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" |
0a02ad93 IM |
13 | |
14 | #include "util/parse-options.h" | |
b1ffe8f3 | 15 | #include "util/trace-event.h" |
0a02ad93 | 16 | |
0a02ad93 IM |
17 | #include "util/debug.h" |
18 | ||
b1ffe8f3 | 19 | #include <sys/prctl.h> |
7b78f136 | 20 | #include <sys/resource.h> |
0a02ad93 | 21 | |
b1ffe8f3 IM |
22 | #include <semaphore.h> |
23 | #include <pthread.h> | |
24 | #include <math.h> | |
419ab0d6 | 25 | |
b1ffe8f3 IM |
26 | #define PR_SET_NAME 15 /* Set process name */ |
27 | #define MAX_CPUS 4096 | |
b1ffe8f3 IM |
28 | #define COMM_LEN 20 |
29 | #define SYM_LEN 129 | |
b1ffe8f3 | 30 | #define MAX_PID 65536 |
ec156764 | 31 | |
39aeb52f | 32 | struct sched_atom; |
ec156764 | 33 | |
b1ffe8f3 IM |
34 | struct task_desc { |
35 | unsigned long nr; | |
36 | unsigned long pid; | |
37 | char comm[COMM_LEN]; | |
ec156764 | 38 | |
b1ffe8f3 IM |
39 | unsigned long nr_events; |
40 | unsigned long curr_event; | |
39aeb52f | 41 | struct sched_atom **atoms; |
b1ffe8f3 IM |
42 | |
43 | pthread_t thread; | |
44 | sem_t sleep_sem; | |
ec156764 | 45 | |
b1ffe8f3 IM |
46 | sem_t ready_for_work; |
47 | sem_t work_done_sem; | |
48 | ||
49 | u64 cpu_usage; | |
50 | }; | |
51 | ||
52 | enum sched_event_type { | |
53 | SCHED_EVENT_RUN, | |
54 | SCHED_EVENT_SLEEP, | |
55 | SCHED_EVENT_WAKEUP, | |
55ffb7a6 | 56 | SCHED_EVENT_MIGRATION, |
b1ffe8f3 IM |
57 | }; |
58 | ||
39aeb52f | 59 | struct sched_atom { |
b1ffe8f3 | 60 | enum sched_event_type type; |
eed05fe7 | 61 | int specific_wait; |
b1ffe8f3 IM |
62 | u64 timestamp; |
63 | u64 duration; | |
64 | unsigned long nr; | |
b1ffe8f3 IM |
65 | sem_t *wait_sem; |
66 | struct task_desc *wakee; | |
67 | }; | |
68 | ||
e936e8e4 | 69 | #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP" |
b1ffe8f3 IM |
70 | |
71 | enum thread_state { | |
72 | THREAD_SLEEPING = 0, | |
73 | THREAD_WAIT_CPU, | |
74 | THREAD_SCHED_IN, | |
75 | THREAD_IGNORE | |
76 | }; | |
77 | ||
78 | struct work_atom { | |
79 | struct list_head list; | |
80 | enum thread_state state; | |
aa1ab9d2 | 81 | u64 sched_out_time; |
b1ffe8f3 IM |
82 | u64 wake_up_time; |
83 | u64 sched_in_time; | |
84 | u64 runtime; | |
85 | }; | |
86 | ||
39aeb52f | 87 | struct work_atoms { |
88 | struct list_head work_list; | |
b1ffe8f3 IM |
89 | struct thread *thread; |
90 | struct rb_node node; | |
91 | u64 max_lat; | |
3786310a | 92 | u64 max_lat_at; |
b1ffe8f3 IM |
93 | u64 total_lat; |
94 | u64 nb_atoms; | |
95 | u64 total_runtime; | |
96 | }; | |
97 | ||
39aeb52f | 98 | typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *); |
b1ffe8f3 | 99 | |
9ec3f4e4 | 100 | struct perf_sched; |
0e9b07e5 | 101 | |
9ec3f4e4 ACM |
102 | struct trace_sched_handler { |
103 | int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel, | |
104 | struct perf_sample *sample, struct machine *machine); | |
0e9b07e5 | 105 | |
9ec3f4e4 ACM |
106 | int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel, |
107 | struct perf_sample *sample, struct machine *machine); | |
0e9b07e5 | 108 | |
9ec3f4e4 ACM |
109 | int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel, |
110 | struct perf_sample *sample, struct machine *machine); | |
0e9b07e5 | 111 | |
cb627505 DA |
112 | /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */ |
113 | int (*fork_event)(struct perf_sched *sched, union perf_event *event, | |
114 | struct machine *machine); | |
0e9b07e5 ACM |
115 | |
116 | int (*migrate_task_event)(struct perf_sched *sched, | |
9ec3f4e4 ACM |
117 | struct perf_evsel *evsel, |
118 | struct perf_sample *sample, | |
119 | struct machine *machine); | |
0e9b07e5 ACM |
120 | }; |
121 | ||
122 | struct perf_sched { | |
123 | struct perf_tool tool; | |
0e9b07e5 ACM |
124 | const char *sort_order; |
125 | unsigned long nr_tasks; | |
126 | struct task_desc *pid_to_task[MAX_PID]; | |
127 | struct task_desc **tasks; | |
128 | const struct trace_sched_handler *tp_handler; | |
129 | pthread_mutex_t start_work_mutex; | |
130 | pthread_mutex_t work_done_wait_mutex; | |
131 | int profile_cpu; | |
132 | /* | |
133 | * Track the current task - that way we can know whether there's any | |
134 | * weird events, such as a task being switched away that is not current. | |
135 | */ | |
136 | int max_cpu; | |
137 | u32 curr_pid[MAX_CPUS]; | |
138 | struct thread *curr_thread[MAX_CPUS]; | |
139 | char next_shortname1; | |
140 | char next_shortname2; | |
141 | unsigned int replay_repeat; | |
142 | unsigned long nr_run_events; | |
143 | unsigned long nr_sleep_events; | |
144 | unsigned long nr_wakeup_events; | |
145 | unsigned long nr_sleep_corrections; | |
146 | unsigned long nr_run_events_optimized; | |
147 | unsigned long targetless_wakeups; | |
148 | unsigned long multitarget_wakeups; | |
149 | unsigned long nr_runs; | |
150 | unsigned long nr_timestamps; | |
151 | unsigned long nr_unordered_timestamps; | |
0e9b07e5 ACM |
152 | unsigned long nr_context_switch_bugs; |
153 | unsigned long nr_events; | |
154 | unsigned long nr_lost_chunks; | |
155 | unsigned long nr_lost_events; | |
156 | u64 run_measurement_overhead; | |
157 | u64 sleep_measurement_overhead; | |
158 | u64 start_time; | |
159 | u64 cpu_usage; | |
160 | u64 runavg_cpu_usage; | |
161 | u64 parent_cpu_usage; | |
162 | u64 runavg_parent_cpu_usage; | |
163 | u64 sum_runtime; | |
164 | u64 sum_fluct; | |
165 | u64 run_avg; | |
166 | u64 all_runtime; | |
167 | u64 all_count; | |
168 | u64 cpu_last_switched[MAX_CPUS]; | |
169 | struct rb_root atom_root, sorted_atom_root; | |
170 | struct list_head sort_list, cmp_pid; | |
171 | }; | |
b1ffe8f3 IM |
172 | |
173 | static u64 get_nsecs(void) | |
ec156764 IM |
174 | { |
175 | struct timespec ts; | |
176 | ||
177 | clock_gettime(CLOCK_MONOTONIC, &ts); | |
178 | ||
179 | return ts.tv_sec * 1000000000ULL + ts.tv_nsec; | |
180 | } | |
181 | ||
0e9b07e5 | 182 | static void burn_nsecs(struct perf_sched *sched, u64 nsecs) |
ec156764 | 183 | { |
b1ffe8f3 | 184 | u64 T0 = get_nsecs(), T1; |
ec156764 IM |
185 | |
186 | do { | |
187 | T1 = get_nsecs(); | |
0e9b07e5 | 188 | } while (T1 + sched->run_measurement_overhead < T0 + nsecs); |
ec156764 IM |
189 | } |
190 | ||
b1ffe8f3 | 191 | static void sleep_nsecs(u64 nsecs) |
ec156764 IM |
192 | { |
193 | struct timespec ts; | |
194 | ||
195 | ts.tv_nsec = nsecs % 999999999; | |
196 | ts.tv_sec = nsecs / 999999999; | |
197 | ||
198 | nanosleep(&ts, NULL); | |
199 | } | |
200 | ||
0e9b07e5 | 201 | static void calibrate_run_measurement_overhead(struct perf_sched *sched) |
ec156764 | 202 | { |
b1ffe8f3 | 203 | u64 T0, T1, delta, min_delta = 1000000000ULL; |
ec156764 IM |
204 | int i; |
205 | ||
206 | for (i = 0; i < 10; i++) { | |
207 | T0 = get_nsecs(); | |
0e9b07e5 | 208 | burn_nsecs(sched, 0); |
ec156764 IM |
209 | T1 = get_nsecs(); |
210 | delta = T1-T0; | |
211 | min_delta = min(min_delta, delta); | |
212 | } | |
0e9b07e5 | 213 | sched->run_measurement_overhead = min_delta; |
ec156764 | 214 | |
9486aa38 | 215 | printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta); |
ec156764 IM |
216 | } |
217 | ||
0e9b07e5 | 218 | static void calibrate_sleep_measurement_overhead(struct perf_sched *sched) |
ec156764 | 219 | { |
b1ffe8f3 | 220 | u64 T0, T1, delta, min_delta = 1000000000ULL; |
ec156764 IM |
221 | int i; |
222 | ||
223 | for (i = 0; i < 10; i++) { | |
224 | T0 = get_nsecs(); | |
225 | sleep_nsecs(10000); | |
226 | T1 = get_nsecs(); | |
227 | delta = T1-T0; | |
228 | min_delta = min(min_delta, delta); | |
229 | } | |
230 | min_delta -= 10000; | |
0e9b07e5 | 231 | sched->sleep_measurement_overhead = min_delta; |
ec156764 | 232 | |
9486aa38 | 233 | printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta); |
ec156764 IM |
234 | } |
235 | ||
39aeb52f | 236 | static struct sched_atom * |
b1ffe8f3 | 237 | get_new_event(struct task_desc *task, u64 timestamp) |
ec156764 | 238 | { |
36479484 | 239 | struct sched_atom *event = zalloc(sizeof(*event)); |
ec156764 IM |
240 | unsigned long idx = task->nr_events; |
241 | size_t size; | |
242 | ||
243 | event->timestamp = timestamp; | |
244 | event->nr = idx; | |
245 | ||
246 | task->nr_events++; | |
39aeb52f | 247 | size = sizeof(struct sched_atom *) * task->nr_events; |
248 | task->atoms = realloc(task->atoms, size); | |
249 | BUG_ON(!task->atoms); | |
ec156764 | 250 | |
39aeb52f | 251 | task->atoms[idx] = event; |
ec156764 IM |
252 | |
253 | return event; | |
254 | } | |
255 | ||
39aeb52f | 256 | static struct sched_atom *last_event(struct task_desc *task) |
ec156764 IM |
257 | { |
258 | if (!task->nr_events) | |
259 | return NULL; | |
260 | ||
39aeb52f | 261 | return task->atoms[task->nr_events - 1]; |
ec156764 IM |
262 | } |
263 | ||
0e9b07e5 ACM |
264 | static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task, |
265 | u64 timestamp, u64 duration) | |
ec156764 | 266 | { |
39aeb52f | 267 | struct sched_atom *event, *curr_event = last_event(task); |
ec156764 IM |
268 | |
269 | /* | |
fbf94829 IM |
270 | * optimize an existing RUN event by merging this one |
271 | * to it: | |
272 | */ | |
ec156764 | 273 | if (curr_event && curr_event->type == SCHED_EVENT_RUN) { |
0e9b07e5 | 274 | sched->nr_run_events_optimized++; |
ec156764 IM |
275 | curr_event->duration += duration; |
276 | return; | |
277 | } | |
278 | ||
279 | event = get_new_event(task, timestamp); | |
280 | ||
281 | event->type = SCHED_EVENT_RUN; | |
282 | event->duration = duration; | |
283 | ||
0e9b07e5 | 284 | sched->nr_run_events++; |
ec156764 IM |
285 | } |
286 | ||
0e9b07e5 ACM |
287 | static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task, |
288 | u64 timestamp, struct task_desc *wakee) | |
ec156764 | 289 | { |
39aeb52f | 290 | struct sched_atom *event, *wakee_event; |
ec156764 IM |
291 | |
292 | event = get_new_event(task, timestamp); | |
293 | event->type = SCHED_EVENT_WAKEUP; | |
294 | event->wakee = wakee; | |
295 | ||
296 | wakee_event = last_event(wakee); | |
297 | if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) { | |
0e9b07e5 | 298 | sched->targetless_wakeups++; |
ec156764 IM |
299 | return; |
300 | } | |
301 | if (wakee_event->wait_sem) { | |
0e9b07e5 | 302 | sched->multitarget_wakeups++; |
ec156764 IM |
303 | return; |
304 | } | |
305 | ||
36479484 | 306 | wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem)); |
ec156764 IM |
307 | sem_init(wakee_event->wait_sem, 0, 0); |
308 | wakee_event->specific_wait = 1; | |
309 | event->wait_sem = wakee_event->wait_sem; | |
310 | ||
0e9b07e5 | 311 | sched->nr_wakeup_events++; |
ec156764 IM |
312 | } |
313 | ||
0e9b07e5 ACM |
314 | static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task, |
315 | u64 timestamp, u64 task_state __maybe_unused) | |
ec156764 | 316 | { |
39aeb52f | 317 | struct sched_atom *event = get_new_event(task, timestamp); |
ec156764 IM |
318 | |
319 | event->type = SCHED_EVENT_SLEEP; | |
320 | ||
0e9b07e5 | 321 | sched->nr_sleep_events++; |
ec156764 IM |
322 | } |
323 | ||
0e9b07e5 ACM |
324 | static struct task_desc *register_pid(struct perf_sched *sched, |
325 | unsigned long pid, const char *comm) | |
ec156764 IM |
326 | { |
327 | struct task_desc *task; | |
328 | ||
329 | BUG_ON(pid >= MAX_PID); | |
330 | ||
0e9b07e5 | 331 | task = sched->pid_to_task[pid]; |
ec156764 IM |
332 | |
333 | if (task) | |
334 | return task; | |
335 | ||
36479484 | 336 | task = zalloc(sizeof(*task)); |
ec156764 | 337 | task->pid = pid; |
0e9b07e5 | 338 | task->nr = sched->nr_tasks; |
ec156764 IM |
339 | strcpy(task->comm, comm); |
340 | /* | |
341 | * every task starts in sleeping state - this gets ignored | |
342 | * if there's no wakeup pointing to this sleep state: | |
343 | */ | |
0e9b07e5 | 344 | add_sched_event_sleep(sched, task, 0, 0); |
ec156764 | 345 | |
0e9b07e5 ACM |
346 | sched->pid_to_task[pid] = task; |
347 | sched->nr_tasks++; | |
348 | sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_task *)); | |
349 | BUG_ON(!sched->tasks); | |
350 | sched->tasks[task->nr] = task; | |
ec156764 | 351 | |
ad236fd2 | 352 | if (verbose) |
0e9b07e5 | 353 | printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm); |
ec156764 IM |
354 | |
355 | return task; | |
356 | } | |
357 | ||
358 | ||
0e9b07e5 | 359 | static void print_task_traces(struct perf_sched *sched) |
ec156764 IM |
360 | { |
361 | struct task_desc *task; | |
362 | unsigned long i; | |
363 | ||
0e9b07e5 ACM |
364 | for (i = 0; i < sched->nr_tasks; i++) { |
365 | task = sched->tasks[i]; | |
ad236fd2 | 366 | printf("task %6ld (%20s:%10ld), nr_events: %ld\n", |
ec156764 IM |
367 | task->nr, task->comm, task->pid, task->nr_events); |
368 | } | |
369 | } | |
370 | ||
0e9b07e5 | 371 | static void add_cross_task_wakeups(struct perf_sched *sched) |
ec156764 IM |
372 | { |
373 | struct task_desc *task1, *task2; | |
374 | unsigned long i, j; | |
375 | ||
0e9b07e5 ACM |
376 | for (i = 0; i < sched->nr_tasks; i++) { |
377 | task1 = sched->tasks[i]; | |
ec156764 | 378 | j = i + 1; |
0e9b07e5 | 379 | if (j == sched->nr_tasks) |
ec156764 | 380 | j = 0; |
0e9b07e5 ACM |
381 | task2 = sched->tasks[j]; |
382 | add_sched_event_wakeup(sched, task1, 0, task2); | |
ec156764 IM |
383 | } |
384 | } | |
385 | ||
0e9b07e5 ACM |
386 | static void perf_sched__process_event(struct perf_sched *sched, |
387 | struct sched_atom *atom) | |
ec156764 IM |
388 | { |
389 | int ret = 0; | |
ec156764 | 390 | |
39aeb52f | 391 | switch (atom->type) { |
ec156764 | 392 | case SCHED_EVENT_RUN: |
0e9b07e5 | 393 | burn_nsecs(sched, atom->duration); |
ec156764 IM |
394 | break; |
395 | case SCHED_EVENT_SLEEP: | |
39aeb52f | 396 | if (atom->wait_sem) |
397 | ret = sem_wait(atom->wait_sem); | |
ec156764 IM |
398 | BUG_ON(ret); |
399 | break; | |
400 | case SCHED_EVENT_WAKEUP: | |
39aeb52f | 401 | if (atom->wait_sem) |
402 | ret = sem_post(atom->wait_sem); | |
ec156764 IM |
403 | BUG_ON(ret); |
404 | break; | |
55ffb7a6 MG |
405 | case SCHED_EVENT_MIGRATION: |
406 | break; | |
ec156764 IM |
407 | default: |
408 | BUG_ON(1); | |
409 | } | |
410 | } | |
411 | ||
b1ffe8f3 | 412 | static u64 get_cpu_usage_nsec_parent(void) |
ec156764 IM |
413 | { |
414 | struct rusage ru; | |
b1ffe8f3 | 415 | u64 sum; |
ec156764 IM |
416 | int err; |
417 | ||
418 | err = getrusage(RUSAGE_SELF, &ru); | |
419 | BUG_ON(err); | |
420 | ||
421 | sum = ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3; | |
422 | sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3; | |
423 | ||
424 | return sum; | |
425 | } | |
426 | ||
c0c9e721 | 427 | static int self_open_counters(void) |
ec156764 | 428 | { |
c0c9e721 XG |
429 | struct perf_event_attr attr; |
430 | int fd; | |
ec156764 | 431 | |
c0c9e721 | 432 | memset(&attr, 0, sizeof(attr)); |
ec156764 | 433 | |
c0c9e721 XG |
434 | attr.type = PERF_TYPE_SOFTWARE; |
435 | attr.config = PERF_COUNT_SW_TASK_CLOCK; | |
ec156764 | 436 | |
c0c9e721 XG |
437 | fd = sys_perf_event_open(&attr, 0, -1, -1, 0); |
438 | ||
439 | if (fd < 0) | |
60b7d14a NK |
440 | pr_err("Error: sys_perf_event_open() syscall returned " |
441 | "with %d (%s)\n", fd, strerror(errno)); | |
c0c9e721 XG |
442 | return fd; |
443 | } | |
444 | ||
445 | static u64 get_cpu_usage_nsec_self(int fd) | |
446 | { | |
447 | u64 runtime; | |
448 | int ret; | |
449 | ||
450 | ret = read(fd, &runtime, sizeof(runtime)); | |
451 | BUG_ON(ret != sizeof(runtime)); | |
452 | ||
453 | return runtime; | |
ec156764 IM |
454 | } |
455 | ||
0e9b07e5 ACM |
456 | struct sched_thread_parms { |
457 | struct task_desc *task; | |
458 | struct perf_sched *sched; | |
459 | }; | |
460 | ||
ec156764 IM |
461 | static void *thread_func(void *ctx) |
462 | { | |
0e9b07e5 ACM |
463 | struct sched_thread_parms *parms = ctx; |
464 | struct task_desc *this_task = parms->task; | |
465 | struct perf_sched *sched = parms->sched; | |
b1ffe8f3 | 466 | u64 cpu_usage_0, cpu_usage_1; |
ec156764 IM |
467 | unsigned long i, ret; |
468 | char comm2[22]; | |
c0c9e721 | 469 | int fd; |
ec156764 | 470 | |
74cf249d | 471 | zfree(&parms); |
0e9b07e5 | 472 | |
ec156764 IM |
473 | sprintf(comm2, ":%s", this_task->comm); |
474 | prctl(PR_SET_NAME, comm2); | |
c0c9e721 | 475 | fd = self_open_counters(); |
a116e05d ACM |
476 | if (fd < 0) |
477 | return NULL; | |
ec156764 IM |
478 | again: |
479 | ret = sem_post(&this_task->ready_for_work); | |
480 | BUG_ON(ret); | |
0e9b07e5 | 481 | ret = pthread_mutex_lock(&sched->start_work_mutex); |
ec156764 | 482 | BUG_ON(ret); |
0e9b07e5 | 483 | ret = pthread_mutex_unlock(&sched->start_work_mutex); |
ec156764 | 484 | BUG_ON(ret); |
ec156764 | 485 | |
c0c9e721 | 486 | cpu_usage_0 = get_cpu_usage_nsec_self(fd); |
ec156764 IM |
487 | |
488 | for (i = 0; i < this_task->nr_events; i++) { | |
489 | this_task->curr_event = i; | |
0e9b07e5 | 490 | perf_sched__process_event(sched, this_task->atoms[i]); |
ec156764 IM |
491 | } |
492 | ||
c0c9e721 | 493 | cpu_usage_1 = get_cpu_usage_nsec_self(fd); |
ec156764 | 494 | this_task->cpu_usage = cpu_usage_1 - cpu_usage_0; |
ec156764 IM |
495 | ret = sem_post(&this_task->work_done_sem); |
496 | BUG_ON(ret); | |
ec156764 | 497 | |
0e9b07e5 | 498 | ret = pthread_mutex_lock(&sched->work_done_wait_mutex); |
ec156764 | 499 | BUG_ON(ret); |
0e9b07e5 | 500 | ret = pthread_mutex_unlock(&sched->work_done_wait_mutex); |
ec156764 | 501 | BUG_ON(ret); |
ec156764 IM |
502 | |
503 | goto again; | |
504 | } | |
505 | ||
0e9b07e5 | 506 | static void create_tasks(struct perf_sched *sched) |
ec156764 IM |
507 | { |
508 | struct task_desc *task; | |
509 | pthread_attr_t attr; | |
510 | unsigned long i; | |
511 | int err; | |
512 | ||
513 | err = pthread_attr_init(&attr); | |
514 | BUG_ON(err); | |
12f7e036 JP |
515 | err = pthread_attr_setstacksize(&attr, |
516 | (size_t) max(16 * 1024, PTHREAD_STACK_MIN)); | |
ec156764 | 517 | BUG_ON(err); |
0e9b07e5 | 518 | err = pthread_mutex_lock(&sched->start_work_mutex); |
ec156764 | 519 | BUG_ON(err); |
0e9b07e5 | 520 | err = pthread_mutex_lock(&sched->work_done_wait_mutex); |
ec156764 | 521 | BUG_ON(err); |
0e9b07e5 ACM |
522 | for (i = 0; i < sched->nr_tasks; i++) { |
523 | struct sched_thread_parms *parms = malloc(sizeof(*parms)); | |
524 | BUG_ON(parms == NULL); | |
525 | parms->task = task = sched->tasks[i]; | |
526 | parms->sched = sched; | |
ec156764 IM |
527 | sem_init(&task->sleep_sem, 0, 0); |
528 | sem_init(&task->ready_for_work, 0, 0); | |
529 | sem_init(&task->work_done_sem, 0, 0); | |
530 | task->curr_event = 0; | |
0e9b07e5 | 531 | err = pthread_create(&task->thread, &attr, thread_func, parms); |
ec156764 IM |
532 | BUG_ON(err); |
533 | } | |
534 | } | |
535 | ||
0e9b07e5 | 536 | static void wait_for_tasks(struct perf_sched *sched) |
ec156764 | 537 | { |
b1ffe8f3 | 538 | u64 cpu_usage_0, cpu_usage_1; |
ec156764 IM |
539 | struct task_desc *task; |
540 | unsigned long i, ret; | |
541 | ||
0e9b07e5 ACM |
542 | sched->start_time = get_nsecs(); |
543 | sched->cpu_usage = 0; | |
544 | pthread_mutex_unlock(&sched->work_done_wait_mutex); | |
ec156764 | 545 | |
0e9b07e5 ACM |
546 | for (i = 0; i < sched->nr_tasks; i++) { |
547 | task = sched->tasks[i]; | |
ec156764 IM |
548 | ret = sem_wait(&task->ready_for_work); |
549 | BUG_ON(ret); | |
550 | sem_init(&task->ready_for_work, 0, 0); | |
551 | } | |
0e9b07e5 | 552 | ret = pthread_mutex_lock(&sched->work_done_wait_mutex); |
ec156764 IM |
553 | BUG_ON(ret); |
554 | ||
555 | cpu_usage_0 = get_cpu_usage_nsec_parent(); | |
556 | ||
0e9b07e5 | 557 | pthread_mutex_unlock(&sched->start_work_mutex); |
ec156764 | 558 | |
0e9b07e5 ACM |
559 | for (i = 0; i < sched->nr_tasks; i++) { |
560 | task = sched->tasks[i]; | |
ec156764 IM |
561 | ret = sem_wait(&task->work_done_sem); |
562 | BUG_ON(ret); | |
563 | sem_init(&task->work_done_sem, 0, 0); | |
0e9b07e5 | 564 | sched->cpu_usage += task->cpu_usage; |
ec156764 IM |
565 | task->cpu_usage = 0; |
566 | } | |
567 | ||
568 | cpu_usage_1 = get_cpu_usage_nsec_parent(); | |
0e9b07e5 ACM |
569 | if (!sched->runavg_cpu_usage) |
570 | sched->runavg_cpu_usage = sched->cpu_usage; | |
571 | sched->runavg_cpu_usage = (sched->runavg_cpu_usage * 9 + sched->cpu_usage) / 10; | |
ec156764 | 572 | |
0e9b07e5 ACM |
573 | sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0; |
574 | if (!sched->runavg_parent_cpu_usage) | |
575 | sched->runavg_parent_cpu_usage = sched->parent_cpu_usage; | |
576 | sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * 9 + | |
577 | sched->parent_cpu_usage)/10; | |
ec156764 | 578 | |
0e9b07e5 | 579 | ret = pthread_mutex_lock(&sched->start_work_mutex); |
ec156764 IM |
580 | BUG_ON(ret); |
581 | ||
0e9b07e5 ACM |
582 | for (i = 0; i < sched->nr_tasks; i++) { |
583 | task = sched->tasks[i]; | |
ec156764 IM |
584 | sem_init(&task->sleep_sem, 0, 0); |
585 | task->curr_event = 0; | |
586 | } | |
587 | } | |
588 | ||
0e9b07e5 | 589 | static void run_one_test(struct perf_sched *sched) |
ec156764 | 590 | { |
fb7d0b3c | 591 | u64 T0, T1, delta, avg_delta, fluct; |
ec156764 IM |
592 | |
593 | T0 = get_nsecs(); | |
0e9b07e5 | 594 | wait_for_tasks(sched); |
ec156764 IM |
595 | T1 = get_nsecs(); |
596 | ||
597 | delta = T1 - T0; | |
0e9b07e5 ACM |
598 | sched->sum_runtime += delta; |
599 | sched->nr_runs++; | |
ec156764 | 600 | |
0e9b07e5 | 601 | avg_delta = sched->sum_runtime / sched->nr_runs; |
ec156764 IM |
602 | if (delta < avg_delta) |
603 | fluct = avg_delta - delta; | |
604 | else | |
605 | fluct = delta - avg_delta; | |
0e9b07e5 ACM |
606 | sched->sum_fluct += fluct; |
607 | if (!sched->run_avg) | |
608 | sched->run_avg = delta; | |
609 | sched->run_avg = (sched->run_avg * 9 + delta) / 10; | |
ec156764 | 610 | |
0e9b07e5 | 611 | printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / 1000000.0); |
ec156764 | 612 | |
0e9b07e5 | 613 | printf("ravg: %0.2f, ", (double)sched->run_avg / 1e6); |
ec156764 | 614 | |
ad236fd2 | 615 | printf("cpu: %0.2f / %0.2f", |
0e9b07e5 | 616 | (double)sched->cpu_usage / 1e6, (double)sched->runavg_cpu_usage / 1e6); |
ec156764 IM |
617 | |
618 | #if 0 | |
619 | /* | |
fbf94829 | 620 | * rusage statistics done by the parent, these are less |
0e9b07e5 | 621 | * accurate than the sched->sum_exec_runtime based statistics: |
fbf94829 | 622 | */ |
ad236fd2 | 623 | printf(" [%0.2f / %0.2f]", |
0e9b07e5 ACM |
624 | (double)sched->parent_cpu_usage/1e6, |
625 | (double)sched->runavg_parent_cpu_usage/1e6); | |
ec156764 IM |
626 | #endif |
627 | ||
ad236fd2 | 628 | printf("\n"); |
ec156764 | 629 | |
0e9b07e5 ACM |
630 | if (sched->nr_sleep_corrections) |
631 | printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections); | |
632 | sched->nr_sleep_corrections = 0; | |
ec156764 IM |
633 | } |
634 | ||
0e9b07e5 | 635 | static void test_calibrations(struct perf_sched *sched) |
ec156764 | 636 | { |
b1ffe8f3 | 637 | u64 T0, T1; |
ec156764 IM |
638 | |
639 | T0 = get_nsecs(); | |
0e9b07e5 | 640 | burn_nsecs(sched, 1e6); |
ec156764 IM |
641 | T1 = get_nsecs(); |
642 | ||
9486aa38 | 643 | printf("the run test took %" PRIu64 " nsecs\n", T1 - T0); |
ec156764 IM |
644 | |
645 | T0 = get_nsecs(); | |
646 | sleep_nsecs(1e6); | |
647 | T1 = get_nsecs(); | |
648 | ||
9486aa38 | 649 | printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0); |
ec156764 IM |
650 | } |
651 | ||
a116e05d | 652 | static int |
0e9b07e5 | 653 | replay_wakeup_event(struct perf_sched *sched, |
9ec3f4e4 ACM |
654 | struct perf_evsel *evsel, struct perf_sample *sample, |
655 | struct machine *machine __maybe_unused) | |
419ab0d6 | 656 | { |
9ec3f4e4 ACM |
657 | const char *comm = perf_evsel__strval(evsel, sample, "comm"); |
658 | const u32 pid = perf_evsel__intval(evsel, sample, "pid"); | |
419ab0d6 | 659 | struct task_desc *waker, *wakee; |
fbf94829 | 660 | |
ad236fd2 | 661 | if (verbose) { |
2b7fcbc5 | 662 | printf("sched_wakeup event %p\n", evsel); |
fbf94829 | 663 | |
9ec3f4e4 | 664 | printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid); |
ad236fd2 | 665 | } |
fbf94829 | 666 | |
2b7fcbc5 | 667 | waker = register_pid(sched, sample->tid, "<unknown>"); |
9ec3f4e4 | 668 | wakee = register_pid(sched, pid, comm); |
fbf94829 | 669 | |
0e9b07e5 | 670 | add_sched_event_wakeup(sched, waker, sample->time, wakee); |
a116e05d | 671 | return 0; |
ec156764 IM |
672 | } |
673 | ||
9ec3f4e4 ACM |
674 | static int replay_switch_event(struct perf_sched *sched, |
675 | struct perf_evsel *evsel, | |
676 | struct perf_sample *sample, | |
677 | struct machine *machine __maybe_unused) | |
ec156764 | 678 | { |
9ec3f4e4 ACM |
679 | const char *prev_comm = perf_evsel__strval(evsel, sample, "prev_comm"), |
680 | *next_comm = perf_evsel__strval(evsel, sample, "next_comm"); | |
681 | const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"), | |
682 | next_pid = perf_evsel__intval(evsel, sample, "next_pid"); | |
683 | const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state"); | |
1d037ca1 | 684 | struct task_desc *prev, __maybe_unused *next; |
7f7f8d0b ACM |
685 | u64 timestamp0, timestamp = sample->time; |
686 | int cpu = sample->cpu; | |
fbf94829 IM |
687 | s64 delta; |
688 | ||
ad236fd2 | 689 | if (verbose) |
2b7fcbc5 | 690 | printf("sched_switch event %p\n", evsel); |
ad236fd2 | 691 | |
fbf94829 | 692 | if (cpu >= MAX_CPUS || cpu < 0) |
a116e05d | 693 | return 0; |
fbf94829 | 694 | |
0e9b07e5 | 695 | timestamp0 = sched->cpu_last_switched[cpu]; |
fbf94829 IM |
696 | if (timestamp0) |
697 | delta = timestamp - timestamp0; | |
698 | else | |
699 | delta = 0; | |
700 | ||
a116e05d | 701 | if (delta < 0) { |
60b7d14a | 702 | pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta); |
a116e05d ACM |
703 | return -1; |
704 | } | |
fbf94829 | 705 | |
9ec3f4e4 ACM |
706 | pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n", |
707 | prev_comm, prev_pid, next_comm, next_pid, delta); | |
fbf94829 | 708 | |
9ec3f4e4 ACM |
709 | prev = register_pid(sched, prev_pid, prev_comm); |
710 | next = register_pid(sched, next_pid, next_comm); | |
fbf94829 | 711 | |
0e9b07e5 | 712 | sched->cpu_last_switched[cpu] = timestamp; |
fbf94829 | 713 | |
0e9b07e5 | 714 | add_sched_event_run(sched, prev, timestamp, delta); |
9ec3f4e4 | 715 | add_sched_event_sleep(sched, prev, timestamp, prev_state); |
a116e05d ACM |
716 | |
717 | return 0; | |
fbf94829 IM |
718 | } |
719 | ||
cb627505 DA |
720 | static int replay_fork_event(struct perf_sched *sched, |
721 | union perf_event *event, | |
722 | struct machine *machine) | |
419ab0d6 | 723 | { |
cb627505 DA |
724 | struct thread *child, *parent; |
725 | ||
314add6b AH |
726 | child = machine__findnew_thread(machine, event->fork.pid, |
727 | event->fork.tid); | |
728 | parent = machine__findnew_thread(machine, event->fork.ppid, | |
729 | event->fork.ptid); | |
cb627505 DA |
730 | |
731 | if (child == NULL || parent == NULL) { | |
732 | pr_debug("thread does not exist on fork event: child %p, parent %p\n", | |
733 | child, parent); | |
734 | return 0; | |
735 | } | |
9ec3f4e4 | 736 | |
419ab0d6 | 737 | if (verbose) { |
cb627505 | 738 | printf("fork event\n"); |
b9c5143a FW |
739 | printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid); |
740 | printf("... child: %s/%d\n", thread__comm_str(child), child->tid); | |
419ab0d6 | 741 | } |
9ec3f4e4 | 742 | |
b9c5143a FW |
743 | register_pid(sched, parent->tid, thread__comm_str(parent)); |
744 | register_pid(sched, child->tid, thread__comm_str(child)); | |
a116e05d | 745 | return 0; |
419ab0d6 | 746 | } |
fbf94829 | 747 | |
b1ffe8f3 IM |
748 | struct sort_dimension { |
749 | const char *name; | |
b5fae128 | 750 | sort_fn_t cmp; |
b1ffe8f3 IM |
751 | struct list_head list; |
752 | }; | |
753 | ||
daa1d7a5 | 754 | static int |
39aeb52f | 755 | thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r) |
daa1d7a5 FW |
756 | { |
757 | struct sort_dimension *sort; | |
758 | int ret = 0; | |
759 | ||
b5fae128 IM |
760 | BUG_ON(list_empty(list)); |
761 | ||
daa1d7a5 FW |
762 | list_for_each_entry(sort, list, list) { |
763 | ret = sort->cmp(l, r); | |
764 | if (ret) | |
765 | return ret; | |
766 | } | |
767 | ||
768 | return ret; | |
769 | } | |
770 | ||
39aeb52f | 771 | static struct work_atoms * |
b5fae128 IM |
772 | thread_atoms_search(struct rb_root *root, struct thread *thread, |
773 | struct list_head *sort_list) | |
774 | { | |
775 | struct rb_node *node = root->rb_node; | |
39aeb52f | 776 | struct work_atoms key = { .thread = thread }; |
b5fae128 IM |
777 | |
778 | while (node) { | |
39aeb52f | 779 | struct work_atoms *atoms; |
b5fae128 IM |
780 | int cmp; |
781 | ||
39aeb52f | 782 | atoms = container_of(node, struct work_atoms, node); |
b5fae128 IM |
783 | |
784 | cmp = thread_lat_cmp(sort_list, &key, atoms); | |
785 | if (cmp > 0) | |
786 | node = node->rb_left; | |
787 | else if (cmp < 0) | |
788 | node = node->rb_right; | |
789 | else { | |
790 | BUG_ON(thread != atoms->thread); | |
791 | return atoms; | |
792 | } | |
793 | } | |
794 | return NULL; | |
795 | } | |
796 | ||
cdce9d73 | 797 | static void |
39aeb52f | 798 | __thread_latency_insert(struct rb_root *root, struct work_atoms *data, |
daa1d7a5 | 799 | struct list_head *sort_list) |
cdce9d73 FW |
800 | { |
801 | struct rb_node **new = &(root->rb_node), *parent = NULL; | |
802 | ||
803 | while (*new) { | |
39aeb52f | 804 | struct work_atoms *this; |
daa1d7a5 | 805 | int cmp; |
cdce9d73 | 806 | |
39aeb52f | 807 | this = container_of(*new, struct work_atoms, node); |
cdce9d73 | 808 | parent = *new; |
daa1d7a5 FW |
809 | |
810 | cmp = thread_lat_cmp(sort_list, data, this); | |
811 | ||
812 | if (cmp > 0) | |
cdce9d73 | 813 | new = &((*new)->rb_left); |
cdce9d73 | 814 | else |
daa1d7a5 | 815 | new = &((*new)->rb_right); |
cdce9d73 FW |
816 | } |
817 | ||
818 | rb_link_node(&data->node, parent, new); | |
819 | rb_insert_color(&data->node, root); | |
820 | } | |
821 | ||
0e9b07e5 | 822 | static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread) |
cdce9d73 | 823 | { |
36479484 | 824 | struct work_atoms *atoms = zalloc(sizeof(*atoms)); |
a116e05d ACM |
825 | if (!atoms) { |
826 | pr_err("No memory at %s\n", __func__); | |
827 | return -1; | |
828 | } | |
cdce9d73 | 829 | |
17562205 | 830 | atoms->thread = thread; |
39aeb52f | 831 | INIT_LIST_HEAD(&atoms->work_list); |
0e9b07e5 | 832 | __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid); |
a116e05d | 833 | return 0; |
cdce9d73 FW |
834 | } |
835 | ||
9ec3f4e4 | 836 | static char sched_out_state(u64 prev_state) |
cdce9d73 FW |
837 | { |
838 | const char *str = TASK_STATE_TO_CHAR_STR; | |
839 | ||
9ec3f4e4 | 840 | return str[prev_state]; |
cdce9d73 FW |
841 | } |
842 | ||
a116e05d | 843 | static int |
39aeb52f | 844 | add_sched_out_event(struct work_atoms *atoms, |
845 | char run_state, | |
846 | u64 timestamp) | |
cdce9d73 | 847 | { |
36479484 | 848 | struct work_atom *atom = zalloc(sizeof(*atom)); |
a116e05d ACM |
849 | if (!atom) { |
850 | pr_err("Non memory at %s", __func__); | |
851 | return -1; | |
852 | } | |
cdce9d73 | 853 | |
aa1ab9d2 FW |
854 | atom->sched_out_time = timestamp; |
855 | ||
39aeb52f | 856 | if (run_state == 'R') { |
b1ffe8f3 | 857 | atom->state = THREAD_WAIT_CPU; |
aa1ab9d2 | 858 | atom->wake_up_time = atom->sched_out_time; |
c6ced611 FW |
859 | } |
860 | ||
39aeb52f | 861 | list_add_tail(&atom->list, &atoms->work_list); |
a116e05d | 862 | return 0; |
cdce9d73 FW |
863 | } |
864 | ||
865 | static void | |
1d037ca1 IT |
866 | add_runtime_event(struct work_atoms *atoms, u64 delta, |
867 | u64 timestamp __maybe_unused) | |
39aeb52f | 868 | { |
869 | struct work_atom *atom; | |
870 | ||
871 | BUG_ON(list_empty(&atoms->work_list)); | |
872 | ||
873 | atom = list_entry(atoms->work_list.prev, struct work_atom, list); | |
874 | ||
875 | atom->runtime += delta; | |
876 | atoms->total_runtime += delta; | |
877 | } | |
878 | ||
879 | static void | |
880 | add_sched_in_event(struct work_atoms *atoms, u64 timestamp) | |
cdce9d73 | 881 | { |
b1ffe8f3 | 882 | struct work_atom *atom; |
66685678 | 883 | u64 delta; |
cdce9d73 | 884 | |
39aeb52f | 885 | if (list_empty(&atoms->work_list)) |
cdce9d73 FW |
886 | return; |
887 | ||
39aeb52f | 888 | atom = list_entry(atoms->work_list.prev, struct work_atom, list); |
cdce9d73 | 889 | |
b1ffe8f3 | 890 | if (atom->state != THREAD_WAIT_CPU) |
cdce9d73 FW |
891 | return; |
892 | ||
b1ffe8f3 IM |
893 | if (timestamp < atom->wake_up_time) { |
894 | atom->state = THREAD_IGNORE; | |
cdce9d73 FW |
895 | return; |
896 | } | |
897 | ||
b1ffe8f3 IM |
898 | atom->state = THREAD_SCHED_IN; |
899 | atom->sched_in_time = timestamp; | |
66685678 | 900 | |
b1ffe8f3 | 901 | delta = atom->sched_in_time - atom->wake_up_time; |
66685678 | 902 | atoms->total_lat += delta; |
3786310a | 903 | if (delta > atoms->max_lat) { |
66685678 | 904 | atoms->max_lat = delta; |
3786310a FW |
905 | atoms->max_lat_at = timestamp; |
906 | } | |
66685678 | 907 | atoms->nb_atoms++; |
cdce9d73 FW |
908 | } |
909 | ||
9ec3f4e4 ACM |
910 | static int latency_switch_event(struct perf_sched *sched, |
911 | struct perf_evsel *evsel, | |
912 | struct perf_sample *sample, | |
913 | struct machine *machine) | |
cdce9d73 | 914 | { |
9ec3f4e4 ACM |
915 | const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"), |
916 | next_pid = perf_evsel__intval(evsel, sample, "next_pid"); | |
917 | const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state"); | |
39aeb52f | 918 | struct work_atoms *out_events, *in_events; |
cdce9d73 | 919 | struct thread *sched_out, *sched_in; |
7f7f8d0b ACM |
920 | u64 timestamp0, timestamp = sample->time; |
921 | int cpu = sample->cpu; | |
ea92ed5a IM |
922 | s64 delta; |
923 | ||
39aeb52f | 924 | BUG_ON(cpu >= MAX_CPUS || cpu < 0); |
ea92ed5a | 925 | |
0e9b07e5 ACM |
926 | timestamp0 = sched->cpu_last_switched[cpu]; |
927 | sched->cpu_last_switched[cpu] = timestamp; | |
ea92ed5a IM |
928 | if (timestamp0) |
929 | delta = timestamp - timestamp0; | |
930 | else | |
931 | delta = 0; | |
932 | ||
a116e05d ACM |
933 | if (delta < 0) { |
934 | pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta); | |
935 | return -1; | |
936 | } | |
cdce9d73 | 937 | |
314add6b AH |
938 | sched_out = machine__findnew_thread(machine, 0, prev_pid); |
939 | sched_in = machine__findnew_thread(machine, 0, next_pid); | |
cdce9d73 | 940 | |
0e9b07e5 | 941 | out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid); |
39aeb52f | 942 | if (!out_events) { |
0e9b07e5 | 943 | if (thread_atoms_insert(sched, sched_out)) |
a116e05d | 944 | return -1; |
0e9b07e5 | 945 | out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid); |
a116e05d ACM |
946 | if (!out_events) { |
947 | pr_err("out-event: Internal tree error"); | |
948 | return -1; | |
949 | } | |
39aeb52f | 950 | } |
9ec3f4e4 | 951 | if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp)) |
a116e05d | 952 | return -1; |
39aeb52f | 953 | |
0e9b07e5 | 954 | in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid); |
39aeb52f | 955 | if (!in_events) { |
0e9b07e5 | 956 | if (thread_atoms_insert(sched, sched_in)) |
a116e05d | 957 | return -1; |
0e9b07e5 | 958 | in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid); |
a116e05d ACM |
959 | if (!in_events) { |
960 | pr_err("in-event: Internal tree error"); | |
961 | return -1; | |
962 | } | |
39aeb52f | 963 | /* |
964 | * Take came in we have not heard about yet, | |
965 | * add in an initial atom in runnable state: | |
966 | */ | |
a116e05d ACM |
967 | if (add_sched_out_event(in_events, 'R', timestamp)) |
968 | return -1; | |
cdce9d73 | 969 | } |
39aeb52f | 970 | add_sched_in_event(in_events, timestamp); |
a116e05d ACM |
971 | |
972 | return 0; | |
39aeb52f | 973 | } |
cdce9d73 | 974 | |
9ec3f4e4 ACM |
975 | static int latency_runtime_event(struct perf_sched *sched, |
976 | struct perf_evsel *evsel, | |
977 | struct perf_sample *sample, | |
978 | struct machine *machine) | |
39aeb52f | 979 | { |
9ec3f4e4 ACM |
980 | const u32 pid = perf_evsel__intval(evsel, sample, "pid"); |
981 | const u64 runtime = perf_evsel__intval(evsel, sample, "runtime"); | |
314add6b | 982 | struct thread *thread = machine__findnew_thread(machine, 0, pid); |
0e9b07e5 | 983 | struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid); |
7f7f8d0b ACM |
984 | u64 timestamp = sample->time; |
985 | int cpu = sample->cpu; | |
39aeb52f | 986 | |
987 | BUG_ON(cpu >= MAX_CPUS || cpu < 0); | |
39aeb52f | 988 | if (!atoms) { |
0e9b07e5 | 989 | if (thread_atoms_insert(sched, thread)) |
a116e05d | 990 | return -1; |
0e9b07e5 | 991 | atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid); |
a116e05d | 992 | if (!atoms) { |
60b7d14a | 993 | pr_err("in-event: Internal tree error"); |
a116e05d ACM |
994 | return -1; |
995 | } | |
996 | if (add_sched_out_event(atoms, 'R', timestamp)) | |
997 | return -1; | |
cdce9d73 FW |
998 | } |
999 | ||
9ec3f4e4 | 1000 | add_runtime_event(atoms, runtime, timestamp); |
a116e05d | 1001 | return 0; |
cdce9d73 FW |
1002 | } |
1003 | ||
9ec3f4e4 ACM |
1004 | static int latency_wakeup_event(struct perf_sched *sched, |
1005 | struct perf_evsel *evsel, | |
1006 | struct perf_sample *sample, | |
1007 | struct machine *machine) | |
cdce9d73 | 1008 | { |
0680ee7d | 1009 | const u32 pid = perf_evsel__intval(evsel, sample, "pid"); |
39aeb52f | 1010 | struct work_atoms *atoms; |
b1ffe8f3 | 1011 | struct work_atom *atom; |
cdce9d73 | 1012 | struct thread *wakee; |
7f7f8d0b | 1013 | u64 timestamp = sample->time; |
cdce9d73 | 1014 | |
314add6b | 1015 | wakee = machine__findnew_thread(machine, 0, pid); |
0e9b07e5 | 1016 | atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid); |
17562205 | 1017 | if (!atoms) { |
0e9b07e5 | 1018 | if (thread_atoms_insert(sched, wakee)) |
a116e05d | 1019 | return -1; |
0e9b07e5 | 1020 | atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid); |
a116e05d | 1021 | if (!atoms) { |
60b7d14a | 1022 | pr_err("wakeup-event: Internal tree error"); |
a116e05d ACM |
1023 | return -1; |
1024 | } | |
1025 | if (add_sched_out_event(atoms, 'S', timestamp)) | |
1026 | return -1; | |
cdce9d73 FW |
1027 | } |
1028 | ||
39aeb52f | 1029 | BUG_ON(list_empty(&atoms->work_list)); |
cdce9d73 | 1030 | |
39aeb52f | 1031 | atom = list_entry(atoms->work_list.prev, struct work_atom, list); |
cdce9d73 | 1032 | |
55ffb7a6 | 1033 | /* |
67d6259d DY |
1034 | * As we do not guarantee the wakeup event happens when |
1035 | * task is out of run queue, also may happen when task is | |
1036 | * on run queue and wakeup only change ->state to TASK_RUNNING, | |
1037 | * then we should not set the ->wake_up_time when wake up a | |
1038 | * task which is on run queue. | |
1039 | * | |
55ffb7a6 MG |
1040 | * You WILL be missing events if you've recorded only |
1041 | * one CPU, or are only looking at only one, so don't | |
67d6259d | 1042 | * skip in this case. |
55ffb7a6 | 1043 | */ |
0e9b07e5 | 1044 | if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING) |
67d6259d | 1045 | return 0; |
cdce9d73 | 1046 | |
0e9b07e5 | 1047 | sched->nr_timestamps++; |
ea57c4f5 | 1048 | if (atom->sched_out_time > timestamp) { |
0e9b07e5 | 1049 | sched->nr_unordered_timestamps++; |
a116e05d | 1050 | return 0; |
ea57c4f5 | 1051 | } |
aa1ab9d2 | 1052 | |
b1ffe8f3 IM |
1053 | atom->state = THREAD_WAIT_CPU; |
1054 | atom->wake_up_time = timestamp; | |
a116e05d | 1055 | return 0; |
cdce9d73 FW |
1056 | } |
1057 | ||
9ec3f4e4 ACM |
1058 | static int latency_migrate_task_event(struct perf_sched *sched, |
1059 | struct perf_evsel *evsel, | |
1060 | struct perf_sample *sample, | |
1061 | struct machine *machine) | |
55ffb7a6 | 1062 | { |
9ec3f4e4 | 1063 | const u32 pid = perf_evsel__intval(evsel, sample, "pid"); |
7f7f8d0b | 1064 | u64 timestamp = sample->time; |
55ffb7a6 MG |
1065 | struct work_atoms *atoms; |
1066 | struct work_atom *atom; | |
1067 | struct thread *migrant; | |
1068 | ||
1069 | /* | |
1070 | * Only need to worry about migration when profiling one CPU. | |
1071 | */ | |
0e9b07e5 | 1072 | if (sched->profile_cpu == -1) |
a116e05d | 1073 | return 0; |
55ffb7a6 | 1074 | |
314add6b | 1075 | migrant = machine__findnew_thread(machine, 0, pid); |
0e9b07e5 | 1076 | atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid); |
55ffb7a6 | 1077 | if (!atoms) { |
0e9b07e5 | 1078 | if (thread_atoms_insert(sched, migrant)) |
a116e05d | 1079 | return -1; |
b9c5143a | 1080 | register_pid(sched, migrant->tid, thread__comm_str(migrant)); |
0e9b07e5 | 1081 | atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid); |
a116e05d | 1082 | if (!atoms) { |
60b7d14a | 1083 | pr_err("migration-event: Internal tree error"); |
a116e05d ACM |
1084 | return -1; |
1085 | } | |
1086 | if (add_sched_out_event(atoms, 'R', timestamp)) | |
1087 | return -1; | |
55ffb7a6 MG |
1088 | } |
1089 | ||
1090 | BUG_ON(list_empty(&atoms->work_list)); | |
1091 | ||
1092 | atom = list_entry(atoms->work_list.prev, struct work_atom, list); | |
1093 | atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp; | |
1094 | ||
0e9b07e5 | 1095 | sched->nr_timestamps++; |
55ffb7a6 MG |
1096 | |
1097 | if (atom->sched_out_time > timestamp) | |
0e9b07e5 | 1098 | sched->nr_unordered_timestamps++; |
a116e05d ACM |
1099 | |
1100 | return 0; | |
55ffb7a6 MG |
1101 | } |
1102 | ||
0e9b07e5 | 1103 | static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list) |
cdce9d73 | 1104 | { |
cdce9d73 FW |
1105 | int i; |
1106 | int ret; | |
66685678 | 1107 | u64 avg; |
cdce9d73 | 1108 | |
39aeb52f | 1109 | if (!work_list->nb_atoms) |
cdce9d73 | 1110 | return; |
ea57c4f5 IM |
1111 | /* |
1112 | * Ignore idle threads: | |
1113 | */ | |
b9c5143a | 1114 | if (!strcmp(thread__comm_str(work_list->thread), "swapper")) |
ea57c4f5 | 1115 | return; |
cdce9d73 | 1116 | |
0e9b07e5 ACM |
1117 | sched->all_runtime += work_list->total_runtime; |
1118 | sched->all_count += work_list->nb_atoms; | |
66685678 | 1119 | |
b9c5143a | 1120 | ret = printf(" %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid); |
cdce9d73 | 1121 | |
08f69e6c | 1122 | for (i = 0; i < 24 - ret; i++) |
cdce9d73 FW |
1123 | printf(" "); |
1124 | ||
39aeb52f | 1125 | avg = work_list->total_lat / work_list->nb_atoms; |
cdce9d73 | 1126 | |
80790e0b | 1127 | printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13.6f s\n", |
39aeb52f | 1128 | (double)work_list->total_runtime / 1e6, |
1129 | work_list->nb_atoms, (double)avg / 1e6, | |
3786310a FW |
1130 | (double)work_list->max_lat / 1e6, |
1131 | (double)work_list->max_lat_at / 1e9); | |
cdce9d73 FW |
1132 | } |
1133 | ||
39aeb52f | 1134 | static int pid_cmp(struct work_atoms *l, struct work_atoms *r) |
daa1d7a5 | 1135 | { |
38051234 | 1136 | if (l->thread->tid < r->thread->tid) |
daa1d7a5 | 1137 | return -1; |
38051234 | 1138 | if (l->thread->tid > r->thread->tid) |
daa1d7a5 FW |
1139 | return 1; |
1140 | ||
1141 | return 0; | |
1142 | } | |
1143 | ||
39aeb52f | 1144 | static int avg_cmp(struct work_atoms *l, struct work_atoms *r) |
daa1d7a5 FW |
1145 | { |
1146 | u64 avgl, avgr; | |
1147 | ||
1148 | if (!l->nb_atoms) | |
1149 | return -1; | |
1150 | ||
1151 | if (!r->nb_atoms) | |
1152 | return 1; | |
1153 | ||
1154 | avgl = l->total_lat / l->nb_atoms; | |
1155 | avgr = r->total_lat / r->nb_atoms; | |
1156 | ||
1157 | if (avgl < avgr) | |
1158 | return -1; | |
1159 | if (avgl > avgr) | |
1160 | return 1; | |
1161 | ||
1162 | return 0; | |
1163 | } | |
1164 | ||
39aeb52f | 1165 | static int max_cmp(struct work_atoms *l, struct work_atoms *r) |
daa1d7a5 FW |
1166 | { |
1167 | if (l->max_lat < r->max_lat) | |
1168 | return -1; | |
1169 | if (l->max_lat > r->max_lat) | |
1170 | return 1; | |
1171 | ||
1172 | return 0; | |
1173 | } | |
1174 | ||
39aeb52f | 1175 | static int switch_cmp(struct work_atoms *l, struct work_atoms *r) |
daa1d7a5 FW |
1176 | { |
1177 | if (l->nb_atoms < r->nb_atoms) | |
1178 | return -1; | |
1179 | if (l->nb_atoms > r->nb_atoms) | |
1180 | return 1; | |
1181 | ||
1182 | return 0; | |
1183 | } | |
1184 | ||
39aeb52f | 1185 | static int runtime_cmp(struct work_atoms *l, struct work_atoms *r) |
daa1d7a5 FW |
1186 | { |
1187 | if (l->total_runtime < r->total_runtime) | |
1188 | return -1; | |
1189 | if (l->total_runtime > r->total_runtime) | |
1190 | return 1; | |
1191 | ||
1192 | return 0; | |
1193 | } | |
1194 | ||
cbef79a8 | 1195 | static int sort_dimension__add(const char *tok, struct list_head *list) |
daa1d7a5 | 1196 | { |
0e9b07e5 ACM |
1197 | size_t i; |
1198 | static struct sort_dimension avg_sort_dimension = { | |
1199 | .name = "avg", | |
1200 | .cmp = avg_cmp, | |
1201 | }; | |
1202 | static struct sort_dimension max_sort_dimension = { | |
1203 | .name = "max", | |
1204 | .cmp = max_cmp, | |
1205 | }; | |
1206 | static struct sort_dimension pid_sort_dimension = { | |
1207 | .name = "pid", | |
1208 | .cmp = pid_cmp, | |
1209 | }; | |
1210 | static struct sort_dimension runtime_sort_dimension = { | |
1211 | .name = "runtime", | |
1212 | .cmp = runtime_cmp, | |
1213 | }; | |
1214 | static struct sort_dimension switch_sort_dimension = { | |
1215 | .name = "switch", | |
1216 | .cmp = switch_cmp, | |
1217 | }; | |
1218 | struct sort_dimension *available_sorts[] = { | |
1219 | &pid_sort_dimension, | |
1220 | &avg_sort_dimension, | |
1221 | &max_sort_dimension, | |
1222 | &switch_sort_dimension, | |
1223 | &runtime_sort_dimension, | |
1224 | }; | |
daa1d7a5 | 1225 | |
0e9b07e5 | 1226 | for (i = 0; i < ARRAY_SIZE(available_sorts); i++) { |
daa1d7a5 FW |
1227 | if (!strcmp(available_sorts[i]->name, tok)) { |
1228 | list_add_tail(&available_sorts[i]->list, list); | |
1229 | ||
1230 | return 0; | |
1231 | } | |
1232 | } | |
1233 | ||
1234 | return -1; | |
1235 | } | |
1236 | ||
0e9b07e5 | 1237 | static void perf_sched__sort_lat(struct perf_sched *sched) |
daa1d7a5 FW |
1238 | { |
1239 | struct rb_node *node; | |
1240 | ||
1241 | for (;;) { | |
39aeb52f | 1242 | struct work_atoms *data; |
0e9b07e5 | 1243 | node = rb_first(&sched->atom_root); |
daa1d7a5 FW |
1244 | if (!node) |
1245 | break; | |
1246 | ||
0e9b07e5 | 1247 | rb_erase(node, &sched->atom_root); |
39aeb52f | 1248 | data = rb_entry(node, struct work_atoms, node); |
0e9b07e5 | 1249 | __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list); |
daa1d7a5 FW |
1250 | } |
1251 | } | |
1252 | ||
0e9b07e5 | 1253 | static int process_sched_wakeup_event(struct perf_tool *tool, |
2b7fcbc5 | 1254 | struct perf_evsel *evsel, |
1d037ca1 | 1255 | struct perf_sample *sample, |
4218e673 | 1256 | struct machine *machine) |
419ab0d6 | 1257 | { |
0e9b07e5 | 1258 | struct perf_sched *sched = container_of(tool, struct perf_sched, tool); |
419ab0d6 | 1259 | |
9ec3f4e4 ACM |
1260 | if (sched->tp_handler->wakeup_event) |
1261 | return sched->tp_handler->wakeup_event(sched, evsel, sample, machine); | |
a116e05d | 1262 | |
2b7fcbc5 | 1263 | return 0; |
419ab0d6 FW |
1264 | } |
1265 | ||
9ec3f4e4 ACM |
1266 | static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel, |
1267 | struct perf_sample *sample, struct machine *machine) | |
0ec04e16 | 1268 | { |
9d372ca5 DY |
1269 | const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid"); |
1270 | struct thread *sched_in; | |
0ec04e16 | 1271 | int new_shortname; |
7f7f8d0b | 1272 | u64 timestamp0, timestamp = sample->time; |
0ec04e16 | 1273 | s64 delta; |
7f7f8d0b | 1274 | int cpu, this_cpu = sample->cpu; |
0ec04e16 IM |
1275 | |
1276 | BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0); | |
1277 | ||
0e9b07e5 ACM |
1278 | if (this_cpu > sched->max_cpu) |
1279 | sched->max_cpu = this_cpu; | |
0ec04e16 | 1280 | |
0e9b07e5 ACM |
1281 | timestamp0 = sched->cpu_last_switched[this_cpu]; |
1282 | sched->cpu_last_switched[this_cpu] = timestamp; | |
0ec04e16 IM |
1283 | if (timestamp0) |
1284 | delta = timestamp - timestamp0; | |
1285 | else | |
1286 | delta = 0; | |
1287 | ||
a116e05d | 1288 | if (delta < 0) { |
60b7d14a | 1289 | pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta); |
a116e05d ACM |
1290 | return -1; |
1291 | } | |
0ec04e16 | 1292 | |
314add6b | 1293 | sched_in = machine__findnew_thread(machine, 0, next_pid); |
0ec04e16 | 1294 | |
0e9b07e5 | 1295 | sched->curr_thread[this_cpu] = sched_in; |
0ec04e16 IM |
1296 | |
1297 | printf(" "); | |
1298 | ||
1299 | new_shortname = 0; | |
1300 | if (!sched_in->shortname[0]) { | |
6bcab4e1 D |
1301 | if (!strcmp(thread__comm_str(sched_in), "swapper")) { |
1302 | /* | |
1303 | * Don't allocate a letter-number for swapper:0 | |
1304 | * as a shortname. Instead, we use '.' for it. | |
1305 | */ | |
1306 | sched_in->shortname[0] = '.'; | |
1307 | sched_in->shortname[1] = ' '; | |
0ec04e16 | 1308 | } else { |
6bcab4e1 D |
1309 | sched_in->shortname[0] = sched->next_shortname1; |
1310 | sched_in->shortname[1] = sched->next_shortname2; | |
1311 | ||
1312 | if (sched->next_shortname1 < 'Z') { | |
1313 | sched->next_shortname1++; | |
0ec04e16 | 1314 | } else { |
6bcab4e1 D |
1315 | sched->next_shortname1 = 'A'; |
1316 | if (sched->next_shortname2 < '9') | |
1317 | sched->next_shortname2++; | |
1318 | else | |
1319 | sched->next_shortname2 = '0'; | |
0ec04e16 IM |
1320 | } |
1321 | } | |
1322 | new_shortname = 1; | |
1323 | } | |
1324 | ||
0e9b07e5 | 1325 | for (cpu = 0; cpu <= sched->max_cpu; cpu++) { |
0ec04e16 IM |
1326 | if (cpu != this_cpu) |
1327 | printf(" "); | |
1328 | else | |
1329 | printf("*"); | |
1330 | ||
6bcab4e1 D |
1331 | if (sched->curr_thread[cpu]) |
1332 | printf("%2s ", sched->curr_thread[cpu]->shortname); | |
1333 | else | |
0ec04e16 IM |
1334 | printf(" "); |
1335 | } | |
1336 | ||
1337 | printf(" %12.6f secs ", (double)timestamp/1e9); | |
1338 | if (new_shortname) { | |
1339 | printf("%s => %s:%d\n", | |
b9c5143a | 1340 | sched_in->shortname, thread__comm_str(sched_in), sched_in->tid); |
0ec04e16 IM |
1341 | } else { |
1342 | printf("\n"); | |
1343 | } | |
a116e05d ACM |
1344 | |
1345 | return 0; | |
0ec04e16 IM |
1346 | } |
1347 | ||
0e9b07e5 | 1348 | static int process_sched_switch_event(struct perf_tool *tool, |
2b7fcbc5 | 1349 | struct perf_evsel *evsel, |
1d037ca1 | 1350 | struct perf_sample *sample, |
4218e673 | 1351 | struct machine *machine) |
419ab0d6 | 1352 | { |
0e9b07e5 | 1353 | struct perf_sched *sched = container_of(tool, struct perf_sched, tool); |
a116e05d | 1354 | int this_cpu = sample->cpu, err = 0; |
2b7fcbc5 ACM |
1355 | u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"), |
1356 | next_pid = perf_evsel__intval(evsel, sample, "next_pid"); | |
419ab0d6 | 1357 | |
0e9b07e5 | 1358 | if (sched->curr_pid[this_cpu] != (u32)-1) { |
c8a37751 IM |
1359 | /* |
1360 | * Are we trying to switch away a PID that is | |
1361 | * not current? | |
1362 | */ | |
2b7fcbc5 | 1363 | if (sched->curr_pid[this_cpu] != prev_pid) |
0e9b07e5 | 1364 | sched->nr_context_switch_bugs++; |
c8a37751 | 1365 | } |
c8a37751 | 1366 | |
9ec3f4e4 ACM |
1367 | if (sched->tp_handler->switch_event) |
1368 | err = sched->tp_handler->switch_event(sched, evsel, sample, machine); | |
2b7fcbc5 ACM |
1369 | |
1370 | sched->curr_pid[this_cpu] = next_pid; | |
a116e05d | 1371 | return err; |
419ab0d6 FW |
1372 | } |
1373 | ||
0e9b07e5 | 1374 | static int process_sched_runtime_event(struct perf_tool *tool, |
2b7fcbc5 | 1375 | struct perf_evsel *evsel, |
1d037ca1 | 1376 | struct perf_sample *sample, |
4218e673 | 1377 | struct machine *machine) |
39aeb52f | 1378 | { |
0e9b07e5 | 1379 | struct perf_sched *sched = container_of(tool, struct perf_sched, tool); |
39aeb52f | 1380 | |
9ec3f4e4 ACM |
1381 | if (sched->tp_handler->runtime_event) |
1382 | return sched->tp_handler->runtime_event(sched, evsel, sample, machine); | |
a116e05d | 1383 | |
2b7fcbc5 | 1384 | return 0; |
39aeb52f | 1385 | } |
1386 | ||
cb627505 DA |
1387 | static int perf_sched__process_fork_event(struct perf_tool *tool, |
1388 | union perf_event *event, | |
1389 | struct perf_sample *sample, | |
1390 | struct machine *machine) | |
fbf94829 | 1391 | { |
0e9b07e5 | 1392 | struct perf_sched *sched = container_of(tool, struct perf_sched, tool); |
46538818 | 1393 | |
cb627505 DA |
1394 | /* run the fork event through the perf machineruy */ |
1395 | perf_event__process_fork(tool, event, sample, machine); | |
1396 | ||
1397 | /* and then run additional processing needed for this command */ | |
9ec3f4e4 | 1398 | if (sched->tp_handler->fork_event) |
cb627505 | 1399 | return sched->tp_handler->fork_event(sched, event, machine); |
a116e05d | 1400 | |
2b7fcbc5 | 1401 | return 0; |
fbf94829 IM |
1402 | } |
1403 | ||
0e9b07e5 | 1404 | static int process_sched_migrate_task_event(struct perf_tool *tool, |
2b7fcbc5 | 1405 | struct perf_evsel *evsel, |
1d037ca1 | 1406 | struct perf_sample *sample, |
4218e673 | 1407 | struct machine *machine) |
55ffb7a6 | 1408 | { |
0e9b07e5 | 1409 | struct perf_sched *sched = container_of(tool, struct perf_sched, tool); |
55ffb7a6 | 1410 | |
9ec3f4e4 ACM |
1411 | if (sched->tp_handler->migrate_task_event) |
1412 | return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine); | |
a116e05d | 1413 | |
2b7fcbc5 | 1414 | return 0; |
55ffb7a6 MG |
1415 | } |
1416 | ||
a116e05d | 1417 | typedef int (*tracepoint_handler)(struct perf_tool *tool, |
2b7fcbc5 | 1418 | struct perf_evsel *evsel, |
a116e05d | 1419 | struct perf_sample *sample, |
4218e673 | 1420 | struct machine *machine); |
ec156764 | 1421 | |
1d037ca1 IT |
1422 | static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused, |
1423 | union perf_event *event __maybe_unused, | |
ee29be62 ACM |
1424 | struct perf_sample *sample, |
1425 | struct perf_evsel *evsel, | |
1426 | struct machine *machine) | |
0a02ad93 | 1427 | { |
a116e05d | 1428 | int err = 0; |
0a02ad93 | 1429 | |
ee29be62 | 1430 | evsel->hists.stats.total_period += sample->period; |
1844dbcb | 1431 | hists__inc_nr_samples(&evsel->hists, true); |
f39cdf25 | 1432 | |
744a9719 ACM |
1433 | if (evsel->handler != NULL) { |
1434 | tracepoint_handler f = evsel->handler; | |
2b7fcbc5 | 1435 | err = f(tool, evsel, sample, machine); |
ee29be62 | 1436 | } |
0a02ad93 | 1437 | |
a116e05d | 1438 | return err; |
0a02ad93 IM |
1439 | } |
1440 | ||
ad9def7c | 1441 | static int perf_sched__read_events(struct perf_sched *sched, |
0e9b07e5 | 1442 | struct perf_session **psession) |
0a02ad93 | 1443 | { |
ee29be62 ACM |
1444 | const struct perf_evsel_str_handler handlers[] = { |
1445 | { "sched:sched_switch", process_sched_switch_event, }, | |
1446 | { "sched:sched_stat_runtime", process_sched_runtime_event, }, | |
1447 | { "sched:sched_wakeup", process_sched_wakeup_event, }, | |
1448 | { "sched:sched_wakeup_new", process_sched_wakeup_event, }, | |
ee29be62 ACM |
1449 | { "sched:sched_migrate_task", process_sched_migrate_task_event, }, |
1450 | }; | |
da378962 | 1451 | struct perf_session *session; |
f5fc1412 JO |
1452 | struct perf_data_file file = { |
1453 | .path = input_name, | |
1454 | .mode = PERF_DATA_MODE_READ, | |
1455 | }; | |
da378962 | 1456 | |
f5fc1412 | 1457 | session = perf_session__new(&file, false, &sched->tool); |
a116e05d ACM |
1458 | if (session == NULL) { |
1459 | pr_debug("No Memory for session\n"); | |
1460 | return -1; | |
1461 | } | |
94c744b6 | 1462 | |
a116e05d ACM |
1463 | if (perf_session__set_tracepoints_handlers(session, handlers)) |
1464 | goto out_delete; | |
ee29be62 | 1465 | |
cee75ac7 | 1466 | if (perf_session__has_traces(session, "record -R")) { |
0e9b07e5 | 1467 | int err = perf_session__process_events(session, &sched->tool); |
a116e05d ACM |
1468 | if (err) { |
1469 | pr_err("Failed to process events, error %d", err); | |
1470 | goto out_delete; | |
1471 | } | |
4c09bafa | 1472 | |
28a6b6aa ACM |
1473 | sched->nr_events = session->stats.nr_events[0]; |
1474 | sched->nr_lost_events = session->stats.total_lost; | |
1475 | sched->nr_lost_chunks = session->stats.nr_events[PERF_RECORD_LOST]; | |
cee75ac7 | 1476 | } |
d549c769 | 1477 | |
4c09bafa JO |
1478 | if (psession) |
1479 | *psession = session; | |
ad9def7c DA |
1480 | else |
1481 | perf_session__delete(session); | |
a116e05d ACM |
1482 | |
1483 | return 0; | |
1484 | ||
1485 | out_delete: | |
1486 | perf_session__delete(session); | |
1487 | return -1; | |
0a02ad93 IM |
1488 | } |
1489 | ||
0e9b07e5 | 1490 | static void print_bad_events(struct perf_sched *sched) |
0ec04e16 | 1491 | { |
0e9b07e5 | 1492 | if (sched->nr_unordered_timestamps && sched->nr_timestamps) { |
0ec04e16 | 1493 | printf(" INFO: %.3f%% unordered timestamps (%ld out of %ld)\n", |
0e9b07e5 ACM |
1494 | (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0, |
1495 | sched->nr_unordered_timestamps, sched->nr_timestamps); | |
0ec04e16 | 1496 | } |
0e9b07e5 | 1497 | if (sched->nr_lost_events && sched->nr_events) { |
0ec04e16 | 1498 | printf(" INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n", |
0e9b07e5 ACM |
1499 | (double)sched->nr_lost_events/(double)sched->nr_events * 100.0, |
1500 | sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks); | |
0ec04e16 | 1501 | } |
0e9b07e5 | 1502 | if (sched->nr_context_switch_bugs && sched->nr_timestamps) { |
0ec04e16 | 1503 | printf(" INFO: %.3f%% context switch bugs (%ld out of %ld)", |
0e9b07e5 ACM |
1504 | (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0, |
1505 | sched->nr_context_switch_bugs, sched->nr_timestamps); | |
1506 | if (sched->nr_lost_events) | |
0ec04e16 IM |
1507 | printf(" (due to lost events?)"); |
1508 | printf("\n"); | |
1509 | } | |
1510 | } | |
1511 | ||
0e9b07e5 | 1512 | static int perf_sched__lat(struct perf_sched *sched) |
0ec04e16 IM |
1513 | { |
1514 | struct rb_node *next; | |
4c09bafa | 1515 | struct perf_session *session; |
0ec04e16 IM |
1516 | |
1517 | setup_pager(); | |
ad9def7c DA |
1518 | |
1519 | /* save session -- references to threads are held in work_list */ | |
1520 | if (perf_sched__read_events(sched, &session)) | |
a116e05d | 1521 | return -1; |
ad9def7c | 1522 | |
0e9b07e5 | 1523 | perf_sched__sort_lat(sched); |
0ec04e16 | 1524 | |
80790e0b RR |
1525 | printf("\n -----------------------------------------------------------------------------------------------------------------\n"); |
1526 | printf(" Task | Runtime ms | Switches | Average delay ms | Maximum delay ms | Maximum delay at |\n"); | |
1527 | printf(" -----------------------------------------------------------------------------------------------------------------\n"); | |
0ec04e16 | 1528 | |
0e9b07e5 | 1529 | next = rb_first(&sched->sorted_atom_root); |
0ec04e16 IM |
1530 | |
1531 | while (next) { | |
1532 | struct work_atoms *work_list; | |
1533 | ||
1534 | work_list = rb_entry(next, struct work_atoms, node); | |
0e9b07e5 | 1535 | output_lat_thread(sched, work_list); |
0ec04e16 IM |
1536 | next = rb_next(next); |
1537 | } | |
1538 | ||
80790e0b | 1539 | printf(" -----------------------------------------------------------------------------------------------------------------\n"); |
9486aa38 | 1540 | printf(" TOTAL: |%11.3f ms |%9" PRIu64 " |\n", |
0e9b07e5 | 1541 | (double)sched->all_runtime / 1e6, sched->all_count); |
0ec04e16 IM |
1542 | |
1543 | printf(" ---------------------------------------------------\n"); | |
1544 | ||
0e9b07e5 | 1545 | print_bad_events(sched); |
0ec04e16 IM |
1546 | printf("\n"); |
1547 | ||
4c09bafa | 1548 | perf_session__delete(session); |
a116e05d | 1549 | return 0; |
0ec04e16 IM |
1550 | } |
1551 | ||
0e9b07e5 | 1552 | static int perf_sched__map(struct perf_sched *sched) |
0ec04e16 | 1553 | { |
0e9b07e5 | 1554 | sched->max_cpu = sysconf(_SC_NPROCESSORS_CONF); |
40749d0f | 1555 | |
0ec04e16 | 1556 | setup_pager(); |
ad9def7c | 1557 | if (perf_sched__read_events(sched, NULL)) |
a116e05d | 1558 | return -1; |
0e9b07e5 | 1559 | print_bad_events(sched); |
a116e05d | 1560 | return 0; |
0ec04e16 IM |
1561 | } |
1562 | ||
0e9b07e5 | 1563 | static int perf_sched__replay(struct perf_sched *sched) |
0ec04e16 IM |
1564 | { |
1565 | unsigned long i; | |
1566 | ||
0e9b07e5 ACM |
1567 | calibrate_run_measurement_overhead(sched); |
1568 | calibrate_sleep_measurement_overhead(sched); | |
0ec04e16 | 1569 | |
0e9b07e5 | 1570 | test_calibrations(sched); |
0ec04e16 | 1571 | |
ad9def7c | 1572 | if (perf_sched__read_events(sched, NULL)) |
a116e05d | 1573 | return -1; |
0ec04e16 | 1574 | |
0e9b07e5 ACM |
1575 | printf("nr_run_events: %ld\n", sched->nr_run_events); |
1576 | printf("nr_sleep_events: %ld\n", sched->nr_sleep_events); | |
1577 | printf("nr_wakeup_events: %ld\n", sched->nr_wakeup_events); | |
0ec04e16 | 1578 | |
0e9b07e5 ACM |
1579 | if (sched->targetless_wakeups) |
1580 | printf("target-less wakeups: %ld\n", sched->targetless_wakeups); | |
1581 | if (sched->multitarget_wakeups) | |
1582 | printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups); | |
1583 | if (sched->nr_run_events_optimized) | |
0ec04e16 | 1584 | printf("run atoms optimized: %ld\n", |
0e9b07e5 | 1585 | sched->nr_run_events_optimized); |
0ec04e16 | 1586 | |
0e9b07e5 ACM |
1587 | print_task_traces(sched); |
1588 | add_cross_task_wakeups(sched); | |
0ec04e16 | 1589 | |
0e9b07e5 | 1590 | create_tasks(sched); |
0ec04e16 | 1591 | printf("------------------------------------------------------------\n"); |
0e9b07e5 ACM |
1592 | for (i = 0; i < sched->replay_repeat; i++) |
1593 | run_one_test(sched); | |
a116e05d ACM |
1594 | |
1595 | return 0; | |
0ec04e16 IM |
1596 | } |
1597 | ||
0e9b07e5 ACM |
1598 | static void setup_sorting(struct perf_sched *sched, const struct option *options, |
1599 | const char * const usage_msg[]) | |
daa1d7a5 | 1600 | { |
0e9b07e5 | 1601 | char *tmp, *tok, *str = strdup(sched->sort_order); |
daa1d7a5 FW |
1602 | |
1603 | for (tok = strtok_r(str, ", ", &tmp); | |
1604 | tok; tok = strtok_r(NULL, ", ", &tmp)) { | |
0e9b07e5 | 1605 | if (sort_dimension__add(tok, &sched->sort_list) < 0) { |
daa1d7a5 | 1606 | error("Unknown --sort key: `%s'", tok); |
0e9b07e5 | 1607 | usage_with_options(usage_msg, options); |
daa1d7a5 FW |
1608 | } |
1609 | } | |
1610 | ||
1611 | free(str); | |
1612 | ||
0e9b07e5 | 1613 | sort_dimension__add("pid", &sched->cmp_pid); |
daa1d7a5 FW |
1614 | } |
1615 | ||
1fc35b29 IM |
1616 | static int __cmd_record(int argc, const char **argv) |
1617 | { | |
1618 | unsigned int rec_argc, i, j; | |
1619 | const char **rec_argv; | |
0e9b07e5 ACM |
1620 | const char * const record_args[] = { |
1621 | "record", | |
1622 | "-a", | |
1623 | "-R", | |
0e9b07e5 ACM |
1624 | "-m", "1024", |
1625 | "-c", "1", | |
1626 | "-e", "sched:sched_switch", | |
1627 | "-e", "sched:sched_stat_wait", | |
1628 | "-e", "sched:sched_stat_sleep", | |
1629 | "-e", "sched:sched_stat_iowait", | |
1630 | "-e", "sched:sched_stat_runtime", | |
0e9b07e5 ACM |
1631 | "-e", "sched:sched_process_fork", |
1632 | "-e", "sched:sched_wakeup", | |
7fff9597 | 1633 | "-e", "sched:sched_wakeup_new", |
0e9b07e5 ACM |
1634 | "-e", "sched:sched_migrate_task", |
1635 | }; | |
1fc35b29 IM |
1636 | |
1637 | rec_argc = ARRAY_SIZE(record_args) + argc - 1; | |
1638 | rec_argv = calloc(rec_argc + 1, sizeof(char *)); | |
1639 | ||
e462dc55 | 1640 | if (rec_argv == NULL) |
ce47dc56 CS |
1641 | return -ENOMEM; |
1642 | ||
1fc35b29 IM |
1643 | for (i = 0; i < ARRAY_SIZE(record_args); i++) |
1644 | rec_argv[i] = strdup(record_args[i]); | |
1645 | ||
1646 | for (j = 1; j < (unsigned int)argc; j++, i++) | |
1647 | rec_argv[i] = argv[j]; | |
1648 | ||
1649 | BUG_ON(i != rec_argc); | |
1650 | ||
1651 | return cmd_record(i, rec_argv, NULL); | |
1652 | } | |
1653 | ||
1d037ca1 | 1654 | int cmd_sched(int argc, const char **argv, const char *prefix __maybe_unused) |
0a02ad93 | 1655 | { |
8a39df8f AH |
1656 | const char default_sort_order[] = "avg, max, switch, runtime"; |
1657 | struct perf_sched sched = { | |
1658 | .tool = { | |
1659 | .sample = perf_sched__process_tracepoint_sample, | |
1660 | .comm = perf_event__process_comm, | |
1661 | .lost = perf_event__process_lost, | |
1662 | .fork = perf_sched__process_fork_event, | |
1663 | .ordered_samples = true, | |
1664 | }, | |
1665 | .cmp_pid = LIST_HEAD_INIT(sched.cmp_pid), | |
1666 | .sort_list = LIST_HEAD_INIT(sched.sort_list), | |
1667 | .start_work_mutex = PTHREAD_MUTEX_INITIALIZER, | |
1668 | .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER, | |
8a39df8f AH |
1669 | .sort_order = default_sort_order, |
1670 | .replay_repeat = 10, | |
1671 | .profile_cpu = -1, | |
1672 | .next_shortname1 = 'A', | |
1673 | .next_shortname2 = '0', | |
1674 | }; | |
0e9b07e5 ACM |
1675 | const struct option latency_options[] = { |
1676 | OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]", | |
1677 | "sort by key(s): runtime, switch, avg, max"), | |
1678 | OPT_INCR('v', "verbose", &verbose, | |
1679 | "be more verbose (show symbol address, etc)"), | |
1680 | OPT_INTEGER('C', "CPU", &sched.profile_cpu, | |
1681 | "CPU to profile on"), | |
1682 | OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, | |
1683 | "dump raw trace in ASCII"), | |
1684 | OPT_END() | |
1685 | }; | |
1686 | const struct option replay_options[] = { | |
1687 | OPT_UINTEGER('r', "repeat", &sched.replay_repeat, | |
1688 | "repeat the workload replay N times (-1: infinite)"), | |
1689 | OPT_INCR('v', "verbose", &verbose, | |
1690 | "be more verbose (show symbol address, etc)"), | |
1691 | OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, | |
1692 | "dump raw trace in ASCII"), | |
1693 | OPT_END() | |
1694 | }; | |
1695 | const struct option sched_options[] = { | |
70cb4e96 | 1696 | OPT_STRING('i', "input", &input_name, "file", |
0e9b07e5 ACM |
1697 | "input file name"), |
1698 | OPT_INCR('v', "verbose", &verbose, | |
1699 | "be more verbose (show symbol address, etc)"), | |
1700 | OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, | |
1701 | "dump raw trace in ASCII"), | |
1702 | OPT_END() | |
1703 | }; | |
1704 | const char * const latency_usage[] = { | |
1705 | "perf sched latency [<options>]", | |
1706 | NULL | |
1707 | }; | |
1708 | const char * const replay_usage[] = { | |
1709 | "perf sched replay [<options>]", | |
1710 | NULL | |
1711 | }; | |
a83edb2d RR |
1712 | const char *const sched_subcommands[] = { "record", "latency", "map", |
1713 | "replay", "script", NULL }; | |
1714 | const char *sched_usage[] = { | |
1715 | NULL, | |
0e9b07e5 ACM |
1716 | NULL |
1717 | }; | |
1718 | struct trace_sched_handler lat_ops = { | |
1719 | .wakeup_event = latency_wakeup_event, | |
1720 | .switch_event = latency_switch_event, | |
1721 | .runtime_event = latency_runtime_event, | |
0e9b07e5 ACM |
1722 | .migrate_task_event = latency_migrate_task_event, |
1723 | }; | |
1724 | struct trace_sched_handler map_ops = { | |
1725 | .switch_event = map_switch_event, | |
1726 | }; | |
1727 | struct trace_sched_handler replay_ops = { | |
1728 | .wakeup_event = replay_wakeup_event, | |
1729 | .switch_event = replay_switch_event, | |
1730 | .fork_event = replay_fork_event, | |
1731 | }; | |
156a2b02 AH |
1732 | unsigned int i; |
1733 | ||
1734 | for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++) | |
1735 | sched.curr_pid[i] = -1; | |
0e9b07e5 | 1736 | |
a83edb2d RR |
1737 | argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands, |
1738 | sched_usage, PARSE_OPT_STOP_AT_NON_OPTION); | |
f2858d8a IM |
1739 | if (!argc) |
1740 | usage_with_options(sched_usage, sched_options); | |
0a02ad93 | 1741 | |
c0777c5a | 1742 | /* |
133dc4c3 | 1743 | * Aliased to 'perf script' for now: |
c0777c5a | 1744 | */ |
133dc4c3 IM |
1745 | if (!strcmp(argv[0], "script")) |
1746 | return cmd_script(argc, argv, prefix); | |
c0777c5a | 1747 | |
75be6cf4 | 1748 | symbol__init(); |
1fc35b29 IM |
1749 | if (!strncmp(argv[0], "rec", 3)) { |
1750 | return __cmd_record(argc, argv); | |
1751 | } else if (!strncmp(argv[0], "lat", 3)) { | |
0e9b07e5 | 1752 | sched.tp_handler = &lat_ops; |
f2858d8a IM |
1753 | if (argc > 1) { |
1754 | argc = parse_options(argc, argv, latency_options, latency_usage, 0); | |
1755 | if (argc) | |
1756 | usage_with_options(latency_usage, latency_options); | |
f2858d8a | 1757 | } |
0e9b07e5 ACM |
1758 | setup_sorting(&sched, latency_options, latency_usage); |
1759 | return perf_sched__lat(&sched); | |
0ec04e16 | 1760 | } else if (!strcmp(argv[0], "map")) { |
0e9b07e5 ACM |
1761 | sched.tp_handler = &map_ops; |
1762 | setup_sorting(&sched, latency_options, latency_usage); | |
1763 | return perf_sched__map(&sched); | |
f2858d8a | 1764 | } else if (!strncmp(argv[0], "rep", 3)) { |
0e9b07e5 | 1765 | sched.tp_handler = &replay_ops; |
f2858d8a IM |
1766 | if (argc) { |
1767 | argc = parse_options(argc, argv, replay_options, replay_usage, 0); | |
1768 | if (argc) | |
1769 | usage_with_options(replay_usage, replay_options); | |
1770 | } | |
0e9b07e5 | 1771 | return perf_sched__replay(&sched); |
f2858d8a IM |
1772 | } else { |
1773 | usage_with_options(sched_usage, sched_options); | |
1774 | } | |
1775 | ||
ec156764 | 1776 | return 0; |
0a02ad93 | 1777 | } |