perf sched: Make it easier to plug in new sub profilers
[linux-3.10.git] / tools / perf / builtin-sched.c
1 #include "builtin.h"
2
3 #include "util/util.h"
4 #include "util/cache.h"
5 #include "util/symbol.h"
6 #include "util/thread.h"
7 #include "util/header.h"
8
9 #include "util/parse-options.h"
10
11 #include "perf.h"
12 #include "util/debug.h"
13
14 #include "util/trace-event.h"
15 #include <sys/types.h>
16
17
18 #define MAX_CPUS 4096
19
20 static char                     const *input_name = "perf.data";
21 static int                      input;
22 static unsigned long            page_size;
23 static unsigned long            mmap_window = 32;
24
25 static unsigned long            total_comm = 0;
26
27 static struct rb_root           threads;
28 static struct thread            *last_match;
29
30 static struct perf_header       *header;
31 static u64                      sample_type;
32
33 static int                      replay_mode;
34
35
36 /*
37  * Scheduler benchmarks
38  */
39 #include <sys/resource.h>
40 #include <sys/types.h>
41 #include <sys/stat.h>
42 #include <sys/time.h>
43 #include <sys/prctl.h>
44
45 #include <linux/unistd.h>
46
47 #include <semaphore.h>
48 #include <pthread.h>
49 #include <signal.h>
50 #include <values.h>
51 #include <string.h>
52 #include <unistd.h>
53 #include <stdlib.h>
54 #include <assert.h>
55 #include <fcntl.h>
56 #include <time.h>
57 #include <math.h>
58
59 #include <stdio.h>
60
61 #define PR_SET_NAME     15               /* Set process name */
62
63 #define BUG_ON(x)       assert(!(x))
64
65 #define DEBUG           0
66
67 typedef unsigned long long nsec_t;
68
69 static nsec_t run_measurement_overhead;
70 static nsec_t sleep_measurement_overhead;
71
72 static nsec_t get_nsecs(void)
73 {
74         struct timespec ts;
75
76         clock_gettime(CLOCK_MONOTONIC, &ts);
77
78         return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
79 }
80
81 static void burn_nsecs(nsec_t nsecs)
82 {
83         nsec_t T0 = get_nsecs(), T1;
84
85         do {
86                 T1 = get_nsecs();
87         } while (T1 + run_measurement_overhead < T0 + nsecs);
88 }
89
90 static void sleep_nsecs(nsec_t nsecs)
91 {
92         struct timespec ts;
93
94         ts.tv_nsec = nsecs % 999999999;
95         ts.tv_sec = nsecs / 999999999;
96
97         nanosleep(&ts, NULL);
98 }
99
100 static void calibrate_run_measurement_overhead(void)
101 {
102         nsec_t T0, T1, delta, min_delta = 1000000000ULL;
103         int i;
104
105         for (i = 0; i < 10; i++) {
106                 T0 = get_nsecs();
107                 burn_nsecs(0);
108                 T1 = get_nsecs();
109                 delta = T1-T0;
110                 min_delta = min(min_delta, delta);
111         }
112         run_measurement_overhead = min_delta;
113
114         printf("run measurement overhead: %Ld nsecs\n", min_delta);
115 }
116
117 static void calibrate_sleep_measurement_overhead(void)
118 {
119         nsec_t T0, T1, delta, min_delta = 1000000000ULL;
120         int i;
121
122         for (i = 0; i < 10; i++) {
123                 T0 = get_nsecs();
124                 sleep_nsecs(10000);
125                 T1 = get_nsecs();
126                 delta = T1-T0;
127                 min_delta = min(min_delta, delta);
128         }
129         min_delta -= 10000;
130         sleep_measurement_overhead = min_delta;
131
132         printf("sleep measurement overhead: %Ld nsecs\n", min_delta);
133 }
134
135 #define COMM_LEN        20
136 #define SYM_LEN         129
137
138 #define MAX_PID         65536
139
140 static unsigned long nr_tasks;
141
142 struct sched_event;
143
144 struct task_desc {
145         unsigned long           nr;
146         unsigned long           pid;
147         char                    comm[COMM_LEN];
148
149         unsigned long           nr_events;
150         unsigned long           curr_event;
151         struct sched_event      **events;
152
153         pthread_t               thread;
154         sem_t                   sleep_sem;
155
156         sem_t                   ready_for_work;
157         sem_t                   work_done_sem;
158
159         nsec_t                  cpu_usage;
160 };
161
162 enum sched_event_type {
163         SCHED_EVENT_RUN,
164         SCHED_EVENT_SLEEP,
165         SCHED_EVENT_WAKEUP,
166 };
167
168 struct sched_event {
169         enum sched_event_type   type;
170         nsec_t                  timestamp;
171         nsec_t                  duration;
172         unsigned long           nr;
173         int                     specific_wait;
174         sem_t                   *wait_sem;
175         struct task_desc        *wakee;
176 };
177
178 static struct task_desc         *pid_to_task[MAX_PID];
179
180 static struct task_desc         **tasks;
181
182 static pthread_mutex_t          start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
183 static nsec_t                   start_time;
184
185 static pthread_mutex_t          work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
186
187 static unsigned long            nr_run_events;
188 static unsigned long            nr_sleep_events;
189 static unsigned long            nr_wakeup_events;
190
191 static unsigned long            nr_sleep_corrections;
192 static unsigned long            nr_run_events_optimized;
193
194 static struct sched_event *
195 get_new_event(struct task_desc *task, nsec_t timestamp)
196 {
197         struct sched_event *event = calloc(1, sizeof(*event));
198         unsigned long idx = task->nr_events;
199         size_t size;
200
201         event->timestamp = timestamp;
202         event->nr = idx;
203
204         task->nr_events++;
205         size = sizeof(struct sched_event *) * task->nr_events;
206         task->events = realloc(task->events, size);
207         BUG_ON(!task->events);
208
209         task->events[idx] = event;
210
211         return event;
212 }
213
214 static struct sched_event *last_event(struct task_desc *task)
215 {
216         if (!task->nr_events)
217                 return NULL;
218
219         return task->events[task->nr_events - 1];
220 }
221
222 static void
223 add_sched_event_run(struct task_desc *task, nsec_t timestamp, u64 duration)
224 {
225         struct sched_event *event, *curr_event = last_event(task);
226
227         /*
228          * optimize an existing RUN event by merging this one
229          * to it:
230          */
231         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
232                 nr_run_events_optimized++;
233                 curr_event->duration += duration;
234                 return;
235         }
236
237         event = get_new_event(task, timestamp);
238
239         event->type = SCHED_EVENT_RUN;
240         event->duration = duration;
241
242         nr_run_events++;
243 }
244
245 static unsigned long targetless_wakeups;
246 static unsigned long multitarget_wakeups;
247
248 static void
249 add_sched_event_wakeup(struct task_desc *task, nsec_t timestamp,
250                        struct task_desc *wakee)
251 {
252         struct sched_event *event, *wakee_event;
253
254         event = get_new_event(task, timestamp);
255         event->type = SCHED_EVENT_WAKEUP;
256         event->wakee = wakee;
257
258         wakee_event = last_event(wakee);
259         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
260                 targetless_wakeups++;
261                 return;
262         }
263         if (wakee_event->wait_sem) {
264                 multitarget_wakeups++;
265                 return;
266         }
267
268         wakee_event->wait_sem = calloc(1, sizeof(*wakee_event->wait_sem));
269         sem_init(wakee_event->wait_sem, 0, 0);
270         wakee_event->specific_wait = 1;
271         event->wait_sem = wakee_event->wait_sem;
272
273         nr_wakeup_events++;
274 }
275
276 static void
277 add_sched_event_sleep(struct task_desc *task, nsec_t timestamp,
278                       u64 task_state __used)
279 {
280         struct sched_event *event = get_new_event(task, timestamp);
281
282         event->type = SCHED_EVENT_SLEEP;
283
284         nr_sleep_events++;
285 }
286
287 static struct task_desc *register_pid(unsigned long pid, const char *comm)
288 {
289         struct task_desc *task;
290
291         BUG_ON(pid >= MAX_PID);
292
293         task = pid_to_task[pid];
294
295         if (task)
296                 return task;
297
298         task = calloc(1, sizeof(*task));
299         task->pid = pid;
300         task->nr = nr_tasks;
301         strcpy(task->comm, comm);
302         /*
303          * every task starts in sleeping state - this gets ignored
304          * if there's no wakeup pointing to this sleep state:
305          */
306         add_sched_event_sleep(task, 0, 0);
307
308         pid_to_task[pid] = task;
309         nr_tasks++;
310         tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
311         BUG_ON(!tasks);
312         tasks[task->nr] = task;
313
314         if (verbose)
315                 printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
316
317         return task;
318 }
319
320
321 static void print_task_traces(void)
322 {
323         struct task_desc *task;
324         unsigned long i;
325
326         for (i = 0; i < nr_tasks; i++) {
327                 task = tasks[i];
328                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
329                         task->nr, task->comm, task->pid, task->nr_events);
330         }
331 }
332
333 static void add_cross_task_wakeups(void)
334 {
335         struct task_desc *task1, *task2;
336         unsigned long i, j;
337
338         for (i = 0; i < nr_tasks; i++) {
339                 task1 = tasks[i];
340                 j = i + 1;
341                 if (j == nr_tasks)
342                         j = 0;
343                 task2 = tasks[j];
344                 add_sched_event_wakeup(task1, 0, task2);
345         }
346 }
347
348 static void
349 process_sched_event(struct task_desc *this_task __used, struct sched_event *event)
350 {
351         int ret = 0;
352         nsec_t now;
353         long long delta;
354
355         now = get_nsecs();
356         delta = start_time + event->timestamp - now;
357
358         switch (event->type) {
359                 case SCHED_EVENT_RUN:
360                         burn_nsecs(event->duration);
361                         break;
362                 case SCHED_EVENT_SLEEP:
363                         if (event->wait_sem)
364                                 ret = sem_wait(event->wait_sem);
365                         BUG_ON(ret);
366                         break;
367                 case SCHED_EVENT_WAKEUP:
368                         if (event->wait_sem)
369                                 ret = sem_post(event->wait_sem);
370                         BUG_ON(ret);
371                         break;
372                 default:
373                         BUG_ON(1);
374         }
375 }
376
377 static nsec_t get_cpu_usage_nsec_parent(void)
378 {
379         struct rusage ru;
380         nsec_t sum;
381         int err;
382
383         err = getrusage(RUSAGE_SELF, &ru);
384         BUG_ON(err);
385
386         sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
387         sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
388
389         return sum;
390 }
391
392 static nsec_t get_cpu_usage_nsec_self(void)
393 {
394         char filename [] = "/proc/1234567890/sched";
395         unsigned long msecs, nsecs;
396         char *line = NULL;
397         nsec_t total = 0;
398         size_t len = 0;
399         ssize_t chars;
400         FILE *file;
401         int ret;
402
403         sprintf(filename, "/proc/%d/sched", getpid());
404         file = fopen(filename, "r");
405         BUG_ON(!file);
406
407         while ((chars = getline(&line, &len, file)) != -1) {
408                 ret = sscanf(line, "se.sum_exec_runtime : %ld.%06ld\n",
409                         &msecs, &nsecs);
410                 if (ret == 2) {
411                         total = msecs*1e6 + nsecs;
412                         break;
413                 }
414         }
415         if (line)
416                 free(line);
417         fclose(file);
418
419         return total;
420 }
421
422 static void *thread_func(void *ctx)
423 {
424         struct task_desc *this_task = ctx;
425         nsec_t cpu_usage_0, cpu_usage_1;
426         unsigned long i, ret;
427         char comm2[22];
428
429         sprintf(comm2, ":%s", this_task->comm);
430         prctl(PR_SET_NAME, comm2);
431
432 again:
433         ret = sem_post(&this_task->ready_for_work);
434         BUG_ON(ret);
435         ret = pthread_mutex_lock(&start_work_mutex);
436         BUG_ON(ret);
437         ret = pthread_mutex_unlock(&start_work_mutex);
438         BUG_ON(ret);
439
440         cpu_usage_0 = get_cpu_usage_nsec_self();
441
442         for (i = 0; i < this_task->nr_events; i++) {
443                 this_task->curr_event = i;
444                 process_sched_event(this_task, this_task->events[i]);
445         }
446
447         cpu_usage_1 = get_cpu_usage_nsec_self();
448         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
449
450         ret = sem_post(&this_task->work_done_sem);
451         BUG_ON(ret);
452
453         ret = pthread_mutex_lock(&work_done_wait_mutex);
454         BUG_ON(ret);
455         ret = pthread_mutex_unlock(&work_done_wait_mutex);
456         BUG_ON(ret);
457
458         goto again;
459 }
460
461 static void create_tasks(void)
462 {
463         struct task_desc *task;
464         pthread_attr_t attr;
465         unsigned long i;
466         int err;
467
468         err = pthread_attr_init(&attr);
469         BUG_ON(err);
470         err = pthread_attr_setstacksize(&attr, (size_t)(16*1024));
471         BUG_ON(err);
472         err = pthread_mutex_lock(&start_work_mutex);
473         BUG_ON(err);
474         err = pthread_mutex_lock(&work_done_wait_mutex);
475         BUG_ON(err);
476         for (i = 0; i < nr_tasks; i++) {
477                 task = tasks[i];
478                 sem_init(&task->sleep_sem, 0, 0);
479                 sem_init(&task->ready_for_work, 0, 0);
480                 sem_init(&task->work_done_sem, 0, 0);
481                 task->curr_event = 0;
482                 err = pthread_create(&task->thread, &attr, thread_func, task);
483                 BUG_ON(err);
484         }
485 }
486
487 static nsec_t cpu_usage;
488 static nsec_t runavg_cpu_usage;
489 static nsec_t parent_cpu_usage;
490 static nsec_t runavg_parent_cpu_usage;
491
492 static void wait_for_tasks(void)
493 {
494         nsec_t cpu_usage_0, cpu_usage_1;
495         struct task_desc *task;
496         unsigned long i, ret;
497
498         start_time = get_nsecs();
499         cpu_usage = 0;
500         pthread_mutex_unlock(&work_done_wait_mutex);
501
502         for (i = 0; i < nr_tasks; i++) {
503                 task = tasks[i];
504                 ret = sem_wait(&task->ready_for_work);
505                 BUG_ON(ret);
506                 sem_init(&task->ready_for_work, 0, 0);
507         }
508         ret = pthread_mutex_lock(&work_done_wait_mutex);
509         BUG_ON(ret);
510
511         cpu_usage_0 = get_cpu_usage_nsec_parent();
512
513         pthread_mutex_unlock(&start_work_mutex);
514
515         for (i = 0; i < nr_tasks; i++) {
516                 task = tasks[i];
517                 ret = sem_wait(&task->work_done_sem);
518                 BUG_ON(ret);
519                 sem_init(&task->work_done_sem, 0, 0);
520                 cpu_usage += task->cpu_usage;
521                 task->cpu_usage = 0;
522         }
523
524         cpu_usage_1 = get_cpu_usage_nsec_parent();
525         if (!runavg_cpu_usage)
526                 runavg_cpu_usage = cpu_usage;
527         runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
528
529         parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
530         if (!runavg_parent_cpu_usage)
531                 runavg_parent_cpu_usage = parent_cpu_usage;
532         runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
533                                    parent_cpu_usage)/10;
534
535         ret = pthread_mutex_lock(&start_work_mutex);
536         BUG_ON(ret);
537
538         for (i = 0; i < nr_tasks; i++) {
539                 task = tasks[i];
540                 sem_init(&task->sleep_sem, 0, 0);
541                 task->curr_event = 0;
542         }
543 }
544
545 static int __cmd_sched(void);
546
547 static void parse_trace(void)
548 {
549         __cmd_sched();
550
551         printf("nr_run_events:        %ld\n", nr_run_events);
552         printf("nr_sleep_events:      %ld\n", nr_sleep_events);
553         printf("nr_wakeup_events:     %ld\n", nr_wakeup_events);
554
555         if (targetless_wakeups)
556                 printf("target-less wakeups:  %ld\n", targetless_wakeups);
557         if (multitarget_wakeups)
558                 printf("multi-target wakeups: %ld\n", multitarget_wakeups);
559         if (nr_run_events_optimized)
560                 printf("run events optimized: %ld\n",
561                         nr_run_events_optimized);
562 }
563
564 static unsigned long nr_runs;
565 static nsec_t sum_runtime;
566 static nsec_t sum_fluct;
567 static nsec_t run_avg;
568
569 static void run_one_test(void)
570 {
571         nsec_t T0, T1, delta, avg_delta, fluct, std_dev;
572
573         T0 = get_nsecs();
574         wait_for_tasks();
575         T1 = get_nsecs();
576
577         delta = T1 - T0;
578         sum_runtime += delta;
579         nr_runs++;
580
581         avg_delta = sum_runtime / nr_runs;
582         if (delta < avg_delta)
583                 fluct = avg_delta - delta;
584         else
585                 fluct = delta - avg_delta;
586         sum_fluct += fluct;
587         std_dev = sum_fluct / nr_runs / sqrt(nr_runs);
588         if (!run_avg)
589                 run_avg = delta;
590         run_avg = (run_avg*9 + delta)/10;
591
592         printf("#%-3ld: %0.3f, ",
593                 nr_runs, (double)delta/1000000.0);
594
595 #if 0
596         printf("%0.2f +- %0.2f, ",
597                 (double)avg_delta/1e6, (double)std_dev/1e6);
598 #endif
599         printf("ravg: %0.2f, ",
600                 (double)run_avg/1e6);
601
602         printf("cpu: %0.2f / %0.2f",
603                 (double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6);
604
605 #if 0
606         /*
607          * rusage statistics done by the parent, these are less
608          * accurate than the sum_exec_runtime based statistics:
609          */
610         printf(" [%0.2f / %0.2f]",
611                 (double)parent_cpu_usage/1e6,
612                 (double)runavg_parent_cpu_usage/1e6);
613 #endif
614
615         printf("\n");
616
617         if (nr_sleep_corrections)
618                 printf(" (%ld sleep corrections)\n", nr_sleep_corrections);
619         nr_sleep_corrections = 0;
620 }
621
622 static void test_calibrations(void)
623 {
624         nsec_t T0, T1;
625
626         T0 = get_nsecs();
627         burn_nsecs(1e6);
628         T1 = get_nsecs();
629
630         printf("the run test took %Ld nsecs\n", T1-T0);
631
632         T0 = get_nsecs();
633         sleep_nsecs(1e6);
634         T1 = get_nsecs();
635
636         printf("the sleep test took %Ld nsecs\n", T1-T0);
637 }
638
639 static int
640 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
641 {
642         struct thread *thread;
643
644         thread = threads__findnew(event->comm.pid, &threads, &last_match);
645
646         dump_printf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
647                 (void *)(offset + head),
648                 (void *)(long)(event->header.size),
649                 event->comm.comm, event->comm.pid);
650
651         if (thread == NULL ||
652             thread__set_comm(thread, event->comm.comm)) {
653                 dump_printf("problem processing PERF_EVENT_COMM, skipping event.\n");
654                 return -1;
655         }
656         total_comm++;
657
658         return 0;
659 }
660
661
662 struct raw_event_sample {
663         u32 size;
664         char data[0];
665 };
666
667 #define FILL_FIELD(ptr, field, event, data)     \
668         ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
669
670 #define FILL_ARRAY(ptr, array, event, data)                     \
671 do {                                                            \
672         void *__array = raw_field_ptr(event, #array, data);     \
673         memcpy(ptr.array, __array, sizeof(ptr.array));  \
674 } while(0)
675
676 #define FILL_COMMON_FIELDS(ptr, event, data)                    \
677 do {                                                            \
678         FILL_FIELD(ptr, common_type, event, data);              \
679         FILL_FIELD(ptr, common_flags, event, data);             \
680         FILL_FIELD(ptr, common_preempt_count, event, data);     \
681         FILL_FIELD(ptr, common_pid, event, data);               \
682         FILL_FIELD(ptr, common_tgid, event, data);              \
683 } while (0)
684
685
686
687 struct trace_switch_event {
688         u32 size;
689
690         u16 common_type;
691         u8 common_flags;
692         u8 common_preempt_count;
693         u32 common_pid;
694         u32 common_tgid;
695
696         char prev_comm[16];
697         u32 prev_pid;
698         u32 prev_prio;
699         u64 prev_state;
700         char next_comm[16];
701         u32 next_pid;
702         u32 next_prio;
703 };
704
705
706 struct trace_wakeup_event {
707         u32 size;
708
709         u16 common_type;
710         u8 common_flags;
711         u8 common_preempt_count;
712         u32 common_pid;
713         u32 common_tgid;
714
715         char comm[16];
716         u32 pid;
717
718         u32 prio;
719         u32 success;
720         u32 cpu;
721 };
722
723 struct trace_fork_event {
724         u32 size;
725
726         u16 common_type;
727         u8 common_flags;
728         u8 common_preempt_count;
729         u32 common_pid;
730         u32 common_tgid;
731
732         char parent_comm[16];
733         u32 parent_pid;
734         char child_comm[16];
735         u32 child_pid;
736 };
737
738 struct trace_sched_handler {
739         void (*switch_event)(struct trace_switch_event *,
740                              struct event *,
741                              int cpu,
742                              u64 timestamp,
743                              struct thread *thread);
744
745         void (*wakeup_event)(struct trace_wakeup_event *,
746                              struct event *,
747                              int cpu,
748                              u64 timestamp,
749                              struct thread *thread);
750
751         void (*fork_event)(struct trace_fork_event *,
752                            struct event *,
753                            int cpu,
754                            u64 timestamp,
755                            struct thread *thread);
756 };
757
758
759 static void
760 replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
761                     struct event *event,
762                     int cpu __used,
763                     u64 timestamp __used,
764                     struct thread *thread __used)
765 {
766         struct task_desc *waker, *wakee;
767
768         if (verbose) {
769                 printf("sched_wakeup event %p\n", event);
770
771                 printf(" ... pid %d woke up %s/%d\n",
772                         wakeup_event->common_pid,
773                         wakeup_event->comm,
774                         wakeup_event->pid);
775         }
776
777         waker = register_pid(wakeup_event->common_pid, "<unknown>");
778         wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
779
780         add_sched_event_wakeup(waker, timestamp, wakee);
781 }
782
783 static unsigned long cpu_last_switched[MAX_CPUS];
784
785 static void
786 replay_switch_event(struct trace_switch_event *switch_event,
787                     struct event *event,
788                     int cpu,
789                     u64 timestamp,
790                     struct thread *thread __used)
791 {
792         struct task_desc *prev, *next;
793         u64 timestamp0;
794         s64 delta;
795
796         if (verbose)
797                 printf("sched_switch event %p\n", event);
798
799         if (cpu >= MAX_CPUS || cpu < 0)
800                 return;
801
802         timestamp0 = cpu_last_switched[cpu];
803         if (timestamp0)
804                 delta = timestamp - timestamp0;
805         else
806                 delta = 0;
807
808         if (delta < 0)
809                 die("hm, delta: %Ld < 0 ?\n", delta);
810
811         if (verbose) {
812                 printf(" ... switch from %s/%d to %s/%d [ran %Ld nsecs]\n",
813                         switch_event->prev_comm, switch_event->prev_pid,
814                         switch_event->next_comm, switch_event->next_pid,
815                         delta);
816         }
817
818         prev = register_pid(switch_event->prev_pid, switch_event->prev_comm);
819         next = register_pid(switch_event->next_pid, switch_event->next_comm);
820
821         cpu_last_switched[cpu] = timestamp;
822
823         add_sched_event_run(prev, timestamp, delta);
824         add_sched_event_sleep(prev, timestamp, switch_event->prev_state);
825 }
826
827
828 static void
829 replay_fork_event(struct trace_fork_event *fork_event,
830                   struct event *event,
831                   int cpu __used,
832                   u64 timestamp __used,
833                   struct thread *thread __used)
834 {
835         if (verbose) {
836                 printf("sched_fork event %p\n", event);
837                 printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid);
838                 printf("...  child: %s/%d\n", fork_event->child_comm, fork_event->child_pid);
839         }
840         register_pid(fork_event->parent_pid, fork_event->parent_comm);
841         register_pid(fork_event->child_pid, fork_event->child_comm);
842 }
843
844 static struct trace_sched_handler replay_ops  = {
845         .wakeup_event = replay_wakeup_event,
846         .switch_event = replay_switch_event,
847         .fork_event = replay_fork_event,
848 };
849
850
851 static struct trace_sched_handler *trace_handler;
852
853 static void
854 process_sched_wakeup_event(struct raw_event_sample *raw,
855                            struct event *event,
856                            int cpu __used,
857                            u64 timestamp __used,
858                            struct thread *thread __used)
859 {
860         struct trace_wakeup_event wakeup_event;
861
862         FILL_COMMON_FIELDS(wakeup_event, event, raw->data);
863
864         FILL_ARRAY(wakeup_event, comm, event, raw->data);
865         FILL_FIELD(wakeup_event, pid, event, raw->data);
866         FILL_FIELD(wakeup_event, prio, event, raw->data);
867         FILL_FIELD(wakeup_event, success, event, raw->data);
868         FILL_FIELD(wakeup_event, cpu, event, raw->data);
869
870         trace_handler->wakeup_event(&wakeup_event, event, cpu, timestamp, thread);
871 }
872
873 static void
874 process_sched_switch_event(struct raw_event_sample *raw,
875                            struct event *event,
876                            int cpu __used,
877                            u64 timestamp __used,
878                            struct thread *thread __used)
879 {
880         struct trace_switch_event switch_event;
881
882         FILL_COMMON_FIELDS(switch_event, event, raw->data);
883
884         FILL_ARRAY(switch_event, prev_comm, event, raw->data);
885         FILL_FIELD(switch_event, prev_pid, event, raw->data);
886         FILL_FIELD(switch_event, prev_prio, event, raw->data);
887         FILL_FIELD(switch_event, prev_state, event, raw->data);
888         FILL_ARRAY(switch_event, next_comm, event, raw->data);
889         FILL_FIELD(switch_event, next_pid, event, raw->data);
890         FILL_FIELD(switch_event, next_prio, event, raw->data);
891
892         trace_handler->switch_event(&switch_event, event, cpu, timestamp, thread);
893 }
894
895 static void
896 process_sched_fork_event(struct raw_event_sample *raw,
897                          struct event *event,
898                          int cpu __used,
899                          u64 timestamp __used,
900                          struct thread *thread __used)
901 {
902         struct trace_fork_event fork_event;
903
904         FILL_COMMON_FIELDS(fork_event, event, raw->data);
905
906         FILL_ARRAY(fork_event, parent_comm, event, raw->data);
907         FILL_FIELD(fork_event, parent_pid, event, raw->data);
908         FILL_ARRAY(fork_event, child_comm, event, raw->data);
909         FILL_FIELD(fork_event, child_pid, event, raw->data);
910
911         trace_handler->fork_event(&fork_event, event, cpu, timestamp, thread);
912 }
913
914 static void
915 process_sched_exit_event(struct event *event,
916                          int cpu __used,
917                          u64 timestamp __used,
918                          struct thread *thread __used)
919 {
920         if (verbose)
921                 printf("sched_exit event %p\n", event);
922 }
923
924 static void
925 process_raw_event(event_t *raw_event __used, void *more_data,
926                   int cpu, u64 timestamp, struct thread *thread)
927 {
928         struct raw_event_sample *raw = more_data;
929         struct event *event;
930         int type;
931
932         type = trace_parse_common_type(raw->data);
933         event = trace_find_event(type);
934
935         if (!strcmp(event->name, "sched_switch"))
936                 process_sched_switch_event(raw, event, cpu, timestamp, thread);
937         if (!strcmp(event->name, "sched_wakeup"))
938                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
939         if (!strcmp(event->name, "sched_wakeup_new"))
940                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
941         if (!strcmp(event->name, "sched_process_fork"))
942                 process_sched_fork_event(raw, event, cpu, timestamp, thread);
943         if (!strcmp(event->name, "sched_process_exit"))
944                 process_sched_exit_event(event, cpu, timestamp, thread);
945 }
946
947 static int
948 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
949 {
950         char level;
951         int show = 0;
952         struct dso *dso = NULL;
953         struct thread *thread;
954         u64 ip = event->ip.ip;
955         u64 timestamp = -1;
956         u32 cpu = -1;
957         u64 period = 1;
958         void *more_data = event->ip.__more_data;
959         int cpumode;
960
961         thread = threads__findnew(event->ip.pid, &threads, &last_match);
962
963         if (sample_type & PERF_SAMPLE_TIME) {
964                 timestamp = *(u64 *)more_data;
965                 more_data += sizeof(u64);
966         }
967
968         if (sample_type & PERF_SAMPLE_CPU) {
969                 cpu = *(u32 *)more_data;
970                 more_data += sizeof(u32);
971                 more_data += sizeof(u32); /* reserved */
972         }
973
974         if (sample_type & PERF_SAMPLE_PERIOD) {
975                 period = *(u64 *)more_data;
976                 more_data += sizeof(u64);
977         }
978
979         dump_printf("%p [%p]: PERF_EVENT_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n",
980                 (void *)(offset + head),
981                 (void *)(long)(event->header.size),
982                 event->header.misc,
983                 event->ip.pid, event->ip.tid,
984                 (void *)(long)ip,
985                 (long long)period);
986
987         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
988
989         if (thread == NULL) {
990                 eprintf("problem processing %d event, skipping it.\n",
991                         event->header.type);
992                 return -1;
993         }
994
995         cpumode = event->header.misc & PERF_EVENT_MISC_CPUMODE_MASK;
996
997         if (cpumode == PERF_EVENT_MISC_KERNEL) {
998                 show = SHOW_KERNEL;
999                 level = 'k';
1000
1001                 dso = kernel_dso;
1002
1003                 dump_printf(" ...... dso: %s\n", dso->name);
1004
1005         } else if (cpumode == PERF_EVENT_MISC_USER) {
1006
1007                 show = SHOW_USER;
1008                 level = '.';
1009
1010         } else {
1011                 show = SHOW_HV;
1012                 level = 'H';
1013
1014                 dso = hypervisor_dso;
1015
1016                 dump_printf(" ...... dso: [hypervisor]\n");
1017         }
1018
1019         if (sample_type & PERF_SAMPLE_RAW)
1020                 process_raw_event(event, more_data, cpu, timestamp, thread);
1021
1022         return 0;
1023 }
1024
1025 static int
1026 process_event(event_t *event, unsigned long offset, unsigned long head)
1027 {
1028         trace_event(event);
1029
1030         switch (event->header.type) {
1031         case PERF_EVENT_MMAP ... PERF_EVENT_LOST:
1032                 return 0;
1033
1034         case PERF_EVENT_COMM:
1035                 return process_comm_event(event, offset, head);
1036
1037         case PERF_EVENT_EXIT ... PERF_EVENT_READ:
1038                 return 0;
1039
1040         case PERF_EVENT_SAMPLE:
1041                 return process_sample_event(event, offset, head);
1042
1043         case PERF_EVENT_MAX:
1044         default:
1045                 return -1;
1046         }
1047
1048         return 0;
1049 }
1050
1051 static int __cmd_sched(void)
1052 {
1053         int ret, rc = EXIT_FAILURE;
1054         unsigned long offset = 0;
1055         unsigned long head = 0;
1056         struct stat perf_stat;
1057         event_t *event;
1058         uint32_t size;
1059         char *buf;
1060
1061         trace_report();
1062         register_idle_thread(&threads, &last_match);
1063
1064         input = open(input_name, O_RDONLY);
1065         if (input < 0) {
1066                 perror("failed to open file");
1067                 exit(-1);
1068         }
1069
1070         ret = fstat(input, &perf_stat);
1071         if (ret < 0) {
1072                 perror("failed to stat file");
1073                 exit(-1);
1074         }
1075
1076         if (!perf_stat.st_size) {
1077                 fprintf(stderr, "zero-sized file, nothing to do!\n");
1078                 exit(0);
1079         }
1080         header = perf_header__read(input);
1081         head = header->data_offset;
1082         sample_type = perf_header__sample_type(header);
1083
1084         if (!(sample_type & PERF_SAMPLE_RAW))
1085                 die("No trace sample to read. Did you call perf record "
1086                     "without -R?");
1087
1088         if (load_kernel() < 0) {
1089                 perror("failed to load kernel symbols");
1090                 return EXIT_FAILURE;
1091         }
1092
1093 remap:
1094         buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1095                            MAP_SHARED, input, offset);
1096         if (buf == MAP_FAILED) {
1097                 perror("failed to mmap file");
1098                 exit(-1);
1099         }
1100
1101 more:
1102         event = (event_t *)(buf + head);
1103
1104         size = event->header.size;
1105         if (!size)
1106                 size = 8;
1107
1108         if (head + event->header.size >= page_size * mmap_window) {
1109                 unsigned long shift = page_size * (head / page_size);
1110                 int res;
1111
1112                 res = munmap(buf, page_size * mmap_window);
1113                 assert(res == 0);
1114
1115                 offset += shift;
1116                 head -= shift;
1117                 goto remap;
1118         }
1119
1120         size = event->header.size;
1121
1122
1123         if (!size || process_event(event, offset, head) < 0) {
1124
1125                 /*
1126                  * assume we lost track of the stream, check alignment, and
1127                  * increment a single u64 in the hope to catch on again 'soon'.
1128                  */
1129
1130                 if (unlikely(head & 7))
1131                         head &= ~7ULL;
1132
1133                 size = 8;
1134         }
1135
1136         head += size;
1137
1138         if (offset + head < (unsigned long)perf_stat.st_size)
1139                 goto more;
1140
1141         rc = EXIT_SUCCESS;
1142         close(input);
1143
1144         return rc;
1145 }
1146
1147 static const char * const annotate_usage[] = {
1148         "perf trace [<options>] <command>",
1149         NULL
1150 };
1151
1152 static const struct option options[] = {
1153         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1154                     "dump raw trace in ASCII"),
1155         OPT_BOOLEAN('r', "replay", &replay_mode,
1156                     "replay sched behaviour from traces"),
1157         OPT_BOOLEAN('v', "verbose", &verbose,
1158                     "be more verbose (show symbol address, etc)"),
1159         OPT_END()
1160 };
1161
1162 int cmd_sched(int argc, const char **argv, const char *prefix __used)
1163 {
1164         long nr_iterations = 10, i;
1165
1166         symbol__init();
1167         page_size = getpagesize();
1168
1169         argc = parse_options(argc, argv, options, annotate_usage, 0);
1170         if (argc) {
1171                 /*
1172                  * Special case: if there's an argument left then assume tha
1173                  * it's a symbol filter:
1174                  */
1175                 if (argc > 1)
1176                         usage_with_options(annotate_usage, options);
1177         }
1178
1179 //      setup_pager();
1180
1181         if (replay_mode)
1182                 trace_handler = &replay_ops;
1183         else /* We may need a default subcommand */
1184                 die("Please select a sub command (-r)\n");
1185
1186         calibrate_run_measurement_overhead();
1187         calibrate_sleep_measurement_overhead();
1188
1189         test_calibrations();
1190
1191         parse_trace();
1192         print_task_traces();
1193         add_cross_task_wakeups();
1194
1195         create_tasks();
1196         printf("------------------------------------------------------------\n");
1197         for (i = 0; i < nr_iterations; i++)
1198                 run_one_test();
1199
1200         return 0;
1201 }