]>
Commit | Line | Data |
---|---|---|
0a02ad93 | 1 | #include "builtin.h" |
b1ffe8f3 | 2 | #include "perf.h" |
0a02ad93 IM |
3 | |
4 | #include "util/util.h" | |
ee29be62 | 5 | #include "util/evlist.h" |
0a02ad93 | 6 | #include "util/cache.h" |
e3f42609 | 7 | #include "util/evsel.h" |
0a02ad93 IM |
8 | #include "util/symbol.h" |
9 | #include "util/thread.h" | |
10 | #include "util/header.h" | |
94c744b6 | 11 | #include "util/session.h" |
45694aa7 | 12 | #include "util/tool.h" |
57480d2c | 13 | #include "util/cloexec.h" |
0a02ad93 IM |
14 | |
15 | #include "util/parse-options.h" | |
b1ffe8f3 | 16 | #include "util/trace-event.h" |
0a02ad93 | 17 | |
0a02ad93 IM |
18 | #include "util/debug.h" |
19 | ||
b1ffe8f3 | 20 | #include <sys/prctl.h> |
7b78f136 | 21 | #include <sys/resource.h> |
0a02ad93 | 22 | |
b1ffe8f3 IM |
23 | #include <semaphore.h> |
24 | #include <pthread.h> | |
25 | #include <math.h> | |
419ab0d6 | 26 | |
b1ffe8f3 IM |
27 | #define PR_SET_NAME 15 /* Set process name */ |
28 | #define MAX_CPUS 4096 | |
b1ffe8f3 IM |
29 | #define COMM_LEN 20 |
30 | #define SYM_LEN 129 | |
b1ffe8f3 | 31 | #define MAX_PID 65536 |
ec156764 | 32 | |
39aeb52f | 33 | struct sched_atom; |
ec156764 | 34 | |
b1ffe8f3 IM |
35 | struct task_desc { |
36 | unsigned long nr; | |
37 | unsigned long pid; | |
38 | char comm[COMM_LEN]; | |
ec156764 | 39 | |
b1ffe8f3 IM |
40 | unsigned long nr_events; |
41 | unsigned long curr_event; | |
39aeb52f | 42 | struct sched_atom **atoms; |
b1ffe8f3 IM |
43 | |
44 | pthread_t thread; | |
45 | sem_t sleep_sem; | |
ec156764 | 46 | |
b1ffe8f3 IM |
47 | sem_t ready_for_work; |
48 | sem_t work_done_sem; | |
49 | ||
50 | u64 cpu_usage; | |
51 | }; | |
52 | ||
53 | enum sched_event_type { | |
54 | SCHED_EVENT_RUN, | |
55 | SCHED_EVENT_SLEEP, | |
56 | SCHED_EVENT_WAKEUP, | |
55ffb7a6 | 57 | SCHED_EVENT_MIGRATION, |
b1ffe8f3 IM |
58 | }; |
59 | ||
39aeb52f | 60 | struct sched_atom { |
b1ffe8f3 | 61 | enum sched_event_type type; |
eed05fe7 | 62 | int specific_wait; |
b1ffe8f3 IM |
63 | u64 timestamp; |
64 | u64 duration; | |
65 | unsigned long nr; | |
b1ffe8f3 IM |
66 | sem_t *wait_sem; |
67 | struct task_desc *wakee; | |
68 | }; | |
69 | ||
e936e8e4 | 70 | #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP" |
b1ffe8f3 IM |
71 | |
72 | enum thread_state { | |
73 | THREAD_SLEEPING = 0, | |
74 | THREAD_WAIT_CPU, | |
75 | THREAD_SCHED_IN, | |
76 | THREAD_IGNORE | |
77 | }; | |
78 | ||
79 | struct work_atom { | |
80 | struct list_head list; | |
81 | enum thread_state state; | |
aa1ab9d2 | 82 | u64 sched_out_time; |
b1ffe8f3 IM |
83 | u64 wake_up_time; |
84 | u64 sched_in_time; | |
85 | u64 runtime; | |
86 | }; | |
87 | ||
39aeb52f | 88 | struct work_atoms { |
89 | struct list_head work_list; | |
b1ffe8f3 IM |
90 | struct thread *thread; |
91 | struct rb_node node; | |
92 | u64 max_lat; | |
3786310a | 93 | u64 max_lat_at; |
b1ffe8f3 IM |
94 | u64 total_lat; |
95 | u64 nb_atoms; | |
96 | u64 total_runtime; | |
97 | }; | |
98 | ||
39aeb52f | 99 | typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *); |
b1ffe8f3 | 100 | |
9ec3f4e4 | 101 | struct perf_sched; |
0e9b07e5 | 102 | |
9ec3f4e4 ACM |
103 | struct trace_sched_handler { |
104 | int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel, | |
105 | struct perf_sample *sample, struct machine *machine); | |
0e9b07e5 | 106 | |
9ec3f4e4 ACM |
107 | int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel, |
108 | struct perf_sample *sample, struct machine *machine); | |
0e9b07e5 | 109 | |
9ec3f4e4 ACM |
110 | int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel, |
111 | struct perf_sample *sample, struct machine *machine); | |
0e9b07e5 | 112 | |
cb627505 DA |
113 | /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */ |
114 | int (*fork_event)(struct perf_sched *sched, union perf_event *event, | |
115 | struct machine *machine); | |
0e9b07e5 ACM |
116 | |
117 | int (*migrate_task_event)(struct perf_sched *sched, | |
9ec3f4e4 ACM |
118 | struct perf_evsel *evsel, |
119 | struct perf_sample *sample, | |
120 | struct machine *machine); | |
0e9b07e5 ACM |
121 | }; |
122 | ||
123 | struct perf_sched { | |
124 | struct perf_tool tool; | |
0e9b07e5 ACM |
125 | const char *sort_order; |
126 | unsigned long nr_tasks; | |
127 | struct task_desc *pid_to_task[MAX_PID]; | |
128 | struct task_desc **tasks; | |
129 | const struct trace_sched_handler *tp_handler; | |
130 | pthread_mutex_t start_work_mutex; | |
131 | pthread_mutex_t work_done_wait_mutex; | |
132 | int profile_cpu; | |
133 | /* | |
134 | * Track the current task - that way we can know whether there's any | |
135 | * weird events, such as a task being switched away that is not current. | |
136 | */ | |
137 | int max_cpu; | |
138 | u32 curr_pid[MAX_CPUS]; | |
139 | struct thread *curr_thread[MAX_CPUS]; | |
140 | char next_shortname1; | |
141 | char next_shortname2; | |
142 | unsigned int replay_repeat; | |
143 | unsigned long nr_run_events; | |
144 | unsigned long nr_sleep_events; | |
145 | unsigned long nr_wakeup_events; | |
146 | unsigned long nr_sleep_corrections; | |
147 | unsigned long nr_run_events_optimized; | |
148 | unsigned long targetless_wakeups; | |
149 | unsigned long multitarget_wakeups; | |
150 | unsigned long nr_runs; | |
151 | unsigned long nr_timestamps; | |
152 | unsigned long nr_unordered_timestamps; | |
0e9b07e5 ACM |
153 | unsigned long nr_context_switch_bugs; |
154 | unsigned long nr_events; | |
155 | unsigned long nr_lost_chunks; | |
156 | unsigned long nr_lost_events; | |
157 | u64 run_measurement_overhead; | |
158 | u64 sleep_measurement_overhead; | |
159 | u64 start_time; | |
160 | u64 cpu_usage; | |
161 | u64 runavg_cpu_usage; | |
162 | u64 parent_cpu_usage; | |
163 | u64 runavg_parent_cpu_usage; | |
164 | u64 sum_runtime; | |
165 | u64 sum_fluct; | |
166 | u64 run_avg; | |
167 | u64 all_runtime; | |
168 | u64 all_count; | |
169 | u64 cpu_last_switched[MAX_CPUS]; | |
170 | struct rb_root atom_root, sorted_atom_root; | |
171 | struct list_head sort_list, cmp_pid; | |
172 | }; | |
b1ffe8f3 IM |
173 | |
174 | static u64 get_nsecs(void) | |
ec156764 IM |
175 | { |
176 | struct timespec ts; | |
177 | ||
178 | clock_gettime(CLOCK_MONOTONIC, &ts); | |
179 | ||
180 | return ts.tv_sec * 1000000000ULL + ts.tv_nsec; | |
181 | } | |
182 | ||
0e9b07e5 | 183 | static void burn_nsecs(struct perf_sched *sched, u64 nsecs) |
ec156764 | 184 | { |
b1ffe8f3 | 185 | u64 T0 = get_nsecs(), T1; |
ec156764 IM |
186 | |
187 | do { | |
188 | T1 = get_nsecs(); | |
0e9b07e5 | 189 | } while (T1 + sched->run_measurement_overhead < T0 + nsecs); |
ec156764 IM |
190 | } |
191 | ||
b1ffe8f3 | 192 | static void sleep_nsecs(u64 nsecs) |
ec156764 IM |
193 | { |
194 | struct timespec ts; | |
195 | ||
196 | ts.tv_nsec = nsecs % 999999999; | |
197 | ts.tv_sec = nsecs / 999999999; | |
198 | ||
199 | nanosleep(&ts, NULL); | |
200 | } | |
201 | ||
0e9b07e5 | 202 | static void calibrate_run_measurement_overhead(struct perf_sched *sched) |
ec156764 | 203 | { |
b1ffe8f3 | 204 | u64 T0, T1, delta, min_delta = 1000000000ULL; |
ec156764 IM |
205 | int i; |
206 | ||
207 | for (i = 0; i < 10; i++) { | |
208 | T0 = get_nsecs(); | |
0e9b07e5 | 209 | burn_nsecs(sched, 0); |
ec156764 IM |
210 | T1 = get_nsecs(); |
211 | delta = T1-T0; | |
212 | min_delta = min(min_delta, delta); | |
213 | } | |
0e9b07e5 | 214 | sched->run_measurement_overhead = min_delta; |
ec156764 | 215 | |
9486aa38 | 216 | printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta); |
ec156764 IM |
217 | } |
218 | ||
0e9b07e5 | 219 | static void calibrate_sleep_measurement_overhead(struct perf_sched *sched) |
ec156764 | 220 | { |
b1ffe8f3 | 221 | u64 T0, T1, delta, min_delta = 1000000000ULL; |
ec156764 IM |
222 | int i; |
223 | ||
224 | for (i = 0; i < 10; i++) { | |
225 | T0 = get_nsecs(); | |
226 | sleep_nsecs(10000); | |
227 | T1 = get_nsecs(); | |
228 | delta = T1-T0; | |
229 | min_delta = min(min_delta, delta); | |
230 | } | |
231 | min_delta -= 10000; | |
0e9b07e5 | 232 | sched->sleep_measurement_overhead = min_delta; |
ec156764 | 233 | |
9486aa38 | 234 | printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta); |
ec156764 IM |
235 | } |
236 | ||
39aeb52f | 237 | static struct sched_atom * |
b1ffe8f3 | 238 | get_new_event(struct task_desc *task, u64 timestamp) |
ec156764 | 239 | { |
36479484 | 240 | struct sched_atom *event = zalloc(sizeof(*event)); |
ec156764 IM |
241 | unsigned long idx = task->nr_events; |
242 | size_t size; | |
243 | ||
244 | event->timestamp = timestamp; | |
245 | event->nr = idx; | |
246 | ||
247 | task->nr_events++; | |
39aeb52f | 248 | size = sizeof(struct sched_atom *) * task->nr_events; |
249 | task->atoms = realloc(task->atoms, size); | |
250 | BUG_ON(!task->atoms); | |
ec156764 | 251 | |
39aeb52f | 252 | task->atoms[idx] = event; |
ec156764 IM |
253 | |
254 | return event; | |
255 | } | |
256 | ||
39aeb52f | 257 | static struct sched_atom *last_event(struct task_desc *task) |
ec156764 IM |
258 | { |
259 | if (!task->nr_events) | |
260 | return NULL; | |
261 | ||
39aeb52f | 262 | return task->atoms[task->nr_events - 1]; |
ec156764 IM |
263 | } |
264 | ||
0e9b07e5 ACM |
265 | static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task, |
266 | u64 timestamp, u64 duration) | |
ec156764 | 267 | { |
39aeb52f | 268 | struct sched_atom *event, *curr_event = last_event(task); |
ec156764 IM |
269 | |
270 | /* | |
fbf94829 IM |
271 | * optimize an existing RUN event by merging this one |
272 | * to it: | |
273 | */ | |
ec156764 | 274 | if (curr_event && curr_event->type == SCHED_EVENT_RUN) { |
0e9b07e5 | 275 | sched->nr_run_events_optimized++; |
ec156764 IM |
276 | curr_event->duration += duration; |
277 | return; | |
278 | } | |
279 | ||
280 | event = get_new_event(task, timestamp); | |
281 | ||
282 | event->type = SCHED_EVENT_RUN; | |
283 | event->duration = duration; | |
284 | ||
0e9b07e5 | 285 | sched->nr_run_events++; |
ec156764 IM |
286 | } |
287 | ||
0e9b07e5 ACM |
288 | static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task, |
289 | u64 timestamp, struct task_desc *wakee) | |
ec156764 | 290 | { |
39aeb52f | 291 | struct sched_atom *event, *wakee_event; |
ec156764 IM |
292 | |
293 | event = get_new_event(task, timestamp); | |
294 | event->type = SCHED_EVENT_WAKEUP; | |
295 | event->wakee = wakee; | |
296 | ||
297 | wakee_event = last_event(wakee); | |
298 | if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) { | |
0e9b07e5 | 299 | sched->targetless_wakeups++; |
ec156764 IM |
300 | return; |
301 | } | |
302 | if (wakee_event->wait_sem) { | |
0e9b07e5 | 303 | sched->multitarget_wakeups++; |
ec156764 IM |
304 | return; |
305 | } | |
306 | ||
36479484 | 307 | wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem)); |
ec156764 IM |
308 | sem_init(wakee_event->wait_sem, 0, 0); |
309 | wakee_event->specific_wait = 1; | |
310 | event->wait_sem = wakee_event->wait_sem; | |
311 | ||
0e9b07e5 | 312 | sched->nr_wakeup_events++; |
ec156764 IM |
313 | } |
314 | ||
0e9b07e5 ACM |
315 | static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task, |
316 | u64 timestamp, u64 task_state __maybe_unused) | |
ec156764 | 317 | { |
39aeb52f | 318 | struct sched_atom *event = get_new_event(task, timestamp); |
ec156764 IM |
319 | |
320 | event->type = SCHED_EVENT_SLEEP; | |
321 | ||
0e9b07e5 | 322 | sched->nr_sleep_events++; |
ec156764 IM |
323 | } |
324 | ||
0e9b07e5 ACM |
325 | static struct task_desc *register_pid(struct perf_sched *sched, |
326 | unsigned long pid, const char *comm) | |
ec156764 IM |
327 | { |
328 | struct task_desc *task; | |
329 | ||
330 | BUG_ON(pid >= MAX_PID); | |
331 | ||
0e9b07e5 | 332 | task = sched->pid_to_task[pid]; |
ec156764 IM |
333 | |
334 | if (task) | |
335 | return task; | |
336 | ||
36479484 | 337 | task = zalloc(sizeof(*task)); |
ec156764 | 338 | task->pid = pid; |
0e9b07e5 | 339 | task->nr = sched->nr_tasks; |
ec156764 IM |
340 | strcpy(task->comm, comm); |
341 | /* | |
342 | * every task starts in sleeping state - this gets ignored | |
343 | * if there's no wakeup pointing to this sleep state: | |
344 | */ | |
0e9b07e5 | 345 | add_sched_event_sleep(sched, task, 0, 0); |
ec156764 | 346 | |
0e9b07e5 ACM |
347 | sched->pid_to_task[pid] = task; |
348 | sched->nr_tasks++; | |
349 | sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_task *)); | |
350 | BUG_ON(!sched->tasks); | |
351 | sched->tasks[task->nr] = task; | |
ec156764 | 352 | |
ad236fd2 | 353 | if (verbose) |
0e9b07e5 | 354 | printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm); |
ec156764 IM |
355 | |
356 | return task; | |
357 | } | |
358 | ||
359 | ||
0e9b07e5 | 360 | static void print_task_traces(struct perf_sched *sched) |
ec156764 IM |
361 | { |
362 | struct task_desc *task; | |
363 | unsigned long i; | |
364 | ||
0e9b07e5 ACM |
365 | for (i = 0; i < sched->nr_tasks; i++) { |
366 | task = sched->tasks[i]; | |
ad236fd2 | 367 | printf("task %6ld (%20s:%10ld), nr_events: %ld\n", |
ec156764 IM |
368 | task->nr, task->comm, task->pid, task->nr_events); |
369 | } | |
370 | } | |
371 | ||
0e9b07e5 | 372 | static void add_cross_task_wakeups(struct perf_sched *sched) |
ec156764 IM |
373 | { |
374 | struct task_desc *task1, *task2; | |
375 | unsigned long i, j; | |
376 | ||
0e9b07e5 ACM |
377 | for (i = 0; i < sched->nr_tasks; i++) { |
378 | task1 = sched->tasks[i]; | |
ec156764 | 379 | j = i + 1; |
0e9b07e5 | 380 | if (j == sched->nr_tasks) |
ec156764 | 381 | j = 0; |
0e9b07e5 ACM |
382 | task2 = sched->tasks[j]; |
383 | add_sched_event_wakeup(sched, task1, 0, task2); | |
ec156764 IM |
384 | } |
385 | } | |
386 | ||
0e9b07e5 ACM |
387 | static void perf_sched__process_event(struct perf_sched *sched, |
388 | struct sched_atom *atom) | |
ec156764 IM |
389 | { |
390 | int ret = 0; | |
ec156764 | 391 | |
39aeb52f | 392 | switch (atom->type) { |
ec156764 | 393 | case SCHED_EVENT_RUN: |
0e9b07e5 | 394 | burn_nsecs(sched, atom->duration); |
ec156764 IM |
395 | break; |
396 | case SCHED_EVENT_SLEEP: | |
39aeb52f | 397 | if (atom->wait_sem) |
398 | ret = sem_wait(atom->wait_sem); | |
ec156764 IM |
399 | BUG_ON(ret); |
400 | break; | |
401 | case SCHED_EVENT_WAKEUP: | |
39aeb52f | 402 | if (atom->wait_sem) |
403 | ret = sem_post(atom->wait_sem); | |
ec156764 IM |
404 | BUG_ON(ret); |
405 | break; | |
55ffb7a6 MG |
406 | case SCHED_EVENT_MIGRATION: |
407 | break; | |
ec156764 IM |
408 | default: |
409 | BUG_ON(1); | |
410 | } | |
411 | } | |
412 | ||
b1ffe8f3 | 413 | static u64 get_cpu_usage_nsec_parent(void) |
ec156764 IM |
414 | { |
415 | struct rusage ru; | |
b1ffe8f3 | 416 | u64 sum; |
ec156764 IM |
417 | int err; |
418 | ||
419 | err = getrusage(RUSAGE_SELF, &ru); | |
420 | BUG_ON(err); | |
421 | ||
422 | sum = ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3; | |
423 | sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3; | |
424 | ||
425 | return sum; | |
426 | } | |
427 | ||
c0c9e721 | 428 | static int self_open_counters(void) |
ec156764 | 429 | { |
c0c9e721 XG |
430 | struct perf_event_attr attr; |
431 | int fd; | |
ec156764 | 432 | |
c0c9e721 | 433 | memset(&attr, 0, sizeof(attr)); |
ec156764 | 434 | |
c0c9e721 XG |
435 | attr.type = PERF_TYPE_SOFTWARE; |
436 | attr.config = PERF_COUNT_SW_TASK_CLOCK; | |
ec156764 | 437 | |
57480d2c YD |
438 | fd = sys_perf_event_open(&attr, 0, -1, -1, |
439 | perf_event_open_cloexec_flag()); | |
c0c9e721 XG |
440 | |
441 | if (fd < 0) | |
60b7d14a NK |
442 | pr_err("Error: sys_perf_event_open() syscall returned " |
443 | "with %d (%s)\n", fd, strerror(errno)); | |
c0c9e721 XG |
444 | return fd; |
445 | } | |
446 | ||
447 | static u64 get_cpu_usage_nsec_self(int fd) | |
448 | { | |
449 | u64 runtime; | |
450 | int ret; | |
451 | ||
452 | ret = read(fd, &runtime, sizeof(runtime)); | |
453 | BUG_ON(ret != sizeof(runtime)); | |
454 | ||
455 | return runtime; | |
ec156764 IM |
456 | } |
457 | ||
0e9b07e5 ACM |
458 | struct sched_thread_parms { |
459 | struct task_desc *task; | |
460 | struct perf_sched *sched; | |
461 | }; | |
462 | ||
ec156764 IM |
463 | static void *thread_func(void *ctx) |
464 | { | |
0e9b07e5 ACM |
465 | struct sched_thread_parms *parms = ctx; |
466 | struct task_desc *this_task = parms->task; | |
467 | struct perf_sched *sched = parms->sched; | |
b1ffe8f3 | 468 | u64 cpu_usage_0, cpu_usage_1; |
ec156764 IM |
469 | unsigned long i, ret; |
470 | char comm2[22]; | |
c0c9e721 | 471 | int fd; |
ec156764 | 472 | |
74cf249d | 473 | zfree(&parms); |
0e9b07e5 | 474 | |
ec156764 IM |
475 | sprintf(comm2, ":%s", this_task->comm); |
476 | prctl(PR_SET_NAME, comm2); | |
c0c9e721 | 477 | fd = self_open_counters(); |
a116e05d ACM |
478 | if (fd < 0) |
479 | return NULL; | |
ec156764 IM |
480 | again: |
481 | ret = sem_post(&this_task->ready_for_work); | |
482 | BUG_ON(ret); | |
0e9b07e5 | 483 | ret = pthread_mutex_lock(&sched->start_work_mutex); |
ec156764 | 484 | BUG_ON(ret); |
0e9b07e5 | 485 | ret = pthread_mutex_unlock(&sched->start_work_mutex); |
ec156764 | 486 | BUG_ON(ret); |
ec156764 | 487 | |
c0c9e721 | 488 | cpu_usage_0 = get_cpu_usage_nsec_self(fd); |
ec156764 IM |
489 | |
490 | for (i = 0; i < this_task->nr_events; i++) { | |
491 | this_task->curr_event = i; | |
0e9b07e5 | 492 | perf_sched__process_event(sched, this_task->atoms[i]); |
ec156764 IM |
493 | } |
494 | ||
c0c9e721 | 495 | cpu_usage_1 = get_cpu_usage_nsec_self(fd); |
ec156764 | 496 | this_task->cpu_usage = cpu_usage_1 - cpu_usage_0; |
ec156764 IM |
497 | ret = sem_post(&this_task->work_done_sem); |
498 | BUG_ON(ret); | |
ec156764 | 499 | |
0e9b07e5 | 500 | ret = pthread_mutex_lock(&sched->work_done_wait_mutex); |
ec156764 | 501 | BUG_ON(ret); |
0e9b07e5 | 502 | ret = pthread_mutex_unlock(&sched->work_done_wait_mutex); |
ec156764 | 503 | BUG_ON(ret); |
ec156764 IM |
504 | |
505 | goto again; | |
506 | } | |
507 | ||
0e9b07e5 | 508 | static void create_tasks(struct perf_sched *sched) |
ec156764 IM |
509 | { |
510 | struct task_desc *task; | |
511 | pthread_attr_t attr; | |
512 | unsigned long i; | |
513 | int err; | |
514 | ||
515 | err = pthread_attr_init(&attr); | |
516 | BUG_ON(err); | |
12f7e036 JP |
517 | err = pthread_attr_setstacksize(&attr, |
518 | (size_t) max(16 * 1024, PTHREAD_STACK_MIN)); | |
ec156764 | 519 | BUG_ON(err); |
0e9b07e5 | 520 | err = pthread_mutex_lock(&sched->start_work_mutex); |
ec156764 | 521 | BUG_ON(err); |
0e9b07e5 | 522 | err = pthread_mutex_lock(&sched->work_done_wait_mutex); |
ec156764 | 523 | BUG_ON(err); |
0e9b07e5 ACM |
524 | for (i = 0; i < sched->nr_tasks; i++) { |
525 | struct sched_thread_parms *parms = malloc(sizeof(*parms)); | |
526 | BUG_ON(parms == NULL); | |
527 | parms->task = task = sched->tasks[i]; | |
528 | parms->sched = sched; | |
ec156764 IM |
529 | sem_init(&task->sleep_sem, 0, 0); |
530 | sem_init(&task->ready_for_work, 0, 0); | |
531 | sem_init(&task->work_done_sem, 0, 0); | |
532 | task->curr_event = 0; | |
0e9b07e5 | 533 | err = pthread_create(&task->thread, &attr, thread_func, parms); |
ec156764 IM |
534 | BUG_ON(err); |
535 | } | |
536 | } | |
537 | ||
0e9b07e5 | 538 | static void wait_for_tasks(struct perf_sched *sched) |
ec156764 | 539 | { |
b1ffe8f3 | 540 | u64 cpu_usage_0, cpu_usage_1; |
ec156764 IM |
541 | struct task_desc *task; |
542 | unsigned long i, ret; | |
543 | ||
0e9b07e5 ACM |
544 | sched->start_time = get_nsecs(); |
545 | sched->cpu_usage = 0; | |
546 | pthread_mutex_unlock(&sched->work_done_wait_mutex); | |
ec156764 | 547 | |
0e9b07e5 ACM |
548 | for (i = 0; i < sched->nr_tasks; i++) { |
549 | task = sched->tasks[i]; | |
ec156764 IM |
550 | ret = sem_wait(&task->ready_for_work); |
551 | BUG_ON(ret); | |
552 | sem_init(&task->ready_for_work, 0, 0); | |
553 | } | |
0e9b07e5 | 554 | ret = pthread_mutex_lock(&sched->work_done_wait_mutex); |
ec156764 IM |
555 | BUG_ON(ret); |
556 | ||
557 | cpu_usage_0 = get_cpu_usage_nsec_parent(); | |
558 | ||
0e9b07e5 | 559 | pthread_mutex_unlock(&sched->start_work_mutex); |
ec156764 | 560 | |
0e9b07e5 ACM |
561 | for (i = 0; i < sched->nr_tasks; i++) { |
562 | task = sched->tasks[i]; | |
ec156764 IM |
563 | ret = sem_wait(&task->work_done_sem); |
564 | BUG_ON(ret); | |
565 | sem_init(&task->work_done_sem, 0, 0); | |
0e9b07e5 | 566 | sched->cpu_usage += task->cpu_usage; |
ec156764 IM |
567 | task->cpu_usage = 0; |
568 | } | |
569 | ||
570 | cpu_usage_1 = get_cpu_usage_nsec_parent(); | |
0e9b07e5 ACM |
571 | if (!sched->runavg_cpu_usage) |
572 | sched->runavg_cpu_usage = sched->cpu_usage; | |
573 | sched->runavg_cpu_usage = (sched->runavg_cpu_usage * 9 + sched->cpu_usage) / 10; | |
ec156764 | 574 | |
0e9b07e5 ACM |
575 | sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0; |
576 | if (!sched->runavg_parent_cpu_usage) | |
577 | sched->runavg_parent_cpu_usage = sched->parent_cpu_usage; | |
578 | sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * 9 + | |
579 | sched->parent_cpu_usage)/10; | |
ec156764 | 580 | |
0e9b07e5 | 581 | ret = pthread_mutex_lock(&sched->start_work_mutex); |
ec156764 IM |
582 | BUG_ON(ret); |
583 | ||
0e9b07e5 ACM |
584 | for (i = 0; i < sched->nr_tasks; i++) { |
585 | task = sched->tasks[i]; | |
ec156764 IM |
586 | sem_init(&task->sleep_sem, 0, 0); |
587 | task->curr_event = 0; | |
588 | } | |
589 | } | |
590 | ||
0e9b07e5 | 591 | static void run_one_test(struct perf_sched *sched) |
ec156764 | 592 | { |
fb7d0b3c | 593 | u64 T0, T1, delta, avg_delta, fluct; |
ec156764 IM |
594 | |
595 | T0 = get_nsecs(); | |
0e9b07e5 | 596 | wait_for_tasks(sched); |
ec156764 IM |
597 | T1 = get_nsecs(); |
598 | ||
599 | delta = T1 - T0; | |
0e9b07e5 ACM |
600 | sched->sum_runtime += delta; |
601 | sched->nr_runs++; | |
ec156764 | 602 | |
0e9b07e5 | 603 | avg_delta = sched->sum_runtime / sched->nr_runs; |
ec156764 IM |
604 | if (delta < avg_delta) |
605 | fluct = avg_delta - delta; | |
606 | else | |
607 | fluct = delta - avg_delta; | |
0e9b07e5 ACM |
608 | sched->sum_fluct += fluct; |
609 | if (!sched->run_avg) | |
610 | sched->run_avg = delta; | |
611 | sched->run_avg = (sched->run_avg * 9 + delta) / 10; | |
ec156764 | 612 | |
0e9b07e5 | 613 | printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / 1000000.0); |
ec156764 | 614 | |
0e9b07e5 | 615 | printf("ravg: %0.2f, ", (double)sched->run_avg / 1e6); |
ec156764 | 616 | |
ad236fd2 | 617 | printf("cpu: %0.2f / %0.2f", |
0e9b07e5 | 618 | (double)sched->cpu_usage / 1e6, (double)sched->runavg_cpu_usage / 1e6); |
ec156764 IM |
619 | |
620 | #if 0 | |
621 | /* | |
fbf94829 | 622 | * rusage statistics done by the parent, these are less |
0e9b07e5 | 623 | * accurate than the sched->sum_exec_runtime based statistics: |
fbf94829 | 624 | */ |
ad236fd2 | 625 | printf(" [%0.2f / %0.2f]", |
0e9b07e5 ACM |
626 | (double)sched->parent_cpu_usage/1e6, |
627 | (double)sched->runavg_parent_cpu_usage/1e6); | |
ec156764 IM |
628 | #endif |
629 | ||
ad236fd2 | 630 | printf("\n"); |
ec156764 | 631 | |
0e9b07e5 ACM |
632 | if (sched->nr_sleep_corrections) |
633 | printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections); | |
634 | sched->nr_sleep_corrections = 0; | |
ec156764 IM |
635 | } |
636 | ||
0e9b07e5 | 637 | static void test_calibrations(struct perf_sched *sched) |
ec156764 | 638 | { |
b1ffe8f3 | 639 | u64 T0, T1; |
ec156764 IM |
640 | |
641 | T0 = get_nsecs(); | |
0e9b07e5 | 642 | burn_nsecs(sched, 1e6); |
ec156764 IM |
643 | T1 = get_nsecs(); |
644 | ||
9486aa38 | 645 | printf("the run test took %" PRIu64 " nsecs\n", T1 - T0); |
ec156764 IM |
646 | |
647 | T0 = get_nsecs(); | |
648 | sleep_nsecs(1e6); | |
649 | T1 = get_nsecs(); | |
650 | ||
9486aa38 | 651 | printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0); |
ec156764 IM |
652 | } |
653 | ||
a116e05d | 654 | static int |
0e9b07e5 | 655 | replay_wakeup_event(struct perf_sched *sched, |
9ec3f4e4 ACM |
656 | struct perf_evsel *evsel, struct perf_sample *sample, |
657 | struct machine *machine __maybe_unused) | |
419ab0d6 | 658 | { |
9ec3f4e4 ACM |
659 | const char *comm = perf_evsel__strval(evsel, sample, "comm"); |
660 | const u32 pid = perf_evsel__intval(evsel, sample, "pid"); | |
419ab0d6 | 661 | struct task_desc *waker, *wakee; |
fbf94829 | 662 | |
ad236fd2 | 663 | if (verbose) { |
2b7fcbc5 | 664 | printf("sched_wakeup event %p\n", evsel); |
fbf94829 | 665 | |
9ec3f4e4 | 666 | printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid); |
ad236fd2 | 667 | } |
fbf94829 | 668 | |
2b7fcbc5 | 669 | waker = register_pid(sched, sample->tid, "<unknown>"); |
9ec3f4e4 | 670 | wakee = register_pid(sched, pid, comm); |
fbf94829 | 671 | |
0e9b07e5 | 672 | add_sched_event_wakeup(sched, waker, sample->time, wakee); |
a116e05d | 673 | return 0; |
ec156764 IM |
674 | } |
675 | ||
9ec3f4e4 ACM |
676 | static int replay_switch_event(struct perf_sched *sched, |
677 | struct perf_evsel *evsel, | |
678 | struct perf_sample *sample, | |
679 | struct machine *machine __maybe_unused) | |
ec156764 | 680 | { |
9ec3f4e4 ACM |
681 | const char *prev_comm = perf_evsel__strval(evsel, sample, "prev_comm"), |
682 | *next_comm = perf_evsel__strval(evsel, sample, "next_comm"); | |
683 | const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"), | |
684 | next_pid = perf_evsel__intval(evsel, sample, "next_pid"); | |
685 | const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state"); | |
1d037ca1 | 686 | struct task_desc *prev, __maybe_unused *next; |
7f7f8d0b ACM |
687 | u64 timestamp0, timestamp = sample->time; |
688 | int cpu = sample->cpu; | |
fbf94829 IM |
689 | s64 delta; |
690 | ||
ad236fd2 | 691 | if (verbose) |
2b7fcbc5 | 692 | printf("sched_switch event %p\n", evsel); |
ad236fd2 | 693 | |
fbf94829 | 694 | if (cpu >= MAX_CPUS || cpu < 0) |
a116e05d | 695 | return 0; |
fbf94829 | 696 | |
0e9b07e5 | 697 | timestamp0 = sched->cpu_last_switched[cpu]; |
fbf94829 IM |
698 | if (timestamp0) |
699 | delta = timestamp - timestamp0; | |
700 | else | |
701 | delta = 0; | |
702 | ||
a116e05d | 703 | if (delta < 0) { |
60b7d14a | 704 | pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta); |
a116e05d ACM |
705 | return -1; |
706 | } | |
fbf94829 | 707 | |
9ec3f4e4 ACM |
708 | pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n", |
709 | prev_comm, prev_pid, next_comm, next_pid, delta); | |
fbf94829 | 710 | |
9ec3f4e4 ACM |
711 | prev = register_pid(sched, prev_pid, prev_comm); |
712 | next = register_pid(sched, next_pid, next_comm); | |
fbf94829 | 713 | |
0e9b07e5 | 714 | sched->cpu_last_switched[cpu] = timestamp; |
fbf94829 | 715 | |
0e9b07e5 | 716 | add_sched_event_run(sched, prev, timestamp, delta); |
9ec3f4e4 | 717 | add_sched_event_sleep(sched, prev, timestamp, prev_state); |
a116e05d ACM |
718 | |
719 | return 0; | |
fbf94829 IM |
720 | } |
721 | ||
cb627505 DA |
722 | static int replay_fork_event(struct perf_sched *sched, |
723 | union perf_event *event, | |
724 | struct machine *machine) | |
419ab0d6 | 725 | { |
cb627505 DA |
726 | struct thread *child, *parent; |
727 | ||
314add6b AH |
728 | child = machine__findnew_thread(machine, event->fork.pid, |
729 | event->fork.tid); | |
730 | parent = machine__findnew_thread(machine, event->fork.ppid, | |
731 | event->fork.ptid); | |
cb627505 DA |
732 | |
733 | if (child == NULL || parent == NULL) { | |
734 | pr_debug("thread does not exist on fork event: child %p, parent %p\n", | |
735 | child, parent); | |
736 | return 0; | |
737 | } | |
9ec3f4e4 | 738 | |
419ab0d6 | 739 | if (verbose) { |
cb627505 | 740 | printf("fork event\n"); |
b9c5143a FW |
741 | printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid); |
742 | printf("... child: %s/%d\n", thread__comm_str(child), child->tid); | |
419ab0d6 | 743 | } |
9ec3f4e4 | 744 | |
b9c5143a FW |
745 | register_pid(sched, parent->tid, thread__comm_str(parent)); |
746 | register_pid(sched, child->tid, thread__comm_str(child)); | |
a116e05d | 747 | return 0; |
419ab0d6 | 748 | } |
fbf94829 | 749 | |
b1ffe8f3 IM |
750 | struct sort_dimension { |
751 | const char *name; | |
b5fae128 | 752 | sort_fn_t cmp; |
b1ffe8f3 IM |
753 | struct list_head list; |
754 | }; | |
755 | ||
daa1d7a5 | 756 | static int |
39aeb52f | 757 | thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r) |
daa1d7a5 FW |
758 | { |
759 | struct sort_dimension *sort; | |
760 | int ret = 0; | |
761 | ||
b5fae128 IM |
762 | BUG_ON(list_empty(list)); |
763 | ||
daa1d7a5 FW |
764 | list_for_each_entry(sort, list, list) { |
765 | ret = sort->cmp(l, r); | |
766 | if (ret) | |
767 | return ret; | |
768 | } | |
769 | ||
770 | return ret; | |
771 | } | |
772 | ||
39aeb52f | 773 | static struct work_atoms * |
b5fae128 IM |
774 | thread_atoms_search(struct rb_root *root, struct thread *thread, |
775 | struct list_head *sort_list) | |
776 | { | |
777 | struct rb_node *node = root->rb_node; | |
39aeb52f | 778 | struct work_atoms key = { .thread = thread }; |
b5fae128 IM |
779 | |
780 | while (node) { | |
39aeb52f | 781 | struct work_atoms *atoms; |
b5fae128 IM |
782 | int cmp; |
783 | ||
39aeb52f | 784 | atoms = container_of(node, struct work_atoms, node); |
b5fae128 IM |
785 | |
786 | cmp = thread_lat_cmp(sort_list, &key, atoms); | |
787 | if (cmp > 0) | |
788 | node = node->rb_left; | |
789 | else if (cmp < 0) | |
790 | node = node->rb_right; | |
791 | else { | |
792 | BUG_ON(thread != atoms->thread); | |
793 | return atoms; | |
794 | } | |
795 | } | |
796 | return NULL; | |
797 | } | |
798 | ||
cdce9d73 | 799 | static void |
39aeb52f | 800 | __thread_latency_insert(struct rb_root *root, struct work_atoms *data, |
daa1d7a5 | 801 | struct list_head *sort_list) |
cdce9d73 FW |
802 | { |
803 | struct rb_node **new = &(root->rb_node), *parent = NULL; | |
804 | ||
805 | while (*new) { | |
39aeb52f | 806 | struct work_atoms *this; |
daa1d7a5 | 807 | int cmp; |
cdce9d73 | 808 | |
39aeb52f | 809 | this = container_of(*new, struct work_atoms, node); |
cdce9d73 | 810 | parent = *new; |
daa1d7a5 FW |
811 | |
812 | cmp = thread_lat_cmp(sort_list, data, this); | |
813 | ||
814 | if (cmp > 0) | |
cdce9d73 | 815 | new = &((*new)->rb_left); |
cdce9d73 | 816 | else |
daa1d7a5 | 817 | new = &((*new)->rb_right); |
cdce9d73 FW |
818 | } |
819 | ||
820 | rb_link_node(&data->node, parent, new); | |
821 | rb_insert_color(&data->node, root); | |
822 | } | |
823 | ||
0e9b07e5 | 824 | static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread) |
cdce9d73 | 825 | { |
36479484 | 826 | struct work_atoms *atoms = zalloc(sizeof(*atoms)); |
a116e05d ACM |
827 | if (!atoms) { |
828 | pr_err("No memory at %s\n", __func__); | |
829 | return -1; | |
830 | } | |
cdce9d73 | 831 | |
17562205 | 832 | atoms->thread = thread; |
39aeb52f | 833 | INIT_LIST_HEAD(&atoms->work_list); |
0e9b07e5 | 834 | __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid); |
a116e05d | 835 | return 0; |
cdce9d73 FW |
836 | } |
837 | ||
9ec3f4e4 | 838 | static char sched_out_state(u64 prev_state) |
cdce9d73 FW |
839 | { |
840 | const char *str = TASK_STATE_TO_CHAR_STR; | |
841 | ||
9ec3f4e4 | 842 | return str[prev_state]; |
cdce9d73 FW |
843 | } |
844 | ||
a116e05d | 845 | static int |
39aeb52f | 846 | add_sched_out_event(struct work_atoms *atoms, |
847 | char run_state, | |
848 | u64 timestamp) | |
cdce9d73 | 849 | { |
36479484 | 850 | struct work_atom *atom = zalloc(sizeof(*atom)); |
a116e05d ACM |
851 | if (!atom) { |
852 | pr_err("Non memory at %s", __func__); | |
853 | return -1; | |
854 | } | |
cdce9d73 | 855 | |
aa1ab9d2 FW |
856 | atom->sched_out_time = timestamp; |
857 | ||
39aeb52f | 858 | if (run_state == 'R') { |
b1ffe8f3 | 859 | atom->state = THREAD_WAIT_CPU; |
aa1ab9d2 | 860 | atom->wake_up_time = atom->sched_out_time; |
c6ced611 FW |
861 | } |
862 | ||
39aeb52f | 863 | list_add_tail(&atom->list, &atoms->work_list); |
a116e05d | 864 | return 0; |
cdce9d73 FW |
865 | } |
866 | ||
867 | static void | |
1d037ca1 IT |
868 | add_runtime_event(struct work_atoms *atoms, u64 delta, |
869 | u64 timestamp __maybe_unused) | |
39aeb52f | 870 | { |
871 | struct work_atom *atom; | |
872 | ||
873 | BUG_ON(list_empty(&atoms->work_list)); | |
874 | ||
875 | atom = list_entry(atoms->work_list.prev, struct work_atom, list); | |
876 | ||
877 | atom->runtime += delta; | |
878 | atoms->total_runtime += delta; | |
879 | } | |
880 | ||
881 | static void | |
882 | add_sched_in_event(struct work_atoms *atoms, u64 timestamp) | |
cdce9d73 | 883 | { |
b1ffe8f3 | 884 | struct work_atom *atom; |
66685678 | 885 | u64 delta; |
cdce9d73 | 886 | |
39aeb52f | 887 | if (list_empty(&atoms->work_list)) |
cdce9d73 FW |
888 | return; |
889 | ||
39aeb52f | 890 | atom = list_entry(atoms->work_list.prev, struct work_atom, list); |
cdce9d73 | 891 | |
b1ffe8f3 | 892 | if (atom->state != THREAD_WAIT_CPU) |
cdce9d73 FW |
893 | return; |
894 | ||
b1ffe8f3 IM |
895 | if (timestamp < atom->wake_up_time) { |
896 | atom->state = THREAD_IGNORE; | |
cdce9d73 FW |
897 | return; |
898 | } | |
899 | ||
b1ffe8f3 IM |
900 | atom->state = THREAD_SCHED_IN; |
901 | atom->sched_in_time = timestamp; | |
66685678 | 902 | |
b1ffe8f3 | 903 | delta = atom->sched_in_time - atom->wake_up_time; |
66685678 | 904 | atoms->total_lat += delta; |
3786310a | 905 | if (delta > atoms->max_lat) { |
66685678 | 906 | atoms->max_lat = delta; |
3786310a FW |
907 | atoms->max_lat_at = timestamp; |
908 | } | |
66685678 | 909 | atoms->nb_atoms++; |
cdce9d73 FW |
910 | } |
911 | ||
9ec3f4e4 ACM |
912 | static int latency_switch_event(struct perf_sched *sched, |
913 | struct perf_evsel *evsel, | |
914 | struct perf_sample *sample, | |
915 | struct machine *machine) | |
cdce9d73 | 916 | { |
9ec3f4e4 ACM |
917 | const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"), |
918 | next_pid = perf_evsel__intval(evsel, sample, "next_pid"); | |
919 | const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state"); | |
39aeb52f | 920 | struct work_atoms *out_events, *in_events; |
cdce9d73 | 921 | struct thread *sched_out, *sched_in; |
7f7f8d0b ACM |
922 | u64 timestamp0, timestamp = sample->time; |
923 | int cpu = sample->cpu; | |
ea92ed5a IM |
924 | s64 delta; |
925 | ||
39aeb52f | 926 | BUG_ON(cpu >= MAX_CPUS || cpu < 0); |
ea92ed5a | 927 | |
0e9b07e5 ACM |
928 | timestamp0 = sched->cpu_last_switched[cpu]; |
929 | sched->cpu_last_switched[cpu] = timestamp; | |
ea92ed5a IM |
930 | if (timestamp0) |
931 | delta = timestamp - timestamp0; | |
932 | else | |
933 | delta = 0; | |
934 | ||
a116e05d ACM |
935 | if (delta < 0) { |
936 | pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta); | |
937 | return -1; | |
938 | } | |
cdce9d73 | 939 | |
1fcb8768 AH |
940 | sched_out = machine__findnew_thread(machine, -1, prev_pid); |
941 | sched_in = machine__findnew_thread(machine, -1, next_pid); | |
cdce9d73 | 942 | |
0e9b07e5 | 943 | out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid); |
39aeb52f | 944 | if (!out_events) { |
0e9b07e5 | 945 | if (thread_atoms_insert(sched, sched_out)) |
a116e05d | 946 | return -1; |
0e9b07e5 | 947 | out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid); |
a116e05d ACM |
948 | if (!out_events) { |
949 | pr_err("out-event: Internal tree error"); | |
950 | return -1; | |
951 | } | |
39aeb52f | 952 | } |
9ec3f4e4 | 953 | if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp)) |
a116e05d | 954 | return -1; |
39aeb52f | 955 | |
0e9b07e5 | 956 | in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid); |
39aeb52f | 957 | if (!in_events) { |
0e9b07e5 | 958 | if (thread_atoms_insert(sched, sched_in)) |
a116e05d | 959 | return -1; |
0e9b07e5 | 960 | in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid); |
a116e05d ACM |
961 | if (!in_events) { |
962 | pr_err("in-event: Internal tree error"); | |
963 | return -1; | |
964 | } | |
39aeb52f | 965 | /* |
966 | * Take came in we have not heard about yet, | |
967 | * add in an initial atom in runnable state: | |
968 | */ | |
a116e05d ACM |
969 | if (add_sched_out_event(in_events, 'R', timestamp)) |
970 | return -1; | |
cdce9d73 | 971 | } |
39aeb52f | 972 | add_sched_in_event(in_events, timestamp); |
a116e05d ACM |
973 | |
974 | return 0; | |
39aeb52f | 975 | } |
cdce9d73 | 976 | |
9ec3f4e4 ACM |
977 | static int latency_runtime_event(struct perf_sched *sched, |
978 | struct perf_evsel *evsel, | |
979 | struct perf_sample *sample, | |
980 | struct machine *machine) | |
39aeb52f | 981 | { |
9ec3f4e4 ACM |
982 | const u32 pid = perf_evsel__intval(evsel, sample, "pid"); |
983 | const u64 runtime = perf_evsel__intval(evsel, sample, "runtime"); | |
1fcb8768 | 984 | struct thread *thread = machine__findnew_thread(machine, -1, pid); |
0e9b07e5 | 985 | struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid); |
7f7f8d0b ACM |
986 | u64 timestamp = sample->time; |
987 | int cpu = sample->cpu; | |
39aeb52f | 988 | |
989 | BUG_ON(cpu >= MAX_CPUS || cpu < 0); | |
39aeb52f | 990 | if (!atoms) { |
0e9b07e5 | 991 | if (thread_atoms_insert(sched, thread)) |
a116e05d | 992 | return -1; |
0e9b07e5 | 993 | atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid); |
a116e05d | 994 | if (!atoms) { |
60b7d14a | 995 | pr_err("in-event: Internal tree error"); |
a116e05d ACM |
996 | return -1; |
997 | } | |
998 | if (add_sched_out_event(atoms, 'R', timestamp)) | |
999 | return -1; | |
cdce9d73 FW |
1000 | } |
1001 | ||
9ec3f4e4 | 1002 | add_runtime_event(atoms, runtime, timestamp); |
a116e05d | 1003 | return 0; |
cdce9d73 FW |
1004 | } |
1005 | ||
9ec3f4e4 ACM |
1006 | static int latency_wakeup_event(struct perf_sched *sched, |
1007 | struct perf_evsel *evsel, | |
1008 | struct perf_sample *sample, | |
1009 | struct machine *machine) | |
cdce9d73 | 1010 | { |
0680ee7d | 1011 | const u32 pid = perf_evsel__intval(evsel, sample, "pid"); |
39aeb52f | 1012 | struct work_atoms *atoms; |
b1ffe8f3 | 1013 | struct work_atom *atom; |
cdce9d73 | 1014 | struct thread *wakee; |
7f7f8d0b | 1015 | u64 timestamp = sample->time; |
cdce9d73 | 1016 | |
1fcb8768 | 1017 | wakee = machine__findnew_thread(machine, -1, pid); |
0e9b07e5 | 1018 | atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid); |
17562205 | 1019 | if (!atoms) { |
0e9b07e5 | 1020 | if (thread_atoms_insert(sched, wakee)) |
a116e05d | 1021 | return -1; |
0e9b07e5 | 1022 | atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid); |
a116e05d | 1023 | if (!atoms) { |
60b7d14a | 1024 | pr_err("wakeup-event: Internal tree error"); |
a116e05d ACM |
1025 | return -1; |
1026 | } | |
1027 | if (add_sched_out_event(atoms, 'S', timestamp)) | |
1028 | return -1; | |
cdce9d73 FW |
1029 | } |
1030 | ||
39aeb52f | 1031 | BUG_ON(list_empty(&atoms->work_list)); |
cdce9d73 | 1032 | |
39aeb52f | 1033 | atom = list_entry(atoms->work_list.prev, struct work_atom, list); |
cdce9d73 | 1034 | |
55ffb7a6 | 1035 | /* |
67d6259d DY |
1036 | * As we do not guarantee the wakeup event happens when |
1037 | * task is out of run queue, also may happen when task is | |
1038 | * on run queue and wakeup only change ->state to TASK_RUNNING, | |
1039 | * then we should not set the ->wake_up_time when wake up a | |
1040 | * task which is on run queue. | |
1041 | * | |
55ffb7a6 MG |
1042 | * You WILL be missing events if you've recorded only |
1043 | * one CPU, or are only looking at only one, so don't | |
67d6259d | 1044 | * skip in this case. |
55ffb7a6 | 1045 | */ |
0e9b07e5 | 1046 | if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING) |
67d6259d | 1047 | return 0; |
cdce9d73 | 1048 | |
0e9b07e5 | 1049 | sched->nr_timestamps++; |
ea57c4f5 | 1050 | if (atom->sched_out_time > timestamp) { |
0e9b07e5 | 1051 | sched->nr_unordered_timestamps++; |
a116e05d | 1052 | return 0; |
ea57c4f5 | 1053 | } |
aa1ab9d2 | 1054 | |
b1ffe8f3 IM |
1055 | atom->state = THREAD_WAIT_CPU; |
1056 | atom->wake_up_time = timestamp; | |
a116e05d | 1057 | return 0; |
cdce9d73 FW |
1058 | } |
1059 | ||
9ec3f4e4 ACM |
1060 | static int latency_migrate_task_event(struct perf_sched *sched, |
1061 | struct perf_evsel *evsel, | |
1062 | struct perf_sample *sample, | |
1063 | struct machine *machine) | |
55ffb7a6 | 1064 | { |
9ec3f4e4 | 1065 | const u32 pid = perf_evsel__intval(evsel, sample, "pid"); |
7f7f8d0b | 1066 | u64 timestamp = sample->time; |
55ffb7a6 MG |
1067 | struct work_atoms *atoms; |
1068 | struct work_atom *atom; | |
1069 | struct thread *migrant; | |
1070 | ||
1071 | /* | |
1072 | * Only need to worry about migration when profiling one CPU. | |
1073 | */ | |
0e9b07e5 | 1074 | if (sched->profile_cpu == -1) |
a116e05d | 1075 | return 0; |
55ffb7a6 | 1076 | |
1fcb8768 | 1077 | migrant = machine__findnew_thread(machine, -1, pid); |
0e9b07e5 | 1078 | atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid); |
55ffb7a6 | 1079 | if (!atoms) { |
0e9b07e5 | 1080 | if (thread_atoms_insert(sched, migrant)) |
a116e05d | 1081 | return -1; |
b9c5143a | 1082 | register_pid(sched, migrant->tid, thread__comm_str(migrant)); |
0e9b07e5 | 1083 | atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid); |
a116e05d | 1084 | if (!atoms) { |
60b7d14a | 1085 | pr_err("migration-event: Internal tree error"); |
a116e05d ACM |
1086 | return -1; |
1087 | } | |
1088 | if (add_sched_out_event(atoms, 'R', timestamp)) | |
1089 | return -1; | |
55ffb7a6 MG |
1090 | } |
1091 | ||
1092 | BUG_ON(list_empty(&atoms->work_list)); | |
1093 | ||
1094 | atom = list_entry(atoms->work_list.prev, struct work_atom, list); | |
1095 | atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp; | |
1096 | ||
0e9b07e5 | 1097 | sched->nr_timestamps++; |
55ffb7a6 MG |
1098 | |
1099 | if (atom->sched_out_time > timestamp) | |
0e9b07e5 | 1100 | sched->nr_unordered_timestamps++; |
a116e05d ACM |
1101 | |
1102 | return 0; | |
55ffb7a6 MG |
1103 | } |
1104 | ||
0e9b07e5 | 1105 | static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list) |
cdce9d73 | 1106 | { |
cdce9d73 FW |
1107 | int i; |
1108 | int ret; | |
66685678 | 1109 | u64 avg; |
cdce9d73 | 1110 | |
39aeb52f | 1111 | if (!work_list->nb_atoms) |
cdce9d73 | 1112 | return; |
ea57c4f5 IM |
1113 | /* |
1114 | * Ignore idle threads: | |
1115 | */ | |
b9c5143a | 1116 | if (!strcmp(thread__comm_str(work_list->thread), "swapper")) |
ea57c4f5 | 1117 | return; |
cdce9d73 | 1118 | |
0e9b07e5 ACM |
1119 | sched->all_runtime += work_list->total_runtime; |
1120 | sched->all_count += work_list->nb_atoms; | |
66685678 | 1121 | |
b9c5143a | 1122 | ret = printf(" %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid); |
cdce9d73 | 1123 | |
08f69e6c | 1124 | for (i = 0; i < 24 - ret; i++) |
cdce9d73 FW |
1125 | printf(" "); |
1126 | ||
39aeb52f | 1127 | avg = work_list->total_lat / work_list->nb_atoms; |
cdce9d73 | 1128 | |
80790e0b | 1129 | printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13.6f s\n", |
39aeb52f | 1130 | (double)work_list->total_runtime / 1e6, |
1131 | work_list->nb_atoms, (double)avg / 1e6, | |
3786310a FW |
1132 | (double)work_list->max_lat / 1e6, |
1133 | (double)work_list->max_lat_at / 1e9); | |
cdce9d73 FW |
1134 | } |
1135 | ||
39aeb52f | 1136 | static int pid_cmp(struct work_atoms *l, struct work_atoms *r) |
daa1d7a5 | 1137 | { |
38051234 | 1138 | if (l->thread->tid < r->thread->tid) |
daa1d7a5 | 1139 | return -1; |
38051234 | 1140 | if (l->thread->tid > r->thread->tid) |
daa1d7a5 FW |
1141 | return 1; |
1142 | ||
1143 | return 0; | |
1144 | } | |
1145 | ||
39aeb52f | 1146 | static int avg_cmp(struct work_atoms *l, struct work_atoms *r) |
daa1d7a5 FW |
1147 | { |
1148 | u64 avgl, avgr; | |
1149 | ||
1150 | if (!l->nb_atoms) | |
1151 | return -1; | |
1152 | ||
1153 | if (!r->nb_atoms) | |
1154 | return 1; | |
1155 | ||
1156 | avgl = l->total_lat / l->nb_atoms; | |
1157 | avgr = r->total_lat / r->nb_atoms; | |
1158 | ||
1159 | if (avgl < avgr) | |
1160 | return -1; | |
1161 | if (avgl > avgr) | |
1162 | return 1; | |
1163 | ||
1164 | return 0; | |
1165 | } | |
1166 | ||
39aeb52f | 1167 | static int max_cmp(struct work_atoms *l, struct work_atoms *r) |
daa1d7a5 FW |
1168 | { |
1169 | if (l->max_lat < r->max_lat) | |
1170 | return -1; | |
1171 | if (l->max_lat > r->max_lat) | |
1172 | return 1; | |
1173 | ||
1174 | return 0; | |
1175 | } | |
1176 | ||
39aeb52f | 1177 | static int switch_cmp(struct work_atoms *l, struct work_atoms *r) |
daa1d7a5 FW |
1178 | { |
1179 | if (l->nb_atoms < r->nb_atoms) | |
1180 | return -1; | |
1181 | if (l->nb_atoms > r->nb_atoms) | |
1182 | return 1; | |
1183 | ||
1184 | return 0; | |
1185 | } | |
1186 | ||
39aeb52f | 1187 | static int runtime_cmp(struct work_atoms *l, struct work_atoms *r) |
daa1d7a5 FW |
1188 | { |
1189 | if (l->total_runtime < r->total_runtime) | |
1190 | return -1; | |
1191 | if (l->total_runtime > r->total_runtime) | |
1192 | return 1; | |
1193 | ||
1194 | return 0; | |
1195 | } | |
1196 | ||
cbef79a8 | 1197 | static int sort_dimension__add(const char *tok, struct list_head *list) |
daa1d7a5 | 1198 | { |
0e9b07e5 ACM |
1199 | size_t i; |
1200 | static struct sort_dimension avg_sort_dimension = { | |
1201 | .name = "avg", | |
1202 | .cmp = avg_cmp, | |
1203 | }; | |
1204 | static struct sort_dimension max_sort_dimension = { | |
1205 | .name = "max", | |
1206 | .cmp = max_cmp, | |
1207 | }; | |
1208 | static struct sort_dimension pid_sort_dimension = { | |
1209 | .name = "pid", | |
1210 | .cmp = pid_cmp, | |
1211 | }; | |
1212 | static struct sort_dimension runtime_sort_dimension = { | |
1213 | .name = "runtime", | |
1214 | .cmp = runtime_cmp, | |
1215 | }; | |
1216 | static struct sort_dimension switch_sort_dimension = { | |
1217 | .name = "switch", | |
1218 | .cmp = switch_cmp, | |
1219 | }; | |
1220 | struct sort_dimension *available_sorts[] = { | |
1221 | &pid_sort_dimension, | |
1222 | &avg_sort_dimension, | |
1223 | &max_sort_dimension, | |
1224 | &switch_sort_dimension, | |
1225 | &runtime_sort_dimension, | |
1226 | }; | |
daa1d7a5 | 1227 | |
0e9b07e5 | 1228 | for (i = 0; i < ARRAY_SIZE(available_sorts); i++) { |
daa1d7a5 FW |
1229 | if (!strcmp(available_sorts[i]->name, tok)) { |
1230 | list_add_tail(&available_sorts[i]->list, list); | |
1231 | ||
1232 | return 0; | |
1233 | } | |
1234 | } | |
1235 | ||
1236 | return -1; | |
1237 | } | |
1238 | ||
0e9b07e5 | 1239 | static void perf_sched__sort_lat(struct perf_sched *sched) |
daa1d7a5 FW |
1240 | { |
1241 | struct rb_node *node; | |
1242 | ||
1243 | for (;;) { | |
39aeb52f | 1244 | struct work_atoms *data; |
0e9b07e5 | 1245 | node = rb_first(&sched->atom_root); |
daa1d7a5 FW |
1246 | if (!node) |
1247 | break; | |
1248 | ||
0e9b07e5 | 1249 | rb_erase(node, &sched->atom_root); |
39aeb52f | 1250 | data = rb_entry(node, struct work_atoms, node); |
0e9b07e5 | 1251 | __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list); |
daa1d7a5 FW |
1252 | } |
1253 | } | |
1254 | ||
0e9b07e5 | 1255 | static int process_sched_wakeup_event(struct perf_tool *tool, |
2b7fcbc5 | 1256 | struct perf_evsel *evsel, |
1d037ca1 | 1257 | struct perf_sample *sample, |
4218e673 | 1258 | struct machine *machine) |
419ab0d6 | 1259 | { |
0e9b07e5 | 1260 | struct perf_sched *sched = container_of(tool, struct perf_sched, tool); |
419ab0d6 | 1261 | |
9ec3f4e4 ACM |
1262 | if (sched->tp_handler->wakeup_event) |
1263 | return sched->tp_handler->wakeup_event(sched, evsel, sample, machine); | |
a116e05d | 1264 | |
2b7fcbc5 | 1265 | return 0; |
419ab0d6 FW |
1266 | } |
1267 | ||
9ec3f4e4 ACM |
1268 | static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel, |
1269 | struct perf_sample *sample, struct machine *machine) | |
0ec04e16 | 1270 | { |
9d372ca5 DY |
1271 | const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid"); |
1272 | struct thread *sched_in; | |
0ec04e16 | 1273 | int new_shortname; |
7f7f8d0b | 1274 | u64 timestamp0, timestamp = sample->time; |
0ec04e16 | 1275 | s64 delta; |
7f7f8d0b | 1276 | int cpu, this_cpu = sample->cpu; |
0ec04e16 IM |
1277 | |
1278 | BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0); | |
1279 | ||
0e9b07e5 ACM |
1280 | if (this_cpu > sched->max_cpu) |
1281 | sched->max_cpu = this_cpu; | |
0ec04e16 | 1282 | |
0e9b07e5 ACM |
1283 | timestamp0 = sched->cpu_last_switched[this_cpu]; |
1284 | sched->cpu_last_switched[this_cpu] = timestamp; | |
0ec04e16 IM |
1285 | if (timestamp0) |
1286 | delta = timestamp - timestamp0; | |
1287 | else | |
1288 | delta = 0; | |
1289 | ||
a116e05d | 1290 | if (delta < 0) { |
60b7d14a | 1291 | pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta); |
a116e05d ACM |
1292 | return -1; |
1293 | } | |
0ec04e16 | 1294 | |
1fcb8768 | 1295 | sched_in = machine__findnew_thread(machine, -1, next_pid); |
0ec04e16 | 1296 | |
0e9b07e5 | 1297 | sched->curr_thread[this_cpu] = sched_in; |
0ec04e16 IM |
1298 | |
1299 | printf(" "); | |
1300 | ||
1301 | new_shortname = 0; | |
1302 | if (!sched_in->shortname[0]) { | |
6bcab4e1 D |
1303 | if (!strcmp(thread__comm_str(sched_in), "swapper")) { |
1304 | /* | |
1305 | * Don't allocate a letter-number for swapper:0 | |
1306 | * as a shortname. Instead, we use '.' for it. | |
1307 | */ | |
1308 | sched_in->shortname[0] = '.'; | |
1309 | sched_in->shortname[1] = ' '; | |
0ec04e16 | 1310 | } else { |
6bcab4e1 D |
1311 | sched_in->shortname[0] = sched->next_shortname1; |
1312 | sched_in->shortname[1] = sched->next_shortname2; | |
1313 | ||
1314 | if (sched->next_shortname1 < 'Z') { | |
1315 | sched->next_shortname1++; | |
0ec04e16 | 1316 | } else { |
6bcab4e1 D |
1317 | sched->next_shortname1 = 'A'; |
1318 | if (sched->next_shortname2 < '9') | |
1319 | sched->next_shortname2++; | |
1320 | else | |
1321 | sched->next_shortname2 = '0'; | |
0ec04e16 IM |
1322 | } |
1323 | } | |
1324 | new_shortname = 1; | |
1325 | } | |
1326 | ||
0e9b07e5 | 1327 | for (cpu = 0; cpu <= sched->max_cpu; cpu++) { |
0ec04e16 IM |
1328 | if (cpu != this_cpu) |
1329 | printf(" "); | |
1330 | else | |
1331 | printf("*"); | |
1332 | ||
6bcab4e1 D |
1333 | if (sched->curr_thread[cpu]) |
1334 | printf("%2s ", sched->curr_thread[cpu]->shortname); | |
1335 | else | |
0ec04e16 IM |
1336 | printf(" "); |
1337 | } | |
1338 | ||
1339 | printf(" %12.6f secs ", (double)timestamp/1e9); | |
1340 | if (new_shortname) { | |
1341 | printf("%s => %s:%d\n", | |
b9c5143a | 1342 | sched_in->shortname, thread__comm_str(sched_in), sched_in->tid); |
0ec04e16 IM |
1343 | } else { |
1344 | printf("\n"); | |
1345 | } | |
a116e05d ACM |
1346 | |
1347 | return 0; | |
0ec04e16 IM |
1348 | } |
1349 | ||
0e9b07e5 | 1350 | static int process_sched_switch_event(struct perf_tool *tool, |
2b7fcbc5 | 1351 | struct perf_evsel *evsel, |
1d037ca1 | 1352 | struct perf_sample *sample, |
4218e673 | 1353 | struct machine *machine) |
419ab0d6 | 1354 | { |
0e9b07e5 | 1355 | struct perf_sched *sched = container_of(tool, struct perf_sched, tool); |
a116e05d | 1356 | int this_cpu = sample->cpu, err = 0; |
2b7fcbc5 ACM |
1357 | u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"), |
1358 | next_pid = perf_evsel__intval(evsel, sample, "next_pid"); | |
419ab0d6 | 1359 | |
0e9b07e5 | 1360 | if (sched->curr_pid[this_cpu] != (u32)-1) { |
c8a37751 IM |
1361 | /* |
1362 | * Are we trying to switch away a PID that is | |
1363 | * not current? | |
1364 | */ | |
2b7fcbc5 | 1365 | if (sched->curr_pid[this_cpu] != prev_pid) |
0e9b07e5 | 1366 | sched->nr_context_switch_bugs++; |
c8a37751 | 1367 | } |
c8a37751 | 1368 | |
9ec3f4e4 ACM |
1369 | if (sched->tp_handler->switch_event) |
1370 | err = sched->tp_handler->switch_event(sched, evsel, sample, machine); | |
2b7fcbc5 ACM |
1371 | |
1372 | sched->curr_pid[this_cpu] = next_pid; | |
a116e05d | 1373 | return err; |
419ab0d6 FW |
1374 | } |
1375 | ||
0e9b07e5 | 1376 | static int process_sched_runtime_event(struct perf_tool *tool, |
2b7fcbc5 | 1377 | struct perf_evsel *evsel, |
1d037ca1 | 1378 | struct perf_sample *sample, |
4218e673 | 1379 | struct machine *machine) |
39aeb52f | 1380 | { |
0e9b07e5 | 1381 | struct perf_sched *sched = container_of(tool, struct perf_sched, tool); |
39aeb52f | 1382 | |
9ec3f4e4 ACM |
1383 | if (sched->tp_handler->runtime_event) |
1384 | return sched->tp_handler->runtime_event(sched, evsel, sample, machine); | |
a116e05d | 1385 | |
2b7fcbc5 | 1386 | return 0; |
39aeb52f | 1387 | } |
1388 | ||
cb627505 DA |
1389 | static int perf_sched__process_fork_event(struct perf_tool *tool, |
1390 | union perf_event *event, | |
1391 | struct perf_sample *sample, | |
1392 | struct machine *machine) | |
fbf94829 | 1393 | { |
0e9b07e5 | 1394 | struct perf_sched *sched = container_of(tool, struct perf_sched, tool); |
46538818 | 1395 | |
cb627505 DA |
1396 | /* run the fork event through the perf machineruy */ |
1397 | perf_event__process_fork(tool, event, sample, machine); | |
1398 | ||
1399 | /* and then run additional processing needed for this command */ | |
9ec3f4e4 | 1400 | if (sched->tp_handler->fork_event) |
cb627505 | 1401 | return sched->tp_handler->fork_event(sched, event, machine); |
a116e05d | 1402 | |
2b7fcbc5 | 1403 | return 0; |
fbf94829 IM |
1404 | } |
1405 | ||
0e9b07e5 | 1406 | static int process_sched_migrate_task_event(struct perf_tool *tool, |
2b7fcbc5 | 1407 | struct perf_evsel *evsel, |
1d037ca1 | 1408 | struct perf_sample *sample, |
4218e673 | 1409 | struct machine *machine) |
55ffb7a6 | 1410 | { |
0e9b07e5 | 1411 | struct perf_sched *sched = container_of(tool, struct perf_sched, tool); |
55ffb7a6 | 1412 | |
9ec3f4e4 ACM |
1413 | if (sched->tp_handler->migrate_task_event) |
1414 | return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine); | |
a116e05d | 1415 | |
2b7fcbc5 | 1416 | return 0; |
55ffb7a6 MG |
1417 | } |
1418 | ||
a116e05d | 1419 | typedef int (*tracepoint_handler)(struct perf_tool *tool, |
2b7fcbc5 | 1420 | struct perf_evsel *evsel, |
a116e05d | 1421 | struct perf_sample *sample, |
4218e673 | 1422 | struct machine *machine); |
ec156764 | 1423 | |
1d037ca1 IT |
1424 | static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused, |
1425 | union perf_event *event __maybe_unused, | |
ee29be62 ACM |
1426 | struct perf_sample *sample, |
1427 | struct perf_evsel *evsel, | |
1428 | struct machine *machine) | |
0a02ad93 | 1429 | { |
a116e05d | 1430 | int err = 0; |
0a02ad93 | 1431 | |
ee29be62 | 1432 | evsel->hists.stats.total_period += sample->period; |
1844dbcb | 1433 | hists__inc_nr_samples(&evsel->hists, true); |
f39cdf25 | 1434 | |
744a9719 ACM |
1435 | if (evsel->handler != NULL) { |
1436 | tracepoint_handler f = evsel->handler; | |
2b7fcbc5 | 1437 | err = f(tool, evsel, sample, machine); |
ee29be62 | 1438 | } |
0a02ad93 | 1439 | |
a116e05d | 1440 | return err; |
0a02ad93 IM |
1441 | } |
1442 | ||
ad9def7c | 1443 | static int perf_sched__read_events(struct perf_sched *sched, |
0e9b07e5 | 1444 | struct perf_session **psession) |
0a02ad93 | 1445 | { |
ee29be62 ACM |
1446 | const struct perf_evsel_str_handler handlers[] = { |
1447 | { "sched:sched_switch", process_sched_switch_event, }, | |
1448 | { "sched:sched_stat_runtime", process_sched_runtime_event, }, | |
1449 | { "sched:sched_wakeup", process_sched_wakeup_event, }, | |
1450 | { "sched:sched_wakeup_new", process_sched_wakeup_event, }, | |
ee29be62 ACM |
1451 | { "sched:sched_migrate_task", process_sched_migrate_task_event, }, |
1452 | }; | |
da378962 | 1453 | struct perf_session *session; |
f5fc1412 JO |
1454 | struct perf_data_file file = { |
1455 | .path = input_name, | |
1456 | .mode = PERF_DATA_MODE_READ, | |
1457 | }; | |
da378962 | 1458 | |
f5fc1412 | 1459 | session = perf_session__new(&file, false, &sched->tool); |
a116e05d ACM |
1460 | if (session == NULL) { |
1461 | pr_debug("No Memory for session\n"); | |
1462 | return -1; | |
1463 | } | |
94c744b6 | 1464 | |
0a7e6d1b | 1465 | symbol__init(&session->header.env); |
04934106 | 1466 | |
a116e05d ACM |
1467 | if (perf_session__set_tracepoints_handlers(session, handlers)) |
1468 | goto out_delete; | |
ee29be62 | 1469 | |
cee75ac7 | 1470 | if (perf_session__has_traces(session, "record -R")) { |
0e9b07e5 | 1471 | int err = perf_session__process_events(session, &sched->tool); |
a116e05d ACM |
1472 | if (err) { |
1473 | pr_err("Failed to process events, error %d", err); | |
1474 | goto out_delete; | |
1475 | } | |
4c09bafa | 1476 | |
28a6b6aa ACM |
1477 | sched->nr_events = session->stats.nr_events[0]; |
1478 | sched->nr_lost_events = session->stats.total_lost; | |
1479 | sched->nr_lost_chunks = session->stats.nr_events[PERF_RECORD_LOST]; | |
cee75ac7 | 1480 | } |
d549c769 | 1481 | |
4c09bafa JO |
1482 | if (psession) |
1483 | *psession = session; | |
ad9def7c DA |
1484 | else |
1485 | perf_session__delete(session); | |
a116e05d ACM |
1486 | |
1487 | return 0; | |
1488 | ||
1489 | out_delete: | |
1490 | perf_session__delete(session); | |
1491 | return -1; | |
0a02ad93 IM |
1492 | } |
1493 | ||
0e9b07e5 | 1494 | static void print_bad_events(struct perf_sched *sched) |
0ec04e16 | 1495 | { |
0e9b07e5 | 1496 | if (sched->nr_unordered_timestamps && sched->nr_timestamps) { |
0ec04e16 | 1497 | printf(" INFO: %.3f%% unordered timestamps (%ld out of %ld)\n", |
0e9b07e5 ACM |
1498 | (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0, |
1499 | sched->nr_unordered_timestamps, sched->nr_timestamps); | |
0ec04e16 | 1500 | } |
0e9b07e5 | 1501 | if (sched->nr_lost_events && sched->nr_events) { |
0ec04e16 | 1502 | printf(" INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n", |
0e9b07e5 ACM |
1503 | (double)sched->nr_lost_events/(double)sched->nr_events * 100.0, |
1504 | sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks); | |
0ec04e16 | 1505 | } |
0e9b07e5 | 1506 | if (sched->nr_context_switch_bugs && sched->nr_timestamps) { |
0ec04e16 | 1507 | printf(" INFO: %.3f%% context switch bugs (%ld out of %ld)", |
0e9b07e5 ACM |
1508 | (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0, |
1509 | sched->nr_context_switch_bugs, sched->nr_timestamps); | |
1510 | if (sched->nr_lost_events) | |
0ec04e16 IM |
1511 | printf(" (due to lost events?)"); |
1512 | printf("\n"); | |
1513 | } | |
1514 | } | |
1515 | ||
0e9b07e5 | 1516 | static int perf_sched__lat(struct perf_sched *sched) |
0ec04e16 IM |
1517 | { |
1518 | struct rb_node *next; | |
4c09bafa | 1519 | struct perf_session *session; |
0ec04e16 IM |
1520 | |
1521 | setup_pager(); | |
ad9def7c DA |
1522 | |
1523 | /* save session -- references to threads are held in work_list */ | |
1524 | if (perf_sched__read_events(sched, &session)) | |
a116e05d | 1525 | return -1; |
ad9def7c | 1526 | |
0e9b07e5 | 1527 | perf_sched__sort_lat(sched); |
0ec04e16 | 1528 | |
80790e0b RR |
1529 | printf("\n -----------------------------------------------------------------------------------------------------------------\n"); |
1530 | printf(" Task | Runtime ms | Switches | Average delay ms | Maximum delay ms | Maximum delay at |\n"); | |
1531 | printf(" -----------------------------------------------------------------------------------------------------------------\n"); | |
0ec04e16 | 1532 | |
0e9b07e5 | 1533 | next = rb_first(&sched->sorted_atom_root); |
0ec04e16 IM |
1534 | |
1535 | while (next) { | |
1536 | struct work_atoms *work_list; | |
1537 | ||
1538 | work_list = rb_entry(next, struct work_atoms, node); | |
0e9b07e5 | 1539 | output_lat_thread(sched, work_list); |
0ec04e16 IM |
1540 | next = rb_next(next); |
1541 | } | |
1542 | ||
80790e0b | 1543 | printf(" -----------------------------------------------------------------------------------------------------------------\n"); |
9486aa38 | 1544 | printf(" TOTAL: |%11.3f ms |%9" PRIu64 " |\n", |
0e9b07e5 | 1545 | (double)sched->all_runtime / 1e6, sched->all_count); |
0ec04e16 IM |
1546 | |
1547 | printf(" ---------------------------------------------------\n"); | |
1548 | ||
0e9b07e5 | 1549 | print_bad_events(sched); |
0ec04e16 IM |
1550 | printf("\n"); |
1551 | ||
4c09bafa | 1552 | perf_session__delete(session); |
a116e05d | 1553 | return 0; |
0ec04e16 IM |
1554 | } |
1555 | ||
0e9b07e5 | 1556 | static int perf_sched__map(struct perf_sched *sched) |
0ec04e16 | 1557 | { |
0e9b07e5 | 1558 | sched->max_cpu = sysconf(_SC_NPROCESSORS_CONF); |
40749d0f | 1559 | |
0ec04e16 | 1560 | setup_pager(); |
ad9def7c | 1561 | if (perf_sched__read_events(sched, NULL)) |
a116e05d | 1562 | return -1; |
0e9b07e5 | 1563 | print_bad_events(sched); |
a116e05d | 1564 | return 0; |
0ec04e16 IM |
1565 | } |
1566 | ||
0e9b07e5 | 1567 | static int perf_sched__replay(struct perf_sched *sched) |
0ec04e16 IM |
1568 | { |
1569 | unsigned long i; | |
1570 | ||
0e9b07e5 ACM |
1571 | calibrate_run_measurement_overhead(sched); |
1572 | calibrate_sleep_measurement_overhead(sched); | |
0ec04e16 | 1573 | |
0e9b07e5 | 1574 | test_calibrations(sched); |
0ec04e16 | 1575 | |
ad9def7c | 1576 | if (perf_sched__read_events(sched, NULL)) |
a116e05d | 1577 | return -1; |
0ec04e16 | 1578 | |
0e9b07e5 ACM |
1579 | printf("nr_run_events: %ld\n", sched->nr_run_events); |
1580 | printf("nr_sleep_events: %ld\n", sched->nr_sleep_events); | |
1581 | printf("nr_wakeup_events: %ld\n", sched->nr_wakeup_events); | |
0ec04e16 | 1582 | |
0e9b07e5 ACM |
1583 | if (sched->targetless_wakeups) |
1584 | printf("target-less wakeups: %ld\n", sched->targetless_wakeups); | |
1585 | if (sched->multitarget_wakeups) | |
1586 | printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups); | |
1587 | if (sched->nr_run_events_optimized) | |
0ec04e16 | 1588 | printf("run atoms optimized: %ld\n", |
0e9b07e5 | 1589 | sched->nr_run_events_optimized); |
0ec04e16 | 1590 | |
0e9b07e5 ACM |
1591 | print_task_traces(sched); |
1592 | add_cross_task_wakeups(sched); | |
0ec04e16 | 1593 | |
0e9b07e5 | 1594 | create_tasks(sched); |
0ec04e16 | 1595 | printf("------------------------------------------------------------\n"); |
0e9b07e5 ACM |
1596 | for (i = 0; i < sched->replay_repeat; i++) |
1597 | run_one_test(sched); | |
a116e05d ACM |
1598 | |
1599 | return 0; | |
0ec04e16 IM |
1600 | } |
1601 | ||
0e9b07e5 ACM |
1602 | static void setup_sorting(struct perf_sched *sched, const struct option *options, |
1603 | const char * const usage_msg[]) | |
daa1d7a5 | 1604 | { |
0e9b07e5 | 1605 | char *tmp, *tok, *str = strdup(sched->sort_order); |
daa1d7a5 FW |
1606 | |
1607 | for (tok = strtok_r(str, ", ", &tmp); | |
1608 | tok; tok = strtok_r(NULL, ", ", &tmp)) { | |
0e9b07e5 | 1609 | if (sort_dimension__add(tok, &sched->sort_list) < 0) { |
daa1d7a5 | 1610 | error("Unknown --sort key: `%s'", tok); |
0e9b07e5 | 1611 | usage_with_options(usage_msg, options); |
daa1d7a5 FW |
1612 | } |
1613 | } | |
1614 | ||
1615 | free(str); | |
1616 | ||
0e9b07e5 | 1617 | sort_dimension__add("pid", &sched->cmp_pid); |
daa1d7a5 FW |
1618 | } |
1619 | ||
1fc35b29 IM |
1620 | static int __cmd_record(int argc, const char **argv) |
1621 | { | |
1622 | unsigned int rec_argc, i, j; | |
1623 | const char **rec_argv; | |
0e9b07e5 ACM |
1624 | const char * const record_args[] = { |
1625 | "record", | |
1626 | "-a", | |
1627 | "-R", | |
0e9b07e5 ACM |
1628 | "-m", "1024", |
1629 | "-c", "1", | |
1630 | "-e", "sched:sched_switch", | |
1631 | "-e", "sched:sched_stat_wait", | |
1632 | "-e", "sched:sched_stat_sleep", | |
1633 | "-e", "sched:sched_stat_iowait", | |
1634 | "-e", "sched:sched_stat_runtime", | |
0e9b07e5 ACM |
1635 | "-e", "sched:sched_process_fork", |
1636 | "-e", "sched:sched_wakeup", | |
7fff9597 | 1637 | "-e", "sched:sched_wakeup_new", |
0e9b07e5 ACM |
1638 | "-e", "sched:sched_migrate_task", |
1639 | }; | |
1fc35b29 IM |
1640 | |
1641 | rec_argc = ARRAY_SIZE(record_args) + argc - 1; | |
1642 | rec_argv = calloc(rec_argc + 1, sizeof(char *)); | |
1643 | ||
e462dc55 | 1644 | if (rec_argv == NULL) |
ce47dc56 CS |
1645 | return -ENOMEM; |
1646 | ||
1fc35b29 IM |
1647 | for (i = 0; i < ARRAY_SIZE(record_args); i++) |
1648 | rec_argv[i] = strdup(record_args[i]); | |
1649 | ||
1650 | for (j = 1; j < (unsigned int)argc; j++, i++) | |
1651 | rec_argv[i] = argv[j]; | |
1652 | ||
1653 | BUG_ON(i != rec_argc); | |
1654 | ||
1655 | return cmd_record(i, rec_argv, NULL); | |
1656 | } | |
1657 | ||
1d037ca1 | 1658 | int cmd_sched(int argc, const char **argv, const char *prefix __maybe_unused) |
0a02ad93 | 1659 | { |
8a39df8f AH |
1660 | const char default_sort_order[] = "avg, max, switch, runtime"; |
1661 | struct perf_sched sched = { | |
1662 | .tool = { | |
1663 | .sample = perf_sched__process_tracepoint_sample, | |
1664 | .comm = perf_event__process_comm, | |
1665 | .lost = perf_event__process_lost, | |
1666 | .fork = perf_sched__process_fork_event, | |
0a8cb85c | 1667 | .ordered_events = true, |
8a39df8f AH |
1668 | }, |
1669 | .cmp_pid = LIST_HEAD_INIT(sched.cmp_pid), | |
1670 | .sort_list = LIST_HEAD_INIT(sched.sort_list), | |
1671 | .start_work_mutex = PTHREAD_MUTEX_INITIALIZER, | |
1672 | .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER, | |
8a39df8f AH |
1673 | .sort_order = default_sort_order, |
1674 | .replay_repeat = 10, | |
1675 | .profile_cpu = -1, | |
1676 | .next_shortname1 = 'A', | |
1677 | .next_shortname2 = '0', | |
1678 | }; | |
0e9b07e5 ACM |
1679 | const struct option latency_options[] = { |
1680 | OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]", | |
1681 | "sort by key(s): runtime, switch, avg, max"), | |
1682 | OPT_INCR('v', "verbose", &verbose, | |
1683 | "be more verbose (show symbol address, etc)"), | |
1684 | OPT_INTEGER('C', "CPU", &sched.profile_cpu, | |
1685 | "CPU to profile on"), | |
1686 | OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, | |
1687 | "dump raw trace in ASCII"), | |
1688 | OPT_END() | |
1689 | }; | |
1690 | const struct option replay_options[] = { | |
1691 | OPT_UINTEGER('r', "repeat", &sched.replay_repeat, | |
1692 | "repeat the workload replay N times (-1: infinite)"), | |
1693 | OPT_INCR('v', "verbose", &verbose, | |
1694 | "be more verbose (show symbol address, etc)"), | |
1695 | OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, | |
1696 | "dump raw trace in ASCII"), | |
1697 | OPT_END() | |
1698 | }; | |
1699 | const struct option sched_options[] = { | |
70cb4e96 | 1700 | OPT_STRING('i', "input", &input_name, "file", |
0e9b07e5 ACM |
1701 | "input file name"), |
1702 | OPT_INCR('v', "verbose", &verbose, | |
1703 | "be more verbose (show symbol address, etc)"), | |
1704 | OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, | |
1705 | "dump raw trace in ASCII"), | |
1706 | OPT_END() | |
1707 | }; | |
1708 | const char * const latency_usage[] = { | |
1709 | "perf sched latency [<options>]", | |
1710 | NULL | |
1711 | }; | |
1712 | const char * const replay_usage[] = { | |
1713 | "perf sched replay [<options>]", | |
1714 | NULL | |
1715 | }; | |
a83edb2d RR |
1716 | const char *const sched_subcommands[] = { "record", "latency", "map", |
1717 | "replay", "script", NULL }; | |
1718 | const char *sched_usage[] = { | |
1719 | NULL, | |
0e9b07e5 ACM |
1720 | NULL |
1721 | }; | |
1722 | struct trace_sched_handler lat_ops = { | |
1723 | .wakeup_event = latency_wakeup_event, | |
1724 | .switch_event = latency_switch_event, | |
1725 | .runtime_event = latency_runtime_event, | |
0e9b07e5 ACM |
1726 | .migrate_task_event = latency_migrate_task_event, |
1727 | }; | |
1728 | struct trace_sched_handler map_ops = { | |
1729 | .switch_event = map_switch_event, | |
1730 | }; | |
1731 | struct trace_sched_handler replay_ops = { | |
1732 | .wakeup_event = replay_wakeup_event, | |
1733 | .switch_event = replay_switch_event, | |
1734 | .fork_event = replay_fork_event, | |
1735 | }; | |
156a2b02 AH |
1736 | unsigned int i; |
1737 | ||
1738 | for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++) | |
1739 | sched.curr_pid[i] = -1; | |
0e9b07e5 | 1740 | |
a83edb2d RR |
1741 | argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands, |
1742 | sched_usage, PARSE_OPT_STOP_AT_NON_OPTION); | |
f2858d8a IM |
1743 | if (!argc) |
1744 | usage_with_options(sched_usage, sched_options); | |
0a02ad93 | 1745 | |
c0777c5a | 1746 | /* |
133dc4c3 | 1747 | * Aliased to 'perf script' for now: |
c0777c5a | 1748 | */ |
133dc4c3 IM |
1749 | if (!strcmp(argv[0], "script")) |
1750 | return cmd_script(argc, argv, prefix); | |
c0777c5a | 1751 | |
1fc35b29 IM |
1752 | if (!strncmp(argv[0], "rec", 3)) { |
1753 | return __cmd_record(argc, argv); | |
1754 | } else if (!strncmp(argv[0], "lat", 3)) { | |
0e9b07e5 | 1755 | sched.tp_handler = &lat_ops; |
f2858d8a IM |
1756 | if (argc > 1) { |
1757 | argc = parse_options(argc, argv, latency_options, latency_usage, 0); | |
1758 | if (argc) | |
1759 | usage_with_options(latency_usage, latency_options); | |
f2858d8a | 1760 | } |
0e9b07e5 ACM |
1761 | setup_sorting(&sched, latency_options, latency_usage); |
1762 | return perf_sched__lat(&sched); | |
0ec04e16 | 1763 | } else if (!strcmp(argv[0], "map")) { |
0e9b07e5 ACM |
1764 | sched.tp_handler = &map_ops; |
1765 | setup_sorting(&sched, latency_options, latency_usage); | |
1766 | return perf_sched__map(&sched); | |
f2858d8a | 1767 | } else if (!strncmp(argv[0], "rep", 3)) { |
0e9b07e5 | 1768 | sched.tp_handler = &replay_ops; |
f2858d8a IM |
1769 | if (argc) { |
1770 | argc = parse_options(argc, argv, replay_options, replay_usage, 0); | |
1771 | if (argc) | |
1772 | usage_with_options(replay_usage, replay_options); | |
1773 | } | |
0e9b07e5 | 1774 | return perf_sched__replay(&sched); |
f2858d8a IM |
1775 | } else { |
1776 | usage_with_options(sched_usage, sched_options); | |
1777 | } | |
1778 | ||
ec156764 | 1779 | return 0; |
0a02ad93 | 1780 | } |