134 struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
7f7f8d0b 1135 u64 timestamp = sample->time;
b91fc39f
ACM		 1136 int cpu = sample->cpu, err = -1;
1137
1138 if (thread == NULL)
1139 return -1;
39aeb52f 1140
1141 BUG_ON(cpu >= MAX_CPUS || cpu < 0);
39aeb52f 1142 if (!atoms) {
0e9b07e5 1143 if (thread_atoms_insert(sched, thread))
b91fc39f 1144 goto out_put;
0e9b07e5 1145 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
a116e05d 1146 if (!atoms) {
60b7d14a 1147 pr_err("in-event: Internal tree error");
b91fc39f 1148 goto out_put;
a116e05d
ACM		 1149 }
1150 if (add_sched_out_event(atoms, 'R', timestamp))
b91fc39f 1151 goto out_put;
cdce9d73
FW		 1152 }
1153
9ec3f4e4 1154 add_runtime_event(atoms, runtime, timestamp);
b91fc39f
ACM		 1155 err = 0;
1156out_put:
1157 thread__put(thread);
1158 return err;
cdce9d73
FW		 1159}
1160
9ec3f4e4
ACM		 1161static int latency_wakeup_event(struct perf_sched *sched,
1162 struct perf_evsel *evsel,
1163 struct perf_sample *sample,
1164 struct machine *machine)
cdce9d73 1165{
0680ee7d 1166 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
39aeb52f 1167 struct work_atoms *atoms;
b1ffe8f3 1168 struct work_atom *atom;
cdce9d73 1169 struct thread *wakee;
7f7f8d0b 1170 u64 timestamp = sample->time;
b91fc39f 1171 int err = -1;
cdce9d73 1172
1fcb8768 1173 wakee = machine__findnew_thread(machine, -1, pid);
b91fc39f
ACM		 1174 if (wakee == NULL)
1175 return -1;
0e9b07e5 1176 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
17562205 1177 if (!atoms) {
0e9b07e5 1178 if (thread_atoms_insert(sched, wakee))
b91fc39f 1179 goto out_put;
0e9b07e5 1180 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
a116e05d 1181 if (!atoms) {
60b7d14a 1182 pr_err("wakeup-event: Internal tree error");
b91fc39f 1183 goto out_put;
a116e05d
ACM		 1184 }
1185 if (add_sched_out_event(atoms, 'S', timestamp))
b91fc39f 1186 goto out_put;
cdce9d73
FW		 1187 }
1188
39aeb52f 1189 BUG_ON(list_empty(&atoms->work_list));
cdce9d73 1190
39aeb52f 1191 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
cdce9d73 1192
55ffb7a6 1193 /*
67d6259d
DY		 1194 * As we do not guarantee the wakeup event happens when
1195 * task is out of run queue, also may happen when task is
1196 * on run queue and wakeup only change ->state to TASK_RUNNING,
1197 * then we should not set the ->wake_up_time when wake up a
1198 * task which is on run queue.
1199 *
55ffb7a6
MG		 1200 * You WILL be missing events if you've recorded only
1201 * one CPU, or are only looking at only one, so don't
67d6259d 1202 * skip in this case.
55ffb7a6 1203 */
0e9b07e5 1204 if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
b91fc39f 1205 goto out_ok;
cdce9d73 1206
0e9b07e5 1207 sched->nr_timestamps++;
ea57c4f5 1208 if (atom->sched_out_time > timestamp) {
0e9b07e5 1209 sched->nr_unordered_timestamps++;
b91fc39f 1210 goto out_ok;
ea57c4f5 1211 }
aa1ab9d2 1212
b1ffe8f3
IM		 1213 atom->state = THREAD_WAIT_CPU;
1214 atom->wake_up_time = timestamp;
b91fc39f
ACM		 1215out_ok:
1216 err = 0;
1217out_put:
1218 thread__put(wakee);
1219 return err;
cdce9d73
FW		 1220}
1221
9ec3f4e4
ACM		 1222static int latency_migrate_task_event(struct perf_sched *sched,
1223 struct perf_evsel *evsel,
1224 struct perf_sample *sample,
1225 struct machine *machine)
55ffb7a6 1226{
9ec3f4e4 1227 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
7f7f8d0b 1228 u64 timestamp = sample->time;
55ffb7a6
MG		 1229 struct work_atoms *atoms;
1230 struct work_atom *atom;
1231 struct thread *migrant;
b91fc39f 1232 int err = -1;
55ffb7a6
MG		 1233
1234 /*
1235 * Only need to worry about migration when profiling one CPU.
1236 */
0e9b07e5 1237 if (sched->profile_cpu == -1)
a116e05d 1238 return 0;
55ffb7a6 1239
1fcb8768 1240 migrant = machine__findnew_thread(machine, -1, pid);
b91fc39f
ACM		 1241 if (migrant == NULL)
1242 return -1;
0e9b07e5 1243 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
55ffb7a6 1244 if (!atoms) {
0e9b07e5 1245 if (thread_atoms_insert(sched, migrant))
b91fc39f 1246 goto out_put;
b9c5143a 1247 register_pid(sched, migrant->tid, thread__comm_str(migrant));
0e9b07e5 1248 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
a116e05d 1249 if (!atoms) {
60b7d14a 1250 pr_err("migration-event: Internal tree error");
b91fc39f 1251 goto out_put;
a116e05d
ACM		 1252 }
1253 if (add_sched_out_event(atoms, 'R', timestamp))
b91fc39f 1254 goto out_put;
55ffb7a6
MG		 1255 }
1256
1257 BUG_ON(list_empty(&atoms->work_list));
1258
1259 atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1260 atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1261
0e9b07e5 1262 sched->nr_timestamps++;
55ffb7a6
MG		 1263
1264 if (atom->sched_out_time > timestamp)
0e9b07e5 1265 sched->nr_unordered_timestamps++;
b91fc39f
ACM		 1266 err = 0;
1267out_put:
1268 thread__put(migrant);
1269 return err;
55ffb7a6
MG		 1270}
1271
0e9b07e5 1272static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
cdce9d73 1273{
cdce9d73
FW		 1274 int i;
1275 int ret;
66685678 1276 u64 avg;
99620a5d 1277 char max_lat_at[32];
cdce9d73 1278
39aeb52f 1279 if (!work_list->nb_atoms)
cdce9d73 1280 return;
ea57c4f5
IM		 1281 /*
1282 * Ignore idle threads:
1283 */
b9c5143a 1284 if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
ea57c4f5 1285 return;
cdce9d73 1286
0e9b07e5
ACM		 1287 sched->all_runtime += work_list->total_runtime;
1288 sched->all_count += work_list->nb_atoms;
66685678 1289
2f80dd44
JB		 1290 if (work_list->num_merged > 1)
1291 ret = printf(" %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1292 else
1293 ret = printf(" %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
cdce9d73 1294
08f69e6c 1295 for (i = 0; i < 24 - ret; i++)
cdce9d73
FW		 1296 printf(" ");
1297
39aeb52f 1298 avg = work_list->total_lat / work_list->nb_atoms;
99620a5d 1299 timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at));
cdce9d73 1300
99620a5d 1301 printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n",
4fc76e49
ACM		 1302 (double)work_list->total_runtime / NSEC_PER_MSEC,
1303 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1304 (double)work_list->max_lat / NSEC_PER_MSEC,
99620a5d 1305 max_lat_at);
cdce9d73
FW		 1306}
1307
39aeb52f 1308static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
daa1d7a5 1309{
0014de17
JO		 1310 if (l->thread == r->thread)
1311 return 0;
38051234 1312 if (l->thread->tid < r->thread->tid)
daa1d7a5 1313 return -1;
38051234 1314 if (l->thread->tid > r->thread->tid)
daa1d7a5 1315 return 1;
0014de17 1316 return (int)(l->thread - r->thread);
daa1d7a5
FW		 1317}
1318
39aeb52f 1319static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
daa1d7a5
FW		 1320{
1321 u64 avgl, avgr;
1322
1323 if (!l->nb_atoms)
1324 return -1;
1325
1326 if (!r->nb_atoms)
1327 return 1;
1328
1329 avgl = l->total_lat / l->nb_atoms;
1330 avgr = r->total_lat / r->nb_atoms;
1331
1332 if (avgl < avgr)
1333 return -1;
1334 if (avgl > avgr)
1335 return 1;
1336
1337 return 0;
1338}
1339
39aeb52f 1340static int max_cmp(struct work_atoms *l, struct work_atoms *r)
daa1d7a5
FW		 1341{
1342 if (l->max_lat < r->max_lat)
1343 return -1;
1344 if (l->max_lat > r->max_lat)
1345 return 1;
1346
1347 return 0;
1348}
1349
39aeb52f 1350static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
daa1d7a5
FW		 1351{
1352 if (l->nb_atoms < r->nb_atoms)
1353 return -1;
1354 if (l->nb_atoms > r->nb_atoms)
1355 return 1;
1356
1357 return 0;
1358}
1359
39aeb52f 1360static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
daa1d7a5
FW		 1361{
1362 if (l->total_runtime < r->total_runtime)
1363 return -1;
1364 if (l->total_runtime > r->total_runtime)
1365 return 1;
1366
1367 return 0;
1368}
1369
cbef79a8 1370static int sort_dimension__add(const char *tok, struct list_head *list)
daa1d7a5 1371{
0e9b07e5
ACM		 1372 size_t i;
1373 static struct sort_dimension avg_sort_dimension = {
1374 .name = "avg",
1375 .cmp = avg_cmp,
1376 };
1377 static struct sort_dimension max_sort_dimension = {
1378 .name = "max",
1379 .cmp = max_cmp,
1380 };
1381 static struct sort_dimension pid_sort_dimension = {
1382 .name = "pid",
1383 .cmp = pid_cmp,
1384 };
1385 static struct sort_dimension runtime_sort_dimension = {
1386 .name = "runtime",
1387 .cmp = runtime_cmp,
1388 };
1389 static struct sort_dimension switch_sort_dimension = {
1390 .name = "switch",
1391 .cmp = switch_cmp,
1392 };
1393 struct sort_dimension *available_sorts[] = {
1394 &pid_sort_dimension,
1395 &avg_sort_dimension,
1396 &max_sort_dimension,
1397 &switch_sort_dimension,
1398 &runtime_sort_dimension,
1399 };
daa1d7a5 1400
0e9b07e5 1401 for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
daa1d7a5
FW		 1402 if (!strcmp(available_sorts[i]->name, tok)) {
1403 list_add_tail(&available_sorts[i]->list, list);
1404
1405 return 0;
1406 }
1407 }
1408
1409 return -1;
1410}
1411
0e9b07e5 1412static void perf_sched__sort_lat(struct perf_sched *sched)
daa1d7a5
FW		 1413{
1414 struct rb_node *node;
2f80dd44
JB		 1415 struct rb_root *root = &sched->atom_root;
1416again:
daa1d7a5 1417 for (;;) {
39aeb52f 1418 struct work_atoms *data;
2f80dd44 1419 node = rb_first(root);
daa1d7a5
FW		 1420 if (!node)
1421 break;
1422
2f80dd44 1423 rb_erase(node, root);
39aeb52f 1424 data = rb_entry(node, struct work_atoms, node);
0e9b07e5 1425 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
daa1d7a5 1426 }
2f80dd44
JB		 1427 if (root == &sched->atom_root) {
1428 root = &sched->merged_atom_root;
1429 goto again;
1430 }
daa1d7a5
FW		 1431}
1432
0e9b07e5 1433static int process_sched_wakeup_event(struct perf_tool *tool,
2b7fcbc5 1434 struct perf_evsel *evsel,
1d037ca1 1435 struct perf_sample *sample,
4218e673 1436 struct machine *machine)
419ab0d6 1437{
0e9b07e5 1438 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
419ab0d6 1439
9ec3f4e4
ACM		 1440 if (sched->tp_handler->wakeup_event)
1441 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
a116e05d 1442
2b7fcbc5 1443 return 0;
419ab0d6
FW		 1444}
1445
a151a37a
JO		 1446union map_priv {
1447 void *ptr;
1448 bool color;
1449};
1450
1451static bool thread__has_color(struct thread *thread)
1452{
1453 union map_priv priv = {
1454 .ptr = thread__priv(thread),
1455 };
1456
1457 return priv.color;
1458}
1459
1460static struct thread*
1461map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1462{
1463 struct thread *thread = machine__findnew_thread(machine, pid, tid);
1464 union map_priv priv = {
1465 .color = false,
1466 };
1467
1468 if (!sched->map.color_pids || !thread || thread__priv(thread))
1469 return thread;
1470
1471 if (thread_map__has(sched->map.color_pids, tid))
1472 priv.color = true;
1473
1474 thread__set_priv(thread, priv.ptr);
1475 return thread;
1476}
1477
9ec3f4e4
ACM		 1478static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1479 struct perf_sample *sample, struct machine *machine)
0ec04e16 1480{
9d372ca5
DY		 1481 const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1482 struct thread *sched_in;
0ec04e16 1483 int new_shortname;
7f7f8d0b 1484 u64 timestamp0, timestamp = sample->time;
0ec04e16 1485 s64 delta;
99623c62
JO		 1486 int i, this_cpu = sample->cpu;
1487 int cpus_nr;
1488 bool new_cpu = false;
8cd91195 1489 const char *color = PERF_COLOR_NORMAL;
99620a5d 1490 char stimestamp[32];
0ec04e16
IM		 1491
1492 BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1493
0e9b07e5
ACM		 1494 if (this_cpu > sched->max_cpu)
1495 sched->max_cpu = this_cpu;
0ec04e16 1496
99623c62
JO		 1497 if (sched->map.comp) {
1498 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1499 if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1500 sched->map.comp_cpus[cpus_nr++] = this_cpu;
1501 new_cpu = true;
1502 }
1503 } else
1504 cpus_nr = sched->max_cpu;
1505
0e9b07e5
ACM		 1506 timestamp0 = sched->cpu_last_switched[this_cpu];
1507 sched->cpu_last_switched[this_cpu] = timestamp;
0ec04e16
IM		 1508 if (timestamp0)
1509 delta = timestamp - timestamp0;
1510 else
1511 delta = 0;
1512
a116e05d 1513 if (delta < 0) {
60b7d14a 1514 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
a116e05d
ACM		 1515 return -1;
1516 }
0ec04e16 1517
a151a37a 1518 sched_in = map__findnew_thread(sched, machine, -1, next_pid);
b91fc39f
ACM		 1519 if (sched_in == NULL)
1520 return -1;
0ec04e16 1521
b91fc39f 1522 sched->curr_thread[this_cpu] = thread__get(sched_in);
0ec04e16
IM		 1523
1524 printf(" ");
1525
1526 new_shortname = 0;
1527 if (!sched_in->shortname[0]) {
6bcab4e1
D		 1528 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1529 /*
1530 * Don't allocate a letter-number for swapper:0
1531 * as a shortname. Instead, we use '.' for it.
1532 */
1533 sched_in->shortname[0] = '.';
1534 sched_in->shortname[1] = ' ';
0ec04e16 1535 } else {
6bcab4e1
D		 1536 sched_in->shortname[0] = sched->next_shortname1;
1537 sched_in->shortname[1] = sched->next_shortname2;
1538
1539 if (sched->next_shortname1 < 'Z') {
1540 sched->next_shortname1++;
0ec04e16 1541 } else {
6bcab4e1
D		 1542 sched->next_shortname1 = 'A';
1543 if (sched->next_shortname2 < '9')
1544 sched->next_shortname2++;
1545 else
1546 sched->next_shortname2 = '0';
0ec04e16
IM		 1547 }
1548 }
1549 new_shortname = 1;
1550 }
1551
99623c62
JO		 1552 for (i = 0; i < cpus_nr; i++) {
1553 int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
a151a37a
JO		 1554 struct thread *curr_thread = sched->curr_thread[cpu];
1555 const char *pid_color = color;
cf294f24 1556 const char *cpu_color = color;
a151a37a
JO		 1557
1558 if (curr_thread && thread__has_color(curr_thread))
1559 pid_color = COLOR_PIDS;
99623c62 1560
73643bb6
JO		 1561 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1562 continue;
1563
cf294f24
JO		 1564 if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1565 cpu_color = COLOR_CPUS;
1566
0ec04e16 1567 if (cpu != this_cpu)
1208bb27 1568 color_fprintf(stdout, color, " ");
0ec04e16 1569 else
cf294f24 1570 color_fprintf(stdout, cpu_color, "*");
0ec04e16 1571
6bcab4e1 1572 if (sched->curr_thread[cpu])
a151a37a 1573 color_fprintf(stdout, pid_color, "%2s ", sched->curr_thread[cpu]->shortname);
6bcab4e1 1574 else
8cd91195 1575 color_fprintf(stdout, color, " ");
0ec04e16
IM		 1576 }
1577
73643bb6
JO		 1578 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1579 goto out;
1580
99620a5d
NK		 1581 timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1582 color_fprintf(stdout, color, " %12s secs ", stimestamp);
bb963e16 1583 if (new_shortname || (verbose > 0 && sched_in->tid)) {
a151a37a
JO		 1584 const char *pid_color = color;
1585
1586 if (thread__has_color(sched_in))
1587 pid_color = COLOR_PIDS;
1588
1589 color_fprintf(stdout, pid_color, "%s => %s:%d",
b9c5143a 1590 sched_in->shortname, thread__comm_str(sched_in), sched_in->tid);
0ec04e16 1591 }
a116e05d 1592
99623c62 1593 if (sched->map.comp && new_cpu)
8cd91195 1594 color_fprintf(stdout, color, " (CPU %d)", this_cpu);
99623c62 1595
73643bb6 1596out:
8cd91195 1597 color_fprintf(stdout, color, "\n");
99623c62 1598
b91fc39f
ACM		 1599 thread__put(sched_in);
1600
a116e05d 1601 return 0;
0ec04e16
IM		 1602}
1603
0e9b07e5 1604static int process_sched_switch_event(struct perf_tool *tool,
2b7fcbc5 1605 struct perf_evsel *evsel,
1d037ca1 1606 struct perf_sample *sample,
4218e673 1607 struct machine *machine)
419ab0d6 1608{
0e9b07e5 1609 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
a116e05d 1610 int this_cpu = sample->cpu, err = 0;
2b7fcbc5
ACM		 1611 u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1612 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
419ab0d6 1613
0e9b07e5 1614 if (sched->curr_pid[this_cpu] != (u32)-1) {
c8a37751
IM		 1615 /*
1616 * Are we trying to switch away a PID that is
1617 * not current?
1618 */
2b7fcbc5 1619 if (sched->curr_pid[this_cpu] != prev_pid)
0e9b07e5 1620 sched->nr_context_switch_bugs++;
c8a37751 1621 }
c8a37751 1622
9ec3f4e4
ACM		 1623 if (sched->tp_handler->switch_event)
1624 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
2b7fcbc5
ACM		 1625
1626 sched->curr_pid[this_cpu] = next_pid;
a116e05d 1627 return err;
419ab0d6
FW		 1628}
1629
0e9b07e5 1630static int process_sched_runtime_event(struct perf_tool *tool,
2b7fcbc5 1631 struct perf_evsel *evsel,
1d037ca1 1632 struct perf_sample *sample,
4218e673 1633 struct machine *machine)
39aeb52f 1634{
0e9b07e5 1635 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
39aeb52f 1636
9ec3f4e4
ACM		 1637 if (sched->tp_handler->runtime_event)
1638 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
a116e05d 1639
2b7fcbc5 1640 return 0;
39aeb52f 1641}
1642
cb627505
DA		 1643static int perf_sched__process_fork_event(struct perf_tool *tool,
1644 union perf_event *event,
1645 struct perf_sample *sample,
1646 struct machine *machine)
fbf94829 1647{
0e9b07e5 1648 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
46538818 1649
cb627505
DA		 1650 /* run the fork event through the perf machineruy */
1651 perf_event__process_fork(tool, event, sample, machine);
1652
1653 /* and then run additional processing needed for this command */
9ec3f4e4 1654 if (sched->tp_handler->fork_event)
cb627505 1655 return sched->tp_handler->fork_event(sched, event, machine);
a116e05d 1656
2b7fcbc5 1657 return 0;
fbf94829
IM		 1658}
1659
0e9b07e5 1660static int process_sched_migrate_task_event(struct perf_tool *tool,
2b7fcbc5 1661 struct perf_evsel *evsel,
1d037ca1 1662 struct perf_sample *sample,
4218e673 1663 struct machine *machine)
55ffb7a6 1664{
0e9b07e5 1665 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
55ffb7a6 1666
9ec3f4e4
ACM		 1667 if (sched->tp_handler->migrate_task_event)
1668 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
a116e05d 1669
2b7fcbc5 1670 return 0;
55ffb7a6
MG		 1671}
1672
a116e05d 1673typedef int (*tracepoint_handler)(struct perf_tool *tool,
2b7fcbc5 1674 struct perf_evsel *evsel,
a116e05d 1675 struct perf_sample *sample,
4218e673 1676 struct machine *machine);
ec156764 1677
1d037ca1
IT		 1678static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1679 union perf_event *event __maybe_unused,
ee29be62
ACM		 1680 struct perf_sample *sample,
1681 struct perf_evsel *evsel,
1682 struct machine *machine)
0a02ad93 1683{
a116e05d 1684 int err = 0;
0a02ad93 1685
744a9719
ACM		 1686 if (evsel->handler != NULL) {
1687 tracepoint_handler f = evsel->handler;
2b7fcbc5 1688 err = f(tool, evsel, sample, machine);
ee29be62 1689 }
0a02ad93 1690
a116e05d 1691 return err;
0a02ad93
IM		 1692}
1693
ae536acf 1694static int perf_sched__read_events(struct perf_sched *sched)
0a02ad93 1695{
ee29be62
ACM		 1696 const struct perf_evsel_str_handler handlers[] = {
1697 { "sched:sched_switch", process_sched_switch_event, },
1698 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1699 { "sched:sched_wakeup", process_sched_wakeup_event, },
1700 { "sched:sched_wakeup_new", process_sched_wakeup_event, },
ee29be62
ACM		 1701 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1702 };
da378962 1703 struct perf_session *session;
f5fc1412
JO		 1704 struct perf_data_file file = {
1705 .path = input_name,
1706 .mode = PERF_DATA_MODE_READ,
f0dd330f 1707 .force = sched->force,
f5fc1412 1708 };
ae536acf 1709 int rc = -1;
da378962 1710
f5fc1412 1711 session = perf_session__new(&file, false, &sched->tool);
a116e05d
ACM		 1712 if (session == NULL) {
1713 pr_debug("No Memory for session\n");
1714 return -1;
1715 }
94c744b6 1716
0a7e6d1b 1717 symbol__init(&session->header.env);
04934106 1718
a116e05d
ACM		 1719 if (perf_session__set_tracepoints_handlers(session, handlers))
1720 goto out_delete;
ee29be62 1721
cee75ac7 1722 if (perf_session__has_traces(session, "record -R")) {
b7b61cbe 1723 int err = perf_session__process_events(session);
a116e05d
ACM		 1724 if (err) {
1725 pr_err("Failed to process events, error %d", err);
1726 goto out_delete;
1727 }
4c09bafa 1728
75be989a
ACM		 1729 sched->nr_events = session->evlist->stats.nr_events[0];
1730 sched->nr_lost_events = session->evlist->stats.total_lost;
1731 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
cee75ac7 1732 }
d549c769 1733
ae536acf 1734 rc = 0;
a116e05d
ACM		 1735out_delete:
1736 perf_session__delete(session);
ae536acf 1737 return rc;
0a02ad93
IM		 1738}
1739
49394a2a
DA		 1740/*
1741 * scheduling times are printed as msec.usec
1742 */
1743static inline void print_sched_time(unsigned long long nsecs, int width)
1744{
1745 unsigned long msecs;
1746 unsigned long usecs;
1747
1748 msecs = nsecs / NSEC_PER_MSEC;
1749 nsecs -= msecs * NSEC_PER_MSEC;
1750 usecs = nsecs / NSEC_PER_USEC;
1751 printf("%*lu.%03lu ", width, msecs, usecs);
1752}
1753
1754/*
1755 * returns runtime data for event, allocating memory for it the
1756 * first time it is used.
1757 */
1758static struct evsel_runtime *perf_evsel__get_runtime(struct perf_evsel *evsel)
1759{
1760 struct evsel_runtime *r = evsel->priv;
1761
1762 if (r == NULL) {
1763 r = zalloc(sizeof(struct evsel_runtime));
1764 evsel->priv = r;
1765 }
1766
1767 return r;
1768}
1769
1770/*
1771 * save last time event was seen per cpu
1772 */
1773static void perf_evsel__save_time(struct perf_evsel *evsel,
1774 u64 timestamp, u32 cpu)
1775{
1776 struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1777
1778 if (r == NULL)
1779 return;
1780
1781 if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1782 int i, n = __roundup_pow_of_two(cpu+1);
1783 void *p = r->last_time;
1784
1785 p = realloc(r->last_time, n * sizeof(u64));
1786 if (!p)
1787 return;
1788
1789 r->last_time = p;
1790 for (i = r->ncpu; i < n; ++i)
1791 r->last_time[i] = (u64) 0;
1792
1793 r->ncpu = n;
1794 }
1795
1796 r->last_time[cpu] = timestamp;
1797}
1798
1799/* returns last time this event was seen on the given cpu */
1800static u64 perf_evsel__get_time(struct perf_evsel *evsel, u32 cpu)
1801{
1802 struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1803
1804 if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1805 return 0;
1806
1807 return r->last_time[cpu];
1808}
1809
9b8087d7 1810static int comm_width = 30;
49394a2a
DA		 1811
1812static char *timehist_get_commstr(struct thread *thread)
1813{
1814 static char str[32];
1815 const char *comm = thread__comm_str(thread);
1816 pid_t tid = thread->tid;
1817 pid_t pid = thread->pid_;
1818 int n;
1819
1820 if (pid == 0)
1821 n = scnprintf(str, sizeof(str), "%s", comm);
1822
1823 else if (tid != pid)
1824 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1825
1826 else
1827 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1828
1829 if (n > comm_width)
1830 comm_width = n;
1831
1832 return str;
1833}
1834
a407b067 1835static void timehist_header(struct perf_sched *sched)
49394a2a 1836{
a407b067
DA		 1837 u32 ncpus = sched->max_cpu + 1;
1838 u32 i, j;
1839
49394a2a
DA		 1840 printf("%15s %6s ", "time", "cpu");
1841
a407b067
DA		 1842 if (sched->show_cpu_visual) {
1843 printf(" ");
1844 for (i = 0, j = 0; i < ncpus; ++i) {
1845 printf("%x", j++);
1846 if (j > 15)
1847 j = 0;
1848 }
1849 printf(" ");
1850 }
1851
0e6758e8 1852 printf(" %-*s %9s %9s %9s", comm_width,
49394a2a
DA		 1853 "task name", "wait time", "sch delay", "run time");
1854
414e050c
NK		 1855 if (sched->show_state)
1856 printf(" %s", "state");
1857
49394a2a
DA		 1858 printf("\n");
1859
1860 /*
1861 * units row
1862 */
1863 printf("%15s %-6s ", "", "");
1864
a407b067
DA		 1865 if (sched->show_cpu_visual)
1866 printf(" %*s ", ncpus, "");
1867
414e050c 1868 printf(" %-*s %9s %9s %9s", comm_width,
0e6758e8 1869 "[tid/pid]", "(msec)", "(msec)", "(msec)");
49394a2a 1870
414e050c
NK		 1871 if (sched->show_state)
1872 printf(" %5s", "");
1873
1874 printf("\n");
1875
49394a2a
DA		 1876 /*
1877 * separator
1878 */
1879 printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1880
a407b067
DA		 1881 if (sched->show_cpu_visual)
1882 printf(" %.*s ", ncpus, graph_dotted_line);
1883
0e6758e8 1884 printf(" %.*s %.9s %.9s %.9s", comm_width,
49394a2a
DA		 1885 graph_dotted_line, graph_dotted_line, graph_dotted_line,
1886 graph_dotted_line);
1887
414e050c
NK		 1888 if (sched->show_state)
1889 printf(" %.5s", graph_dotted_line);
1890
49394a2a
DA		 1891 printf("\n");
1892}
1893
414e050c
NK		 1894static char task_state_char(struct thread *thread, int state)
1895{
1896 static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
1897 unsigned bit = state ? ffs(state) : 0;
1898
1899 /* 'I' for idle */
1900 if (thread->tid == 0)
1901 return 'I';
1902
1903 return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
1904}
1905
fc1469f1 1906static void timehist_print_sample(struct perf_sched *sched,
292c4a8f 1907 struct perf_evsel *evsel,
fc1469f1 1908 struct perf_sample *sample,
6c973c90 1909 struct addr_location *al,
853b7407 1910 struct thread *thread,
414e050c 1911 u64 t, int state)
49394a2a
DA		 1912{
1913 struct thread_runtime *tr = thread__priv(thread);
292c4a8f
BG		 1914 const char *next_comm = perf_evsel__strval(evsel, sample, "next_comm");
1915 const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
a407b067 1916 u32 max_cpus = sched->max_cpu + 1;
49394a2a 1917 char tstr[64];
292c4a8f 1918 char nstr[30];
941bdea7 1919 u64 wait_time;
49394a2a 1920
853b7407 1921 timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
49394a2a
DA		 1922 printf("%15s [%04d] ", tstr, sample->cpu);
1923
a407b067
DA		 1924 if (sched->show_cpu_visual) {
1925 u32 i;
1926 char c;
1927
1928 printf(" ");
1929 for (i = 0; i < max_cpus; ++i) {
1930 /* flag idle times with 'i'; others are sched events */
1931 if (i == sample->cpu)
1932 c = (thread->tid == 0) ? 'i' : 's';
1933 else
1934 c = ' ';
1935 printf("%c", c);
1936 }
1937 printf(" ");
1938 }
1939
49394a2a
DA		 1940 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
1941
941bdea7
NK		 1942 wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
1943 print_sched_time(wait_time, 6);
1944
49394a2a
DA		 1945 print_sched_time(tr->dt_delay, 6);
1946 print_sched_time(tr->dt_run, 6);
fc1469f1 1947
414e050c
NK		 1948 if (sched->show_state)
1949 printf(" %5c ", task_state_char(thread, state));
1950
292c4a8f
BG		 1951 if (sched->show_next) {
1952 snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
1953 printf(" %-*s", comm_width, nstr);
1954 }
1955
1956 if (sched->show_wakeups && !sched->show_next)
fc1469f1
DA		 1957 printf(" %-*s", comm_width, "");
1958
6c973c90
DA		 1959 if (thread->tid == 0)
1960 goto out;
1961
1962 if (sched->show_callchain)
1963 printf(" ");
1964
1965 sample__fprintf_sym(sample, al, 0,
1966 EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2d9bbf6e
NK		 1967 EVSEL__PRINT_CALLCHAIN_ARROW |
1968 EVSEL__PRINT_SKIP_IGNORED,
6c973c90
DA		 1969 &callchain_cursor, stdout);
1970
1971out:
49394a2a
DA		 1972 printf("\n");
1973}
1974
1975/*
1976 * Explanation of delta-time stats:
1977 *
1978 * t = time of current schedule out event
1979 * tprev = time of previous sched out event
1980 * also time of schedule-in event for current task
1981 * last_time = time of last sched change event for current task
1982 * (i.e, time process was last scheduled out)
1983 * ready_to_run = time of wakeup for current task
1984 *
1985 * -----|------------|------------|------------|------
1986 * last ready tprev t
1987 * time to run
1988 *
1989 * |-------- dt_wait --------|
1990 * |- dt_delay -|-- dt_run --|
1991 *
1992 * dt_run = run time of current task
1993 * dt_wait = time between last schedule out event for task and tprev
1994 * represents time spent off the cpu
1995 * dt_delay = time between wakeup and schedule-in of task
1996 */
1997
1998static void timehist_update_runtime_stats(struct thread_runtime *r,
1999 u64 t, u64 tprev)
2000{
2001 r->dt_delay = 0;
941bdea7
NK		 2002 r->dt_sleep = 0;
2003 r->dt_iowait = 0;
2004 r->dt_preempt = 0;
49394a2a 2005 r->dt_run = 0;
941bdea7 2006
49394a2a
DA		 2007 if (tprev) {
2008 r->dt_run = t - tprev;
2009 if (r->ready_to_run) {
2010 if (r->ready_to_run > tprev)
2011 pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2012 else
2013 r->dt_delay = tprev - r->ready_to_run;
2014 }
2015
2016 if (r->last_time > tprev)
2017 pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
941bdea7
NK		 2018 else if (r->last_time) {
2019 u64 dt_wait = tprev - r->last_time;
2020
2021 if (r->last_state == TASK_RUNNING)
2022 r->dt_preempt = dt_wait;
2023 else if (r->last_state == TASK_UNINTERRUPTIBLE)
2024 r->dt_iowait = dt_wait;
2025 else
2026 r->dt_sleep = dt_wait;
2027 }
49394a2a
DA		 2028 }
2029
2030 update_stats(&r->run_stats, r->dt_run);
587782c5
NK		 2031
2032 r->total_run_time += r->dt_run;
2033 r->total_delay_time += r->dt_delay;
2034 r->total_sleep_time += r->dt_sleep;
2035 r->total_iowait_time += r->dt_iowait;
2036 r->total_preempt_time += r->dt_preempt;
49394a2a
DA		 2037}
2038
96039c7c
NK		 2039static bool is_idle_sample(struct perf_sample *sample,
2040 struct perf_evsel *evsel)
49394a2a
DA		 2041{
2042 /* pid 0 == swapper == idle task */
96039c7c
NK		 2043 if (strcmp(perf_evsel__name(evsel), "sched:sched_switch") == 0)
2044 return perf_evsel__intval(evsel, sample, "prev_pid") == 0;
49394a2a 2045
96039c7c
NK		 2046 return sample->pid == 0;
2047}
2048
2049static void save_task_callchain(struct perf_sched *sched,
2050 struct perf_sample *sample,
2051 struct perf_evsel *evsel,
2052 struct machine *machine)
2053{
2054 struct callchain_cursor *cursor = &callchain_cursor;
2055 struct thread *thread;
6c973c90
DA		 2056
2057 /* want main thread for process - has maps */
2058 thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2059 if (thread == NULL) {
2060 pr_debug("Failed to get thread for pid %d.\n", sample->pid);
96039c7c 2061 return;
6c973c90
DA		 2062 }
2063
2064 if (!symbol_conf.use_callchain || sample->callchain == NULL)
96039c7c 2065 return;
6c973c90
DA		 2066
2067 if (thread__resolve_callchain(thread, cursor, evsel, sample,
8388deb3 2068 NULL, NULL, sched->max_stack + 2) != 0) {
bb963e16 2069 if (verbose > 0)
62d94b00 2070 pr_err("Failed to resolve callchain. Skipping\n");
6c973c90 2071
96039c7c 2072 return;
6c973c90 2073 }
cdeb01bf 2074
6c973c90 2075 callchain_cursor_commit(cursor);
cdeb01bf
NK		 2076
2077 while (true) {
2078 struct callchain_cursor_node *node;
2079 struct symbol *sym;
2080
2081 node = callchain_cursor_current(cursor);
2082 if (node == NULL)
2083 break;
2084
2085 sym = node->sym;
a7c3899c 2086 if (sym) {
cdeb01bf
NK		 2087 if (!strcmp(sym->name, "schedule") ||
2088 !strcmp(sym->name, "__schedule") ||
2089 !strcmp(sym->name, "preempt_schedule"))
2090 sym->ignore = 1;
2091 }
2092
2093 callchain_cursor_advance(cursor);
2094 }
49394a2a
DA		 2095}
2096
3bc2fa9c
NK		 2097static int init_idle_thread(struct thread *thread)
2098{
2099 struct idle_thread_runtime *itr;
2100
2101 thread__set_comm(thread, idle_comm, 0);
2102
2103 itr = zalloc(sizeof(*itr));
2104 if (itr == NULL)
2105 return -ENOMEM;
2106
2107 init_stats(&itr->tr.run_stats);
2108 callchain_init(&itr->callchain);
2109 callchain_cursor_reset(&itr->cursor);
2110 thread__set_priv(thread, itr);
2111
2112 return 0;
2113}
2114
49394a2a
DA		 2115/*
2116 * Track idle stats per cpu by maintaining a local thread
2117 * struct for the idle task on each cpu.
2118 */
2119static int init_idle_threads(int ncpu)
2120{
3bc2fa9c 2121 int i, ret;
49394a2a
DA		 2122
2123 idle_threads = zalloc(ncpu * sizeof(struct thread *));
2124 if (!idle_threads)
2125 return -ENOMEM;
2126
b336352b 2127 idle_max_cpu = ncpu;
49394a2a
DA		 2128
2129 /* allocate the actual thread struct if needed */
2130 for (i = 0; i < ncpu; ++i) {
2131 idle_threads[i] = thread__new(0, 0);
2132 if (idle_threads[i] == NULL)
2133 return -ENOMEM;
2134
3bc2fa9c
NK		 2135 ret = init_idle_thread(idle_threads[i]);
2136 if (ret < 0)
2137 return ret;
49394a2a
DA		 2138 }
2139
2140 return 0;
2141}
2142
2143static void free_idle_threads(void)
2144{
2145 int i;
2146
2147 if (idle_threads == NULL)
2148 return;
2149
b336352b 2150 for (i = 0; i < idle_max_cpu; ++i) {
49394a2a
DA		 2151 if ((idle_threads[i]))
2152 thread__delete(idle_threads[i]);
2153 }
2154
2155 free(idle_threads);
2156}
2157
2158static struct thread *get_idle_thread(int cpu)
2159{
2160 /*
2161 * expand/allocate array of pointers to local thread
2162 * structs if needed
2163 */
2164 if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2165 int i, j = __roundup_pow_of_two(cpu+1);
2166 void *p;
2167
2168 p = realloc(idle_threads, j * sizeof(struct thread *));
2169 if (!p)
2170 return NULL;
2171
2172 idle_threads = (struct thread **) p;
b336352b 2173 for (i = idle_max_cpu; i < j; ++i)
49394a2a
DA		 2174 idle_threads[i] = NULL;
2175
2176 idle_max_cpu = j;
2177 }
2178
2179 /* allocate a new thread struct if needed */
2180 if (idle_threads[cpu] == NULL) {
2181 idle_threads[cpu] = thread__new(0, 0);
2182
2183 if (idle_threads[cpu]) {
3bc2fa9c
NK		 2184 if (init_idle_thread(idle_threads[cpu]) < 0)
2185 return NULL;
49394a2a
DA		 2186 }
2187 }
2188
2189 return idle_threads[cpu];
2190}
2191
699b5b92
NK		 2192static void save_idle_callchain(struct idle_thread_runtime *itr,
2193 struct perf_sample *sample)
2194{
2195 if (!symbol_conf.use_callchain || sample->callchain == NULL)
2196 return;
2197
2198 callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2199}
2200
49394a2a
DA		 2201/*
2202 * handle runtime stats saved per thread
2203 */
2204static struct thread_runtime *thread__init_runtime(struct thread *thread)
2205{
2206 struct thread_runtime *r;
2207
2208 r = zalloc(sizeof(struct thread_runtime));
2209 if (!r)
2210 return NULL;
2211
2212 init_stats(&r->run_stats);
2213 thread__set_priv(thread, r);
2214
2215 return r;
2216}
2217
2218static struct thread_runtime *thread__get_runtime(struct thread *thread)
2219{
2220 struct thread_runtime *tr;
2221
2222 tr = thread__priv(thread);
2223 if (tr == NULL) {
2224 tr = thread__init_runtime(thread);
2225 if (tr == NULL)
2226 pr_debug("Failed to malloc memory for runtime data.\n");
2227 }
2228
2229 return tr;
2230}
2231
6c973c90
DA		 2232static struct thread *timehist_get_thread(struct perf_sched *sched,
2233 struct perf_sample *sample,
49394a2a
DA		 2234 struct machine *machine,
2235 struct perf_evsel *evsel)
2236{
2237 struct thread *thread;
2238
96039c7c 2239 if (is_idle_sample(sample, evsel)) {
49394a2a
DA		 2240 thread = get_idle_thread(sample->cpu);
2241 if (thread == NULL)
2242 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2243
2244 } else {
5d92d96a
NK		 2245 /* there were samples with tid 0 but non-zero pid */
2246 thread = machine__findnew_thread(machine, sample->pid,
2247 sample->tid ?: sample->pid);
49394a2a
DA		 2248 if (thread == NULL) {
2249 pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2250 sample->tid);
2251 }
96039c7c
NK		 2252
2253 save_task_callchain(sched, sample, evsel, machine);
699b5b92
NK		 2254 if (sched->idle_hist) {
2255 struct thread *idle;
2256 struct idle_thread_runtime *itr;
2257
2258 idle = get_idle_thread(sample->cpu);
2259 if (idle == NULL) {
2260 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2261 return NULL;
2262 }
2263
2264 itr = thread__priv(idle);
2265 if (itr == NULL)
2266 return NULL;
2267
2268 itr->last_thread = thread;
2269
2270 /* copy task callchain when entering to idle */
2271 if (perf_evsel__intval(evsel, sample, "next_pid") == 0)
2272 save_idle_callchain(itr, sample);
2273 }
49394a2a
DA		 2274 }
2275
2276 return thread;
2277}
2278
52df138c 2279static bool timehist_skip_sample(struct perf_sched *sched,
a4b2b6f5
NK		 2280 struct thread *thread,
2281 struct perf_evsel *evsel,
2282 struct perf_sample *sample)
49394a2a
DA		 2283{
2284 bool rc = false;
2285
52df138c 2286 if (thread__is_filtered(thread)) {
49394a2a 2287 rc = true;
52df138c
DA		 2288 sched->skipped_samples++;
2289 }
49394a2a 2290
a4b2b6f5
NK		 2291 if (sched->idle_hist) {
2292 if (strcmp(perf_evsel__name(evsel), "sched:sched_switch"))
2293 rc = true;
2294 else if (perf_evsel__intval(evsel, sample, "prev_pid") != 0 &&
2295 perf_evsel__intval(evsel, sample, "next_pid") != 0)
2296 rc = true;
2297 }
2298
49394a2a
DA		 2299 return rc;
2300}
2301
fc1469f1 2302static void timehist_print_wakeup_event(struct perf_sched *sched,
a4b2b6f5 2303 struct perf_evsel *evsel,
fc1469f1
DA		 2304 struct perf_sample *sample,
2305 struct machine *machine,
2306 struct thread *awakened)
2307{
2308 struct thread *thread;
2309 char tstr[64];
2310
2311 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2312 if (thread == NULL)
2313 return;
2314
2315 /* show wakeup unless both awakee and awaker are filtered */
a4b2b6f5
NK		 2316 if (timehist_skip_sample(sched, thread, evsel, sample) &&
2317 timehist_skip_sample(sched, awakened, evsel, sample)) {
fc1469f1
DA		 2318 return;
2319 }
2320
2321 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2322 printf("%15s [%04d] ", tstr, sample->cpu);
a407b067
DA		 2323 if (sched->show_cpu_visual)
2324 printf(" %*s ", sched->max_cpu + 1, "");
fc1469f1
DA		 2325
2326 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2327
2328 /* dt spacer */
2329 printf(" %9s %9s %9s ", "", "", "");
2330
2331 printf("awakened: %s", timehist_get_commstr(awakened));
2332
2333 printf("\n");
2334}
2335
2336static int timehist_sched_wakeup_event(struct perf_tool *tool,
49394a2a
DA		 2337 union perf_event *event __maybe_unused,
2338 struct perf_evsel *evsel,
2339 struct perf_sample *sample,
2340 struct machine *machine)
2341{
fc1469f1 2342 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
49394a2a
DA		 2343 struct thread *thread;
2344 struct thread_runtime *tr = NULL;
2345 /* want pid of awakened task not pid in sample */
2346 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2347
2348 thread = machine__findnew_thread(machine, 0, pid);
2349 if (thread == NULL)
2350 return -1;
2351
2352 tr = thread__get_runtime(thread);
2353 if (tr == NULL)
2354 return -1;
2355
2356 if (tr->ready_to_run == 0)
2357 tr->ready_to_run = sample->time;
2358
fc1469f1 2359 /* show wakeups if requested */
853b7407
DA		 2360 if (sched->show_wakeups &&
2361 !perf_time__skip_sample(&sched->ptime, sample->time))
a4b2b6f5 2362 timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
fc1469f1 2363
49394a2a
DA		 2364 return 0;
2365}
2366
350f54fa
DA		 2367static void timehist_print_migration_event(struct perf_sched *sched,
2368 struct perf_evsel *evsel,
2369 struct perf_sample *sample,
2370 struct machine *machine,
2371 struct thread *migrated)
2372{
2373 struct thread *thread;
2374 char tstr[64];
2375 u32 max_cpus = sched->max_cpu + 1;
2376 u32 ocpu, dcpu;
2377
2378 if (sched->summary_only)
2379 return;
2380
2381 max_cpus = sched->max_cpu + 1;
2382 ocpu = perf_evsel__intval(evsel, sample, "orig_cpu");
2383 dcpu = perf_evsel__intval(evsel, sample, "dest_cpu");
2384
2385 thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2386 if (thread == NULL)
2387 return;
2388
a4b2b6f5
NK		 2389 if (timehist_skip_sample(sched, thread, evsel, sample) &&
2390 timehist_skip_sample(sched, migrated, evsel, sample)) {
350f54fa
DA		 2391 return;
2392 }
2393
2394 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2395 printf("%15s [%04d] ", tstr, sample->cpu);
2396
2397 if (sched->show_cpu_visual) {
2398 u32 i;
2399 char c;
2400
2401 printf(" ");
2402 for (i = 0; i < max_cpus; ++i) {
2403 c = (i == sample->cpu) ? 'm' : ' ';
2404 printf("%c", c);
2405 }
2406 printf(" ");
2407 }
2408
2409 printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2410
2411 /* dt spacer */
2412 printf(" %9s %9s %9s ", "", "", "");
2413
2414 printf("migrated: %s", timehist_get_commstr(migrated));
2415 printf(" cpu %d => %d", ocpu, dcpu);
2416
2417 printf("\n");
2418}
2419
2420static int timehist_migrate_task_event(struct perf_tool *tool,
2421 union perf_event *event __maybe_unused,
2422 struct perf_evsel *evsel,
2423 struct perf_sample *sample,
2424 struct machine *machine)
2425{
2426 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2427 struct thread *thread;
2428 struct thread_runtime *tr = NULL;
2429 /* want pid of migrated task not pid in sample */
2430 const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2431
2432 thread = machine__findnew_thread(machine, 0, pid);
2433 if (thread == NULL)
2434 return -1;
2435
2436 tr = thread__get_runtime(thread);
2437 if (tr == NULL)
2438 return -1;
2439
2440 tr->migrations++;
2441
2442 /* show migrations if requested */
2443 timehist_print_migration_event(sched, evsel, sample, machine, thread);
2444
2445 return 0;
2446}
2447
52df138c 2448static int timehist_sched_change_event(struct perf_tool *tool,
49394a2a
DA		 2449 union perf_event *event,
2450 struct perf_evsel *evsel,
2451 struct perf_sample *sample,
2452 struct machine *machine)
2453{
fc1469f1 2454 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
853b7407 2455 struct perf_time_interval *ptime = &sched->ptime;
49394a2a
DA		 2456 struct addr_location al;
2457 struct thread *thread;
2458 struct thread_runtime *tr = NULL;
853b7407 2459 u64 tprev, t = sample->time;
49394a2a 2460 int rc = 0;
414e050c
NK		 2461 int state = perf_evsel__intval(evsel, sample, "prev_state");
2462
49394a2a
DA		 2463
2464 if (machine__resolve(machine, &al, sample) < 0) {
2465 pr_err("problem processing %d event. skipping it\n",
2466 event->header.type);
2467 rc = -1;
2468 goto out;
2469 }
2470
6c973c90 2471 thread = timehist_get_thread(sched, sample, machine, evsel);
49394a2a
DA		 2472 if (thread == NULL) {
2473 rc = -1;
2474 goto out;
2475 }
2476
a4b2b6f5 2477 if (timehist_skip_sample(sched, thread, evsel, sample))
49394a2a
DA		 2478 goto out;
2479
2480 tr = thread__get_runtime(thread);
2481 if (tr == NULL) {
2482 rc = -1;
2483 goto out;
2484 }
2485
2486 tprev = perf_evsel__get_time(evsel, sample->cpu);
2487
853b7407
DA		 2488 /*
2489 * If start time given:
2490 * - sample time is under window user cares about - skip sample
2491 * - tprev is under window user cares about - reset to start of window
2492 */
2493 if (ptime->start && ptime->start > t)
2494 goto out;
2495
bdd75729 2496 if (tprev && ptime->start > tprev)
853b7407
DA		 2497 tprev = ptime->start;
2498
2499 /*
2500 * If end time given:
2501 * - previous sched event is out of window - we are done
2502 * - sample time is beyond window user cares about - reset it
2503 * to close out stats for time window interest
2504 */
2505 if (ptime->end) {
2506 if (tprev > ptime->end)
2507 goto out;
2508
2509 if (t > ptime->end)
2510 t = ptime->end;
2511 }
2512
07235f84
NK		 2513 if (!sched->idle_hist || thread->tid == 0) {
2514 timehist_update_runtime_stats(tr, t, tprev);
2515
2516 if (sched->idle_hist) {
2517 struct idle_thread_runtime *itr = (void *)tr;
2518 struct thread_runtime *last_tr;
2519
2520 BUG_ON(thread->tid != 0);
2521
2522 if (itr->last_thread == NULL)
2523 goto out;
2524
2525 /* add current idle time as last thread's runtime */
2526 last_tr = thread__get_runtime(itr->last_thread);
2527 if (last_tr == NULL)
2528 goto out;
2529
2530 timehist_update_runtime_stats(last_tr, t, tprev);
2531 /*
2532 * remove delta time of last thread as it's not updated
2533 * and otherwise it will show an invalid value next
2534 * time. we only care total run time and run stat.
2535 */
2536 last_tr->dt_run = 0;
07235f84 2537 last_tr->dt_delay = 0;
941bdea7
NK		 2538 last_tr->dt_sleep = 0;
2539 last_tr->dt_iowait = 0;
2540 last_tr->dt_preempt = 0;
07235f84 2541
ba957ebb
NK		 2542 if (itr->cursor.nr)
2543 callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2544
07235f84
NK		 2545 itr->last_thread = NULL;
2546 }
2547 }
853b7407 2548
52df138c 2549 if (!sched->summary_only)
292c4a8f 2550 timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
49394a2a
DA		 2551
2552out:
9396c9cb
NK		 2553 if (sched->hist_time.start == 0 && t >= ptime->start)
2554 sched->hist_time.start = t;
2555 if (ptime->end == 0 || t <= ptime->end)
2556 sched->hist_time.end = t;
2557
49394a2a
DA		 2558 if (tr) {
2559 /* time of this sched_switch event becomes last time task seen */
2560 tr->last_time = sample->time;
2561
941bdea7 2562 /* last state is used to determine where to account wait time */
414e050c 2563 tr->last_state = state;
941bdea7 2564
49394a2a
DA		 2565 /* sched out event for task so reset ready to run time */
2566 tr->ready_to_run = 0;
2567 }
2568
2569 perf_evsel__save_time(evsel, sample->time, sample->cpu);
2570
2571 return rc;
2572}
2573
2574static int timehist_sched_switch_event(struct perf_tool *tool,
2575 union perf_event *event,
2576 struct perf_evsel *evsel,
2577 struct perf_sample *sample,
2578 struct machine *machine __maybe_unused)
2579{
2580 return timehist_sched_change_event(tool, event, evsel, sample, machine);
2581}
2582
2583static int process_lost(struct perf_tool *tool __maybe_unused,
2584 union perf_event *event,
2585 struct perf_sample *sample,
2586 struct machine *machine __maybe_unused)
2587{
2588 char tstr[64];
2589
2590 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2591 printf("%15s ", tstr);
2592 printf("lost %" PRIu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2593
2594 return 0;
2595}
2596
2597
52df138c
DA		 2598static void print_thread_runtime(struct thread *t,
2599 struct thread_runtime *r)
2600{
2601 double mean = avg_stats(&r->run_stats);
2602 float stddev;
2603
2604 printf("%*s %5d %9" PRIu64 " ",
2605 comm_width, timehist_get_commstr(t), t->ppid,
2606 (u64) r->run_stats.n);
2607
2608 print_sched_time(r->total_run_time, 8);
2609 stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2610 print_sched_time(r->run_stats.min, 6);
2611 printf(" ");
2612 print_sched_time((u64) mean, 6);
2613 printf(" ");
2614 print_sched_time(r->run_stats.max, 6);
2615 printf(" ");
2616 printf("%5.2f", stddev);
350f54fa 2617 printf(" %5" PRIu64, r->migrations);
52df138c
DA		 2618 printf("\n");
2619}
2620
587782c5
NK		 2621static void print_thread_waittime(struct thread *t,
2622 struct thread_runtime *r)
2623{
2624 printf("%*s %5d %9" PRIu64 " ",
2625 comm_width, timehist_get_commstr(t), t->ppid,
2626 (u64) r->run_stats.n);
2627
2628 print_sched_time(r->total_run_time, 8);
2629 print_sched_time(r->total_sleep_time, 6);
2630 printf(" ");
2631 print_sched_time(r->total_iowait_time, 6);
2632 printf(" ");
2633 print_sched_time(r->total_preempt_time, 6);
2634 printf(" ");
2635 print_sched_time(r->total_delay_time, 6);
2636 printf("\n");
2637}
2638
52df138c 2639struct total_run_stats {
587782c5 2640 struct perf_sched *sched;
52df138c
DA		 2641 u64 sched_count;
2642 u64 task_count;
2643 u64 total_run_time;
2644};
2645
2646static int __show_thread_runtime(struct thread *t, void *priv)
2647{
2648 struct total_run_stats *stats = priv;
2649 struct thread_runtime *r;
2650
2651 if (thread__is_filtered(t))
2652 return 0;
2653
2654 r = thread__priv(t);
2655 if (r && r->run_stats.n) {
2656 stats->task_count++;
2657 stats->sched_count += r->run_stats.n;
2658 stats->total_run_time += r->total_run_time;
587782c5
NK		 2659
2660 if (stats->sched->show_state)
2661 print_thread_waittime(t, r);
2662 else
2663 print_thread_runtime(t, r);
52df138c
DA		 2664 }
2665
2666 return 0;
2667}
2668
2669static int show_thread_runtime(struct thread *t, void *priv)
2670{
2671 if (t->dead)
2672 return 0;
2673
2674 return __show_thread_runtime(t, priv);
2675}
2676
2677static int show_deadthread_runtime(struct thread *t, void *priv)
2678{
2679 if (!t->dead)
2680 return 0;
2681
2682 return __show_thread_runtime(t, priv);
2683}
2684
ba957ebb
NK		 2685static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2686{
2687 const char *sep = " <- ";
2688 struct callchain_list *chain;
2689 size_t ret = 0;
2690 char bf[1024];
2691 bool first;
2692
2693 if (node == NULL)
2694 return 0;
2695
2696 ret = callchain__fprintf_folded(fp, node->parent);
2697 first = (ret == 0);
2698
2699 list_for_each_entry(chain, &node->val, list) {
2700 if (chain->ip >= PERF_CONTEXT_MAX)
2701 continue;
2702 if (chain->ms.sym && chain->ms.sym->ignore)
2703 continue;
2704 ret += fprintf(fp, "%s%s", first ? "" : sep,
2705 callchain_list__sym_name(chain, bf, sizeof(bf),
2706 false));
2707 first = false;
2708 }
2709
2710 return ret;
2711}
2712
2713static size_t timehist_print_idlehist_callchain(struct rb_root *root)
2714{
2715 size_t ret = 0;
2716 FILE *fp = stdout;
2717 struct callchain_node *chain;
2718 struct rb_node *rb_node = rb_first(root);
2719
2720 printf(" %16s %8s %s\n", "Idle time (msec)", "Count", "Callchains");
2721 printf(" %.16s %.8s %.50s\n", graph_dotted_line, graph_dotted_line,
2722 graph_dotted_line);
2723
2724 while (rb_node) {
2725 chain = rb_entry(rb_node, struct callchain_node, rb_node);
2726 rb_node = rb_next(rb_node);
2727
2728 ret += fprintf(fp, " ");
2729 print_sched_time(chain->hit, 12);
2730 ret += 16; /* print_sched_time returns 2nd arg + 4 */
2731 ret += fprintf(fp, " %8d ", chain->count);
2732 ret += callchain__fprintf_folded(fp, chain);
2733 ret += fprintf(fp, "\n");
2734 }
2735
2736 return ret;
2737}
2738
52df138c
DA		 2739static void timehist_print_summary(struct perf_sched *sched,
2740 struct perf_session *session)
2741{
2742 struct machine *m = &session->machines.host;
2743 struct total_run_stats totals;
2744 u64 task_count;
2745 struct thread *t;
2746 struct thread_runtime *r;
2747 int i;
9396c9cb 2748 u64 hist_time = sched->hist_time.end - sched->hist_time.start;
52df138c
DA		 2749
2750 memset(&totals, 0, sizeof(totals));
587782c5 2751 totals.sched = sched;
52df138c 2752
07235f84
NK		 2753 if (sched->idle_hist) {
2754 printf("\nIdle-time summary\n");
2755 printf("%*s parent sched-out ", comm_width, "comm");
2756 printf(" idle-time min-idle avg-idle max-idle stddev migrations\n");
587782c5
NK		 2757 } else if (sched->show_state) {
2758 printf("\nWait-time summary\n");
2759 printf("%*s parent sched-in ", comm_width, "comm");
2760 printf(" run-time sleep iowait preempt delay\n");
07235f84
NK		 2761 } else {
2762 printf("\nRuntime summary\n");
2763 printf("%*s parent sched-in ", comm_width, "comm");
2764 printf(" run-time min-run avg-run max-run stddev migrations\n");
2765 }
52df138c 2766 printf("%*s (count) ", comm_width, "");
587782c5
NK		 2767 printf(" (msec) (msec) (msec) (msec) %s\n",
2768 sched->show_state ? "(msec)" : "%");
350f54fa 2769 printf("%.117s\n", graph_dotted_line);
52df138c
DA		 2770
2771 machine__for_each_thread(m, show_thread_runtime, &totals);
2772 task_count = totals.task_count;
2773 if (!task_count)
2774 printf("<no still running tasks>\n");
2775
2776 printf("\nTerminated tasks:\n");
2777 machine__for_each_thread(m, show_deadthread_runtime, &totals);
2778 if (task_count == totals.task_count)
2779 printf("<no terminated tasks>\n");
2780
2781 /* CPU idle stats not tracked when samples were skipped */
07235f84 2782 if (sched->skipped_samples && !sched->idle_hist)
52df138c
DA		 2783 return;
2784
2785 printf("\nIdle stats:\n");
b336352b 2786 for (i = 0; i < idle_max_cpu; ++i) {
52df138c
DA		 2787 t = idle_threads[i];
2788 if (!t)
2789 continue;
2790
2791 r = thread__priv(t);
2792 if (r && r->run_stats.n) {
2793 totals.sched_count += r->run_stats.n;
2794 printf(" CPU %2d idle for ", i);
2795 print_sched_time(r->total_run_time, 6);
9396c9cb 2796 printf(" msec (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
52df138c
DA		 2797 } else
2798 printf(" CPU %2d idle entire time window\n", i);
2799 }
2800
ba957ebb
NK		 2801 if (sched->idle_hist && symbol_conf.use_callchain) {
2802 callchain_param.mode = CHAIN_FOLDED;
2803 callchain_param.value = CCVAL_PERIOD;
2804
2805 callchain_register_param(&callchain_param);
2806
2807 printf("\nIdle stats by callchain:\n");
2808 for (i = 0; i < idle_max_cpu; ++i) {
2809 struct idle_thread_runtime *itr;
2810
2811 t = idle_threads[i];
2812 if (!t)
2813 continue;
2814
2815 itr = thread__priv(t);
2816 if (itr == NULL)
2817 continue;
2818
2819 callchain_param.sort(&itr->sorted_root, &itr->callchain,
2820 0, &callchain_param);
2821
2822 printf(" CPU %2d:", i);
2823 print_sched_time(itr->tr.total_run_time, 6);
2824 printf(" msec\n");
2825 timehist_print_idlehist_callchain(&itr->sorted_root);
2826 printf("\n");
2827 }
2828 }
2829
52df138c
DA		 2830 printf("\n"
2831 " Total number of unique tasks: %" PRIu64 "\n"
9396c9cb 2832 "Total number of context switches: %" PRIu64 "\n",
52df138c
DA		 2833 totals.task_count, totals.sched_count);
2834
9396c9cb 2835 printf(" Total run time (msec): ");
52df138c
DA		 2836 print_sched_time(totals.total_run_time, 2);
2837 printf("\n");
9396c9cb
NK		 2838
2839 printf(" Total scheduling time (msec): ");
2840 print_sched_time(hist_time, 2);
2841 printf(" (x %d)\n", sched->max_cpu);
52df138c
DA		 2842}
2843
49394a2a
DA		 2844typedef int (*sched_handler)(struct perf_tool *tool,
2845 union perf_event *event,
2846 struct perf_evsel *evsel,
2847 struct perf_sample *sample,
2848 struct machine *machine);
2849
2850static int perf_timehist__process_sample(struct perf_tool *tool,
2851 union perf_event *event,
2852 struct perf_sample *sample,
2853 struct perf_evsel *evsel,
2854 struct machine *machine)
2855{
2856 struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2857 int err = 0;
2858 int this_cpu = sample->cpu;
2859
2860 if (this_cpu > sched->max_cpu)
2861 sched->max_cpu = this_cpu;
2862
2863 if (evsel->handler != NULL) {
2864 sched_handler f = evsel->handler;
2865
2866 err = f(tool, event, evsel, sample, machine);
2867 }
2868
2869 return err;
2870}
2871
6c973c90
DA		 2872static int timehist_check_attr(struct perf_sched *sched,
2873 struct perf_evlist *evlist)
2874{
2875 struct perf_evsel *evsel;
2876 struct evsel_runtime *er;
2877
2878 list_for_each_entry(evsel, &evlist->entries, node) {
2879 er = perf_evsel__get_runtime(evsel);
2880 if (er == NULL) {
2881 pr_err("Failed to allocate memory for evsel runtime data\n");
2882 return -1;
2883 }
2884
2885 if (sched->show_callchain &&
2886 !(evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN)) {
2887 pr_info("Samples do not have callchains.\n");
2888 sched->show_callchain = 0;
2889 symbol_conf.use_callchain = 0;
2890 }
2891 }
2892
2893 return 0;
2894}
2895
49394a2a
DA		 2896static int perf_sched__timehist(struct perf_sched *sched)
2897{
2898 const struct perf_evsel_str_handler handlers[] = {
2899 { "sched:sched_switch", timehist_sched_switch_event, },
2900 { "sched:sched_wakeup", timehist_sched_wakeup_event, },
2901 { "sched:sched_wakeup_new", timehist_sched_wakeup_event, },
2902 };
350f54fa
DA		 2903 const struct perf_evsel_str_handler migrate_handlers[] = {
2904 { "sched:sched_migrate_task", timehist_migrate_task_event, },
2905 };
49394a2a
DA		 2906 struct perf_data_file file = {
2907 .path = input_name,
2908 .mode = PERF_DATA_MODE_READ,
6fa94258 2909 .force = sched->force,
49394a2a
DA		 2910 };
2911
2912 struct perf_session *session;
52df138c 2913 struct perf_evlist *evlist;
49394a2a
DA		 2914 int err = -1;
2915
2916 /*
2917 * event handlers for timehist option
2918 */
2919 sched->tool.sample = perf_timehist__process_sample;
2920 sched->tool.mmap = perf_event__process_mmap;
2921 sched->tool.comm = perf_event__process_comm;
2922 sched->tool.exit = perf_event__process_exit;
2923 sched->tool.fork = perf_event__process_fork;
2924 sched->tool.lost = process_lost;
2925 sched->tool.attr = perf_event__process_attr;
2926 sched->tool.tracing_data = perf_event__process_tracing_data;
2927 sched->tool.build_id = perf_event__process_build_id;
2928
2929 sched->tool.ordered_events = true;
2930 sched->tool.ordering_requires_timestamps = true;
2931
6c973c90
DA		 2932 symbol_conf.use_callchain = sched->show_callchain;
2933
49394a2a
DA		 2934 session = perf_session__new(&file, false, &sched->tool);
2935 if (session == NULL)
2936 return -ENOMEM;
2937
52df138c
DA		 2938 evlist = session->evlist;
2939
49394a2a
DA		 2940 symbol__init(&session->header.env);
2941
853b7407
DA		 2942 if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
2943 pr_err("Invalid time string\n");
2944 return -EINVAL;
2945 }
2946
6c973c90
DA		 2947 if (timehist_check_attr(sched, evlist) != 0)
2948 goto out;
2949
49394a2a
DA		 2950 setup_pager();
2951
2952 /* setup per-evsel handlers */
2953 if (perf_session__set_tracepoints_handlers(session, handlers))
2954 goto out;
2955
f45bf8d3
DA		 2956 /* sched_switch event at a minimum needs to exist */
2957 if (!perf_evlist__find_tracepoint_by_name(session->evlist,
2958 "sched:sched_switch")) {
2959 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
49394a2a 2960 goto out;
f45bf8d3 2961 }
49394a2a 2962
350f54fa
DA		 2963 if (sched->show_migrations &&
2964 perf_session__set_tracepoints_handlers(session, migrate_handlers))
2965 goto out;
2966
49394a2a
DA		 2967 /* pre-allocate struct for per-CPU idle stats */
2968 sched->max_cpu = session->header.env.nr_cpus_online;
2969 if (sched->max_cpu == 0)
2970 sched->max_cpu = 4;
2971 if (init_idle_threads(sched->max_cpu))
2972 goto out;
2973
52df138c
DA		 2974 /* summary_only implies summary option, but don't overwrite summary if set */
2975 if (sched->summary_only)
2976 sched->summary = sched->summary_only;
2977
2978 if (!sched->summary_only)
a407b067 2979 timehist_header(sched);
49394a2a
DA		 2980
2981 err = perf_session__process_events(session);
2982 if (err) {
2983 pr_err("Failed to process events, error %d", err);
2984 goto out;
2985 }
2986
52df138c
DA		 2987 sched->nr_events = evlist->stats.nr_events[0];
2988 sched->nr_lost_events = evlist->stats.total_lost;
2989 sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
2990
2991 if (sched->summary)
2992 timehist_print_summary(sched, session);
2993
49394a2a
DA		 2994out:
2995 free_idle_threads();
2996 perf_session__delete(session);
2997
2998 return err;
2999}
3000
3001
0e9b07e5 3002static void print_bad_events(struct perf_sched *sched)
0ec04e16 3003{
0e9b07e5 3004 if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
0ec04e16 3005 printf(" INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
0e9b07e5
ACM		 3006 (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3007 sched->nr_unordered_timestamps, sched->nr_timestamps);
0ec04e16 3008 }
0e9b07e5 3009 if (sched->nr_lost_events && sched->nr_events) {
0ec04e16 3010 printf(" INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
0e9b07e5
ACM		 3011 (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3012 sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
0ec04e16 3013 }
0e9b07e5 3014 if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
0ec04e16 3015 printf(" INFO: %.3f%% context switch bugs (%ld out of %ld)",
0e9b07e5
ACM		 3016 (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3017 sched->nr_context_switch_bugs, sched->nr_timestamps);
3018 if (sched->nr_lost_events)
0ec04e16
IM		 3019 printf(" (due to lost events?)");
3020 printf("\n");
3021 }
3022}
3023
2f80dd44
JB		 3024static void __merge_work_atoms(struct rb_root *root, struct work_atoms *data)
3025{
3026 struct rb_node **new = &(root->rb_node), *parent = NULL;
3027 struct work_atoms *this;
3028 const char *comm = thread__comm_str(data->thread), *this_comm;
3029
3030 while (*new) {
3031 int cmp;
3032
3033 this = container_of(*new, struct work_atoms, node);
3034 parent = *new;
3035
3036 this_comm = thread__comm_str(this->thread);
3037 cmp = strcmp(comm, this_comm);
3038 if (cmp > 0) {
3039 new = &((*new)->rb_left);
3040 } else if (cmp < 0) {
3041 new = &((*new)->rb_right);
3042 } else {
3043 this->num_merged++;
3044 this->total_runtime += data->total_runtime;
3045 this->nb_atoms += data->nb_atoms;
3046 this->total_lat += data->total_lat;
3047 list_splice(&data->work_list, &this->work_list);
3048 if (this->max_lat < data->max_lat) {
3049 this->max_lat = data->max_lat;
3050 this->max_lat_at = data->max_lat_at;
3051 }
3052 zfree(&data);
3053 return;
3054 }
3055 }
3056
3057 data->num_merged++;
3058 rb_link_node(&data->node, parent, new);
3059 rb_insert_color(&data->node, root);
3060}
3061
3062static void perf_sched__merge_lat(struct perf_sched *sched)
3063{
3064 struct work_atoms *data;
3065 struct rb_node *node;
3066
3067 if (sched->skip_merge)
3068 return;
3069
3070 while ((node = rb_first(&sched->atom_root))) {
3071 rb_erase(node, &sched->atom_root);
3072 data = rb_entry(node, struct work_atoms, node);
3073 __merge_work_atoms(&sched->merged_atom_root, data);
3074 }
3075}
3076
0e9b07e5 3077static int perf_sched__lat(struct perf_sched *sched)
0ec04e16
IM		 3078{
3079 struct rb_node *next;
3080
3081 setup_pager();
ad9def7c 3082
ae536acf 3083 if (perf_sched__read_events(sched))
a116e05d 3084 return -1;
ad9def7c 3085
2f80dd44 3086 perf_sched__merge_lat(sched);
0e9b07e5 3087 perf_sched__sort_lat(sched);
0ec04e16 3088
80790e0b
RR		 3089 printf("\n -----------------------------------------------------------------------------------------------------------------\n");
3090 printf(" Task | Runtime ms | Switches | Average delay ms | Maximum delay ms | Maximum delay at |\n");
3091 printf(" -----------------------------------------------------------------------------------------------------------------\n");
0ec04e16 3092
0e9b07e5 3093 next = rb_first(&sched->sorted_atom_root);
0ec04e16
IM		 3094
3095 while (next) {
3096 struct work_atoms *work_list;
3097
3098 work_list = rb_entry(next, struct work_atoms, node);
0e9b07e5 3099 output_lat_thread(sched, work_list);
0ec04e16 3100 next = rb_next(next);
ae536acf 3101 thread__zput(work_list->thread);
0ec04e16
IM		 3102 }
3103
80790e0b 3104 printf(" -----------------------------------------------------------------------------------------------------------------\n");
9486aa38 3105 printf(" TOTAL: |%11.3f ms |%9" PRIu64 " |\n",
4fc76e49 3106 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
0ec04e16
IM		 3107
3108 printf(" ---------------------------------------------------\n");
3109
0e9b07e5 3110 print_bad_events(sched);
0ec04e16
IM		 3111 printf("\n");
3112
a116e05d 3113 return 0;
0ec04e16
IM		 3114}
3115
99623c62
JO		 3116static int setup_map_cpus(struct perf_sched *sched)
3117{
73643bb6
JO		 3118 struct cpu_map *map;
3119
99623c62
JO		 3120 sched->max_cpu = sysconf(_SC_NPROCESSORS_CONF);
3121
3122 if (sched->map.comp) {
3123 sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
cf294f24
JO		 3124 if (!sched->map.comp_cpus)
3125 return -1;
99623c62
JO		 3126 }
3127
73643bb6
JO		 3128 if (!sched->map.cpus_str)
3129 return 0;
3130
3131 map = cpu_map__new(sched->map.cpus_str);
3132 if (!map) {
3133 pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3134 return -1;
3135 }
3136
3137 sched->map.cpus = map;
99623c62
JO		 3138 return 0;
3139}
3140
a151a37a
JO		 3141static int setup_color_pids(struct perf_sched *sched)
3142{
3143 struct thread_map *map;
3144
3145 if (!sched->map.color_pids_str)
3146 return 0;
3147
3148 map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3149 if (!map) {
3150 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3151 return -1;
3152 }
3153
3154 sched->map.color_pids = map;
3155 return 0;
3156}
3157
cf294f24
JO		 3158static int setup_color_cpus(struct perf_sched *sched)
3159{
3160 struct cpu_map *map;
3161
3162 if (!sched->map.color_cpus_str)
3163 return 0;
3164
3165 map = cpu_map__new(sched->map.color_cpus_str);
3166 if (!map) {
3167 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3168 return -1;
3169 }
3170
3171 sched->map.color_cpus = map;
3172 return 0;
3173}
3174
0e9b07e5 3175static int perf_sched__map(struct perf_sched *sched)
0ec04e16 3176{
99623c62
JO		 3177 if (setup_map_cpus(sched))
3178 return -1;
40749d0f 3179
a151a37a
JO		 3180 if (setup_color_pids(sched))
3181 return -1;
3182
cf294f24
JO		 3183 if (setup_color_cpus(sched))
3184 return -1;
3185
0ec04e16 3186 setup_pager();
ae536acf 3187 if (perf_sched__read_events(sched))
a116e05d 3188 return -1;
0e9b07e5 3189 print_bad_events(sched);
a116e05d 3190 return 0;
0ec04e16
IM		 3191}
3192
0e9b07e5 3193static int perf_sched__replay(struct perf_sched *sched)
0ec04e16
IM		 3194{
3195 unsigned long i;
3196
0e9b07e5
ACM		 3197 calibrate_run_measurement_overhead(sched);
3198 calibrate_sleep_measurement_overhead(sched);
0ec04e16 3199
0e9b07e5 3200 test_calibrations(sched);
0ec04e16 3201
ae536acf 3202 if (perf_sched__read_events(sched))
a116e05d 3203 return -1;
0ec04e16 3204
0e9b07e5
ACM		 3205 printf("nr_run_events: %ld\n", sched->nr_run_events);
3206 printf("nr_sleep_events: %ld\n", sched->nr_sleep_events);
3207 printf("nr_wakeup_events: %ld\n", sched->nr_wakeup_events);
0ec04e16 3208
0e9b07e5
ACM		 3209 if (sched->targetless_wakeups)
3210 printf("target-less wakeups: %ld\n", sched->targetless_wakeups);
3211 if (sched->multitarget_wakeups)
3212 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3213 if (sched->nr_run_events_optimized)
0ec04e16 3214 printf("run atoms optimized: %ld\n",
0e9b07e5 3215 sched->nr_run_events_optimized);
0ec04e16 3216
0e9b07e5
ACM		 3217 print_task_traces(sched);
3218 add_cross_task_wakeups(sched);
0ec04e16 3219
0e9b07e5 3220 create_tasks(sched);
0ec04e16 3221 printf("------------------------------------------------------------\n");
0e9b07e5
ACM		 3222 for (i = 0; i < sched->replay_repeat; i++)
3223 run_one_test(sched);
a116e05d
ACM		 3224
3225 return 0;
0ec04e16
IM		 3226}
3227
0e9b07e5
ACM		 3228static void setup_sorting(struct perf_sched *sched, const struct option *options,
3229 const char * const usage_msg[])
daa1d7a5 3230{
0e9b07e5 3231 char *tmp, *tok, *str = strdup(sched->sort_order);
daa1d7a5
FW		 3232
3233 for (tok = strtok_r(str, ", ", &tmp);
3234 tok; tok = strtok_r(NULL, ", ", &tmp)) {
0e9b07e5 3235 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
c7118369
NK		 3236 usage_with_options_msg(usage_msg, options,
3237 "Unknown --sort key: `%s'", tok);
daa1d7a5
FW		 3238 }
3239 }
3240
3241 free(str);
3242
0e9b07e5 3243 sort_dimension__add("pid", &sched->cmp_pid);
daa1d7a5
FW		 3244}
3245
1fc35b29
IM		 3246static int __cmd_record(int argc, const char **argv)
3247{
3248 unsigned int rec_argc, i, j;
3249 const char **rec_argv;
0e9b07e5
ACM		 3250 const char * const record_args[] = {
3251 "record",
3252 "-a",
3253 "-R",
0e9b07e5
ACM		 3254 "-m", "1024",
3255 "-c", "1",
3256 "-e", "sched:sched_switch",
3257 "-e", "sched:sched_stat_wait",
3258 "-e", "sched:sched_stat_sleep",
3259 "-e", "sched:sched_stat_iowait",
3260 "-e", "sched:sched_stat_runtime",
0e9b07e5
ACM		 3261 "-e", "sched:sched_process_fork",
3262 "-e", "sched:sched_wakeup",
7fff9597 3263 "-e", "sched:sched_wakeup_new",
0e9b07e5
ACM		 3264 "-e", "sched:sched_migrate_task",
3265 };
1fc35b29
IM		 3266
3267 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
3268 rec_argv = calloc(rec_argc + 1, sizeof(char *));
3269
e462dc55 3270 if (rec_argv == NULL)
ce47dc56
CS		 3271 return -ENOMEM;
3272
1fc35b29
IM		 3273 for (i = 0; i < ARRAY_SIZE(record_args); i++)
3274 rec_argv[i] = strdup(record_args[i]);
3275
3276 for (j = 1; j < (unsigned int)argc; j++, i++)
3277 rec_argv[i] = argv[j];
3278
3279 BUG_ON(i != rec_argc);
3280
b0ad8ea6 3281 return cmd_record(i, rec_argv);
1fc35b29
IM		 3282}
3283
b0ad8ea6 3284int cmd_sched(int argc, const char **argv)
0a02ad93 3285{
8a39df8f
AH		 3286 const char default_sort_order[] = "avg, max, switch, runtime";
3287 struct perf_sched sched = {
3288 .tool = {
3289 .sample = perf_sched__process_tracepoint_sample,
3290 .comm = perf_event__process_comm,
f3b3614a 3291 .namespaces = perf_event__process_namespaces,
8a39df8f
AH		 3292 .lost = perf_event__process_lost,
3293 .fork = perf_sched__process_fork_event,
0a8cb85c 3294 .ordered_events = true,
8a39df8f
AH		 3295 },
3296 .cmp_pid = LIST_HEAD_INIT(sched.cmp_pid),
3297 .sort_list = LIST_HEAD_INIT(sched.sort_list),
3298 .start_work_mutex = PTHREAD_MUTEX_INITIALIZER,
3299 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
8a39df8f
AH		 3300 .sort_order = default_sort_order,
3301 .replay_repeat = 10,
3302 .profile_cpu = -1,
3303 .next_shortname1 = 'A',
3304 .next_shortname2 = '0',
2f80dd44 3305 .skip_merge = 0,
6c973c90
DA		 3306 .show_callchain = 1,
3307 .max_stack = 5,
8a39df8f 3308 };
77f02f44
NK		 3309 const struct option sched_options[] = {
3310 OPT_STRING('i', "input", &input_name, "file",
3311 "input file name"),
3312 OPT_INCR('v', "verbose", &verbose,
3313 "be more verbose (show symbol address, etc)"),
3314 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3315 "dump raw trace in ASCII"),
6fa94258 3316 OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
77f02f44
NK		 3317 OPT_END()
3318 };
0e9b07e5
ACM		 3319 const struct option latency_options[] = {
3320 OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3321 "sort by key(s): runtime, switch, avg, max"),
0e9b07e5
ACM		 3322 OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3323 "CPU to profile on"),
2f80dd44
JB		 3324 OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3325 "latency stats per pid instead of per comm"),
77f02f44 3326 OPT_PARENT(sched_options)
0e9b07e5
ACM		 3327 };
3328 const struct option replay_options[] = {
3329 OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3330 "repeat the workload replay N times (-1: infinite)"),
77f02f44 3331 OPT_PARENT(sched_options)
0e9b07e5 3332 };
99623c62
JO		 3333 const struct option map_options[] = {
3334 OPT_BOOLEAN(0, "compact", &sched.map.comp,
3335 "map output in compact mode"),
a151a37a
JO		 3336 OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3337 "highlight given pids in map"),
cf294f24
JO		 3338 OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3339 "highlight given CPUs in map"),
73643bb6
JO		 3340 OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3341 "display given CPUs in map"),
77f02f44 3342 OPT_PARENT(sched_options)
99623c62 3343 };
49394a2a
DA		 3344 const struct option timehist_options[] = {
3345 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3346 "file", "vmlinux pathname"),
3347 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3348 "file", "kallsyms pathname"),
6c973c90
DA		 3349 OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3350 "Display call chains if present (default on)"),
3351 OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3352 "Maximum number of functions to display backtrace."),
49394a2a
DA		 3353 OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3354 "Look for files with symbols relative to this directory"),
52df138c
DA		 3355 OPT_BOOLEAN('s', "summary", &sched.summary_only,
3356 "Show only syscall summary with statistics"),
3357 OPT_BOOLEAN('S', "with-summary", &sched.summary,
3358 "Show all syscalls and summary with statistics"),
fc1469f1 3359 OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
292c4a8f 3360 OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
350f54fa 3361 OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
a407b067 3362 OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
07235f84 3363 OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
853b7407
DA		 3364 OPT_STRING(0, "time", &sched.time_str, "str",
3365 "Time span for analysis (start,stop)"),
414e050c 3366 OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
49394a2a
DA		 3367 OPT_PARENT(sched_options)
3368 };
3369
0e9b07e5
ACM		 3370 const char * const latency_usage[] = {
3371 "perf sched latency [<options>]",
3372 NULL
3373 };
3374 const char * const replay_usage[] = {
3375 "perf sched replay [<options>]",
3376 NULL
3377 };
99623c62
JO		 3378 const char * const map_usage[] = {
3379 "perf sched map [<options>]",
3380 NULL
3381 };
49394a2a
DA		 3382 const char * const timehist_usage[] = {
3383 "perf sched timehist [<options>]",
3384 NULL
3385 };
a83edb2d 3386 const char *const sched_subcommands[] = { "record", "latency", "map",
49394a2a
DA		 3387 "replay", "script",
3388 "timehist", NULL };
a83edb2d
RR		 3389 const char *sched_usage[] = {
3390 NULL,
0e9b07e5
ACM		 3391 NULL
3392 };
3393 struct trace_sched_handler lat_ops = {
3394 .wakeup_event = latency_wakeup_event,
3395 .switch_event = latency_switch_event,
3396 .runtime_event = latency_runtime_event,
0e9b07e5
ACM		 3397 .migrate_task_event = latency_migrate_task_event,
3398 };
3399 struct trace_sched_handler map_ops = {
3400 .switch_event = map_switch_event,
3401 };
3402 struct trace_sched_handler replay_ops = {
3403 .wakeup_event = replay_wakeup_event,
3404 .switch_event = replay_switch_event,
3405 .fork_event = replay_fork_event,
3406 };
156a2b02
AH		 3407 unsigned int i;
3408
3409 for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3410 sched.curr_pid[i] = -1;
0e9b07e5 3411
a83edb2d
RR		 3412 argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3413 sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
f2858d8a
IM		 3414 if (!argc)
3415 usage_with_options(sched_usage, sched_options);
0a02ad93 3416
c0777c5a 3417 /*
133dc4c3 3418 * Aliased to 'perf script' for now:
c0777c5a 3419 */
133dc4c3 3420 if (!strcmp(argv[0], "script"))
b0ad8ea6 3421 return cmd_script(argc, argv);
c0777c5a 3422
1fc35b29
IM		 3423 if (!strncmp(argv[0], "rec", 3)) {
3424 return __cmd_record(argc, argv);
3425 } else if (!strncmp(argv[0], "lat", 3)) {
0e9b07e5 3426 sched.tp_handler = &lat_ops;
f2858d8a
IM		 3427 if (argc > 1) {
3428 argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3429 if (argc)
3430 usage_with_options(latency_usage, latency_options);
f2858d8a 3431 }
0e9b07e5
ACM		 3432 setup_sorting(&sched, latency_options, latency_usage);
3433 return perf_sched__lat(&sched);
0ec04e16 3434 } else if (!strcmp(argv[0], "map")) {
99623c62 3435 if (argc) {
a151a37a 3436 argc = parse_options(argc, argv, map_options, map_usage, 0);
99623c62
JO		 3437 if (argc)
3438 usage_with_options(map_usage, map_options);
3439 }
0e9b07e5
ACM		 3440 sched.tp_handler = &map_ops;
3441 setup_sorting(&sched, latency_options, latency_usage);
3442 return perf_sched__map(&sched);
f2858d8a 3443 } else if (!strncmp(argv[0], "rep", 3)) {
0e9b07e5 3444 sched.tp_handler = &replay_ops;
f2858d8a
IM		 3445 if (argc) {
3446 argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3447 if (argc)
3448 usage_with_options(replay_usage, replay_options);
3449 }
0e9b07e5 3450 return perf_sched__replay(&sched);
49394a2a
DA		 3451 } else if (!strcmp(argv[0], "timehist")) {
3452 if (argc) {
3453 argc = parse_options(argc, argv, timehist_options,
3454 timehist_usage, 0);
3455 if (argc)
3456 usage_with_options(timehist_usage, timehist_options);
3457 }
292c4a8f
BG		 3458 if ((sched.show_wakeups || sched.show_next) &&
3459 sched.summary_only) {
3460 pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
fc1469f1 3461 parse_options_usage(timehist_usage, timehist_options, "s", true);
292c4a8f
BG		 3462 if (sched.show_wakeups)
3463 parse_options_usage(NULL, timehist_options, "w", true);
3464 if (sched.show_next)
3465 parse_options_usage(NULL, timehist_options, "n", true);
fc1469f1
DA		 3466 return -EINVAL;
3467 }
3468
49394a2a 3469 return perf_sched__timehist(&sched);
f2858d8a
IM		 3470 } else {
3471 usage_with_options(sched_usage, sched_options);
3472 }
3473
ec156764 3474 return 0;
0a02ad93 3475}
Linux 6.x block layer and io_uring tree(s)