]>
Commit | Line | Data |
---|---|---|
0a02ad93 IM |
1 | #include "builtin.h" |
2 | ||
3 | #include "util/util.h" | |
4 | #include "util/cache.h" | |
5 | #include "util/symbol.h" | |
6 | #include "util/thread.h" | |
7 | #include "util/header.h" | |
8 | ||
9 | #include "util/parse-options.h" | |
10 | ||
11 | #include "perf.h" | |
12 | #include "util/debug.h" | |
13 | ||
14 | #include "util/trace-event.h" | |
ec156764 | 15 | #include <sys/types.h> |
0a02ad93 | 16 | |
ec156764 IM |
17 | static char const *input_name = "perf.data"; |
18 | static int input; | |
19 | static unsigned long page_size; | |
20 | static unsigned long mmap_window = 32; | |
0a02ad93 | 21 | |
ec156764 | 22 | static unsigned long total_comm = 0; |
0a02ad93 | 23 | |
ec156764 IM |
24 | static struct rb_root threads; |
25 | static struct thread *last_match; | |
0a02ad93 | 26 | |
ec156764 IM |
27 | static struct perf_header *header; |
28 | static u64 sample_type; | |
0a02ad93 IM |
29 | |
30 | ||
ec156764 IM |
31 | /* |
32 | * Scheduler benchmarks | |
33 | */ | |
34 | #include <sys/resource.h> | |
35 | #include <sys/types.h> | |
36 | #include <sys/stat.h> | |
37 | #include <sys/time.h> | |
38 | #include <sys/prctl.h> | |
39 | ||
40 | #include <linux/unistd.h> | |
41 | ||
42 | #include <semaphore.h> | |
43 | #include <pthread.h> | |
44 | #include <signal.h> | |
45 | #include <values.h> | |
46 | #include <string.h> | |
47 | #include <unistd.h> | |
48 | #include <stdlib.h> | |
49 | #include <assert.h> | |
50 | #include <fcntl.h> | |
51 | #include <time.h> | |
52 | #include <math.h> | |
53 | ||
54 | #include <stdio.h> | |
55 | ||
56 | #define PR_SET_NAME 15 /* Set process name */ | |
57 | ||
58 | #define BUG_ON(x) assert(!(x)) | |
59 | ||
fbf94829 | 60 | #define DEBUG 0 |
ec156764 IM |
61 | |
62 | typedef unsigned long long nsec_t; | |
63 | ||
ec156764 IM |
64 | static nsec_t run_measurement_overhead; |
65 | static nsec_t sleep_measurement_overhead; | |
66 | ||
67 | static nsec_t get_nsecs(void) | |
68 | { | |
69 | struct timespec ts; | |
70 | ||
71 | clock_gettime(CLOCK_MONOTONIC, &ts); | |
72 | ||
73 | return ts.tv_sec * 1000000000ULL + ts.tv_nsec; | |
74 | } | |
75 | ||
76 | static void burn_nsecs(nsec_t nsecs) | |
77 | { | |
78 | nsec_t T0 = get_nsecs(), T1; | |
79 | ||
80 | do { | |
81 | T1 = get_nsecs(); | |
82 | } while (T1 + run_measurement_overhead < T0 + nsecs); | |
83 | } | |
84 | ||
85 | static void sleep_nsecs(nsec_t nsecs) | |
86 | { | |
87 | struct timespec ts; | |
88 | ||
89 | ts.tv_nsec = nsecs % 999999999; | |
90 | ts.tv_sec = nsecs / 999999999; | |
91 | ||
92 | nanosleep(&ts, NULL); | |
93 | } | |
94 | ||
95 | static void calibrate_run_measurement_overhead(void) | |
96 | { | |
97 | nsec_t T0, T1, delta, min_delta = 1000000000ULL; | |
98 | int i; | |
99 | ||
100 | for (i = 0; i < 10; i++) { | |
101 | T0 = get_nsecs(); | |
102 | burn_nsecs(0); | |
103 | T1 = get_nsecs(); | |
104 | delta = T1-T0; | |
105 | min_delta = min(min_delta, delta); | |
106 | } | |
107 | run_measurement_overhead = min_delta; | |
108 | ||
ad236fd2 | 109 | printf("run measurement overhead: %Ld nsecs\n", min_delta); |
ec156764 IM |
110 | } |
111 | ||
112 | static void calibrate_sleep_measurement_overhead(void) | |
113 | { | |
114 | nsec_t T0, T1, delta, min_delta = 1000000000ULL; | |
115 | int i; | |
116 | ||
117 | for (i = 0; i < 10; i++) { | |
118 | T0 = get_nsecs(); | |
119 | sleep_nsecs(10000); | |
120 | T1 = get_nsecs(); | |
121 | delta = T1-T0; | |
122 | min_delta = min(min_delta, delta); | |
123 | } | |
124 | min_delta -= 10000; | |
125 | sleep_measurement_overhead = min_delta; | |
126 | ||
ad236fd2 | 127 | printf("sleep measurement overhead: %Ld nsecs\n", min_delta); |
ec156764 IM |
128 | } |
129 | ||
130 | #define COMM_LEN 20 | |
131 | #define SYM_LEN 129 | |
132 | ||
133 | #define MAX_PID 65536 | |
134 | ||
135 | static unsigned long nr_tasks; | |
136 | ||
137 | struct sched_event; | |
138 | ||
139 | struct task_desc { | |
140 | unsigned long nr; | |
141 | unsigned long pid; | |
142 | char comm[COMM_LEN]; | |
143 | ||
144 | unsigned long nr_events; | |
145 | unsigned long curr_event; | |
146 | struct sched_event **events; | |
147 | ||
148 | pthread_t thread; | |
149 | sem_t sleep_sem; | |
150 | ||
151 | sem_t ready_for_work; | |
152 | sem_t work_done_sem; | |
153 | ||
154 | nsec_t cpu_usage; | |
155 | }; | |
156 | ||
157 | enum sched_event_type { | |
158 | SCHED_EVENT_RUN, | |
159 | SCHED_EVENT_SLEEP, | |
160 | SCHED_EVENT_WAKEUP, | |
161 | }; | |
162 | ||
163 | struct sched_event { | |
164 | enum sched_event_type type; | |
165 | nsec_t timestamp; | |
166 | nsec_t duration; | |
167 | unsigned long nr; | |
168 | int specific_wait; | |
169 | sem_t *wait_sem; | |
170 | struct task_desc *wakee; | |
171 | }; | |
172 | ||
173 | static struct task_desc *pid_to_task[MAX_PID]; | |
174 | ||
175 | static struct task_desc **tasks; | |
176 | ||
177 | static pthread_mutex_t start_work_mutex = PTHREAD_MUTEX_INITIALIZER; | |
178 | static nsec_t start_time; | |
179 | ||
180 | static pthread_mutex_t work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER; | |
181 | ||
182 | static unsigned long nr_run_events; | |
183 | static unsigned long nr_sleep_events; | |
184 | static unsigned long nr_wakeup_events; | |
185 | ||
186 | static unsigned long nr_sleep_corrections; | |
187 | static unsigned long nr_run_events_optimized; | |
188 | ||
189 | static struct sched_event * | |
190 | get_new_event(struct task_desc *task, nsec_t timestamp) | |
191 | { | |
192 | struct sched_event *event = calloc(1, sizeof(*event)); | |
193 | unsigned long idx = task->nr_events; | |
194 | size_t size; | |
195 | ||
196 | event->timestamp = timestamp; | |
197 | event->nr = idx; | |
198 | ||
199 | task->nr_events++; | |
200 | size = sizeof(struct sched_event *) * task->nr_events; | |
201 | task->events = realloc(task->events, size); | |
202 | BUG_ON(!task->events); | |
203 | ||
204 | task->events[idx] = event; | |
205 | ||
206 | return event; | |
207 | } | |
208 | ||
209 | static struct sched_event *last_event(struct task_desc *task) | |
210 | { | |
211 | if (!task->nr_events) | |
212 | return NULL; | |
213 | ||
214 | return task->events[task->nr_events - 1]; | |
215 | } | |
216 | ||
217 | static void | |
fbf94829 | 218 | add_sched_event_run(struct task_desc *task, nsec_t timestamp, u64 duration) |
ec156764 IM |
219 | { |
220 | struct sched_event *event, *curr_event = last_event(task); | |
221 | ||
222 | /* | |
fbf94829 IM |
223 | * optimize an existing RUN event by merging this one |
224 | * to it: | |
225 | */ | |
ec156764 IM |
226 | if (curr_event && curr_event->type == SCHED_EVENT_RUN) { |
227 | nr_run_events_optimized++; | |
228 | curr_event->duration += duration; | |
229 | return; | |
230 | } | |
231 | ||
232 | event = get_new_event(task, timestamp); | |
233 | ||
234 | event->type = SCHED_EVENT_RUN; | |
235 | event->duration = duration; | |
236 | ||
237 | nr_run_events++; | |
238 | } | |
239 | ||
240 | static unsigned long targetless_wakeups; | |
241 | static unsigned long multitarget_wakeups; | |
242 | ||
243 | static void | |
244 | add_sched_event_wakeup(struct task_desc *task, nsec_t timestamp, | |
245 | struct task_desc *wakee) | |
246 | { | |
247 | struct sched_event *event, *wakee_event; | |
248 | ||
249 | event = get_new_event(task, timestamp); | |
250 | event->type = SCHED_EVENT_WAKEUP; | |
251 | event->wakee = wakee; | |
252 | ||
253 | wakee_event = last_event(wakee); | |
254 | if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) { | |
255 | targetless_wakeups++; | |
256 | return; | |
257 | } | |
258 | if (wakee_event->wait_sem) { | |
259 | multitarget_wakeups++; | |
260 | return; | |
261 | } | |
262 | ||
263 | wakee_event->wait_sem = calloc(1, sizeof(*wakee_event->wait_sem)); | |
264 | sem_init(wakee_event->wait_sem, 0, 0); | |
265 | wakee_event->specific_wait = 1; | |
266 | event->wait_sem = wakee_event->wait_sem; | |
267 | ||
268 | nr_wakeup_events++; | |
269 | } | |
270 | ||
271 | static void | |
272 | add_sched_event_sleep(struct task_desc *task, nsec_t timestamp, | |
ad236fd2 | 273 | u64 task_state __used) |
ec156764 IM |
274 | { |
275 | struct sched_event *event = get_new_event(task, timestamp); | |
276 | ||
277 | event->type = SCHED_EVENT_SLEEP; | |
278 | ||
279 | nr_sleep_events++; | |
280 | } | |
281 | ||
282 | static struct task_desc *register_pid(unsigned long pid, const char *comm) | |
283 | { | |
284 | struct task_desc *task; | |
285 | ||
286 | BUG_ON(pid >= MAX_PID); | |
287 | ||
288 | task = pid_to_task[pid]; | |
289 | ||
290 | if (task) | |
291 | return task; | |
292 | ||
293 | task = calloc(1, sizeof(*task)); | |
294 | task->pid = pid; | |
295 | task->nr = nr_tasks; | |
296 | strcpy(task->comm, comm); | |
297 | /* | |
298 | * every task starts in sleeping state - this gets ignored | |
299 | * if there's no wakeup pointing to this sleep state: | |
300 | */ | |
301 | add_sched_event_sleep(task, 0, 0); | |
302 | ||
303 | pid_to_task[pid] = task; | |
304 | nr_tasks++; | |
305 | tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *)); | |
306 | BUG_ON(!tasks); | |
307 | tasks[task->nr] = task; | |
308 | ||
ad236fd2 IM |
309 | if (verbose) |
310 | printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm); | |
ec156764 IM |
311 | |
312 | return task; | |
313 | } | |
314 | ||
315 | ||
ec156764 IM |
316 | static void print_task_traces(void) |
317 | { | |
318 | struct task_desc *task; | |
319 | unsigned long i; | |
320 | ||
321 | for (i = 0; i < nr_tasks; i++) { | |
322 | task = tasks[i]; | |
ad236fd2 | 323 | printf("task %6ld (%20s:%10ld), nr_events: %ld\n", |
ec156764 IM |
324 | task->nr, task->comm, task->pid, task->nr_events); |
325 | } | |
326 | } | |
327 | ||
328 | static void add_cross_task_wakeups(void) | |
329 | { | |
330 | struct task_desc *task1, *task2; | |
331 | unsigned long i, j; | |
332 | ||
333 | for (i = 0; i < nr_tasks; i++) { | |
334 | task1 = tasks[i]; | |
335 | j = i + 1; | |
336 | if (j == nr_tasks) | |
337 | j = 0; | |
338 | task2 = tasks[j]; | |
339 | add_sched_event_wakeup(task1, 0, task2); | |
340 | } | |
341 | } | |
342 | ||
343 | static void | |
fbf94829 | 344 | process_sched_event(struct task_desc *this_task __used, struct sched_event *event) |
ec156764 IM |
345 | { |
346 | int ret = 0; | |
347 | nsec_t now; | |
348 | long long delta; | |
349 | ||
350 | now = get_nsecs(); | |
351 | delta = start_time + event->timestamp - now; | |
352 | ||
ec156764 IM |
353 | switch (event->type) { |
354 | case SCHED_EVENT_RUN: | |
ec156764 IM |
355 | burn_nsecs(event->duration); |
356 | break; | |
357 | case SCHED_EVENT_SLEEP: | |
ec156764 IM |
358 | if (event->wait_sem) |
359 | ret = sem_wait(event->wait_sem); | |
360 | BUG_ON(ret); | |
361 | break; | |
362 | case SCHED_EVENT_WAKEUP: | |
ec156764 IM |
363 | if (event->wait_sem) |
364 | ret = sem_post(event->wait_sem); | |
365 | BUG_ON(ret); | |
366 | break; | |
367 | default: | |
368 | BUG_ON(1); | |
369 | } | |
370 | } | |
371 | ||
372 | static nsec_t get_cpu_usage_nsec_parent(void) | |
373 | { | |
374 | struct rusage ru; | |
375 | nsec_t sum; | |
376 | int err; | |
377 | ||
378 | err = getrusage(RUSAGE_SELF, &ru); | |
379 | BUG_ON(err); | |
380 | ||
381 | sum = ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3; | |
382 | sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3; | |
383 | ||
384 | return sum; | |
385 | } | |
386 | ||
387 | static nsec_t get_cpu_usage_nsec_self(void) | |
388 | { | |
389 | char filename [] = "/proc/1234567890/sched"; | |
390 | unsigned long msecs, nsecs; | |
391 | char *line = NULL; | |
392 | nsec_t total = 0; | |
393 | size_t len = 0; | |
394 | ssize_t chars; | |
395 | FILE *file; | |
396 | int ret; | |
397 | ||
398 | sprintf(filename, "/proc/%d/sched", getpid()); | |
399 | file = fopen(filename, "r"); | |
400 | BUG_ON(!file); | |
401 | ||
402 | while ((chars = getline(&line, &len, file)) != -1) { | |
ec156764 IM |
403 | ret = sscanf(line, "se.sum_exec_runtime : %ld.%06ld\n", |
404 | &msecs, &nsecs); | |
405 | if (ret == 2) { | |
406 | total = msecs*1e6 + nsecs; | |
ec156764 IM |
407 | break; |
408 | } | |
409 | } | |
410 | if (line) | |
411 | free(line); | |
412 | fclose(file); | |
413 | ||
414 | return total; | |
415 | } | |
416 | ||
417 | static void *thread_func(void *ctx) | |
418 | { | |
419 | struct task_desc *this_task = ctx; | |
420 | nsec_t cpu_usage_0, cpu_usage_1; | |
421 | unsigned long i, ret; | |
422 | char comm2[22]; | |
423 | ||
ec156764 IM |
424 | sprintf(comm2, ":%s", this_task->comm); |
425 | prctl(PR_SET_NAME, comm2); | |
426 | ||
427 | again: | |
428 | ret = sem_post(&this_task->ready_for_work); | |
429 | BUG_ON(ret); | |
ec156764 IM |
430 | ret = pthread_mutex_lock(&start_work_mutex); |
431 | BUG_ON(ret); | |
432 | ret = pthread_mutex_unlock(&start_work_mutex); | |
433 | BUG_ON(ret); | |
ec156764 IM |
434 | |
435 | cpu_usage_0 = get_cpu_usage_nsec_self(); | |
436 | ||
437 | for (i = 0; i < this_task->nr_events; i++) { | |
438 | this_task->curr_event = i; | |
439 | process_sched_event(this_task, this_task->events[i]); | |
440 | } | |
441 | ||
442 | cpu_usage_1 = get_cpu_usage_nsec_self(); | |
443 | this_task->cpu_usage = cpu_usage_1 - cpu_usage_0; | |
444 | ||
ec156764 IM |
445 | ret = sem_post(&this_task->work_done_sem); |
446 | BUG_ON(ret); | |
ec156764 IM |
447 | |
448 | ret = pthread_mutex_lock(&work_done_wait_mutex); | |
449 | BUG_ON(ret); | |
450 | ret = pthread_mutex_unlock(&work_done_wait_mutex); | |
451 | BUG_ON(ret); | |
ec156764 IM |
452 | |
453 | goto again; | |
454 | } | |
455 | ||
456 | static void create_tasks(void) | |
457 | { | |
458 | struct task_desc *task; | |
459 | pthread_attr_t attr; | |
460 | unsigned long i; | |
461 | int err; | |
462 | ||
463 | err = pthread_attr_init(&attr); | |
464 | BUG_ON(err); | |
465 | err = pthread_attr_setstacksize(&attr, (size_t)(16*1024)); | |
466 | BUG_ON(err); | |
467 | err = pthread_mutex_lock(&start_work_mutex); | |
468 | BUG_ON(err); | |
469 | err = pthread_mutex_lock(&work_done_wait_mutex); | |
470 | BUG_ON(err); | |
471 | for (i = 0; i < nr_tasks; i++) { | |
472 | task = tasks[i]; | |
473 | sem_init(&task->sleep_sem, 0, 0); | |
474 | sem_init(&task->ready_for_work, 0, 0); | |
475 | sem_init(&task->work_done_sem, 0, 0); | |
476 | task->curr_event = 0; | |
477 | err = pthread_create(&task->thread, &attr, thread_func, task); | |
478 | BUG_ON(err); | |
479 | } | |
480 | } | |
481 | ||
482 | static nsec_t cpu_usage; | |
483 | static nsec_t runavg_cpu_usage; | |
484 | static nsec_t parent_cpu_usage; | |
485 | static nsec_t runavg_parent_cpu_usage; | |
486 | ||
487 | static void wait_for_tasks(void) | |
488 | { | |
489 | nsec_t cpu_usage_0, cpu_usage_1; | |
490 | struct task_desc *task; | |
491 | unsigned long i, ret; | |
492 | ||
ec156764 | 493 | start_time = get_nsecs(); |
ec156764 IM |
494 | cpu_usage = 0; |
495 | pthread_mutex_unlock(&work_done_wait_mutex); | |
496 | ||
497 | for (i = 0; i < nr_tasks; i++) { | |
498 | task = tasks[i]; | |
499 | ret = sem_wait(&task->ready_for_work); | |
500 | BUG_ON(ret); | |
501 | sem_init(&task->ready_for_work, 0, 0); | |
502 | } | |
503 | ret = pthread_mutex_lock(&work_done_wait_mutex); | |
504 | BUG_ON(ret); | |
505 | ||
506 | cpu_usage_0 = get_cpu_usage_nsec_parent(); | |
507 | ||
508 | pthread_mutex_unlock(&start_work_mutex); | |
509 | ||
ec156764 IM |
510 | for (i = 0; i < nr_tasks; i++) { |
511 | task = tasks[i]; | |
512 | ret = sem_wait(&task->work_done_sem); | |
513 | BUG_ON(ret); | |
514 | sem_init(&task->work_done_sem, 0, 0); | |
515 | cpu_usage += task->cpu_usage; | |
516 | task->cpu_usage = 0; | |
517 | } | |
518 | ||
519 | cpu_usage_1 = get_cpu_usage_nsec_parent(); | |
520 | if (!runavg_cpu_usage) | |
521 | runavg_cpu_usage = cpu_usage; | |
522 | runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10; | |
523 | ||
524 | parent_cpu_usage = cpu_usage_1 - cpu_usage_0; | |
525 | if (!runavg_parent_cpu_usage) | |
526 | runavg_parent_cpu_usage = parent_cpu_usage; | |
527 | runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 + | |
528 | parent_cpu_usage)/10; | |
529 | ||
530 | ret = pthread_mutex_lock(&start_work_mutex); | |
531 | BUG_ON(ret); | |
532 | ||
533 | for (i = 0; i < nr_tasks; i++) { | |
534 | task = tasks[i]; | |
535 | sem_init(&task->sleep_sem, 0, 0); | |
536 | task->curr_event = 0; | |
537 | } | |
538 | } | |
539 | ||
540 | static int __cmd_sched(void); | |
541 | ||
542 | static void parse_trace(void) | |
543 | { | |
544 | __cmd_sched(); | |
545 | ||
ad236fd2 IM |
546 | printf("nr_run_events: %ld\n", nr_run_events); |
547 | printf("nr_sleep_events: %ld\n", nr_sleep_events); | |
548 | printf("nr_wakeup_events: %ld\n", nr_wakeup_events); | |
ec156764 IM |
549 | |
550 | if (targetless_wakeups) | |
ad236fd2 | 551 | printf("target-less wakeups: %ld\n", targetless_wakeups); |
ec156764 | 552 | if (multitarget_wakeups) |
ad236fd2 | 553 | printf("multi-target wakeups: %ld\n", multitarget_wakeups); |
ec156764 | 554 | if (nr_run_events_optimized) |
ad236fd2 | 555 | printf("run events optimized: %ld\n", |
ec156764 IM |
556 | nr_run_events_optimized); |
557 | } | |
558 | ||
559 | static unsigned long nr_runs; | |
560 | static nsec_t sum_runtime; | |
561 | static nsec_t sum_fluct; | |
562 | static nsec_t run_avg; | |
563 | ||
564 | static void run_one_test(void) | |
565 | { | |
566 | nsec_t T0, T1, delta, avg_delta, fluct, std_dev; | |
567 | ||
568 | T0 = get_nsecs(); | |
569 | wait_for_tasks(); | |
570 | T1 = get_nsecs(); | |
571 | ||
572 | delta = T1 - T0; | |
573 | sum_runtime += delta; | |
574 | nr_runs++; | |
575 | ||
576 | avg_delta = sum_runtime / nr_runs; | |
577 | if (delta < avg_delta) | |
578 | fluct = avg_delta - delta; | |
579 | else | |
580 | fluct = delta - avg_delta; | |
581 | sum_fluct += fluct; | |
582 | std_dev = sum_fluct / nr_runs / sqrt(nr_runs); | |
583 | if (!run_avg) | |
584 | run_avg = delta; | |
585 | run_avg = (run_avg*9 + delta)/10; | |
586 | ||
ad236fd2 | 587 | printf("#%-3ld: %0.3f, ", |
ec156764 IM |
588 | nr_runs, (double)delta/1000000.0); |
589 | ||
590 | #if 0 | |
ad236fd2 | 591 | printf("%0.2f +- %0.2f, ", |
ec156764 IM |
592 | (double)avg_delta/1e6, (double)std_dev/1e6); |
593 | #endif | |
ad236fd2 | 594 | printf("ravg: %0.2f, ", |
ec156764 IM |
595 | (double)run_avg/1e6); |
596 | ||
ad236fd2 | 597 | printf("cpu: %0.2f / %0.2f", |
ec156764 IM |
598 | (double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6); |
599 | ||
600 | #if 0 | |
601 | /* | |
fbf94829 IM |
602 | * rusage statistics done by the parent, these are less |
603 | * accurate than the sum_exec_runtime based statistics: | |
604 | */ | |
ad236fd2 | 605 | printf(" [%0.2f / %0.2f]", |
ec156764 IM |
606 | (double)parent_cpu_usage/1e6, |
607 | (double)runavg_parent_cpu_usage/1e6); | |
608 | #endif | |
609 | ||
ad236fd2 | 610 | printf("\n"); |
ec156764 IM |
611 | |
612 | if (nr_sleep_corrections) | |
ad236fd2 | 613 | printf(" (%ld sleep corrections)\n", nr_sleep_corrections); |
ec156764 IM |
614 | nr_sleep_corrections = 0; |
615 | } | |
616 | ||
617 | static void test_calibrations(void) | |
618 | { | |
619 | nsec_t T0, T1; | |
620 | ||
621 | T0 = get_nsecs(); | |
622 | burn_nsecs(1e6); | |
623 | T1 = get_nsecs(); | |
624 | ||
ad236fd2 | 625 | printf("the run test took %Ld nsecs\n", T1-T0); |
ec156764 IM |
626 | |
627 | T0 = get_nsecs(); | |
628 | sleep_nsecs(1e6); | |
629 | T1 = get_nsecs(); | |
630 | ||
ad236fd2 | 631 | printf("the sleep test took %Ld nsecs\n", T1-T0); |
ec156764 IM |
632 | } |
633 | ||
0a02ad93 IM |
634 | static int |
635 | process_comm_event(event_t *event, unsigned long offset, unsigned long head) | |
636 | { | |
637 | struct thread *thread; | |
638 | ||
639 | thread = threads__findnew(event->comm.pid, &threads, &last_match); | |
640 | ||
641 | dump_printf("%p [%p]: PERF_EVENT_COMM: %s:%d\n", | |
642 | (void *)(offset + head), | |
643 | (void *)(long)(event->header.size), | |
644 | event->comm.comm, event->comm.pid); | |
645 | ||
646 | if (thread == NULL || | |
647 | thread__set_comm(thread, event->comm.comm)) { | |
648 | dump_printf("problem processing PERF_EVENT_COMM, skipping event.\n"); | |
649 | return -1; | |
650 | } | |
651 | total_comm++; | |
652 | ||
653 | return 0; | |
654 | } | |
655 | ||
fbf94829 IM |
656 | struct trace_wakeup_event { |
657 | u32 size; | |
658 | ||
659 | u16 common_type; | |
660 | u8 common_flags; | |
661 | u8 common_preempt_count; | |
662 | u32 common_pid; | |
663 | u32 common_tgid; | |
664 | ||
665 | char comm[16]; | |
666 | u32 pid; | |
667 | ||
668 | u32 prio; | |
669 | u32 success; | |
670 | u32 cpu; | |
671 | }; | |
672 | ||
673 | static void | |
674 | process_sched_wakeup_event(struct trace_wakeup_event *wakeup_event, struct event *event, | |
ec156764 IM |
675 | int cpu __used, u64 timestamp __used, struct thread *thread __used) |
676 | { | |
fbf94829 IM |
677 | struct task_desc *waker, *wakee; |
678 | ||
ad236fd2 IM |
679 | if (verbose) { |
680 | printf("sched_wakeup event %p\n", event); | |
fbf94829 | 681 | |
ad236fd2 IM |
682 | printf(" ... pid %d woke up %s/%d\n", |
683 | wakeup_event->common_pid, | |
684 | wakeup_event->comm, | |
685 | wakeup_event->pid); | |
686 | } | |
fbf94829 IM |
687 | |
688 | waker = register_pid(wakeup_event->common_pid, "<unknown>"); | |
689 | wakee = register_pid(wakeup_event->pid, wakeup_event->comm); | |
690 | ||
691 | add_sched_event_wakeup(waker, timestamp, wakee); | |
ec156764 IM |
692 | } |
693 | ||
fbf94829 IM |
694 | struct trace_switch_event { |
695 | u32 size; | |
696 | ||
697 | u16 common_type; | |
698 | u8 common_flags; | |
699 | u8 common_preempt_count; | |
700 | u32 common_pid; | |
701 | u32 common_tgid; | |
702 | ||
703 | char prev_comm[16]; | |
704 | u32 prev_pid; | |
705 | u32 prev_prio; | |
706 | u64 prev_state; | |
707 | char next_comm[16]; | |
708 | u32 next_pid; | |
709 | u32 next_prio; | |
710 | }; | |
711 | ||
712 | #define MAX_CPUS 4096 | |
713 | ||
714 | unsigned long cpu_last_switched[MAX_CPUS]; | |
715 | ||
716 | static void | |
717 | process_sched_switch_event(struct trace_switch_event *switch_event, struct event *event, | |
ec156764 IM |
718 | int cpu __used, u64 timestamp __used, struct thread *thread __used) |
719 | { | |
fbf94829 IM |
720 | struct task_desc *prev, *next; |
721 | u64 timestamp0; | |
722 | s64 delta; | |
723 | ||
ad236fd2 IM |
724 | if (verbose) |
725 | printf("sched_switch event %p\n", event); | |
726 | ||
fbf94829 IM |
727 | if (cpu >= MAX_CPUS || cpu < 0) |
728 | return; | |
729 | ||
730 | timestamp0 = cpu_last_switched[cpu]; | |
731 | if (timestamp0) | |
732 | delta = timestamp - timestamp0; | |
733 | else | |
734 | delta = 0; | |
735 | ||
736 | if (delta < 0) | |
737 | die("hm, delta: %Ld < 0 ?\n", delta); | |
738 | ||
ad236fd2 IM |
739 | if (verbose) { |
740 | printf(" ... switch from %s/%d to %s/%d [ran %Ld nsecs]\n", | |
741 | switch_event->prev_comm, switch_event->prev_pid, | |
742 | switch_event->next_comm, switch_event->next_pid, | |
743 | delta); | |
744 | } | |
fbf94829 IM |
745 | |
746 | prev = register_pid(switch_event->prev_pid, switch_event->prev_comm); | |
747 | next = register_pid(switch_event->next_pid, switch_event->next_comm); | |
748 | ||
749 | cpu_last_switched[cpu] = timestamp; | |
750 | ||
751 | add_sched_event_run(prev, timestamp, delta); | |
ad236fd2 | 752 | add_sched_event_sleep(prev, timestamp, switch_event->prev_state); |
fbf94829 IM |
753 | } |
754 | ||
755 | struct trace_fork_event { | |
756 | u32 size; | |
757 | ||
758 | u16 common_type; | |
759 | u8 common_flags; | |
760 | u8 common_preempt_count; | |
761 | u32 common_pid; | |
762 | u32 common_tgid; | |
763 | ||
764 | char parent_comm[16]; | |
765 | u32 parent_pid; | |
766 | char child_comm[16]; | |
767 | u32 child_pid; | |
768 | }; | |
769 | ||
770 | static void | |
771 | process_sched_fork_event(struct trace_fork_event *fork_event, struct event *event, | |
772 | int cpu __used, u64 timestamp __used, struct thread *thread __used) | |
773 | { | |
ad236fd2 IM |
774 | if (verbose) { |
775 | printf("sched_fork event %p\n", event); | |
776 | printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid); | |
777 | printf("... child: %s/%d\n", fork_event->child_comm, fork_event->child_pid); | |
778 | } | |
fbf94829 IM |
779 | register_pid(fork_event->parent_pid, fork_event->parent_comm); |
780 | register_pid(fork_event->child_pid, fork_event->child_comm); | |
781 | } | |
782 | ||
783 | static void process_sched_exit_event(struct event *event, | |
784 | int cpu __used, u64 timestamp __used, struct thread *thread __used) | |
785 | { | |
ad236fd2 IM |
786 | if (verbose) |
787 | printf("sched_exit event %p\n", event); | |
ec156764 IM |
788 | } |
789 | ||
790 | static void | |
ad236fd2 | 791 | process_raw_event(event_t *raw_event __used, void *more_data, |
ec156764 IM |
792 | int cpu, u64 timestamp, struct thread *thread) |
793 | { | |
794 | struct { | |
795 | u32 size; | |
796 | char data[0]; | |
797 | } *raw = more_data; | |
798 | struct event *event; | |
799 | int type; | |
800 | ||
801 | type = trace_parse_common_type(raw->data); | |
802 | event = trace_find_event(type); | |
803 | ||
ec156764 | 804 | if (!strcmp(event->name, "sched_switch")) |
fbf94829 | 805 | process_sched_switch_event(more_data, event, cpu, timestamp, thread); |
ec156764 | 806 | if (!strcmp(event->name, "sched_wakeup")) |
fbf94829 IM |
807 | process_sched_wakeup_event(more_data, event, cpu, timestamp, thread); |
808 | if (!strcmp(event->name, "sched_wakeup_new")) | |
809 | process_sched_wakeup_event(more_data, event, cpu, timestamp, thread); | |
810 | if (!strcmp(event->name, "sched_process_fork")) | |
811 | process_sched_fork_event(more_data, event, cpu, timestamp, thread); | |
812 | if (!strcmp(event->name, "sched_process_exit")) | |
813 | process_sched_exit_event(event, cpu, timestamp, thread); | |
ec156764 IM |
814 | } |
815 | ||
0a02ad93 IM |
816 | static int |
817 | process_sample_event(event_t *event, unsigned long offset, unsigned long head) | |
818 | { | |
819 | char level; | |
820 | int show = 0; | |
821 | struct dso *dso = NULL; | |
822 | struct thread *thread; | |
823 | u64 ip = event->ip.ip; | |
824 | u64 timestamp = -1; | |
825 | u32 cpu = -1; | |
826 | u64 period = 1; | |
827 | void *more_data = event->ip.__more_data; | |
828 | int cpumode; | |
829 | ||
830 | thread = threads__findnew(event->ip.pid, &threads, &last_match); | |
831 | ||
832 | if (sample_type & PERF_SAMPLE_TIME) { | |
833 | timestamp = *(u64 *)more_data; | |
834 | more_data += sizeof(u64); | |
835 | } | |
836 | ||
837 | if (sample_type & PERF_SAMPLE_CPU) { | |
838 | cpu = *(u32 *)more_data; | |
839 | more_data += sizeof(u32); | |
840 | more_data += sizeof(u32); /* reserved */ | |
841 | } | |
842 | ||
843 | if (sample_type & PERF_SAMPLE_PERIOD) { | |
844 | period = *(u64 *)more_data; | |
845 | more_data += sizeof(u64); | |
846 | } | |
847 | ||
848 | dump_printf("%p [%p]: PERF_EVENT_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n", | |
849 | (void *)(offset + head), | |
850 | (void *)(long)(event->header.size), | |
851 | event->header.misc, | |
852 | event->ip.pid, event->ip.tid, | |
853 | (void *)(long)ip, | |
854 | (long long)period); | |
855 | ||
856 | dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid); | |
857 | ||
858 | if (thread == NULL) { | |
859 | eprintf("problem processing %d event, skipping it.\n", | |
860 | event->header.type); | |
861 | return -1; | |
862 | } | |
863 | ||
864 | cpumode = event->header.misc & PERF_EVENT_MISC_CPUMODE_MASK; | |
865 | ||
866 | if (cpumode == PERF_EVENT_MISC_KERNEL) { | |
867 | show = SHOW_KERNEL; | |
868 | level = 'k'; | |
869 | ||
870 | dso = kernel_dso; | |
871 | ||
872 | dump_printf(" ...... dso: %s\n", dso->name); | |
873 | ||
874 | } else if (cpumode == PERF_EVENT_MISC_USER) { | |
875 | ||
876 | show = SHOW_USER; | |
877 | level = '.'; | |
878 | ||
879 | } else { | |
880 | show = SHOW_HV; | |
881 | level = 'H'; | |
882 | ||
883 | dso = hypervisor_dso; | |
884 | ||
885 | dump_printf(" ...... dso: [hypervisor]\n"); | |
886 | } | |
887 | ||
ec156764 IM |
888 | if (sample_type & PERF_SAMPLE_RAW) |
889 | process_raw_event(event, more_data, cpu, timestamp, thread); | |
0a02ad93 IM |
890 | |
891 | return 0; | |
892 | } | |
893 | ||
894 | static int | |
895 | process_event(event_t *event, unsigned long offset, unsigned long head) | |
896 | { | |
897 | trace_event(event); | |
898 | ||
899 | switch (event->header.type) { | |
900 | case PERF_EVENT_MMAP ... PERF_EVENT_LOST: | |
901 | return 0; | |
902 | ||
903 | case PERF_EVENT_COMM: | |
904 | return process_comm_event(event, offset, head); | |
905 | ||
906 | case PERF_EVENT_EXIT ... PERF_EVENT_READ: | |
907 | return 0; | |
908 | ||
909 | case PERF_EVENT_SAMPLE: | |
910 | return process_sample_event(event, offset, head); | |
911 | ||
912 | case PERF_EVENT_MAX: | |
913 | default: | |
914 | return -1; | |
915 | } | |
916 | ||
917 | return 0; | |
918 | } | |
919 | ||
920 | static int __cmd_sched(void) | |
921 | { | |
922 | int ret, rc = EXIT_FAILURE; | |
923 | unsigned long offset = 0; | |
924 | unsigned long head = 0; | |
925 | struct stat perf_stat; | |
926 | event_t *event; | |
927 | uint32_t size; | |
928 | char *buf; | |
929 | ||
930 | trace_report(); | |
931 | register_idle_thread(&threads, &last_match); | |
932 | ||
933 | input = open(input_name, O_RDONLY); | |
934 | if (input < 0) { | |
935 | perror("failed to open file"); | |
936 | exit(-1); | |
937 | } | |
938 | ||
939 | ret = fstat(input, &perf_stat); | |
940 | if (ret < 0) { | |
941 | perror("failed to stat file"); | |
942 | exit(-1); | |
943 | } | |
944 | ||
945 | if (!perf_stat.st_size) { | |
946 | fprintf(stderr, "zero-sized file, nothing to do!\n"); | |
947 | exit(0); | |
948 | } | |
949 | header = perf_header__read(input); | |
950 | head = header->data_offset; | |
951 | sample_type = perf_header__sample_type(header); | |
952 | ||
953 | if (!(sample_type & PERF_SAMPLE_RAW)) | |
954 | die("No trace sample to read. Did you call perf record " | |
955 | "without -R?"); | |
956 | ||
957 | if (load_kernel() < 0) { | |
958 | perror("failed to load kernel symbols"); | |
959 | return EXIT_FAILURE; | |
960 | } | |
961 | ||
962 | remap: | |
963 | buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ, | |
964 | MAP_SHARED, input, offset); | |
965 | if (buf == MAP_FAILED) { | |
966 | perror("failed to mmap file"); | |
967 | exit(-1); | |
968 | } | |
969 | ||
970 | more: | |
971 | event = (event_t *)(buf + head); | |
972 | ||
973 | size = event->header.size; | |
974 | if (!size) | |
975 | size = 8; | |
976 | ||
977 | if (head + event->header.size >= page_size * mmap_window) { | |
978 | unsigned long shift = page_size * (head / page_size); | |
979 | int res; | |
980 | ||
981 | res = munmap(buf, page_size * mmap_window); | |
982 | assert(res == 0); | |
983 | ||
984 | offset += shift; | |
985 | head -= shift; | |
986 | goto remap; | |
987 | } | |
988 | ||
989 | size = event->header.size; | |
990 | ||
991 | ||
992 | if (!size || process_event(event, offset, head) < 0) { | |
993 | ||
994 | /* | |
995 | * assume we lost track of the stream, check alignment, and | |
996 | * increment a single u64 in the hope to catch on again 'soon'. | |
997 | */ | |
998 | ||
999 | if (unlikely(head & 7)) | |
1000 | head &= ~7ULL; | |
1001 | ||
1002 | size = 8; | |
1003 | } | |
1004 | ||
1005 | head += size; | |
1006 | ||
1007 | if (offset + head < (unsigned long)perf_stat.st_size) | |
1008 | goto more; | |
1009 | ||
1010 | rc = EXIT_SUCCESS; | |
1011 | close(input); | |
1012 | ||
1013 | return rc; | |
1014 | } | |
1015 | ||
1016 | static const char * const annotate_usage[] = { | |
1017 | "perf trace [<options>] <command>", | |
1018 | NULL | |
1019 | }; | |
1020 | ||
1021 | static const struct option options[] = { | |
1022 | OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, | |
1023 | "dump raw trace in ASCII"), | |
1024 | OPT_BOOLEAN('v', "verbose", &verbose, | |
1025 | "be more verbose (show symbol address, etc)"), | |
1026 | OPT_END() | |
1027 | }; | |
1028 | ||
1029 | int cmd_sched(int argc, const char **argv, const char *prefix __used) | |
1030 | { | |
fbf94829 | 1031 | long nr_iterations = 10, i; |
ec156764 | 1032 | |
0a02ad93 IM |
1033 | symbol__init(); |
1034 | page_size = getpagesize(); | |
1035 | ||
1036 | argc = parse_options(argc, argv, options, annotate_usage, 0); | |
1037 | if (argc) { | |
1038 | /* | |
1039 | * Special case: if there's an argument left then assume tha | |
1040 | * it's a symbol filter: | |
1041 | */ | |
1042 | if (argc > 1) | |
1043 | usage_with_options(annotate_usage, options); | |
1044 | } | |
1045 | ||
fbf94829 | 1046 | // setup_pager(); |
0a02ad93 | 1047 | |
ec156764 IM |
1048 | calibrate_run_measurement_overhead(); |
1049 | calibrate_sleep_measurement_overhead(); | |
1050 | ||
1051 | test_calibrations(); | |
1052 | ||
1053 | parse_trace(); | |
1054 | print_task_traces(); | |
1055 | add_cross_task_wakeups(); | |
1056 | ||
1057 | create_tasks(); | |
ad236fd2 | 1058 | printf("------------------------------------------------------------\n"); |
ec156764 IM |
1059 | for (i = 0; i < nr_iterations; i++) |
1060 | run_one_test(); | |
1061 | ||
1062 | return 0; | |
0a02ad93 | 1063 | } |